Pileus Git - ~andy/linux/blob - drivers/block/mtip32xx/mtip32xx.c

   1 /*
   2  * Driver for the Micron P320 SSD
   3  *   Copyright (C) 2011 Micron Technology, Inc.
   4  *
   5  * Portions of this code were derived from works subjected to the
   6  * following copyright:
   7  *    Copyright (C) 2009 Integrated Device Technology, Inc.
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  */
  20
  21 #include <linux/pci.h>
  22 #include <linux/interrupt.h>
  23 #include <linux/ata.h>
  24 #include <linux/delay.h>
  25 #include <linux/hdreg.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/random.h>
  28 #include <linux/smp.h>
  29 #include <linux/compat.h>
  30 #include <linux/fs.h>
  31 #include <linux/module.h>
  32 #include <linux/genhd.h>
  33 #include <linux/blkdev.h>
  34 #include <linux/bio.h>
  35 #include <linux/dma-mapping.h>
  36 #include <linux/idr.h>
  37 #include <linux/kthread.h>
  38 #include <../drivers/ata/ahci.h>
  39 #include <linux/export.h>
  40 #include <linux/debugfs.h>
  41 #include "mtip32xx.h"
  42
  43 #define HW_CMD_SLOT_SZ          (MTIP_MAX_COMMAND_SLOTS * 32)
  44
  45 /* DMA region containing RX Fis, Identify, RLE10, and SMART buffers */
  46 #define AHCI_RX_FIS_SZ          0x100
  47 #define AHCI_RX_FIS_OFFSET      0x0
  48 #define AHCI_IDFY_SZ            ATA_SECT_SIZE
  49 #define AHCI_IDFY_OFFSET        0x400
  50 #define AHCI_SECTBUF_SZ         ATA_SECT_SIZE
  51 #define AHCI_SECTBUF_OFFSET     0x800
  52 #define AHCI_SMARTBUF_SZ        ATA_SECT_SIZE
  53 #define AHCI_SMARTBUF_OFFSET    0xC00
  54 /* 0x100 + 0x200 + 0x200 + 0x200 is smaller than 4k but we pad it out */
  55 #define BLOCK_DMA_ALLOC_SZ      4096
  56
  57 /* DMA region containing command table (should be 8192 bytes) */
  58 #define AHCI_CMD_SLOT_SZ        sizeof(struct mtip_cmd_hdr)
  59 #define AHCI_CMD_TBL_SZ         (MTIP_MAX_COMMAND_SLOTS * AHCI_CMD_SLOT_SZ)
  60 #define AHCI_CMD_TBL_OFFSET     0x0
  61
  62 /* DMA region per command (contains header and SGL) */
  63 #define AHCI_CMD_TBL_HDR_SZ     0x80
  64 #define AHCI_CMD_TBL_HDR_OFFSET 0x0
  65 #define AHCI_CMD_TBL_SGL_SZ     (MTIP_MAX_SG * sizeof(struct mtip_cmd_sg))
  66 #define AHCI_CMD_TBL_SGL_OFFSET AHCI_CMD_TBL_HDR_SZ
  67 #define CMD_DMA_ALLOC_SZ        (AHCI_CMD_TBL_SGL_SZ + AHCI_CMD_TBL_HDR_SZ)
  68
  69
  70 #define HOST_CAP_NZDMA          (1 << 19)
  71 #define HOST_HSORG              0xFC
  72 #define HSORG_DISABLE_SLOTGRP_INTR (1<<24)
  73 #define HSORG_DISABLE_SLOTGRP_PXIS (1<<16)
  74 #define HSORG_HWREV             0xFF00
  75 #define HSORG_STYLE             0x8
  76 #define HSORG_SLOTGROUPS        0x7
  77
  78 #define PORT_COMMAND_ISSUE      0x38
  79 #define PORT_SDBV               0x7C
  80
  81 #define PORT_OFFSET             0x100
  82 #define PORT_MEM_SIZE           0x80
  83
  84 #define PORT_IRQ_ERR \
  85         (PORT_IRQ_HBUS_ERR | PORT_IRQ_IF_ERR | PORT_IRQ_CONNECT | \
  86          PORT_IRQ_PHYRDY | PORT_IRQ_UNK_FIS | PORT_IRQ_BAD_PMP | \
  87          PORT_IRQ_TF_ERR | PORT_IRQ_HBUS_DATA_ERR | PORT_IRQ_IF_NONFATAL | \
  88          PORT_IRQ_OVERFLOW)
  89 #define PORT_IRQ_LEGACY \
  90         (PORT_IRQ_PIOS_FIS | PORT_IRQ_D2H_REG_FIS)
  91 #define PORT_IRQ_HANDLED \
  92         (PORT_IRQ_SDB_FIS | PORT_IRQ_LEGACY | \
  93          PORT_IRQ_TF_ERR | PORT_IRQ_IF_ERR | \
  94          PORT_IRQ_CONNECT | PORT_IRQ_PHYRDY)
  95 #define DEF_PORT_IRQ \
  96         (PORT_IRQ_ERR | PORT_IRQ_LEGACY | PORT_IRQ_SDB_FIS)
  97
  98 /* product numbers */
  99 #define MTIP_PRODUCT_UNKNOWN    0x00
 100 #define MTIP_PRODUCT_ASICFPGA   0x11
 101
 102 /* Device instance number, incremented each time a device is probed. */
 103 static int instance;
 104
 105 struct list_head online_list;
 106 struct list_head removing_list;
 107 spinlock_t dev_lock;
 108
 109 /*
 110  * Global variable used to hold the major block device number
 111  * allocated in mtip_init().
 112  */
 113 static int mtip_major;
 114 static struct dentry *dfs_parent;
 115 static struct dentry *dfs_device_status;
 116
 117 static u32 cpu_use[NR_CPUS];
 118
 119 static DEFINE_SPINLOCK(rssd_index_lock);
 120 static DEFINE_IDA(rssd_index_ida);
 121
 122 static int mtip_block_initialize(struct driver_data *dd);
 123
 124 #ifdef CONFIG_COMPAT
 125 struct mtip_compat_ide_task_request_s {
 126         __u8            io_ports[8];
 127         __u8            hob_ports[8];
 128         ide_reg_valid_t out_flags;
 129         ide_reg_valid_t in_flags;
 130         int             data_phase;
 131         int             req_cmd;
 132         compat_ulong_t  out_size;
 133         compat_ulong_t  in_size;
 134 };
 135 #endif
 136
 137 /*
 138  * This function check_for_surprise_removal is called
 139  * while card is removed from the system and it will
 140  * read the vendor id from the configration space
 141  *
 142  * @pdev Pointer to the pci_dev structure.
 143  *
 144  * return value
 145  *       true if device removed, else false
 146  */
 147 static bool mtip_check_surprise_removal(struct pci_dev *pdev)
 148 {
 149         u16 vendor_id = 0;
 150         struct driver_data *dd = pci_get_drvdata(pdev);
 151
 152         if (dd->sr)
 153                 return true;
 154
 155        /* Read the vendorID from the configuration space */
 156         pci_read_config_word(pdev, 0x00, &vendor_id);
 157         if (vendor_id == 0xFFFF) {
 158                 dd->sr = true;
 159                 if (dd->queue)
 160                         set_bit(QUEUE_FLAG_DEAD, &dd->queue->queue_flags);
 161                 else
 162                         dev_warn(&dd->pdev->dev,
 163                                 "%s: dd->queue is NULL\n", __func__);
 164                 if (dd->port) {
 165                         set_bit(MTIP_PF_SR_CLEANUP_BIT, &dd->port->flags);
 166                         wake_up_interruptible(&dd->port->svc_wait);
 167                 } else
 168                         dev_warn(&dd->pdev->dev,
 169                                 "%s: dd->port is NULL\n", __func__);
 170                 return true; /* device removed */
 171         }
 172
 173         return false; /* device present */
 174 }
 175
 176 /*
 177  * Obtain an empty command slot.
 178  *
 179  * This function needs to be reentrant since it could be called
 180  * at the same time on multiple CPUs. The allocation of the
 181  * command slot must be atomic.
 182  *
 183  * @port Pointer to the port data structure.
 184  *
 185  * return value
 186  *      >= 0    Index of command slot obtained.
 187  *      -1      No command slots available.
 188  */
 189 static int get_slot(struct mtip_port *port)
 190 {
 191         int slot, i;
 192         unsigned int num_command_slots = port->dd->slot_groups * 32;
 193
 194         /*
 195          * Try 10 times, because there is a small race here.
 196          *  that's ok, because it's still cheaper than a lock.
 197          *
 198          * Race: Since this section is not protected by lock, same bit
 199          * could be chosen by different process contexts running in
 200          * different processor. So instead of costly lock, we are going
 201          * with loop.
 202          */
 203         for (i = 0; i < 10; i++) {
 204                 slot = find_next_zero_bit(port->allocated,
 205                                          num_command_slots, 1);
 206                 if ((slot < num_command_slots) &&
 207                     (!test_and_set_bit(slot, port->allocated)))
 208                         return slot;
 209         }
 210         dev_warn(&port->dd->pdev->dev, "Failed to get a tag.\n");
 211
 212         mtip_check_surprise_removal(port->dd->pdev);
 213         return -1;
 214 }
 215
 216 /*
 217  * Release a command slot.
 218  *
 219  * @port Pointer to the port data structure.
 220  * @tag  Tag of command to release
 221  *
 222  * return value
 223  *      None
 224  */
 225 static inline void release_slot(struct mtip_port *port, int tag)
 226 {
 227         smp_mb__before_clear_bit();
 228         clear_bit(tag, port->allocated);
 229         smp_mb__after_clear_bit();
 230 }
 231
 232 /*
 233  * IO completion function.
 234  *
 235  * This completion function is called by the driver ISR when a
 236  * command that was issued by the kernel completes. It first calls the
 237  * asynchronous completion function which normally calls back into the block
 238  * layer passing the asynchronous callback data, then unmaps the
 239  * scatter list associated with the completed command, and finally
 240  * clears the allocated bit associated with the completed command.
 241  *
 242  * @port   Pointer to the port data structure.
 243  * @tag    Tag of the command.
 244  * @data   Pointer to driver_data.
 245  * @status Completion status.
 246  *
 247  * return value
 248  *      None
 249  */
 250 static void mtip_async_complete(struct mtip_port *port,
 251                                 int tag,
 252                                 void *data,
 253                                 int status)
 254 {
 255         struct mtip_cmd *command;
 256         struct driver_data *dd = data;
 257         int cb_status = status ? -EIO : 0;
 258
 259         if (unlikely(!dd) || unlikely(!port))
 260                 return;
 261
 262         command = &port->commands[tag];
 263
 264         if (unlikely(status == PORT_IRQ_TF_ERR)) {
 265                 dev_warn(&port->dd->pdev->dev,
 266                         "Command tag %d failed due to TFE\n", tag);
 267         }
 268
 269         /* Upper layer callback */
 270         if (likely(command->async_callback))
 271                 command->async_callback(command->async_data, cb_status);
 272
 273         command->async_callback = NULL;
 274         command->comp_func = NULL;
 275
 276         /* Unmap the DMA scatter list entries */
 277         dma_unmap_sg(&dd->pdev->dev,
 278                 command->sg,
 279                 command->scatter_ents,
 280                 command->direction);
 281
 282         /* Clear the allocated and active bits for the command */
 283         atomic_set(&port->commands[tag].active, 0);
 284         release_slot(port, tag);
 285
 286         up(&port->cmd_slot);
 287 }
 288
 289 /*
 290  * This function is called for clean the pending command in the
 291  * command slot during the surprise removal of device and return
 292  * error to the upper layer.
 293  *
 294  * @dd Pointer to the DRIVER_DATA structure.
 295  *
 296  * return value
 297  *      None
 298  */
 299 static void mtip_command_cleanup(struct driver_data *dd)
 300 {
 301         int tag = 0;
 302         struct mtip_cmd *cmd;
 303         struct mtip_port *port = dd->port;
 304         unsigned int num_cmd_slots = dd->slot_groups * 32;
 305
 306         if (!test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag))
 307                 return;
 308
 309         if (!port)
 310                 return;
 311
 312         cmd = &port->commands[MTIP_TAG_INTERNAL];
 313         if (atomic_read(&cmd->active))
 314                 if (readl(port->cmd_issue[MTIP_TAG_INTERNAL]) &
 315                                         (1 << MTIP_TAG_INTERNAL))
 316                         if (cmd->comp_func)
 317                                 cmd->comp_func(port, MTIP_TAG_INTERNAL,
 318                                          cmd->comp_data, -ENODEV);
 319
 320         while (1) {
 321                 tag = find_next_bit(port->allocated, num_cmd_slots, tag);
 322                 if (tag >= num_cmd_slots)
 323                         break;
 324
 325                 cmd = &port->commands[tag];
 326                 if (atomic_read(&cmd->active))
 327                         mtip_async_complete(port, tag, dd, -ENODEV);
 328         }
 329
 330         set_bit(MTIP_DDF_CLEANUP_BIT, &dd->dd_flag);
 331 }
 332
 333 /*
 334  * Reset the HBA (without sleeping)
 335  *
 336  * @dd Pointer to the driver data structure.
 337  *
 338  * return value
 339  *      0       The reset was successful.
 340  *      -1      The HBA Reset bit did not clear.
 341  */
 342 static int mtip_hba_reset(struct driver_data *dd)
 343 {
 344         unsigned long timeout;
 345
 346         /* Set the reset bit */
 347         writel(HOST_RESET, dd->mmio + HOST_CTL);
 348
 349         /* Flush */
 350         readl(dd->mmio + HOST_CTL);
 351
 352         /* Spin for up to 2 seconds, waiting for reset acknowledgement */
 353         timeout = jiffies + msecs_to_jiffies(2000);
 354         do {
 355                 mdelay(10);
 356                 if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))
 357                         return -1;
 358
 359         } while ((readl(dd->mmio + HOST_CTL) & HOST_RESET)
 360                  && time_before(jiffies, timeout));
 361
 362         if (readl(dd->mmio + HOST_CTL) & HOST_RESET)
 363                 return -1;
 364
 365         return 0;
 366 }
 367
 368 /*
 369  * Issue a command to the hardware.
 370  *
 371  * Set the appropriate bit in the s_active and Command Issue hardware
 372  * registers, causing hardware command processing to begin.
 373  *
 374  * @port Pointer to the port structure.
 375  * @tag  The tag of the command to be issued.
 376  *
 377  * return value
 378  *      None
 379  */
 380 static inline void mtip_issue_ncq_command(struct mtip_port *port, int tag)
 381 {
 382         int group = tag >> 5;
 383
 384         atomic_set(&port->commands[tag].active, 1);
 385
 386         /* guard SACT and CI registers */
 387         spin_lock(&port->cmd_issue_lock[group]);
 388         writel((1 << MTIP_TAG_BIT(tag)),
 389                         port->s_active[MTIP_TAG_INDEX(tag)]);
 390         writel((1 << MTIP_TAG_BIT(tag)),
 391                         port->cmd_issue[MTIP_TAG_INDEX(tag)]);
 392         spin_unlock(&port->cmd_issue_lock[group]);
 393
 394         /* Set the command's timeout value.*/
 395         port->commands[tag].comp_time = jiffies + msecs_to_jiffies(
 396                                         MTIP_NCQ_COMMAND_TIMEOUT_MS);
 397 }
 398
 399 /*
 400  * Enable/disable the reception of FIS
 401  *
 402  * @port   Pointer to the port data structure
 403  * @enable 1 to enable, 0 to disable
 404  *
 405  * return value
 406  *      Previous state: 1 enabled, 0 disabled
 407  */
 408 static int mtip_enable_fis(struct mtip_port *port, int enable)
 409 {
 410         u32 tmp;
 411
 412         /* enable FIS reception */
 413         tmp = readl(port->mmio + PORT_CMD);
 414         if (enable)
 415                 writel(tmp | PORT_CMD_FIS_RX, port->mmio + PORT_CMD);
 416         else
 417                 writel(tmp & ~PORT_CMD_FIS_RX, port->mmio + PORT_CMD);
 418
 419         /* Flush */
 420         readl(port->mmio + PORT_CMD);
 421
 422         return (((tmp & PORT_CMD_FIS_RX) == PORT_CMD_FIS_RX));
 423 }
 424
 425 /*
 426  * Enable/disable the DMA engine
 427  *
 428  * @port   Pointer to the port data structure
 429  * @enable 1 to enable, 0 to disable
 430  *
 431  * return value
 432  *      Previous state: 1 enabled, 0 disabled.
 433  */
 434 static int mtip_enable_engine(struct mtip_port *port, int enable)
 435 {
 436         u32 tmp;
 437
 438         /* enable FIS reception */
 439         tmp = readl(port->mmio + PORT_CMD);
 440         if (enable)
 441                 writel(tmp | PORT_CMD_START, port->mmio + PORT_CMD);
 442         else
 443                 writel(tmp & ~PORT_CMD_START, port->mmio + PORT_CMD);
 444
 445         readl(port->mmio + PORT_CMD);
 446         return (((tmp & PORT_CMD_START) == PORT_CMD_START));
 447 }
 448
 449 /*
 450  * Enables the port DMA engine and FIS reception.
 451  *
 452  * return value
 453  *      None
 454  */
 455 static inline void mtip_start_port(struct mtip_port *port)
 456 {
 457         /* Enable FIS reception */
 458         mtip_enable_fis(port, 1);
 459
 460         /* Enable the DMA engine */
 461         mtip_enable_engine(port, 1);
 462 }
 463
 464 /*
 465  * Deinitialize a port by disabling port interrupts, the DMA engine,
 466  * and FIS reception.
 467  *
 468  * @port Pointer to the port structure
 469  *
 470  * return value
 471  *      None
 472  */
 473 static inline void mtip_deinit_port(struct mtip_port *port)
 474 {
 475         /* Disable interrupts on this port */
 476         writel(0, port->mmio + PORT_IRQ_MASK);
 477
 478         /* Disable the DMA engine */
 479         mtip_enable_engine(port, 0);
 480
 481         /* Disable FIS reception */
 482         mtip_enable_fis(port, 0);
 483 }
 484
 485 /*
 486  * Initialize a port.
 487  *
 488  * This function deinitializes the port by calling mtip_deinit_port() and
 489  * then initializes it by setting the command header and RX FIS addresses,
 490  * clearing the SError register and any pending port interrupts before
 491  * re-enabling the default set of port interrupts.
 492  *
 493  * @port Pointer to the port structure.
 494  *
 495  * return value
 496  *      None
 497  */
 498 static void mtip_init_port(struct mtip_port *port)
 499 {
 500         int i;
 501         mtip_deinit_port(port);
 502
 503         /* Program the command list base and FIS base addresses */
 504         if (readl(port->dd->mmio + HOST_CAP) & HOST_CAP_64) {
 505                 writel((port->command_list_dma >> 16) >> 16,
 506                          port->mmio + PORT_LST_ADDR_HI);
 507                 writel((port->rxfis_dma >> 16) >> 16,
 508                          port->mmio + PORT_FIS_ADDR_HI);
 509         }
 510
 511         writel(port->command_list_dma & 0xFFFFFFFF,
 512                         port->mmio + PORT_LST_ADDR);
 513         writel(port->rxfis_dma & 0xFFFFFFFF, port->mmio + PORT_FIS_ADDR);
 514
 515         /* Clear SError */
 516         writel(readl(port->mmio + PORT_SCR_ERR), port->mmio + PORT_SCR_ERR);
 517
 518         /* reset the completed registers.*/
 519         for (i = 0; i < port->dd->slot_groups; i++)
 520                 writel(0xFFFFFFFF, port->completed[i]);
 521
 522         /* Clear any pending interrupts for this port */
 523         writel(readl(port->mmio + PORT_IRQ_STAT), port->mmio + PORT_IRQ_STAT);
 524
 525         /* Clear any pending interrupts on the HBA. */
 526         writel(readl(port->dd->mmio + HOST_IRQ_STAT),
 527                                         port->dd->mmio + HOST_IRQ_STAT);
 528
 529         /* Enable port interrupts */
 530         writel(DEF_PORT_IRQ, port->mmio + PORT_IRQ_MASK);
 531 }
 532
 533 /*
 534  * Restart a port
 535  *
 536  * @port Pointer to the port data structure.
 537  *
 538  * return value
 539  *      None
 540  */
 541 static void mtip_restart_port(struct mtip_port *port)
 542 {
 543         unsigned long timeout;
 544
 545         /* Disable the DMA engine */
 546         mtip_enable_engine(port, 0);
 547
 548         /* Chip quirk: wait up to 500ms for PxCMD.CR == 0 */
 549         timeout = jiffies + msecs_to_jiffies(500);
 550         while ((readl(port->mmio + PORT_CMD) & PORT_CMD_LIST_ON)
 551                  && time_before(jiffies, timeout))
 552                 ;
 553
 554         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 555                 return;
 556
 557         /*
 558          * Chip quirk: escalate to hba reset if
 559          * PxCMD.CR not clear after 500 ms
 560          */
 561         if (readl(port->mmio + PORT_CMD) & PORT_CMD_LIST_ON) {
 562                 dev_warn(&port->dd->pdev->dev,
 563                         "PxCMD.CR not clear, escalating reset\n");
 564
 565                 if (mtip_hba_reset(port->dd))
 566                         dev_err(&port->dd->pdev->dev,
 567                                 "HBA reset escalation failed.\n");
 568
 569                 /* 30 ms delay before com reset to quiesce chip */
 570                 mdelay(30);
 571         }
 572
 573         dev_warn(&port->dd->pdev->dev, "Issuing COM reset\n");
 574
 575         /* Set PxSCTL.DET */
 576         writel(readl(port->mmio + PORT_SCR_CTL) |
 577                          1, port->mmio + PORT_SCR_CTL);
 578         readl(port->mmio + PORT_SCR_CTL);
 579
 580         /* Wait 1 ms to quiesce chip function */
 581         timeout = jiffies + msecs_to_jiffies(1);
 582         while (time_before(jiffies, timeout))
 583                 ;
 584
 585         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 586                 return;
 587
 588         /* Clear PxSCTL.DET */
 589         writel(readl(port->mmio + PORT_SCR_CTL) & ~1,
 590                          port->mmio + PORT_SCR_CTL);
 591         readl(port->mmio + PORT_SCR_CTL);
 592
 593         /* Wait 500 ms for bit 0 of PORT_SCR_STS to be set */
 594         timeout = jiffies + msecs_to_jiffies(500);
 595         while (((readl(port->mmio + PORT_SCR_STAT) & 0x01) == 0)
 596                          && time_before(jiffies, timeout))
 597                 ;
 598
 599         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 600                 return;
 601
 602         if ((readl(port->mmio + PORT_SCR_STAT) & 0x01) == 0)
 603                 dev_warn(&port->dd->pdev->dev,
 604                         "COM reset failed\n");
 605
 606         mtip_init_port(port);
 607         mtip_start_port(port);
 608
 609 }
 610
 611 static int mtip_device_reset(struct driver_data *dd)
 612 {
 613         int rv = 0;
 614
 615         if (mtip_check_surprise_removal(dd->pdev))
 616                 return 0;
 617
 618         if (mtip_hba_reset(dd) < 0)
 619                 rv = -EFAULT;
 620
 621         mdelay(1);
 622         mtip_init_port(dd->port);
 623         mtip_start_port(dd->port);
 624
 625         /* Enable interrupts on the HBA. */
 626         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
 627                                         dd->mmio + HOST_CTL);
 628         return rv;
 629 }
 630
 631 /*
 632  * Helper function for tag logging
 633  */
 634 static void print_tags(struct driver_data *dd,
 635                         char *msg,
 636                         unsigned long *tagbits,
 637                         int cnt)
 638 {
 639         unsigned char tagmap[128];
 640         int group, tagmap_len = 0;
 641
 642         memset(tagmap, 0, sizeof(tagmap));
 643         for (group = SLOTBITS_IN_LONGS; group > 0; group--)
 644                 tagmap_len = sprintf(tagmap + tagmap_len, "%016lX ",
 645                                                 tagbits[group-1]);
 646         dev_warn(&dd->pdev->dev,
 647                         "%d command(s) %s: tagmap [%s]", cnt, msg, tagmap);
 648 }
 649
 650 /*
 651  * Called periodically to see if any read/write commands are
 652  * taking too long to complete.
 653  *
 654  * @data Pointer to the PORT data structure.
 655  *
 656  * return value
 657  *      None
 658  */
 659 static void mtip_timeout_function(unsigned long int data)
 660 {
 661         struct mtip_port *port = (struct mtip_port *) data;
 662         struct host_to_dev_fis *fis;
 663         struct mtip_cmd *command;
 664         int tag, cmdto_cnt = 0;
 665         unsigned int bit, group;
 666         unsigned int num_command_slots;
 667         unsigned long to, tagaccum[SLOTBITS_IN_LONGS];
 668
 669         if (unlikely(!port))
 670                 return;
 671
 672         if (unlikely(port->dd->sr))
 673                 return;
 674
 675         if (test_bit(MTIP_DDF_RESUME_BIT, &port->dd->dd_flag)) {
 676                 mod_timer(&port->cmd_timer,
 677                         jiffies + msecs_to_jiffies(30000));
 678                 return;
 679         }
 680         /* clear the tag accumulator */
 681         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 682         num_command_slots = port->dd->slot_groups * 32;
 683
 684         for (tag = 0; tag < num_command_slots; tag++) {
 685                 /*
 686                  * Skip internal command slot as it has
 687                  * its own timeout mechanism
 688                  */
 689                 if (tag == MTIP_TAG_INTERNAL)
 690                         continue;
 691
 692                 if (atomic_read(&port->commands[tag].active) &&
 693                    (time_after(jiffies, port->commands[tag].comp_time))) {
 694                         group = tag >> 5;
 695                         bit = tag & 0x1F;
 696
 697                         command = &port->commands[tag];
 698                         fis = (struct host_to_dev_fis *) command->command;
 699
 700                         set_bit(tag, tagaccum);
 701                         cmdto_cnt++;
 702                         if (cmdto_cnt == 1)
 703                                 set_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
 704
 705                         /*
 706                          * Clear the completed bit. This should prevent
 707                          *  any interrupt handlers from trying to retire
 708                          *  the command.
 709                          */
 710                         writel(1 << bit, port->completed[group]);
 711
 712                         /* Call the async completion callback. */
 713                         if (likely(command->async_callback))
 714                                 command->async_callback(command->async_data,
 715                                                          -EIO);
 716                         command->async_callback = NULL;
 717                         command->comp_func = NULL;
 718
 719                         /* Unmap the DMA scatter list entries */
 720                         dma_unmap_sg(&port->dd->pdev->dev,
 721                                         command->sg,
 722                                         command->scatter_ents,
 723                                         command->direction);
 724
 725                         /*
 726                          * Clear the allocated bit and active tag for the
 727                          * command.
 728                          */
 729                         atomic_set(&port->commands[tag].active, 0);
 730                         release_slot(port, tag);
 731
 732                         up(&port->cmd_slot);
 733                 }
 734         }
 735
 736         if (cmdto_cnt) {
 737                 print_tags(port->dd, "timed out", tagaccum, cmdto_cnt);
 738                 if (!test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags)) {
 739                         mtip_device_reset(port->dd);
 740                         wake_up_interruptible(&port->svc_wait);
 741                 }
 742                 clear_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
 743         }
 744
 745         if (port->ic_pause_timer) {
 746                 to  = port->ic_pause_timer + msecs_to_jiffies(1000);
 747                 if (time_after(jiffies, to)) {
 748                         if (!test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags)) {
 749                                 port->ic_pause_timer = 0;
 750                                 clear_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
 751                                 clear_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
 752                                 clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
 753                                 wake_up_interruptible(&port->svc_wait);
 754                         }
 755
 756
 757                 }
 758         }
 759
 760         /* Restart the timer */
 761         mod_timer(&port->cmd_timer,
 762                 jiffies + msecs_to_jiffies(MTIP_TIMEOUT_CHECK_PERIOD));
 763 }
 764
 765 /*
 766  * Internal command completion callback function.
 767  *
 768  * This function is normally called by the driver ISR when an internal
 769  * command completed. This function signals the command completion by
 770  * calling complete().
 771  *
 772  * @port   Pointer to the port data structure.
 773  * @tag    Tag of the command that has completed.
 774  * @data   Pointer to a completion structure.
 775  * @status Completion status.
 776  *
 777  * return value
 778  *      None
 779  */
 780 static void mtip_completion(struct mtip_port *port,
 781                             int tag,
 782                             void *data,
 783                             int status)
 784 {
 785         struct mtip_cmd *command = &port->commands[tag];
 786         struct completion *waiting = data;
 787         if (unlikely(status == PORT_IRQ_TF_ERR))
 788                 dev_warn(&port->dd->pdev->dev,
 789                         "Internal command %d completed with TFE\n", tag);
 790
 791         command->async_callback = NULL;
 792         command->comp_func = NULL;
 793
 794         complete(waiting);
 795 }
 796
 797 static void mtip_null_completion(struct mtip_port *port,
 798                             int tag,
 799                             void *data,
 800                             int status)
 801 {
 802         return;
 803 }
 804
 805 static int mtip_read_log_page(struct mtip_port *port, u8 page, u16 *buffer,
 806                                 dma_addr_t buffer_dma, unsigned int sectors);
 807 static int mtip_get_smart_attr(struct mtip_port *port, unsigned int id,
 808                                                 struct smart_attr *attrib);
 809 /*
 810  * Handle an error.
 811  *
 812  * @dd Pointer to the DRIVER_DATA structure.
 813  *
 814  * return value
 815  *      None
 816  */
 817 static void mtip_handle_tfe(struct driver_data *dd)
 818 {
 819         int group, tag, bit, reissue, rv;
 820         struct mtip_port *port;
 821         struct mtip_cmd  *cmd;
 822         u32 completed;
 823         struct host_to_dev_fis *fis;
 824         unsigned long tagaccum[SLOTBITS_IN_LONGS];
 825         unsigned int cmd_cnt = 0;
 826         unsigned char *buf;
 827         char *fail_reason = NULL;
 828         int fail_all_ncq_write = 0, fail_all_ncq_cmds = 0;
 829
 830         dev_warn(&dd->pdev->dev, "Taskfile error\n");
 831
 832         port = dd->port;
 833
 834         /* Stop the timer to prevent command timeouts. */
 835         del_timer(&port->cmd_timer);
 836         set_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
 837
 838         if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags) &&
 839                         test_bit(MTIP_TAG_INTERNAL, port->allocated)) {
 840                 cmd = &port->commands[MTIP_TAG_INTERNAL];
 841                 dbg_printk(MTIP_DRV_NAME " TFE for the internal command\n");
 842
 843                 atomic_inc(&cmd->active); /* active > 1 indicates error */
 844                 if (cmd->comp_data && cmd->comp_func) {
 845                         cmd->comp_func(port, MTIP_TAG_INTERNAL,
 846                                         cmd->comp_data, PORT_IRQ_TF_ERR);
 847                 }
 848                 goto handle_tfe_exit;
 849         }
 850
 851         /* clear the tag accumulator */
 852         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 853
 854         /* Loop through all the groups */
 855         for (group = 0; group < dd->slot_groups; group++) {
 856                 completed = readl(port->completed[group]);
 857
 858                 /* clear completed status register in the hardware.*/
 859                 writel(completed, port->completed[group]);
 860
 861                 /* Process successfully completed commands */
 862                 for (bit = 0; bit < 32 && completed; bit++) {
 863                         if (!(completed & (1<<bit)))
 864                                 continue;
 865                         tag = (group << 5) + bit;
 866
 867                         /* Skip the internal command slot */
 868                         if (tag == MTIP_TAG_INTERNAL)
 869                                 continue;
 870
 871                         cmd = &port->commands[tag];
 872                         if (likely(cmd->comp_func)) {
 873                                 set_bit(tag, tagaccum);
 874                                 cmd_cnt++;
 875                                 atomic_set(&cmd->active, 0);
 876                                 cmd->comp_func(port,
 877                                          tag,
 878                                          cmd->comp_data,
 879                                          0);
 880                         } else {
 881                                 dev_err(&port->dd->pdev->dev,
 882                                         "Missing completion func for tag %d",
 883                                         tag);
 884                                 if (mtip_check_surprise_removal(dd->pdev)) {
 885                                         /* don't proceed further */
 886                                         return;
 887                                 }
 888                         }
 889                 }
 890         }
 891
 892         print_tags(dd, "completed (TFE)", tagaccum, cmd_cnt);
 893
 894         /* Restart the port */
 895         mdelay(20);
 896         mtip_restart_port(port);
 897
 898         /* Trying to determine the cause of the error */
 899         rv = mtip_read_log_page(dd->port, ATA_LOG_SATA_NCQ,
 900                                 dd->port->log_buf,
 901                                 dd->port->log_buf_dma, 1);
 902         if (rv) {
 903                 dev_warn(&dd->pdev->dev,
 904                         "Error in READ LOG EXT (10h) command\n");
 905                 /* non-critical error, don't fail the load */
 906         } else {
 907                 buf = (unsigned char *)dd->port->log_buf;
 908                 if (buf[259] & 0x1) {
 909                         dev_info(&dd->pdev->dev,
 910                                 "Write protect bit is set.\n");
 911                         set_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag);
 912                         fail_all_ncq_write = 1;
 913                         fail_reason = "write protect";
 914                 }
 915                 if (buf[288] == 0xF7) {
 916                         dev_info(&dd->pdev->dev,
 917                                 "Exceeded Tmax, drive in thermal shutdown.\n");
 918                         set_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag);
 919                         fail_all_ncq_cmds = 1;
 920                         fail_reason = "thermal shutdown";
 921                 }
 922                 if (buf[288] == 0xBF) {
 923                         set_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag);
 924                         dev_info(&dd->pdev->dev,
 925                                 "Drive indicates rebuild has failed. Secure erase required.\n");
 926                         fail_all_ncq_cmds = 1;
 927                         fail_reason = "rebuild failed";
 928                 }
 929         }
 930
 931         /* clear the tag accumulator */
 932         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 933
 934         /* Loop through all the groups */
 935         for (group = 0; group < dd->slot_groups; group++) {
 936                 for (bit = 0; bit < 32; bit++) {
 937                         reissue = 1;
 938                         tag = (group << 5) + bit;
 939                         cmd = &port->commands[tag];
 940
 941                         /* If the active bit is set re-issue the command */
 942                         if (atomic_read(&cmd->active) == 0)
 943                                 continue;
 944
 945                         fis = (struct host_to_dev_fis *)cmd->command;
 946
 947                         /* Should re-issue? */
 948                         if (tag == MTIP_TAG_INTERNAL ||
 949                             fis->command == ATA_CMD_SET_FEATURES)
 950                                 reissue = 0;
 951                         else {
 952                                 if (fail_all_ncq_cmds ||
 953                                         (fail_all_ncq_write &&
 954                                         fis->command == ATA_CMD_FPDMA_WRITE)) {
 955                                         dev_warn(&dd->pdev->dev,
 956                                         "  Fail: %s w/tag %d [%s].\n",
 957                                         fis->command == ATA_CMD_FPDMA_WRITE ?
 958                                                 "write" : "read",
 959                                         tag,
 960                                         fail_reason != NULL ?
 961                                                 fail_reason : "unknown");
 962                                         atomic_set(&cmd->active, 0);
 963                                         if (cmd->comp_func) {
 964                                                 cmd->comp_func(port, tag,
 965                                                         cmd->comp_data,
 966                                                         -ENODATA);
 967                                         }
 968                                         continue;
 969                                 }
 970                         }
 971
 972                         /*
 973                          * First check if this command has
 974                          *  exceeded its retries.
 975                          */
 976                         if (reissue && (cmd->retries-- > 0)) {
 977
 978                                 set_bit(tag, tagaccum);
 979
 980                                 /* Re-issue the command. */
 981                                 mtip_issue_ncq_command(port, tag);
 982
 983                                 continue;
 984                         }
 985
 986                         /* Retire a command that will not be reissued */
 987                         dev_warn(&port->dd->pdev->dev,
 988                                 "retiring tag %d\n", tag);
 989                         atomic_set(&cmd->active, 0);
 990
 991                         if (cmd->comp_func)
 992                                 cmd->comp_func(
 993                                         port,
 994                                         tag,
 995                                         cmd->comp_data,
 996                                         PORT_IRQ_TF_ERR);
 997                         else
 998                                 dev_warn(&port->dd->pdev->dev,
 999                                         "Bad completion for tag %d\n",
1000                                         tag);
1001                 }
1002         }
1003         print_tags(dd, "reissued (TFE)", tagaccum, cmd_cnt);
1004
1005 handle_tfe_exit:
1006         /* clear eh_active */
1007         clear_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
1008         wake_up_interruptible(&port->svc_wait);
1009
1010         mod_timer(&port->cmd_timer,
1011                  jiffies + msecs_to_jiffies(MTIP_TIMEOUT_CHECK_PERIOD));
1012 }
1013
1014 /*
1015  * Handle a set device bits interrupt
1016  */
1017 static inline void mtip_workq_sdbfx(struct mtip_port *port, int group,
1018                                                         u32 completed)
1019 {
1020         struct driver_data *dd = port->dd;
1021         int tag, bit;
1022         struct mtip_cmd *command;
1023
1024         if (!completed) {
1025                 WARN_ON_ONCE(!completed);
1026                 return;
1027         }
1028         /* clear completed status register in the hardware.*/
1029         writel(completed, port->completed[group]);
1030
1031         /* Process completed commands. */
1032         for (bit = 0; (bit < 32) && completed; bit++) {
1033                 if (completed & 0x01) {
1034                         tag = (group << 5) | bit;
1035
1036                         /* skip internal command slot. */
1037                         if (unlikely(tag == MTIP_TAG_INTERNAL))
1038                                 continue;
1039
1040                         command = &port->commands[tag];
1041                         /* make internal callback */
1042                         if (likely(command->comp_func)) {
1043                                 command->comp_func(
1044                                         port,
1045                                         tag,
1046                                         command->comp_data,
1047                                         0);
1048                         } else {
1049                                 dev_dbg(&dd->pdev->dev,
1050                                         "Null completion for tag %d",
1051                                         tag);
1052
1053                                 if (mtip_check_surprise_removal(
1054                                         dd->pdev)) {
1055                                         return;
1056                                 }
1057                         }
1058                 }
1059                 completed >>= 1;
1060         }
1061
1062         /* If last, re-enable interrupts */
1063         if (atomic_dec_return(&dd->irq_workers_active) == 0)
1064                 writel(0xffffffff, dd->mmio + HOST_IRQ_STAT);
1065 }
1066
1067 /*
1068  * Process legacy pio and d2h interrupts
1069  */
1070 static inline void mtip_process_legacy(struct driver_data *dd, u32 port_stat)
1071 {
1072         struct mtip_port *port = dd->port;
1073         struct mtip_cmd *cmd = &port->commands[MTIP_TAG_INTERNAL];
1074
1075         if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags) &&
1076             (cmd != NULL) && !(readl(port->cmd_issue[MTIP_TAG_INTERNAL])
1077                 & (1 << MTIP_TAG_INTERNAL))) {
1078                 if (cmd->comp_func) {
1079                         cmd->comp_func(port,
1080                                 MTIP_TAG_INTERNAL,
1081                                 cmd->comp_data,
1082                                 0);
1083                         return;
1084                 }
1085         }
1086
1087         return;
1088 }
1089
1090 /*
1091  * Demux and handle errors
1092  */
1093 static inline void mtip_process_errors(struct driver_data *dd, u32 port_stat)
1094 {
1095         if (likely(port_stat & (PORT_IRQ_TF_ERR | PORT_IRQ_IF_ERR)))
1096                 mtip_handle_tfe(dd);
1097
1098         if (unlikely(port_stat & PORT_IRQ_CONNECT)) {
1099                 dev_warn(&dd->pdev->dev,
1100                         "Clearing PxSERR.DIAG.x\n");
1101                 writel((1 << 26), dd->port->mmio + PORT_SCR_ERR);
1102         }
1103
1104         if (unlikely(port_stat & PORT_IRQ_PHYRDY)) {
1105                 dev_warn(&dd->pdev->dev,
1106                         "Clearing PxSERR.DIAG.n\n");
1107                 writel((1 << 16), dd->port->mmio + PORT_SCR_ERR);
1108         }
1109
1110         if (unlikely(port_stat & ~PORT_IRQ_HANDLED)) {
1111                 dev_warn(&dd->pdev->dev,
1112                         "Port stat errors %x unhandled\n",
1113                         (port_stat & ~PORT_IRQ_HANDLED));
1114         }
1115 }
1116
1117 static inline irqreturn_t mtip_handle_irq(struct driver_data *data)
1118 {
1119         struct driver_data *dd = (struct driver_data *) data;
1120         struct mtip_port *port = dd->port;
1121         u32 hba_stat, port_stat;
1122         int rv = IRQ_NONE;
1123         int do_irq_enable = 1, i, workers;
1124         struct mtip_work *twork;
1125
1126         hba_stat = readl(dd->mmio + HOST_IRQ_STAT);
1127         if (hba_stat) {
1128                 rv = IRQ_HANDLED;
1129
1130                 /* Acknowledge the interrupt status on the port.*/
1131                 port_stat = readl(port->mmio + PORT_IRQ_STAT);
1132                 writel(port_stat, port->mmio + PORT_IRQ_STAT);
1133
1134                 /* Demux port status */
1135                 if (likely(port_stat & PORT_IRQ_SDB_FIS)) {
1136                         do_irq_enable = 0;
1137                         WARN_ON_ONCE(atomic_read(&dd->irq_workers_active) != 0);
1138
1139                         /* Start at 1: group zero is always local? */
1140                         for (i = 0, workers = 0; i < MTIP_MAX_SLOT_GROUPS;
1141                                                                         i++) {
1142                                 twork = &dd->work[i];
1143                                 twork->completed = readl(port->completed[i]);
1144                                 if (twork->completed)
1145                                         workers++;
1146                         }
1147
1148                         atomic_set(&dd->irq_workers_active, workers);
1149                         if (workers) {
1150                                 for (i = 1; i < MTIP_MAX_SLOT_GROUPS; i++) {
1151                                         twork = &dd->work[i];
1152                                         if (twork->completed)
1153                                                 queue_work_on(
1154                                                         twork->cpu_binding,
1155                                                         dd->isr_workq,
1156                                                         &twork->work);
1157                                 }
1158
1159                                 if (likely(dd->work[0].completed))
1160                                         mtip_workq_sdbfx(port, 0,
1161                                                         dd->work[0].completed);
1162
1163                         } else {
1164                                 /*
1165                                  * Chip quirk: SDB interrupt but nothing
1166                                  * to complete
1167                                  */
1168                                 do_irq_enable = 1;
1169                         }
1170                 }
1171
1172                 if (unlikely(port_stat & PORT_IRQ_ERR)) {
1173                         if (unlikely(mtip_check_surprise_removal(dd->pdev))) {
1174                                 /* don't proceed further */
1175                                 return IRQ_HANDLED;
1176                         }
1177                         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
1178                                                         &dd->dd_flag))
1179                                 return rv;
1180
1181                         mtip_process_errors(dd, port_stat & PORT_IRQ_ERR);
1182                 }
1183
1184                 if (unlikely(port_stat & PORT_IRQ_LEGACY))
1185                         mtip_process_legacy(dd, port_stat & PORT_IRQ_LEGACY);
1186         }
1187
1188         /* acknowledge interrupt */
1189         if (unlikely(do_irq_enable))
1190                 writel(hba_stat, dd->mmio + HOST_IRQ_STAT);
1191
1192         return rv;
1193 }
1194
1195 /*
1196  * HBA interrupt subroutine.
1197  *
1198  * @irq         IRQ number.
1199  * @instance    Pointer to the driver data structure.
1200  *
1201  * return value
1202  *      IRQ_HANDLED     A HBA interrupt was pending and handled.
1203  *      IRQ_NONE        This interrupt was not for the HBA.
1204  */
1205 static irqreturn_t mtip_irq_handler(int irq, void *instance)
1206 {
1207         struct driver_data *dd = instance;
1208
1209         return mtip_handle_irq(dd);
1210 }
1211
1212 static void mtip_issue_non_ncq_command(struct mtip_port *port, int tag)
1213 {
1214         atomic_set(&port->commands[tag].active, 1);
1215         writel(1 << MTIP_TAG_BIT(tag),
1216                 port->cmd_issue[MTIP_TAG_INDEX(tag)]);
1217 }
1218
1219 static bool mtip_pause_ncq(struct mtip_port *port,
1220                                 struct host_to_dev_fis *fis)
1221 {
1222         struct host_to_dev_fis *reply;
1223         unsigned long task_file_data;
1224
1225         reply = port->rxfis + RX_FIS_D2H_REG;
1226         task_file_data = readl(port->mmio+PORT_TFDATA);
1227
1228         if (fis->command == ATA_CMD_SEC_ERASE_UNIT)
1229                 clear_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1230
1231         if ((task_file_data & 1))
1232                 return false;
1233
1234         if (fis->command == ATA_CMD_SEC_ERASE_PREP) {
1235                 set_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
1236                 set_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1237                 port->ic_pause_timer = jiffies;
1238                 return true;
1239         } else if ((fis->command == ATA_CMD_DOWNLOAD_MICRO) &&
1240                                         (fis->features == 0x03)) {
1241                 set_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
1242                 port->ic_pause_timer = jiffies;
1243                 return true;
1244         } else if ((fis->command == ATA_CMD_SEC_ERASE_UNIT) ||
1245                 ((fis->command == 0xFC) &&
1246                         (fis->features == 0x27 || fis->features == 0x72 ||
1247                          fis->features == 0x62 || fis->features == 0x26))) {
1248                 /* Com reset after secure erase or lowlevel format */
1249                 mtip_restart_port(port);
1250                 return false;
1251         }
1252
1253         return false;
1254 }
1255
1256 /*
1257  * Wait for port to quiesce
1258  *
1259  * @port    Pointer to port data structure
1260  * @timeout Max duration to wait (ms)
1261  *
1262  * return value
1263  *      0       Success
1264  *      -EBUSY  Commands still active
1265  */
1266 static int mtip_quiesce_io(struct mtip_port *port, unsigned long timeout)
1267 {
1268         unsigned long to;
1269         unsigned int n;
1270         unsigned int active = 1;
1271
1272         to = jiffies + msecs_to_jiffies(timeout);
1273         do {
1274                 if (test_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags) &&
1275                         test_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags)) {
1276                         msleep(20);
1277                         continue; /* svc thd is actively issuing commands */
1278                 }
1279                 if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
1280                         return -EFAULT;
1281                 /*
1282                  * Ignore s_active bit 0 of array element 0.
1283                  * This bit will always be set
1284                  */
1285                 active = readl(port->s_active[0]) & 0xFFFFFFFE;
1286                 for (n = 1; n < port->dd->slot_groups; n++)
1287                         active |= readl(port->s_active[n]);
1288
1289                 if (!active)
1290                         break;
1291
1292                 msleep(20);
1293         } while (time_before(jiffies, to));
1294
1295         return active ? -EBUSY : 0;
1296 }
1297
1298 /*
1299  * Execute an internal command and wait for the completion.
1300  *
1301  * @port    Pointer to the port data structure.
1302  * @fis     Pointer to the FIS that describes the command.
1303  * @fis_len  Length in WORDS of the FIS.
1304  * @buffer  DMA accessible for command data.
1305  * @buf_len  Length, in bytes, of the data buffer.
1306  * @opts    Command header options, excluding the FIS length
1307  *             and the number of PRD entries.
1308  * @timeout Time in ms to wait for the command to complete.
1309  *
1310  * return value
1311  *      0        Command completed successfully.
1312  *      -EFAULT  The buffer address is not correctly aligned.
1313  *      -EBUSY   Internal command or other IO in progress.
1314  *      -EAGAIN  Time out waiting for command to complete.
1315  */
1316 static int mtip_exec_internal_command(struct mtip_port *port,
1317                                         struct host_to_dev_fis *fis,
1318                                         int fis_len,
1319                                         dma_addr_t buffer,
1320                                         int buf_len,
1321                                         u32 opts,
1322                                         gfp_t atomic,
1323                                         unsigned long timeout)
1324 {
1325         struct mtip_cmd_sg *command_sg;
1326         DECLARE_COMPLETION_ONSTACK(wait);
1327         int rv = 0, ready2go = 1;
1328         struct mtip_cmd *int_cmd = &port->commands[MTIP_TAG_INTERNAL];
1329         unsigned long to;
1330         struct driver_data *dd = port->dd;
1331
1332         /* Make sure the buffer is 8 byte aligned. This is asic specific. */
1333         if (buffer & 0x00000007) {
1334                 dev_err(&dd->pdev->dev, "SG buffer is not 8 byte aligned\n");
1335                 return -EFAULT;
1336         }
1337
1338         to = jiffies + msecs_to_jiffies(timeout);
1339         do {
1340                 ready2go = !test_and_set_bit(MTIP_TAG_INTERNAL,
1341                                                 port->allocated);
1342                 if (ready2go)
1343                         break;
1344                 mdelay(100);
1345         } while (time_before(jiffies, to));
1346         if (!ready2go) {
1347                 dev_warn(&dd->pdev->dev,
1348                         "Internal cmd active. new cmd [%02X]\n", fis->command);
1349                 return -EBUSY;
1350         }
1351         set_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1352         port->ic_pause_timer = 0;
1353
1354         clear_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
1355         clear_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
1356
1357         if (atomic == GFP_KERNEL) {
1358                 if (fis->command != ATA_CMD_STANDBYNOW1) {
1359                         /* wait for io to complete if non atomic */
1360                         if (mtip_quiesce_io(port, 5000) < 0) {
1361                                 dev_warn(&dd->pdev->dev,
1362                                         "Failed to quiesce IO\n");
1363                                 release_slot(port, MTIP_TAG_INTERNAL);
1364                                 clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1365                                 wake_up_interruptible(&port->svc_wait);
1366                                 return -EBUSY;
1367                         }
1368                 }
1369
1370                 /* Set the completion function and data for the command. */
1371                 int_cmd->comp_data = &wait;
1372                 int_cmd->comp_func = mtip_completion;
1373
1374         } else {
1375                 /* Clear completion - we're going to poll */
1376                 int_cmd->comp_data = NULL;
1377                 int_cmd->comp_func = mtip_null_completion;
1378         }
1379
1380         /* Copy the command to the command table */
1381         memcpy(int_cmd->command, fis, fis_len*4);
1382
1383         /* Populate the SG list */
1384         int_cmd->command_header->opts =
1385                  __force_bit2int cpu_to_le32(opts | fis_len);
1386         if (buf_len) {
1387                 command_sg = int_cmd->command + AHCI_CMD_TBL_HDR_SZ;
1388
1389                 command_sg->info =
1390                         __force_bit2int cpu_to_le32((buf_len-1) & 0x3FFFFF);
1391                 command_sg->dba =
1392                         __force_bit2int cpu_to_le32(buffer & 0xFFFFFFFF);
1393                 command_sg->dba_upper =
1394                         __force_bit2int cpu_to_le32((buffer >> 16) >> 16);
1395
1396                 int_cmd->command_header->opts |=
1397                         __force_bit2int cpu_to_le32((1 << 16));
1398         }
1399
1400         /* Populate the command header */
1401         int_cmd->command_header->byte_count = 0;
1402
1403         /* Issue the command to the hardware */
1404         mtip_issue_non_ncq_command(port, MTIP_TAG_INTERNAL);
1405
1406         if (atomic == GFP_KERNEL) {
1407                 /* Wait for the command to complete or timeout. */
1408                 if (wait_for_completion_interruptible_timeout(
1409                                 &wait,
1410                                 msecs_to_jiffies(timeout)) <= 0) {
1411                         if (rv == -ERESTARTSYS) { /* interrupted */
1412                                 dev_err(&dd->pdev->dev,
1413                                         "Internal command [%02X] was interrupted after %lu ms\n",
1414                                         fis->command, timeout);
1415                                 rv = -EINTR;
1416                                 goto exec_ic_exit;
1417                         } else if (rv == 0) /* timeout */
1418                                 dev_err(&dd->pdev->dev,
1419                                         "Internal command did not complete [%02X] within timeout of  %lu ms\n",
1420                                         fis->command, timeout);
1421                         else
1422                                 dev_err(&dd->pdev->dev,
1423                                         "Internal command [%02X] wait returned code [%d] after %lu ms - unhandled\n",
1424                                         fis->command, rv, timeout);
1425
1426                         if (mtip_check_surprise_removal(dd->pdev) ||
1427                                 test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
1428                                                 &dd->dd_flag)) {
1429                                 dev_err(&dd->pdev->dev,
1430                                         "Internal command [%02X] wait returned due to SR\n",
1431                                         fis->command);
1432                                 rv = -ENXIO;
1433                                 goto exec_ic_exit;
1434                         }
1435                         mtip_device_reset(dd); /* recover from timeout issue */
1436                         rv = -EAGAIN;
1437                         goto exec_ic_exit;
1438                 }
1439         } else {
1440                 u32 hba_stat, port_stat;
1441
1442                 /* Spin for <timeout> checking if command still outstanding */
1443                 timeout = jiffies + msecs_to_jiffies(timeout);
1444                 while ((readl(port->cmd_issue[MTIP_TAG_INTERNAL])
1445                                 & (1 << MTIP_TAG_INTERNAL))
1446                                 && time_before(jiffies, timeout)) {
1447                         if (mtip_check_surprise_removal(dd->pdev)) {
1448                                 rv = -ENXIO;
1449                                 goto exec_ic_exit;
1450                         }
1451                         if ((fis->command != ATA_CMD_STANDBYNOW1) &&
1452                                 test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
1453                                                 &dd->dd_flag)) {
1454                                 rv = -ENXIO;
1455                                 goto exec_ic_exit;
1456                         }
1457                         port_stat = readl(port->mmio + PORT_IRQ_STAT);
1458                         if (!port_stat)
1459                                 continue;
1460
1461                         if (port_stat & PORT_IRQ_ERR) {
1462                                 dev_err(&dd->pdev->dev,
1463                                         "Internal command [%02X] failed\n",
1464                                         fis->command);
1465                                 mtip_device_reset(dd);
1466                                 rv = -EIO;
1467                                 goto exec_ic_exit;
1468                         } else {
1469                                 writel(port_stat, port->mmio + PORT_IRQ_STAT);
1470                                 hba_stat = readl(dd->mmio + HOST_IRQ_STAT);
1471                                 if (hba_stat)
1472                                         writel(hba_stat,
1473                                                 dd->mmio + HOST_IRQ_STAT);
1474                         }
1475                         break;
1476                 }
1477         }
1478
1479         if (readl(port->cmd_issue[MTIP_TAG_INTERNAL])
1480                         & (1 << MTIP_TAG_INTERNAL)) {
1481                 rv = -ENXIO;
1482                 if (!test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)) {
1483                         mtip_device_reset(dd);
1484                         rv = -EAGAIN;
1485                 }
1486         }
1487 exec_ic_exit:
1488         /* Clear the allocated and active bits for the internal command. */
1489         atomic_set(&int_cmd->active, 0);
1490         release_slot(port, MTIP_TAG_INTERNAL);
1491         if (rv >= 0 && mtip_pause_ncq(port, fis)) {
1492                 /* NCQ paused */
1493                 return rv;
1494         }
1495         clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1496         wake_up_interruptible(&port->svc_wait);
1497
1498         return rv;
1499 }
1500
1501 /*
1502  * Byte-swap ATA ID strings.
1503  *
1504  * ATA identify data contains strings in byte-swapped 16-bit words.
1505  * They must be swapped (on all architectures) to be usable as C strings.
1506  * This function swaps bytes in-place.
1507  *
1508  * @buf The buffer location of the string
1509  * @len The number of bytes to swap
1510  *
1511  * return value
1512  *      None
1513  */
1514 static inline void ata_swap_string(u16 *buf, unsigned int len)
1515 {
1516         int i;
1517         for (i = 0; i < (len/2); i++)
1518                 be16_to_cpus(&buf[i]);
1519 }
1520
1521 /*
1522  * Request the device identity information.
1523  *
1524  * If a user space buffer is not specified, i.e. is NULL, the
1525  * identify information is still read from the drive and placed
1526  * into the identify data buffer (@e port->identify) in the
1527  * port data structure.
1528  * When the identify buffer contains valid identify information @e
1529  * port->identify_valid is non-zero.
1530  *
1531  * @port         Pointer to the port structure.
1532  * @user_buffer  A user space buffer where the identify data should be
1533  *                    copied.
1534  *
1535  * return value
1536  *      0       Command completed successfully.
1537  *      -EFAULT An error occurred while coping data to the user buffer.
1538  *      -1      Command failed.
1539  */
1540 static int mtip_get_identify(struct mtip_port *port, void __user *user_buffer)
1541 {
1542         int rv = 0;
1543         struct host_to_dev_fis fis;
1544
1545         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
1546                 return -EFAULT;
1547
1548         /* Build the FIS. */
1549         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1550         fis.type        = 0x27;
1551         fis.opts        = 1 << 7;
1552         fis.command     = ATA_CMD_ID_ATA;
1553
1554         /* Set the identify information as invalid. */
1555         port->identify_valid = 0;
1556
1557         /* Clear the identify information. */
1558         memset(port->identify, 0, sizeof(u16) * ATA_ID_WORDS);
1559
1560         /* Execute the command. */
1561         if (mtip_exec_internal_command(port,
1562                                 &fis,
1563                                 5,
1564                                 port->identify_dma,
1565                                 sizeof(u16) * ATA_ID_WORDS,
1566                                 0,
1567                                 GFP_KERNEL,
1568                                 MTIP_INTERNAL_COMMAND_TIMEOUT_MS)
1569                                 < 0) {
1570                 rv = -1;
1571                 goto out;
1572         }
1573
1574         /*
1575          * Perform any necessary byte-swapping.  Yes, the kernel does in fact
1576          * perform field-sensitive swapping on the string fields.
1577          * See the kernel use of ata_id_string() for proof of this.
1578          */
1579 #ifdef __LITTLE_ENDIAN
1580         ata_swap_string(port->identify + 27, 40);  /* model string*/
1581         ata_swap_string(port->identify + 23, 8);   /* firmware string*/
1582         ata_swap_string(port->identify + 10, 20);  /* serial# string*/
1583 #else
1584         {
1585                 int i;
1586                 for (i = 0; i < ATA_ID_WORDS; i++)
1587                         port->identify[i] = le16_to_cpu(port->identify[i]);
1588         }
1589 #endif
1590
1591         /* Check security locked state */
1592         if (port->identify[128] & 0x4)
1593                 set_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1594         else
1595                 clear_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1596
1597 #ifdef MTIP_TRIM /* Disabling TRIM support temporarily */
1598         /* Demux ID.DRAT & ID.RZAT to determine trim support */
1599         if (port->identify[69] & (1 << 14) && port->identify[69] & (1 << 5))
1600                 port->dd->trim_supp = true;
1601         else
1602 #endif
1603                 port->dd->trim_supp = false;
1604
1605         /* Set the identify buffer as valid. */
1606         port->identify_valid = 1;
1607
1608         if (user_buffer) {
1609                 if (copy_to_user(
1610                         user_buffer,
1611                         port->identify,
1612                         ATA_ID_WORDS * sizeof(u16))) {
1613                         rv = -EFAULT;
1614                         goto out;
1615                 }
1616         }
1617
1618 out:
1619         return rv;
1620 }
1621
1622 /*
1623  * Issue a standby immediate command to the device.
1624  *
1625  * @port Pointer to the port structure.
1626  *
1627  * return value
1628  *      0       Command was executed successfully.
1629  *      -1      An error occurred while executing the command.
1630  */
1631 static int mtip_standby_immediate(struct mtip_port *port)
1632 {
1633         int rv;
1634         struct host_to_dev_fis  fis;
1635         unsigned long start;
1636
1637         /* Build the FIS. */
1638         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1639         fis.type        = 0x27;
1640         fis.opts        = 1 << 7;
1641         fis.command     = ATA_CMD_STANDBYNOW1;
1642
1643         start = jiffies;
1644         rv = mtip_exec_internal_command(port,
1645                                         &fis,
1646                                         5,
1647                                         0,
1648                                         0,
1649                                         0,
1650                                         GFP_ATOMIC,
1651                                         15000);
1652         dbg_printk(MTIP_DRV_NAME "Time taken to complete standby cmd: %d ms\n",
1653                         jiffies_to_msecs(jiffies - start));
1654         if (rv)
1655                 dev_warn(&port->dd->pdev->dev,
1656                         "STANDBY IMMEDIATE command failed.\n");
1657
1658         return rv;
1659 }
1660
1661 /*
1662  * Issue a READ LOG EXT command to the device.
1663  *
1664  * @port        pointer to the port structure.
1665  * @page        page number to fetch
1666  * @buffer      pointer to buffer
1667  * @buffer_dma  dma address corresponding to @buffer
1668  * @sectors     page length to fetch, in sectors
1669  *
1670  * return value
1671  *      @rv     return value from mtip_exec_internal_command()
1672  */
1673 static int mtip_read_log_page(struct mtip_port *port, u8 page, u16 *buffer,
1674                                 dma_addr_t buffer_dma, unsigned int sectors)
1675 {
1676         struct host_to_dev_fis fis;
1677
1678         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1679         fis.type        = 0x27;
1680         fis.opts        = 1 << 7;
1681         fis.command     = ATA_CMD_READ_LOG_EXT;
1682         fis.sect_count  = sectors & 0xFF;
1683         fis.sect_cnt_ex = (sectors >> 8) & 0xFF;
1684         fis.lba_low     = page;
1685         fis.lba_mid     = 0;
1686         fis.device      = ATA_DEVICE_OBS;
1687
1688         memset(buffer, 0, sectors * ATA_SECT_SIZE);
1689
1690         return mtip_exec_internal_command(port,
1691                                         &fis,
1692                                         5,
1693                                         buffer_dma,
1694                                         sectors * ATA_SECT_SIZE,
1695                                         0,
1696                                         GFP_ATOMIC,
1697                                         MTIP_INTERNAL_COMMAND_TIMEOUT_MS);
1698 }
1699
1700 /*
1701  * Issue a SMART READ DATA command to the device.
1702  *
1703  * @port        pointer to the port structure.
1704  * @buffer      pointer to buffer
1705  * @buffer_dma  dma address corresponding to @buffer
1706  *
1707  * return value
1708  *      @rv     return value from mtip_exec_internal_command()
1709  */
1710 static int mtip_get_smart_data(struct mtip_port *port, u8 *buffer,
1711                                         dma_addr_t buffer_dma)
1712 {
1713         struct host_to_dev_fis fis;
1714
1715         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1716         fis.type        = 0x27;
1717         fis.opts        = 1 << 7;
1718         fis.command     = ATA_CMD_SMART;
1719         fis.features    = 0xD0;
1720         fis.sect_count  = 1;
1721         fis.lba_mid     = 0x4F;
1722         fis.lba_hi      = 0xC2;
1723         fis.device      = ATA_DEVICE_OBS;
1724
1725         return mtip_exec_internal_command(port,
1726                                         &fis,
1727                                         5,
1728                                         buffer_dma,
1729                                         ATA_SECT_SIZE,
1730                                         0,
1731                                         GFP_ATOMIC,
1732                                         15000);
1733 }
1734
1735 /*
1736  * Get the value of a smart attribute
1737  *
1738  * @port        pointer to the port structure
1739  * @id          attribute number
1740  * @attrib      pointer to return attrib information corresponding to @id
1741  *
1742  * return value
1743  *      -EINVAL NULL buffer passed or unsupported attribute @id.
1744  *      -EPERM  Identify data not valid, SMART not supported or not enabled
1745  */
1746 static int mtip_get_smart_attr(struct mtip_port *port, unsigned int id,
1747                                                 struct smart_attr *attrib)
1748 {
1749         int rv, i;
1750         struct smart_attr *pattr;
1751
1752         if (!attrib)
1753                 return -EINVAL;
1754
1755         if (!port->identify_valid) {
1756                 dev_warn(&port->dd->pdev->dev, "IDENTIFY DATA not valid\n");
1757                 return -EPERM;
1758         }
1759         if (!(port->identify[82] & 0x1)) {
1760                 dev_warn(&port->dd->pdev->dev, "SMART not supported\n");
1761                 return -EPERM;
1762         }
1763         if (!(port->identify[85] & 0x1)) {
1764                 dev_warn(&port->dd->pdev->dev, "SMART not enabled\n");
1765                 return -EPERM;
1766         }
1767
1768         memset(port->smart_buf, 0, ATA_SECT_SIZE);
1769         rv = mtip_get_smart_data(port, port->smart_buf, port->smart_buf_dma);
1770         if (rv) {
1771                 dev_warn(&port->dd->pdev->dev, "Failed to ge SMART data\n");
1772                 return rv;
1773         }
1774
1775         pattr = (struct smart_attr *)(port->smart_buf + 2);
1776         for (i = 0; i < 29; i++, pattr++)
1777                 if (pattr->attr_id == id) {
1778                         memcpy(attrib, pattr, sizeof(struct smart_attr));
1779                         break;
1780                 }
1781
1782         if (i == 29) {
1783                 dev_warn(&port->dd->pdev->dev,
1784                         "Query for invalid SMART attribute ID\n");
1785                 rv = -EINVAL;
1786         }
1787
1788         return rv;
1789 }
1790
1791 /*
1792  * Trim unused sectors
1793  *
1794  * @dd          pointer to driver_data structure
1795  * @lba         starting lba
1796  * @len         # of 512b sectors to trim
1797  *
1798  * return value
1799  *      -ENOMEM         Out of dma memory
1800  *      -EINVAL         Invalid parameters passed in, trim not supported
1801  *      -EIO            Error submitting trim request to hw
1802  */
1803 static int mtip_send_trim(struct driver_data *dd, unsigned int lba,
1804                                 unsigned int len)
1805 {
1806         int i, rv = 0;
1807         u64 tlba, tlen, sect_left;
1808         struct mtip_trim_entry *buf;
1809         dma_addr_t dma_addr;
1810         struct host_to_dev_fis fis;
1811
1812         if (!len || dd->trim_supp == false)
1813                 return -EINVAL;
1814
1815         /* Trim request too big */
1816         WARN_ON(len > (MTIP_MAX_TRIM_ENTRY_LEN * MTIP_MAX_TRIM_ENTRIES));
1817
1818         /* Trim request not aligned on 4k boundary */
1819         WARN_ON(len % 8 != 0);
1820
1821         /* Warn if vu_trim structure is too big */
1822         WARN_ON(sizeof(struct mtip_trim) > ATA_SECT_SIZE);
1823
1824         /* Allocate a DMA buffer for the trim structure */
1825         buf = dmam_alloc_coherent(&dd->pdev->dev, ATA_SECT_SIZE, &dma_addr,
1826                                                                 GFP_KERNEL);
1827         if (!buf)
1828                 return -ENOMEM;
1829         memset(buf, 0, ATA_SECT_SIZE);
1830
1831         for (i = 0, sect_left = len, tlba = lba;
1832                         i < MTIP_MAX_TRIM_ENTRIES && sect_left;
1833                         i++) {
1834                 tlen = (sect_left >= MTIP_MAX_TRIM_ENTRY_LEN ?
1835                                         MTIP_MAX_TRIM_ENTRY_LEN :
1836                                         sect_left);
1837                 buf[i].lba = __force_bit2int cpu_to_le32(tlba);
1838                 buf[i].range = __force_bit2int cpu_to_le16(tlen);
1839                 tlba += tlen;
1840                 sect_left -= tlen;
1841         }
1842         WARN_ON(sect_left != 0);
1843
1844         /* Build the fis */
1845         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1846         fis.type       = 0x27;
1847         fis.opts       = 1 << 7;
1848         fis.command    = 0xfb;
1849         fis.features   = 0x60;
1850         fis.sect_count = 1;
1851         fis.device     = ATA_DEVICE_OBS;
1852
1853         if (mtip_exec_internal_command(dd->port,
1854                                         &fis,
1855                                         5,
1856                                         dma_addr,
1857                                         ATA_SECT_SIZE,
1858                                         0,
1859                                         GFP_KERNEL,
1860                                         MTIP_TRIM_TIMEOUT_MS) < 0)
1861                 rv = -EIO;
1862
1863         dmam_free_coherent(&dd->pdev->dev, ATA_SECT_SIZE, buf, dma_addr);
1864         return rv;
1865 }
1866
1867 /*
1868  * Get the drive capacity.
1869  *
1870  * @dd      Pointer to the device data structure.
1871  * @sectors Pointer to the variable that will receive the sector count.
1872  *
1873  * return value
1874  *      1 Capacity was returned successfully.
1875  *      0 The identify information is invalid.
1876  */
1877 static bool mtip_hw_get_capacity(struct driver_data *dd, sector_t *sectors)
1878 {
1879         struct mtip_port *port = dd->port;
1880         u64 total, raw0, raw1, raw2, raw3;
1881         raw0 = port->identify[100];
1882         raw1 = port->identify[101];
1883         raw2 = port->identify[102];
1884         raw3 = port->identify[103];
1885         total = raw0 | raw1<<16 | raw2<<32 | raw3<<48;
1886         *sectors = total;
1887         return (bool) !!port->identify_valid;
1888 }
1889
1890 /*
1891  * Display the identify command data.
1892  *
1893  * @port Pointer to the port data structure.
1894  *
1895  * return value
1896  *      None
1897  */
1898 static void mtip_dump_identify(struct mtip_port *port)
1899 {
1900         sector_t sectors;
1901         unsigned short revid;
1902         char cbuf[42];
1903
1904         if (!port->identify_valid)
1905                 return;
1906
1907         strlcpy(cbuf, (char *)(port->identify+10), 21);
1908         dev_info(&port->dd->pdev->dev,
1909                 "Serial No.: %s\n", cbuf);
1910
1911         strlcpy(cbuf, (char *)(port->identify+23), 9);
1912         dev_info(&port->dd->pdev->dev,
1913                 "Firmware Ver.: %s\n", cbuf);
1914
1915         strlcpy(cbuf, (char *)(port->identify+27), 41);
1916         dev_info(&port->dd->pdev->dev, "Model: %s\n", cbuf);
1917
1918         dev_info(&port->dd->pdev->dev, "Security: %04x %s\n",
1919                 port->identify[128],
1920                 port->identify[128] & 0x4 ? "(LOCKED)" : "");
1921
1922         if (mtip_hw_get_capacity(port->dd, &sectors))
1923                 dev_info(&port->dd->pdev->dev,
1924                         "Capacity: %llu sectors (%llu MB)\n",
1925                          (u64)sectors,
1926                          ((u64)sectors) * ATA_SECT_SIZE >> 20);
1927
1928         pci_read_config_word(port->dd->pdev, PCI_REVISION_ID, &revid);
1929         switch (revid & 0xFF) {
1930         case 0x1:
1931                 strlcpy(cbuf, "A0", 3);
1932                 break;
1933         case 0x3:
1934                 strlcpy(cbuf, "A2", 3);
1935                 break;
1936         default:
1937                 strlcpy(cbuf, "?", 2);
1938                 break;
1939         }
1940         dev_info(&port->dd->pdev->dev,
1941                 "Card Type: %s\n", cbuf);
1942 }
1943
1944 /*
1945  * Map the commands scatter list into the command table.
1946  *
1947  * @command Pointer to the command.
1948  * @nents Number of scatter list entries.
1949  *
1950  * return value
1951  *      None
1952  */
1953 static inline void fill_command_sg(struct driver_data *dd,
1954                                 struct mtip_cmd *command,
1955                                 int nents)
1956 {
1957         int n;
1958         unsigned int dma_len;
1959         struct mtip_cmd_sg *command_sg;
1960         struct scatterlist *sg = command->sg;
1961
1962         command_sg = command->command + AHCI_CMD_TBL_HDR_SZ;
1963
1964         for (n = 0; n < nents; n++) {
1965                 dma_len = sg_dma_len(sg);
1966                 if (dma_len > 0x400000)
1967                         dev_err(&dd->pdev->dev,
1968                                 "DMA segment length truncated\n");
1969                 command_sg->info = __force_bit2int
1970                         cpu_to_le32((dma_len-1) & 0x3FFFFF);
1971                 command_sg->dba = __force_bit2int
1972                         cpu_to_le32(sg_dma_address(sg));
1973                 command_sg->dba_upper = __force_bit2int
1974                         cpu_to_le32((sg_dma_address(sg) >> 16) >> 16);
1975                 command_sg++;
1976                 sg++;
1977         }
1978 }
1979
1980 /*
1981  * @brief Execute a drive command.
1982  *
1983  * return value 0 The command completed successfully.
1984  * return value -1 An error occurred while executing the command.
1985  */
1986 static int exec_drive_task(struct mtip_port *port, u8 *command)
1987 {
1988         struct host_to_dev_fis  fis;
1989         struct host_to_dev_fis *reply = (port->rxfis + RX_FIS_D2H_REG);
1990
1991         /* Build the FIS. */
1992         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1993         fis.type        = 0x27;
1994         fis.opts        = 1 << 7;
1995         fis.command     = command[0];
1996         fis.features    = command[1];
1997         fis.sect_count  = command[2];
1998         fis.sector      = command[3];
1999         fis.cyl_low     = command[4];
2000         fis.cyl_hi      = command[5];
2001         fis.device      = command[6] & ~0x10; /* Clear the dev bit*/
2002
2003         dbg_printk(MTIP_DRV_NAME " %s: User Command: cmd %x, feat %x, nsect %x, sect %x, lcyl %x, hcyl %x, sel %x\n",
2004                 __func__,
2005                 command[0],
2006                 command[1],
2007                 command[2],
2008                 command[3],
2009                 command[4],
2010                 command[5],
2011                 command[6]);
2012
2013         /* Execute the command. */
2014         if (mtip_exec_internal_command(port,
2015                                  &fis,
2016                                  5,
2017                                  0,
2018                                  0,
2019                                  0,
2020                                  GFP_KERNEL,
2021                                  MTIP_IOCTL_COMMAND_TIMEOUT_MS) < 0) {
2022                 return -1;
2023         }
2024
2025         command[0] = reply->command; /* Status*/
2026         command[1] = reply->features; /* Error*/
2027         command[4] = reply->cyl_low;
2028         command[5] = reply->cyl_hi;
2029
2030         dbg_printk(MTIP_DRV_NAME " %s: Completion Status: stat %x, err %x , cyl_lo %x cyl_hi %x\n",
2031                 __func__,
2032                 command[0],
2033                 command[1],
2034                 command[4],
2035                 command[5]);
2036
2037         return 0;
2038 }
2039
2040 /*
2041  * @brief Execute a drive command.
2042  *
2043  * @param port Pointer to the port data structure.
2044  * @param command Pointer to the user specified command parameters.
2045  * @param user_buffer Pointer to the user space buffer where read sector
2046  *                   data should be copied.
2047  *
2048  * return value 0 The command completed successfully.
2049  * return value -EFAULT An error occurred while copying the completion
2050  *                 data to the user space buffer.
2051  * return value -1 An error occurred while executing the command.
2052  */
2053 static int exec_drive_command(struct mtip_port *port, u8 *command,
2054                                 void __user *user_buffer)
2055 {
2056         struct host_to_dev_fis  fis;
2057         struct host_to_dev_fis *reply;
2058         u8 *buf = NULL;
2059         dma_addr_t dma_addr = 0;
2060         int rv = 0, xfer_sz = command[3];
2061
2062         if (xfer_sz) {
2063                 if (!user_buffer)
2064                         return -EFAULT;
2065
2066                 buf = dmam_alloc_coherent(&port->dd->pdev->dev,
2067                                 ATA_SECT_SIZE * xfer_sz,
2068                                 &dma_addr,
2069                                 GFP_KERNEL);
2070                 if (!buf) {
2071                         dev_err(&port->dd->pdev->dev,
2072                                 "Memory allocation failed (%d bytes)\n",
2073                                 ATA_SECT_SIZE * xfer_sz);
2074                         return -ENOMEM;
2075                 }
2076                 memset(buf, 0, ATA_SECT_SIZE * xfer_sz);
2077         }
2078
2079         /* Build the FIS. */
2080         memset(&fis, 0, sizeof(struct host_to_dev_fis));
2081         fis.type        = 0x27;
2082         fis.opts        = 1 << 7;
2083         fis.command     = command[0];
2084         fis.features    = command[2];
2085         fis.sect_count  = command[3];
2086         if (fis.command == ATA_CMD_SMART) {
2087                 fis.sector      = command[1];
2088                 fis.cyl_low     = 0x4F;
2089                 fis.cyl_hi      = 0xC2;
2090         }
2091
2092         if (xfer_sz)
2093                 reply = (port->rxfis + RX_FIS_PIO_SETUP);
2094         else
2095                 reply = (port->rxfis + RX_FIS_D2H_REG);
2096
2097         dbg_printk(MTIP_DRV_NAME
2098                 " %s: User Command: cmd %x, sect %x, "
2099                 "feat %x, sectcnt %x\n",
2100                 __func__,
2101                 command[0],
2102                 command[1],
2103                 command[2],
2104                 command[3]);
2105
2106         /* Execute the command. */
2107         if (mtip_exec_internal_command(port,
2108                                 &fis,
2109                                  5,
2110                                  (xfer_sz ? dma_addr : 0),
2111                                  (xfer_sz ? ATA_SECT_SIZE * xfer_sz : 0),
2112                                  0,
2113                                  GFP_KERNEL,
2114                                  MTIP_IOCTL_COMMAND_TIMEOUT_MS)
2115                                  < 0) {
2116                 rv = -EFAULT;
2117                 goto exit_drive_command;
2118         }
2119
2120         /* Collect the completion status. */
2121         command[0] = reply->command; /* Status*/
2122         command[1] = reply->features; /* Error*/
2123         command[2] = reply->sect_count;
2124
2125         dbg_printk(MTIP_DRV_NAME
2126                 " %s: Completion Status: stat %x, "
2127                 "err %x, nsect %x\n",
2128                 __func__,
2129                 command[0],
2130                 command[1],
2131                 command[2]);
2132
2133         if (xfer_sz) {
2134                 if (copy_to_user(user_buffer,
2135                                  buf,
2136                                  ATA_SECT_SIZE * command[3])) {
2137                         rv = -EFAULT;
2138                         goto exit_drive_command;
2139                 }
2140         }
2141 exit_drive_command:
2142         if (buf)
2143                 dmam_free_coherent(&port->dd->pdev->dev,
2144                                 ATA_SECT_SIZE * xfer_sz, buf, dma_addr);
2145         return rv;
2146 }
2147
2148 /*
2149  *  Indicates whether a command has a single sector payload.
2150  *
2151  *  @command passed to the device to perform the certain event.
2152  *  @features passed to the device to perform the certain event.
2153  *
2154  *  return value
2155  *      1       command is one that always has a single sector payload,
2156  *              regardless of the value in the Sector Count field.
2157  *      0       otherwise
2158  *
2159  */
2160 static unsigned int implicit_sector(unsigned char command,
2161                                     unsigned char features)
2162 {
2163         unsigned int rv = 0;
2164
2165         /* list of commands that have an implicit sector count of 1 */
2166         switch (command) {
2167         case ATA_CMD_SEC_SET_PASS:
2168         case ATA_CMD_SEC_UNLOCK:
2169         case ATA_CMD_SEC_ERASE_PREP:
2170         case ATA_CMD_SEC_ERASE_UNIT:
2171         case ATA_CMD_SEC_FREEZE_LOCK:
2172         case ATA_CMD_SEC_DISABLE_PASS:
2173         case ATA_CMD_PMP_READ:
2174         case ATA_CMD_PMP_WRITE:
2175                 rv = 1;
2176                 break;
2177         case ATA_CMD_SET_MAX:
2178                 if (features == ATA_SET_MAX_UNLOCK)
2179                         rv = 1;
2180                 break;
2181         case ATA_CMD_SMART:
2182                 if ((features == ATA_SMART_READ_VALUES) ||
2183                                 (features == ATA_SMART_READ_THRESHOLDS))
2184                         rv = 1;
2185                 break;
2186         case ATA_CMD_CONF_OVERLAY:
2187                 if ((features == ATA_DCO_IDENTIFY) ||
2188                                 (features == ATA_DCO_SET))
2189                         rv = 1;
2190                 break;
2191         }
2192         return rv;
2193 }
2194 static void mtip_set_timeout(struct driver_data *dd,
2195                                         struct host_to_dev_fis *fis,
2196                                         unsigned int *timeout, u8 erasemode)
2197 {
2198         switch (fis->command) {
2199         case ATA_CMD_DOWNLOAD_MICRO:
2200                 *timeout = 120000; /* 2 minutes */
2201                 break;
2202         case ATA_CMD_SEC_ERASE_UNIT:
2203         case 0xFC:
2204                 if (erasemode)
2205                         *timeout = ((*(dd->port->identify + 90) * 2) * 60000);
2206                 else
2207                         *timeout = ((*(dd->port->identify + 89) * 2) * 60000);
2208                 break;
2209         case ATA_CMD_STANDBYNOW1:
2210                 *timeout = 120000;  /* 2 minutes */
2211                 break;
2212         case 0xF7:
2213         case 0xFA:
2214                 *timeout = 60000;  /* 60 seconds */
2215                 break;
2216         case ATA_CMD_SMART:
2217                 *timeout = 15000;  /* 15 seconds */
2218                 break;
2219         default:
2220                 *timeout = MTIP_IOCTL_COMMAND_TIMEOUT_MS;
2221                 break;
2222         }
2223 }
2224
2225 /*
2226  * Executes a taskfile
2227  * See ide_taskfile_ioctl() for derivation
2228  */
2229 static int exec_drive_taskfile(struct driver_data *dd,
2230                                void __user *buf,
2231                                ide_task_request_t *req_task,
2232                                int outtotal)
2233 {
2234         struct host_to_dev_fis  fis;
2235         struct host_to_dev_fis *reply;
2236         u8 *outbuf = NULL;
2237         u8 *inbuf = NULL;
2238         dma_addr_t outbuf_dma = 0;
2239         dma_addr_t inbuf_dma = 0;
2240         dma_addr_t dma_buffer = 0;
2241         int err = 0;
2242         unsigned int taskin = 0;
2243         unsigned int taskout = 0;
2244         u8 nsect = 0;
2245         unsigned int timeout;
2246         unsigned int force_single_sector;
2247         unsigned int transfer_size;
2248         unsigned long task_file_data;
2249         int intotal = outtotal + req_task->out_size;
2250         int erasemode = 0;
2251
2252         taskout = req_task->out_size;
2253         taskin = req_task->in_size;
2254         /* 130560 = 512 * 0xFF*/
2255         if (taskin > 130560 || taskout > 130560) {
2256                 err = -EINVAL;
2257                 goto abort;
2258         }
2259
2260         if (taskout) {
2261                 outbuf = kzalloc(taskout, GFP_KERNEL);
2262                 if (outbuf == NULL) {
2263                         err = -ENOMEM;
2264                         goto abort;
2265                 }
2266                 if (copy_from_user(outbuf, buf + outtotal, taskout)) {
2267                         err = -EFAULT;
2268                         goto abort;
2269                 }
2270                 outbuf_dma = pci_map_single(dd->pdev,
2271                                          outbuf,
2272                                          taskout,
2273                                          DMA_TO_DEVICE);
2274                 if (outbuf_dma == 0) {
2275                         err = -ENOMEM;
2276                         goto abort;
2277                 }
2278                 dma_buffer = outbuf_dma;
2279         }
2280
2281         if (taskin) {
2282                 inbuf = kzalloc(taskin, GFP_KERNEL);
2283                 if (inbuf == NULL) {
2284                         err = -ENOMEM;
2285                         goto abort;
2286                 }
2287
2288                 if (copy_from_user(inbuf, buf + intotal, taskin)) {
2289                         err = -EFAULT;
2290                         goto abort;
2291                 }
2292                 inbuf_dma = pci_map_single(dd->pdev,
2293                                          inbuf,
2294                                          taskin, DMA_FROM_DEVICE);
2295                 if (inbuf_dma == 0) {
2296                         err = -ENOMEM;
2297                         goto abort;
2298                 }
2299                 dma_buffer = inbuf_dma;
2300         }
2301
2302         /* only supports PIO and non-data commands from this ioctl. */
2303         switch (req_task->data_phase) {
2304         case TASKFILE_OUT:
2305                 nsect = taskout / ATA_SECT_SIZE;
2306                 reply = (dd->port->rxfis + RX_FIS_PIO_SETUP);
2307                 break;
2308         case TASKFILE_IN:
2309                 reply = (dd->port->rxfis + RX_FIS_PIO_SETUP);
2310                 break;
2311         case TASKFILE_NO_DATA:
2312                 reply = (dd->port->rxfis + RX_FIS_D2H_REG);
2313                 break;
2314         default:
2315                 err = -EINVAL;
2316                 goto abort;
2317         }
2318
2319         /* Build the FIS. */
2320         memset(&fis, 0, sizeof(struct host_to_dev_fis));
2321
2322         fis.type        = 0x27;
2323         fis.opts        = 1 << 7;
2324         fis.command     = req_task->io_ports[7];
2325         fis.features    = req_task->io_ports[1];
2326         fis.sect_count  = req_task->io_ports[2];
2327         fis.lba_low     = req_task->io_ports[3];
2328         fis.lba_mid     = req_task->io_ports[4];
2329         fis.lba_hi      = req_task->io_ports[5];
2330          /* Clear the dev bit*/
2331         fis.device      = req_task->io_ports[6] & ~0x10;
2332
2333         if ((req_task->in_flags.all == 0) && (req_task->out_flags.all & 1)) {
2334                 req_task->in_flags.all  =
2335                         IDE_TASKFILE_STD_IN_FLAGS |
2336                         (IDE_HOB_STD_IN_FLAGS << 8);
2337                 fis.lba_low_ex          = req_task->hob_ports[3];
2338                 fis.lba_mid_ex          = req_task->hob_ports[4];
2339                 fis.lba_hi_ex           = req_task->hob_ports[5];
2340                 fis.features_ex         = req_task->hob_ports[1];
2341                 fis.sect_cnt_ex         = req_task->hob_ports[2];
2342
2343         } else {
2344                 req_task->in_flags.all = IDE_TASKFILE_STD_IN_FLAGS;
2345         }
2346
2347         force_single_sector = implicit_sector(fis.command, fis.features);
2348
2349         if ((taskin || taskout) && (!fis.sect_count)) {
2350                 if (nsect)
2351                         fis.sect_count = nsect;
2352                 else {
2353                         if (!force_single_sector) {
2354                                 dev_warn(&dd->pdev->dev,
2355                                         "data movement but "
2356                                         "sect_count is 0\n");
2357                                         err = -EINVAL;
2358                                         goto abort;
2359                         }
2360                 }
2361         }
2362
2363         dbg_printk(MTIP_DRV_NAME
2364                 " %s: cmd %x, feat %x, nsect %x,"
2365                 " sect/lbal %x, lcyl/lbam %x, hcyl/lbah %x,"
2366                 " head/dev %x\n",
2367                 __func__,
2368                 fis.command,
2369                 fis.features,
2370                 fis.sect_count,
2371                 fis.lba_low,
2372                 fis.lba_mid,
2373                 fis.lba_hi,
2374                 fis.device);
2375
2376         /* check for erase mode support during secure erase.*/
2377         if ((fis.command == ATA_CMD_SEC_ERASE_UNIT) && outbuf &&
2378                                         (outbuf[0] & MTIP_SEC_ERASE_MODE)) {
2379                 erasemode = 1;
2380         }
2381
2382         mtip_set_timeout(dd, &fis, &timeout, erasemode);
2383
2384         /* Determine the correct transfer size.*/
2385         if (force_single_sector)
2386                 transfer_size = ATA_SECT_SIZE;
2387         else
2388                 transfer_size = ATA_SECT_SIZE * fis.sect_count;
2389
2390         /* Execute the command.*/
2391         if (mtip_exec_internal_command(dd->port,
2392                                  &fis,
2393                                  5,
2394                                  dma_buffer,
2395                                  transfer_size,
2396                                  0,
2397                                  GFP_KERNEL,
2398                                  timeout) < 0) {
2399                 err = -EIO;
2400                 goto abort;
2401         }
2402
2403         task_file_data = readl(dd->port->mmio+PORT_TFDATA);
2404
2405         if ((req_task->data_phase == TASKFILE_IN) && !(task_file_data & 1)) {
2406                 reply = dd->port->rxfis + RX_FIS_PIO_SETUP;
2407                 req_task->io_ports[7] = reply->control;
2408         } else {
2409                 reply = dd->port->rxfis + RX_FIS_D2H_REG;
2410                 req_task->io_ports[7] = reply->command;
2411         }
2412
2413         /* reclaim the DMA buffers.*/
2414         if (inbuf_dma)
2415                 pci_unmap_single(dd->pdev, inbuf_dma,
2416                         taskin, DMA_FROM_DEVICE);
2417         if (outbuf_dma)
2418                 pci_unmap_single(dd->pdev, outbuf_dma,
2419                         taskout, DMA_TO_DEVICE);
2420         inbuf_dma  = 0;
2421         outbuf_dma = 0;
2422
2423         /* return the ATA registers to the caller.*/
2424         req_task->io_ports[1] = reply->features;
2425         req_task->io_ports[2] = reply->sect_count;
2426         req_task->io_ports[3] = reply->lba_low;
2427         req_task->io_ports[4] = reply->lba_mid;
2428         req_task->io_ports[5] = reply->lba_hi;
2429         req_task->io_ports[6] = reply->device;
2430
2431         if (req_task->out_flags.all & 1)  {
2432
2433                 req_task->hob_ports[3] = reply->lba_low_ex;
2434                 req_task->hob_ports[4] = reply->lba_mid_ex;
2435                 req_task->hob_ports[5] = reply->lba_hi_ex;
2436                 req_task->hob_ports[1] = reply->features_ex;
2437                 req_task->hob_ports[2] = reply->sect_cnt_ex;
2438         }
2439         dbg_printk(MTIP_DRV_NAME
2440                 " %s: Completion: stat %x,"
2441                 "err %x, sect_cnt %x, lbalo %x,"
2442                 "lbamid %x, lbahi %x, dev %x\n",
2443                 __func__,
2444                 req_task->io_ports[7],
2445                 req_task->io_ports[1],
2446                 req_task->io_ports[2],
2447                 req_task->io_ports[3],
2448                 req_task->io_ports[4],
2449                 req_task->io_ports[5],
2450                 req_task->io_ports[6]);
2451
2452         if (taskout) {
2453                 if (copy_to_user(buf + outtotal, outbuf, taskout)) {
2454                         err = -EFAULT;
2455                         goto abort;
2456                 }
2457         }
2458         if (taskin) {
2459                 if (copy_to_user(buf + intotal, inbuf, taskin)) {
2460                         err = -EFAULT;
2461                         goto abort;
2462                 }
2463         }
2464 abort:
2465         if (inbuf_dma)
2466                 pci_unmap_single(dd->pdev, inbuf_dma,
2467                                         taskin, DMA_FROM_DEVICE);
2468         if (outbuf_dma)
2469                 pci_unmap_single(dd->pdev, outbuf_dma,
2470                                         taskout, DMA_TO_DEVICE);
2471         kfree(outbuf);
2472         kfree(inbuf);
2473
2474         return err;
2475 }
2476
2477 /*
2478  * Handle IOCTL calls from the Block Layer.
2479  *
2480  * This function is called by the Block Layer when it receives an IOCTL
2481  * command that it does not understand. If the IOCTL command is not supported
2482  * this function returns -ENOTTY.
2483  *
2484  * @dd  Pointer to the driver data structure.
2485  * @cmd IOCTL command passed from the Block Layer.
2486  * @arg IOCTL argument passed from the Block Layer.
2487  *
2488  * return value
2489  *      0       The IOCTL completed successfully.
2490  *      -ENOTTY The specified command is not supported.
2491  *      -EFAULT An error occurred copying data to a user space buffer.
2492  *      -EIO    An error occurred while executing the command.
2493  */
2494 static int mtip_hw_ioctl(struct driver_data *dd, unsigned int cmd,
2495                          unsigned long arg)
2496 {
2497         switch (cmd) {
2498         case HDIO_GET_IDENTITY:
2499         {
2500                 if (copy_to_user((void __user *)arg, dd->port->identify,
2501                                                 sizeof(u16) * ATA_ID_WORDS))
2502                         return -EFAULT;
2503                 break;
2504         }
2505         case HDIO_DRIVE_CMD:
2506         {
2507                 u8 drive_command[4];
2508
2509                 /* Copy the user command info to our buffer. */
2510                 if (copy_from_user(drive_command,
2511                                          (void __user *) arg,
2512                                          sizeof(drive_command)))
2513                         return -EFAULT;
2514
2515                 /* Execute the drive command. */
2516                 if (exec_drive_command(dd->port,
2517                                          drive_command,
2518                                          (void __user *) (arg+4)))
2519                         return -EIO;
2520
2521                 /* Copy the status back to the users buffer. */
2522                 if (copy_to_user((void __user *) arg,
2523                                          drive_command,
2524                                          sizeof(drive_command)))
2525                         return -EFAULT;
2526
2527                 break;
2528         }
2529         case HDIO_DRIVE_TASK:
2530         {
2531                 u8 drive_command[7];
2532
2533                 /* Copy the user command info to our buffer. */
2534                 if (copy_from_user(drive_command,
2535                                          (void __user *) arg,
2536                                          sizeof(drive_command)))
2537                         return -EFAULT;
2538
2539                 /* Execute the drive command. */
2540                 if (exec_drive_task(dd->port, drive_command))
2541                         return -EIO;
2542
2543                 /* Copy the status back to the users buffer. */
2544                 if (copy_to_user((void __user *) arg,
2545                                          drive_command,
2546                                          sizeof(drive_command)))
2547                         return -EFAULT;
2548
2549                 break;
2550         }
2551         case HDIO_DRIVE_TASKFILE: {
2552                 ide_task_request_t req_task;
2553                 int ret, outtotal;
2554
2555                 if (copy_from_user(&req_task, (void __user *) arg,
2556                                         sizeof(req_task)))
2557                         return -EFAULT;
2558
2559                 outtotal = sizeof(req_task);
2560
2561                 ret = exec_drive_taskfile(dd, (void __user *) arg,
2562                                                 &req_task, outtotal);
2563
2564                 if (copy_to_user((void __user *) arg, &req_task,
2565                                                         sizeof(req_task)))
2566                         return -EFAULT;
2567
2568                 return ret;
2569         }
2570
2571         default:
2572                 return -EINVAL;
2573         }
2574         return 0;
2575 }
2576
2577 /*
2578  * Submit an IO to the hw
2579  *
2580  * This function is called by the block layer to issue an io
2581  * to the device. Upon completion, the callback function will
2582  * be called with the data parameter passed as the callback data.
2583  *
2584  * @dd       Pointer to the driver data structure.
2585  * @start    First sector to read.
2586  * @nsect    Number of sectors to read.
2587  * @nents    Number of entries in scatter list for the read command.
2588  * @tag      The tag of this read command.
2589  * @callback Pointer to the function that should be called
2590  *           when the read completes.
2591  * @data     Callback data passed to the callback function
2592  *           when the read completes.
2593  * @dir      Direction (read or write)
2594  *
2595  * return value
2596  *      None
2597  */
2598 static void mtip_hw_submit_io(struct driver_data *dd, sector_t sector,
2599                               int nsect, int nents, int tag, void *callback,
2600                               void *data, int dir, int unaligned)
2601 {
2602         struct host_to_dev_fis  *fis;
2603         struct mtip_port *port = dd->port;
2604         struct mtip_cmd *command = &port->commands[tag];
2605         int dma_dir = (dir == READ) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
2606         u64 start = sector;
2607
2608         /* Map the scatter list for DMA access */
2609         nents = dma_map_sg(&dd->pdev->dev, command->sg, nents, dma_dir);
2610
2611         command->scatter_ents = nents;
2612
2613         command->unaligned = unaligned;
2614         /*
2615          * The number of retries for this command before it is
2616          * reported as a failure to the upper layers.
2617          */
2618         command->retries = MTIP_MAX_RETRIES;
2619
2620         /* Fill out fis */
2621         fis = command->command;
2622         fis->type        = 0x27;
2623         fis->opts        = 1 << 7;
2624         fis->command     =
2625                 (dir == READ ? ATA_CMD_FPDMA_READ : ATA_CMD_FPDMA_WRITE);
2626         fis->lba_low     = start & 0xFF;
2627         fis->lba_mid     = (start >> 8) & 0xFF;
2628         fis->lba_hi      = (start >> 16) & 0xFF;
2629         fis->lba_low_ex  = (start >> 24) & 0xFF;
2630         fis->lba_mid_ex  = (start >> 32) & 0xFF;
2631         fis->lba_hi_ex   = (start >> 40) & 0xFF;
2632         fis->device      = 1 << 6;
2633         fis->features    = nsect & 0xFF;
2634         fis->features_ex = (nsect >> 8) & 0xFF;
2635         fis->sect_count  = ((tag << 3) | (tag >> 5));
2636         fis->sect_cnt_ex = 0;
2637         fis->control     = 0;
2638         fis->res2        = 0;
2639         fis->res3        = 0;
2640         fill_command_sg(dd, command, nents);
2641
2642         if (unaligned)
2643                 fis->device |= 1 << 7;
2644
2645         /* Populate the command header */
2646         command->command_header->opts =
2647                         __force_bit2int cpu_to_le32(
2648                                 (nents << 16) | 5 | AHCI_CMD_PREFETCH);
2649         command->command_header->byte_count = 0;
2650
2651         /*
2652          * Set the completion function and data for the command
2653          * within this layer.
2654          */
2655         command->comp_data = dd;
2656         command->comp_func = mtip_async_complete;
2657         command->direction = dma_dir;
2658
2659         /*
2660          * Set the completion function and data for the command passed
2661          * from the upper layer.
2662          */
2663         command->async_data = data;
2664         command->async_callback = callback;
2665
2666         /*
2667          * To prevent this command from being issued
2668          * if an internal command is in progress or error handling is active.
2669          */
2670         if (port->flags & MTIP_PF_PAUSE_IO) {
2671                 set_bit(tag, port->cmds_to_issue);
2672                 set_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags);
2673                 return;
2674         }
2675
2676         /* Issue the command to the hardware */
2677         mtip_issue_ncq_command(port, tag);
2678
2679         return;
2680 }
2681
2682 /*
2683  * Release a command slot.
2684  *
2685  * @dd  Pointer to the driver data structure.
2686  * @tag Slot tag
2687  *
2688  * return value
2689  *      None
2690  */
2691 static void mtip_hw_release_scatterlist(struct driver_data *dd, int tag,
2692                                                                 int unaligned)
2693 {
2694         struct semaphore *sem = unaligned ? &dd->port->cmd_slot_unal :
2695                                                         &dd->port->cmd_slot;
2696         release_slot(dd->port, tag);
2697         up(sem);
2698 }
2699
2700 /*
2701  * Obtain a command slot and return its associated scatter list.
2702  *
2703  * @dd  Pointer to the driver data structure.
2704  * @tag Pointer to an int that will receive the allocated command
2705  *            slot tag.
2706  *
2707  * return value
2708  *      Pointer to the scatter list for the allocated command slot
2709  *      or NULL if no command slots are available.
2710  */
2711 static struct scatterlist *mtip_hw_get_scatterlist(struct driver_data *dd,
2712                                                    int *tag, int unaligned)
2713 {
2714         struct semaphore *sem = unaligned ? &dd->port->cmd_slot_unal :
2715                                                         &dd->port->cmd_slot;
2716
2717         /*
2718          * It is possible that, even with this semaphore, a thread
2719          * may think that no command slots are available. Therefore, we
2720          * need to make an attempt to get_slot().
2721          */
2722         down(sem);
2723         *tag = get_slot(dd->port);
2724
2725         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))) {
2726                 up(sem);
2727                 return NULL;
2728         }
2729         if (unlikely(*tag < 0)) {
2730                 up(sem);
2731                 return NULL;
2732         }
2733
2734         return dd->port->commands[*tag].sg;
2735 }
2736
2737 /*
2738  * Sysfs status dump.
2739  *
2740  * @dev  Pointer to the device structure, passed by the kernrel.
2741  * @attr Pointer to the device_attribute structure passed by the kernel.
2742  * @buf  Pointer to the char buffer that will receive the stats info.
2743  *
2744  * return value
2745  *      The size, in bytes, of the data copied into buf.
2746  */
2747 static ssize_t mtip_hw_show_status(struct device *dev,
2748                                 struct device_attribute *attr,
2749                                 char *buf)
2750 {
2751         struct driver_data *dd = dev_to_disk(dev)->private_data;
2752         int size = 0;
2753
2754         if (test_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag))
2755                 size += sprintf(buf, "%s", "thermal_shutdown\n");
2756         else if (test_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag))
2757                 size += sprintf(buf, "%s", "write_protect\n");
2758         else
2759                 size += sprintf(buf, "%s", "online\n");
2760
2761         return size;
2762 }
2763
2764 static DEVICE_ATTR(status, S_IRUGO, mtip_hw_show_status, NULL);
2765
2766 /* debugsfs entries */
2767
2768 static ssize_t show_device_status(struct device_driver *drv, char *buf)
2769 {
2770         int size = 0;
2771         struct driver_data *dd, *tmp;
2772         unsigned long flags;
2773         char id_buf[42];
2774         u16 status = 0;
2775
2776         spin_lock_irqsave(&dev_lock, flags);
2777         size += sprintf(&buf[size], "Devices Present:\n");
2778         list_for_each_entry_safe(dd, tmp, &online_list, online_list) {
2779                 if (dd->pdev) {
2780                         if (dd->port &&
2781                             dd->port->identify &&
2782                             dd->port->identify_valid) {
2783                                 strlcpy(id_buf,
2784                                         (char *) (dd->port->identify + 10), 21);
2785                                 status = *(dd->port->identify + 141);
2786                         } else {
2787                                 memset(id_buf, 0, 42);
2788                                 status = 0;
2789                         }
2790
2791                         if (dd->port &&
2792                             test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags)) {
2793                                 size += sprintf(&buf[size],
2794                                         " device %s %s (ftl rebuild %d %%)\n",
2795                                         dev_name(&dd->pdev->dev),
2796                                         id_buf,
2797                                         status);
2798                         } else {
2799                                 size += sprintf(&buf[size],
2800                                         " device %s %s\n",
2801                                         dev_name(&dd->pdev->dev),
2802                                         id_buf);
2803                         }
2804                 }
2805         }
2806
2807         size += sprintf(&buf[size], "Devices Being Removed:\n");
2808         list_for_each_entry_safe(dd, tmp, &removing_list, remove_list) {
2809                 if (dd->pdev) {
2810                         if (dd->port &&
2811                             dd->port->identify &&
2812                             dd->port->identify_valid) {
2813                                 strlcpy(id_buf,
2814                                         (char *) (dd->port->identify+10), 21);
2815                                 status = *(dd->port->identify + 141);
2816                         } else {
2817                                 memset(id_buf, 0, 42);
2818                                 status = 0;
2819                         }
2820
2821                         if (dd->port &&
2822                             test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags)) {
2823                                 size += sprintf(&buf[size],
2824                                         " device %s %s (ftl rebuild %d %%)\n",
2825                                         dev_name(&dd->pdev->dev),
2826                                         id_buf,
2827                                         status);
2828                         } else {
2829                                 size += sprintf(&buf[size],
2830                                         " device %s %s\n",
2831                                         dev_name(&dd->pdev->dev),
2832                                         id_buf);
2833                         }
2834                 }
2835         }
2836         spin_unlock_irqrestore(&dev_lock, flags);
2837
2838         return size;
2839 }
2840
2841 static ssize_t mtip_hw_read_device_status(struct file *f, char __user *ubuf,
2842                                                 size_t len, loff_t *offset)
2843 {
2844         struct driver_data *dd =  (struct driver_data *)f->private_data;
2845         int size = *offset;
2846         char *buf;
2847         int rv = 0;
2848
2849         if (!len || *offset)
2850                 return 0;
2851
2852         buf = kzalloc(MTIP_DFS_MAX_BUF_SIZE, GFP_KERNEL);
2853         if (!buf) {
2854                 dev_err(&dd->pdev->dev,
2855                         "Memory allocation: status buffer\n");
2856                 return -ENOMEM;
2857         }
2858
2859         size += show_device_status(NULL, buf);
2860
2861         *offset = size <= len ? size : len;
2862         size = copy_to_user(ubuf, buf, *offset);
2863         if (size)
2864                 rv = -EFAULT;
2865
2866         kfree(buf);
2867         return rv ? rv : *offset;
2868 }
2869
2870 static ssize_t mtip_hw_read_registers(struct file *f, char __user *ubuf,
2871                                   size_t len, loff_t *offset)
2872 {
2873         struct driver_data *dd =  (struct driver_data *)f->private_data;
2874         char *buf;
2875         u32 group_allocated;
2876         int size = *offset;
2877         int n, rv = 0;
2878
2879         if (!len || size)
2880                 return 0;
2881
2882         buf = kzalloc(MTIP_DFS_MAX_BUF_SIZE, GFP_KERNEL);
2883         if (!buf) {
2884                 dev_err(&dd->pdev->dev,
2885                         "Memory allocation: register buffer\n");
2886                 return -ENOMEM;
2887         }
2888
2889         size += sprintf(&buf[size], "H/ S ACTive      : [ 0x");
2890
2891         for (n = dd->slot_groups-1; n >= 0; n--)
2892                 size += sprintf(&buf[size], "%08X ",
2893                                          readl(dd->port->s_active[n]));
2894
2895         size += sprintf(&buf[size], "]\n");
2896         size += sprintf(&buf[size], "H/ Command Issue : [ 0x");
2897
2898         for (n = dd->slot_groups-1; n >= 0; n--)
2899                 size += sprintf(&buf[size], "%08X ",
2900                                         readl(dd->port->cmd_issue[n]));
2901
2902         size += sprintf(&buf[size], "]\n");
2903         size += sprintf(&buf[size], "H/ Completed     : [ 0x");
2904
2905         for (n = dd->slot_groups-1; n >= 0; n--)
2906                 size += sprintf(&buf[size], "%08X ",
2907                                 readl(dd->port->completed[n]));
2908
2909         size += sprintf(&buf[size], "]\n");
2910         size += sprintf(&buf[size], "H/ PORT IRQ STAT : [ 0x%08X ]\n",
2911                                 readl(dd->port->mmio + PORT_IRQ_STAT));
2912         size += sprintf(&buf[size], "H/ HOST IRQ STAT : [ 0x%08X ]\n",
2913                                 readl(dd->mmio + HOST_IRQ_STAT));
2914         size += sprintf(&buf[size], "\n");
2915
2916         size += sprintf(&buf[size], "L/ Allocated     : [ 0x");
2917
2918         for (n = dd->slot_groups-1; n >= 0; n--) {
2919                 if (sizeof(long) > sizeof(u32))
2920                         group_allocated =
2921                                 dd->port->allocated[n/2] >> (32*(n&1));
2922                 else
2923                         group_allocated = dd->port->allocated[n];
2924                 size += sprintf(&buf[size], "%08X ", group_allocated);
2925         }
2926         size += sprintf(&buf[size], "]\n");
2927
2928         size += sprintf(&buf[size], "L/ Commands in Q : [ 0x");
2929
2930         for (n = dd->slot_groups-1; n >= 0; n--) {
2931                 if (sizeof(long) > sizeof(u32))
2932                         group_allocated =
2933                                 dd->port->cmds_to_issue[n/2] >> (32*(n&1));
2934                 else
2935                         group_allocated = dd->port->cmds_to_issue[n];
2936                 size += sprintf(&buf[size], "%08X ", group_allocated);
2937         }
2938         size += sprintf(&buf[size], "]\n");
2939
2940         *offset = size <= len ? size : len;
2941         size = copy_to_user(ubuf, buf, *offset);
2942         if (size)
2943                 rv = -EFAULT;
2944
2945         kfree(buf);
2946         return rv ? rv : *offset;
2947 }
2948
2949 static ssize_t mtip_hw_read_flags(struct file *f, char __user *ubuf,
2950                                   size_t len, loff_t *offset)
2951 {
2952         struct driver_data *dd =  (struct driver_data *)f->private_data;
2953         char *buf;
2954         int size = *offset;
2955         int rv = 0;
2956
2957         if (!len || size)
2958                 return 0;
2959
2960         buf = kzalloc(MTIP_DFS_MAX_BUF_SIZE, GFP_KERNEL);
2961         if (!buf) {
2962                 dev_err(&dd->pdev->dev,
2963                         "Memory allocation: flag buffer\n");
2964                 return -ENOMEM;
2965         }
2966
2967         size += sprintf(&buf[size], "Flag-port : [ %08lX ]\n",
2968                                                         dd->port->flags);
2969         size += sprintf(&buf[size], "Flag-dd   : [ %08lX ]\n",
2970                                                         dd->dd_flag);
2971
2972         *offset = size <= len ? size : len;
2973         size = copy_to_user(ubuf, buf, *offset);
2974         if (size)
2975                 rv = -EFAULT;
2976
2977         kfree(buf);
2978         return rv ? rv : *offset;
2979 }
2980
2981 static const struct file_operations mtip_device_status_fops = {
2982         .owner  = THIS_MODULE,
2983         .open   = simple_open,
2984         .read   = mtip_hw_read_device_status,
2985         .llseek = no_llseek,
2986 };
2987
2988 static const struct file_operations mtip_regs_fops = {
2989         .owner  = THIS_MODULE,
2990         .open   = simple_open,
2991         .read   = mtip_hw_read_registers,
2992         .llseek = no_llseek,
2993 };
2994
2995 static const struct file_operations mtip_flags_fops = {
2996         .owner  = THIS_MODULE,
2997         .open   = simple_open,
2998         .read   = mtip_hw_read_flags,
2999         .llseek = no_llseek,
3000 };
3001
3002 /*
3003  * Create the sysfs related attributes.
3004  *
3005  * @dd   Pointer to the driver data structure.
3006  * @kobj Pointer to the kobj for the block device.
3007  *
3008  * return value
3009  *      0       Operation completed successfully.
3010  *      -EINVAL Invalid parameter.
3011  */
3012 static int mtip_hw_sysfs_init(struct driver_data *dd, struct kobject *kobj)
3013 {
3014         if (!kobj || !dd)
3015                 return -EINVAL;
3016
3017         if (sysfs_create_file(kobj, &dev_attr_status.attr))
3018                 dev_warn(&dd->pdev->dev,
3019                         "Error creating 'status' sysfs entry\n");
3020         return 0;
3021 }
3022
3023 /*
3024  * Remove the sysfs related attributes.
3025  *
3026  * @dd   Pointer to the driver data structure.
3027  * @kobj Pointer to the kobj for the block device.
3028  *
3029  * return value
3030  *      0       Operation completed successfully.
3031  *      -EINVAL Invalid parameter.
3032  */
3033 static int mtip_hw_sysfs_exit(struct driver_data *dd, struct kobject *kobj)
3034 {
3035         if (!kobj || !dd)
3036                 return -EINVAL;
3037
3038         sysfs_remove_file(kobj, &dev_attr_status.attr);
3039
3040         return 0;
3041 }
3042
3043 static int mtip_hw_debugfs_init(struct driver_data *dd)
3044 {
3045         if (!dfs_parent)
3046                 return -1;
3047
3048         dd->dfs_node = debugfs_create_dir(dd->disk->disk_name, dfs_parent);
3049         if (IS_ERR_OR_NULL(dd->dfs_node)) {
3050                 dev_warn(&dd->pdev->dev,
3051                         "Error creating node %s under debugfs\n",
3052                                                 dd->disk->disk_name);
3053                 dd->dfs_node = NULL;
3054                 return -1;
3055         }
3056
3057         debugfs_create_file("flags", S_IRUGO, dd->dfs_node, dd,
3058                                                         &mtip_flags_fops);
3059         debugfs_create_file("registers", S_IRUGO, dd->dfs_node, dd,
3060                                                         &mtip_regs_fops);
3061
3062         return 0;
3063 }
3064
3065 static void mtip_hw_debugfs_exit(struct driver_data *dd)
3066 {
3067         if (dd->dfs_node)
3068                 debugfs_remove_recursive(dd->dfs_node);
3069 }
3070
3071 static int mtip_free_orphan(struct driver_data *dd)
3072 {
3073         struct kobject *kobj;
3074
3075         if (dd->bdev) {
3076                 if (dd->bdev->bd_holders >= 1)
3077                         return -2;
3078
3079                 bdput(dd->bdev);
3080                 dd->bdev = NULL;
3081         }
3082
3083         mtip_hw_debugfs_exit(dd);
3084
3085         spin_lock(&rssd_index_lock);
3086         ida_remove(&rssd_index_ida, dd->index);
3087         spin_unlock(&rssd_index_lock);
3088
3089         if (!test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag) &&
3090                         test_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag)) {
3091                 put_disk(dd->disk);
3092         } else {
3093                 if (dd->disk) {
3094                         kobj = kobject_get(&disk_to_dev(dd->disk)->kobj);
3095                         if (kobj) {
3096                                 mtip_hw_sysfs_exit(dd, kobj);
3097                                 kobject_put(kobj);
3098                         }
3099                         del_gendisk(dd->disk);
3100                         dd->disk = NULL;
3101                 }
3102                 if (dd->queue) {
3103                         dd->queue->queuedata = NULL;
3104                         blk_cleanup_queue(dd->queue);
3105                         dd->queue = NULL;
3106                 }
3107         }
3108         kfree(dd);
3109         return 0;
3110 }
3111
3112 /*
3113  * Perform any init/resume time hardware setup
3114  *
3115  * @dd Pointer to the driver data structure.
3116  *
3117  * return value
3118  *      None
3119  */
3120 static inline void hba_setup(struct driver_data *dd)
3121 {
3122         u32 hwdata;
3123         hwdata = readl(dd->mmio + HOST_HSORG);
3124
3125         /* interrupt bug workaround: use only 1 IS bit.*/
3126         writel(hwdata |
3127                 HSORG_DISABLE_SLOTGRP_INTR |
3128                 HSORG_DISABLE_SLOTGRP_PXIS,
3129                 dd->mmio + HOST_HSORG);
3130 }
3131
3132 static int mtip_device_unaligned_constrained(struct driver_data *dd)
3133 {
3134         return (dd->pdev->device == P420M_DEVICE_ID ? 1 : 0);
3135 }
3136
3137 /*
3138  * Detect the details of the product, and store anything needed
3139  * into the driver data structure.  This includes product type and
3140  * version and number of slot groups.
3141  *
3142  * @dd Pointer to the driver data structure.
3143  *
3144  * return value
3145  *      None
3146  */
3147 static void mtip_detect_product(struct driver_data *dd)
3148 {
3149         u32 hwdata;
3150         unsigned int rev, slotgroups;
3151
3152         /*
3153          * HBA base + 0xFC [15:0] - vendor-specific hardware interface
3154          * info register:
3155          * [15:8] hardware/software interface rev#
3156          * [   3] asic-style interface
3157          * [ 2:0] number of slot groups, minus 1 (only valid for asic-style).
3158          */
3159         hwdata = readl(dd->mmio + HOST_HSORG);
3160
3161         dd->product_type = MTIP_PRODUCT_UNKNOWN;
3162         dd->slot_groups = 1;
3163
3164         if (hwdata & 0x8) {
3165                 dd->product_type = MTIP_PRODUCT_ASICFPGA;
3166                 rev = (hwdata & HSORG_HWREV) >> 8;
3167                 slotgroups = (hwdata & HSORG_SLOTGROUPS) + 1;
3168                 dev_info(&dd->pdev->dev,
3169                         "ASIC-FPGA design, HS rev 0x%x, "
3170                         "%i slot groups [%i slots]\n",
3171                          rev,
3172                          slotgroups,
3173                          slotgroups * 32);
3174
3175                 if (slotgroups > MTIP_MAX_SLOT_GROUPS) {
3176                         dev_warn(&dd->pdev->dev,
3177                                 "Warning: driver only supports "
3178                                 "%i slot groups.\n", MTIP_MAX_SLOT_GROUPS);
3179                         slotgroups = MTIP_MAX_SLOT_GROUPS;
3180                 }
3181                 dd->slot_groups = slotgroups;
3182                 return;
3183         }
3184
3185         dev_warn(&dd->pdev->dev, "Unrecognized product id\n");
3186 }
3187
3188 /*
3189  * Blocking wait for FTL rebuild to complete
3190  *
3191  * @dd Pointer to the DRIVER_DATA structure.
3192  *
3193  * return value
3194  *      0       FTL rebuild completed successfully
3195  *      -EFAULT FTL rebuild error/timeout/interruption
3196  */
3197 static int mtip_ftl_rebuild_poll(struct driver_data *dd)
3198 {
3199         unsigned long timeout, cnt = 0, start;
3200
3201         dev_warn(&dd->pdev->dev,
3202                 "FTL rebuild in progress. Polling for completion.\n");
3203
3204         start = jiffies;
3205         timeout = jiffies + msecs_to_jiffies(MTIP_FTL_REBUILD_TIMEOUT_MS);
3206
3207         do {
3208                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
3209                                 &dd->dd_flag)))
3210                         return -EFAULT;
3211                 if (mtip_check_surprise_removal(dd->pdev))
3212                         return -EFAULT;
3213
3214                 if (mtip_get_identify(dd->port, NULL) < 0)
3215                         return -EFAULT;
3216
3217                 if (*(dd->port->identify + MTIP_FTL_REBUILD_OFFSET) ==
3218                         MTIP_FTL_REBUILD_MAGIC) {
3219                         ssleep(1);
3220                         /* Print message every 3 minutes */
3221                         if (cnt++ >= 180) {
3222                                 dev_warn(&dd->pdev->dev,
3223                                 "FTL rebuild in progress (%d secs).\n",
3224                                 jiffies_to_msecs(jiffies - start) / 1000);
3225                                 cnt = 0;
3226                         }
3227                 } else {
3228                         dev_warn(&dd->pdev->dev,
3229                                 "FTL rebuild complete (%d secs).\n",
3230                         jiffies_to_msecs(jiffies - start) / 1000);
3231                         mtip_block_initialize(dd);
3232                         return 0;
3233                 }
3234                 ssleep(10);
3235         } while (time_before(jiffies, timeout));
3236
3237         /* Check for timeout */
3238         dev_err(&dd->pdev->dev,
3239                 "Timed out waiting for FTL rebuild to complete (%d secs).\n",
3240                 jiffies_to_msecs(jiffies - start) / 1000);
3241         return -EFAULT;
3242 }
3243
3244 /*
3245  * service thread to issue queued commands
3246  *
3247  * @data Pointer to the driver data structure.
3248  *
3249  * return value
3250  *      0
3251  */
3252
3253 static int mtip_service_thread(void *data)
3254 {
3255         struct driver_data *dd = (struct driver_data *)data;
3256         unsigned long slot, slot_start, slot_wrap;
3257         unsigned int num_cmd_slots = dd->slot_groups * 32;
3258         struct mtip_port *port = dd->port;
3259         int ret;
3260
3261         while (1) {
3262                 /*
3263                  * the condition is to check neither an internal command is
3264                  * is in progress nor error handling is active
3265                  */
3266                 wait_event_interruptible(port->svc_wait, (port->flags) &&
3267                         !(port->flags & MTIP_PF_PAUSE_IO));
3268
3269                 if (kthread_should_stop())
3270                         goto st_out;
3271
3272                 set_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags);
3273
3274                 /* If I am an orphan, start self cleanup */
3275                 if (test_bit(MTIP_PF_SR_CLEANUP_BIT, &port->flags))
3276                         break;
3277
3278                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
3279                                 &dd->dd_flag)))
3280                         goto st_out;
3281
3282                 if (test_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags)) {
3283                         slot = 1;
3284                         /* used to restrict the loop to one iteration */
3285                         slot_start = num_cmd_slots;
3286                         slot_wrap = 0;
3287                         while (1) {
3288                                 slot = find_next_bit(port->cmds_to_issue,
3289                                                 num_cmd_slots, slot);
3290                                 if (slot_wrap == 1) {
3291                                         if ((slot_start >= slot) ||
3292                                                 (slot >= num_cmd_slots))
3293                                                 break;
3294                                 }
3295                                 if (unlikely(slot_start == num_cmd_slots))
3296                                         slot_start = slot;
3297
3298                                 if (unlikely(slot == num_cmd_slots)) {
3299                                         slot = 1;
3300                                         slot_wrap = 1;
3301                                         continue;
3302                                 }
3303
3304                                 /* Issue the command to the hardware */
3305                                 mtip_issue_ncq_command(port, slot);
3306
3307                                 clear_bit(slot, port->cmds_to_issue);
3308                         }
3309
3310                         clear_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags);
3311                 } else if (test_bit(MTIP_PF_REBUILD_BIT, &port->flags)) {
3312                         if (mtip_ftl_rebuild_poll(dd) < 0)
3313                                 set_bit(MTIP_DDF_REBUILD_FAILED_BIT,
3314                                                         &dd->dd_flag);
3315                         clear_bit(MTIP_PF_REBUILD_BIT, &port->flags);
3316                 }
3317                 clear_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags);
3318
3319                 if (test_bit(MTIP_PF_SVC_THD_STOP_BIT, &port->flags))
3320                         goto st_out;
3321         }
3322
3323         /* wait for pci remove to exit */
3324         while (1) {
3325                 if (test_bit(MTIP_DDF_REMOVE_DONE_BIT, &dd->dd_flag))
3326                         break;
3327                 msleep_interruptible(1000);
3328                 if (kthread_should_stop())
3329                         goto st_out;
3330         }
3331
3332         while (1) {
3333                 ret = mtip_free_orphan(dd);
3334                 if (!ret) {
3335                         /* NOTE: All data structures are invalid, do not
3336                          * access any here */
3337                         return 0;
3338                 }
3339                 msleep_interruptible(1000);
3340                 if (kthread_should_stop())
3341                         goto st_out;
3342         }
3343 st_out:
3344         return 0;
3345 }
3346
3347 /*
3348  * DMA region teardown
3349  *
3350  * @dd Pointer to driver_data structure
3351  *
3352  * return value
3353  *      None
3354  */
3355 static void mtip_dma_free(struct driver_data *dd)
3356 {
3357         int i;
3358         struct mtip_port *port = dd->port;
3359
3360         if (port->block1)
3361                 dmam_free_coherent(&dd->pdev->dev, BLOCK_DMA_ALLOC_SZ,
3362                                         port->block1, port->block1_dma);
3363
3364         if (port->command_list) {
3365                 dmam_free_coherent(&dd->pdev->dev, AHCI_CMD_TBL_SZ,
3366                                 port->command_list, port->command_list_dma);
3367         }
3368
3369         for (i = 0; i < MTIP_MAX_COMMAND_SLOTS; i++) {
3370                 if (port->commands[i].command)
3371                         dmam_free_coherent(&dd->pdev->dev, CMD_DMA_ALLOC_SZ,
3372                                 port->commands[i].command,
3373                                 port->commands[i].command_dma);
3374         }
3375 }
3376
3377 /*
3378  * DMA region setup
3379  *
3380  * @dd Pointer to driver_data structure
3381  *
3382  * return value
3383  *      -ENOMEM Not enough free DMA region space to initialize driver
3384  */
3385 static int mtip_dma_alloc(struct driver_data *dd)
3386 {
3387         struct mtip_port *port = dd->port;
3388         int i, rv = 0;
3389         u32 host_cap_64 = readl(dd->mmio + HOST_CAP) & HOST_CAP_64;
3390
3391         /* Allocate dma memory for RX Fis, Identify, and Sector Bufffer */
3392         port->block1 =
3393                 dmam_alloc_coherent(&dd->pdev->dev, BLOCK_DMA_ALLOC_SZ,
3394                                         &port->block1_dma, GFP_KERNEL);
3395         if (!port->block1)
3396                 return -ENOMEM;
3397         memset(port->block1, 0, BLOCK_DMA_ALLOC_SZ);
3398
3399         /* Allocate dma memory for command list */
3400         port->command_list =
3401                 dmam_alloc_coherent(&dd->pdev->dev, AHCI_CMD_TBL_SZ,
3402                                         &port->command_list_dma, GFP_KERNEL);
3403         if (!port->command_list) {
3404                 dmam_free_coherent(&dd->pdev->dev, BLOCK_DMA_ALLOC_SZ,
3405                                         port->block1, port->block1_dma);
3406                 port->block1 = NULL;
3407                 port->block1_dma = 0;
3408                 return -ENOMEM;
3409         }
3410         memset(port->command_list, 0, AHCI_CMD_TBL_SZ);
3411
3412         /* Setup all pointers into first DMA region */
3413         port->rxfis         = port->block1 + AHCI_RX_FIS_OFFSET;
3414         port->rxfis_dma     = port->block1_dma + AHCI_RX_FIS_OFFSET;
3415         port->identify      = port->block1 + AHCI_IDFY_OFFSET;
3416         port->identify_dma  = port->block1_dma + AHCI_IDFY_OFFSET;
3417         port->log_buf       = port->block1 + AHCI_SECTBUF_OFFSET;
3418         port->log_buf_dma   = port->block1_dma + AHCI_SECTBUF_OFFSET;
3419         port->smart_buf     = port->block1 + AHCI_SMARTBUF_OFFSET;
3420         port->smart_buf_dma = port->block1_dma + AHCI_SMARTBUF_OFFSET;
3421
3422         /* Setup per command SGL DMA region */
3423
3424         /* Point the command headers at the command tables */
3425         for (i = 0; i < MTIP_MAX_COMMAND_SLOTS; i++) {
3426                 port->commands[i].command =
3427                         dmam_alloc_coherent(&dd->pdev->dev, CMD_DMA_ALLOC_SZ,
3428                                 &port->commands[i].command_dma, GFP_KERNEL);
3429                 if (!port->commands[i].command) {
3430                         rv = -ENOMEM;
3431                         mtip_dma_free(dd);
3432                         return rv;
3433                 }
3434                 memset(port->commands[i].command, 0, CMD_DMA_ALLOC_SZ);
3435
3436                 port->commands[i].command_header = port->command_list +
3437                                         (sizeof(struct mtip_cmd_hdr) * i);
3438                 port->commands[i].command_header_dma =
3439                                         dd->port->command_list_dma +
3440                                         (sizeof(struct mtip_cmd_hdr) * i);
3441
3442                 if (host_cap_64)
3443                         port->commands[i].command_header->ctbau =
3444                                 __force_bit2int cpu_to_le32(
3445                                 (port->commands[i].command_dma >> 16) >> 16);
3446
3447                 port->commands[i].command_header->ctba =
3448                                 __force_bit2int cpu_to_le32(
3449                                 port->commands[i].command_dma & 0xFFFFFFFF);
3450
3451                 sg_init_table(port->commands[i].sg, MTIP_MAX_SG);
3452
3453                 /* Mark command as currently inactive */
3454                 atomic_set(&dd->port->commands[i].active, 0);
3455         }
3456         return 0;
3457 }
3458
3459 /*
3460  * Called once for each card.
3461  *
3462  * @dd Pointer to the driver data structure.
3463  *
3464  * return value
3465  *      0 on success, else an error code.
3466  */
3467 static int mtip_hw_init(struct driver_data *dd)
3468 {
3469         int i;
3470         int rv;
3471         unsigned int num_command_slots;
3472         unsigned long timeout, timetaken;
3473         unsigned char *buf;
3474         struct smart_attr attr242;
3475
3476         dd->mmio = pcim_iomap_table(dd->pdev)[MTIP_ABAR];
3477
3478         mtip_detect_product(dd);
3479         if (dd->product_type == MTIP_PRODUCT_UNKNOWN) {
3480                 rv = -EIO;
3481                 goto out1;
3482         }
3483         num_command_slots = dd->slot_groups * 32;
3484
3485         hba_setup(dd);
3486
3487         dd->port = kzalloc_node(sizeof(struct mtip_port), GFP_KERNEL,
3488                                 dd->numa_node);
3489         if (!dd->port) {
3490                 dev_err(&dd->pdev->dev,
3491                         "Memory allocation: port structure\n");
3492                 return -ENOMEM;
3493         }
3494
3495         /* Continue workqueue setup */
3496         for (i = 0; i < MTIP_MAX_SLOT_GROUPS; i++)
3497                 dd->work[i].port = dd->port;
3498
3499         /* Enable unaligned IO constraints for some devices */
3500         if (mtip_device_unaligned_constrained(dd))
3501                 dd->unal_qdepth = MTIP_MAX_UNALIGNED_SLOTS;
3502         else
3503                 dd->unal_qdepth = 0;
3504
3505         /* Counting semaphore to track command slot usage */
3506         sema_init(&dd->port->cmd_slot, num_command_slots - 1 - dd->unal_qdepth);
3507         sema_init(&dd->port->cmd_slot_unal, dd->unal_qdepth);
3508
3509         /* Spinlock to prevent concurrent issue */
3510         for (i = 0; i < MTIP_MAX_SLOT_GROUPS; i++)
3511                 spin_lock_init(&dd->port->cmd_issue_lock[i]);
3512
3513         /* Set the port mmio base address. */
3514         dd->port->mmio  = dd->mmio + PORT_OFFSET;
3515         dd->port->dd    = dd;
3516
3517         /* DMA allocations */
3518         rv = mtip_dma_alloc(dd);
3519         if (rv < 0)
3520                 goto out1;
3521
3522         /* Setup the pointers to the extended s_active and CI registers. */
3523         for (i = 0; i < dd->slot_groups; i++) {
3524                 dd->port->s_active[i] =
3525                         dd->port->mmio + i*0x80 + PORT_SCR_ACT;
3526                 dd->port->cmd_issue[i] =
3527                         dd->port->mmio + i*0x80 + PORT_COMMAND_ISSUE;
3528                 dd->port->completed[i] =
3529                         dd->port->mmio + i*0x80 + PORT_SDBV;
3530         }
3531
3532         timetaken = jiffies;
3533         timeout = jiffies + msecs_to_jiffies(30000);
3534         while (((readl(dd->port->mmio + PORT_SCR_STAT) & 0x0F) != 0x03) &&
3535                  time_before(jiffies, timeout)) {
3536                 mdelay(100);
3537         }
3538         if (unlikely(mtip_check_surprise_removal(dd->pdev))) {
3539                 timetaken = jiffies - timetaken;
3540                 dev_warn(&dd->pdev->dev,
3541                         "Surprise removal detected at %u ms\n",
3542                         jiffies_to_msecs(timetaken));
3543                 rv = -ENODEV;
3544                 goto out2 ;
3545         }
3546         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))) {
3547                 timetaken = jiffies - timetaken;
3548                 dev_warn(&dd->pdev->dev,
3549                         "Removal detected at %u ms\n",
3550                         jiffies_to_msecs(timetaken));
3551                 rv = -EFAULT;
3552                 goto out2;
3553         }
3554
3555         /* Conditionally reset the HBA. */
3556         if (!(readl(dd->mmio + HOST_CAP) & HOST_CAP_NZDMA)) {
3557                 if (mtip_hba_reset(dd) < 0) {
3558                         dev_err(&dd->pdev->dev,
3559                                 "Card did not reset within timeout\n");
3560                         rv = -EIO;
3561                         goto out2;
3562                 }
3563         } else {
3564                 /* Clear any pending interrupts on the HBA */
3565                 writel(readl(dd->mmio + HOST_IRQ_STAT),
3566                         dd->mmio + HOST_IRQ_STAT);
3567         }
3568
3569         mtip_init_port(dd->port);
3570         mtip_start_port(dd->port);
3571
3572         /* Setup the ISR and enable interrupts. */
3573         rv = devm_request_irq(&dd->pdev->dev,
3574                                 dd->pdev->irq,
3575                                 mtip_irq_handler,
3576                                 IRQF_SHARED,
3577                                 dev_driver_string(&dd->pdev->dev),
3578                                 dd);
3579
3580         if (rv) {
3581                 dev_err(&dd->pdev->dev,
3582                         "Unable to allocate IRQ %d\n", dd->pdev->irq);
3583                 goto out2;
3584         }
3585         irq_set_affinity_hint(dd->pdev->irq, get_cpu_mask(dd->isr_binding));
3586
3587         /* Enable interrupts on the HBA. */
3588         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
3589                                         dd->mmio + HOST_CTL);
3590
3591         init_timer(&dd->port->cmd_timer);
3592         init_waitqueue_head(&dd->port->svc_wait);
3593
3594         dd->port->cmd_timer.data = (unsigned long int) dd->port;
3595         dd->port->cmd_timer.function = mtip_timeout_function;
3596         mod_timer(&dd->port->cmd_timer,
3597                 jiffies + msecs_to_jiffies(MTIP_TIMEOUT_CHECK_PERIOD));
3598
3599
3600         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)) {
3601                 rv = -EFAULT;
3602                 goto out3;
3603         }
3604
3605         if (mtip_get_identify(dd->port, NULL) < 0) {
3606                 rv = -EFAULT;
3607                 goto out3;
3608         }
3609         mtip_dump_identify(dd->port);
3610
3611         if (*(dd->port->identify + MTIP_FTL_REBUILD_OFFSET) ==
3612                 MTIP_FTL_REBUILD_MAGIC) {
3613                 set_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags);
3614                 return MTIP_FTL_REBUILD_MAGIC;
3615         }
3616
3617         /* check write protect, over temp and rebuild statuses */
3618         rv = mtip_read_log_page(dd->port, ATA_LOG_SATA_NCQ,
3619                                 dd->port->log_buf,
3620                                 dd->port->log_buf_dma, 1);
3621         if (rv) {
3622                 dev_warn(&dd->pdev->dev,
3623                         "Error in READ LOG EXT (10h) command\n");
3624                 /* non-critical error, don't fail the load */
3625         } else {
3626                 buf = (unsigned char *)dd->port->log_buf;
3627                 if (buf[259] & 0x1) {
3628                         dev_info(&dd->pdev->dev,
3629                                 "Write protect bit is set.\n");
3630                         set_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag);
3631                 }
3632                 if (buf[288] == 0xF7) {
3633                         dev_info(&dd->pdev->dev,
3634                                 "Exceeded Tmax, drive in thermal shutdown.\n");
3635                         set_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag);
3636                 }
3637                 if (buf[288] == 0xBF) {
3638                         dev_info(&dd->pdev->dev,
3639                                 "Drive is in security locked state.\n");
3640                         set_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag);
3641                 }
3642         }
3643
3644         /* get write protect progess */
3645         memset(&attr242, 0, sizeof(struct smart_attr));
3646         if (mtip_get_smart_attr(dd->port, 242, &attr242))
3647                 dev_warn(&dd->pdev->dev,
3648                                 "Unable to check write protect progress\n");
3649         else
3650                 dev_info(&dd->pdev->dev,
3651                                 "Write protect progress: %u%% (%u blocks)\n",
3652                                 attr242.cur, le32_to_cpu(attr242.data));
3653         return rv;
3654
3655 out3:
3656         del_timer_sync(&dd->port->cmd_timer);
3657
3658         /* Disable interrupts on the HBA. */
3659         writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3660                         dd->mmio + HOST_CTL);
3661
3662         /* Release the IRQ. */
3663         irq_set_affinity_hint(dd->pdev->irq, NULL);
3664         devm_free_irq(&dd->pdev->dev, dd->pdev->irq, dd);
3665
3666 out2:
3667         mtip_deinit_port(dd->port);
3668         mtip_dma_free(dd);
3669
3670 out1:
3671         /* Free the memory allocated for the for structure. */
3672         kfree(dd->port);
3673
3674         return rv;
3675 }
3676
3677 /*
3678  * Called to deinitialize an interface.
3679  *
3680  * @dd Pointer to the driver data structure.
3681  *
3682  * return value
3683  *      0
3684  */
3685 static int mtip_hw_exit(struct driver_data *dd)
3686 {
3687         /*
3688          * Send standby immediate (E0h) to the drive so that it
3689          * saves its state.
3690          */
3691         if (!dd->sr) {
3692                 if (!test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags) &&
3693                     !test_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag))
3694                         if (mtip_standby_immediate(dd->port))
3695                                 dev_warn(&dd->pdev->dev,
3696                                         "STANDBY IMMEDIATE failed\n");
3697
3698                 /* de-initialize the port. */
3699                 mtip_deinit_port(dd->port);
3700
3701                 /* Disable interrupts on the HBA. */
3702                 writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3703                                 dd->mmio + HOST_CTL);
3704         }
3705
3706         del_timer_sync(&dd->port->cmd_timer);
3707
3708         /* Release the IRQ. */
3709         irq_set_affinity_hint(dd->pdev->irq, NULL);
3710         devm_free_irq(&dd->pdev->dev, dd->pdev->irq, dd);
3711
3712         /* Free dma regions */
3713         mtip_dma_free(dd);
3714
3715         /* Free the memory allocated for the for structure. */
3716         kfree(dd->port);
3717         dd->port = NULL;
3718
3719         return 0;
3720 }
3721
3722 /*
3723  * Issue a Standby Immediate command to the device.
3724  *
3725  * This function is called by the Block Layer just before the
3726  * system powers off during a shutdown.
3727  *
3728  * @dd Pointer to the driver data structure.
3729  *
3730  * return value
3731  *      0
3732  */
3733 static int mtip_hw_shutdown(struct driver_data *dd)
3734 {
3735         /*
3736          * Send standby immediate (E0h) to the drive so that it
3737          * saves its state.
3738          */
3739         if (!dd->sr && dd->port)
3740                 mtip_standby_immediate(dd->port);
3741
3742         return 0;
3743 }
3744
3745 /*
3746  * Suspend function
3747  *
3748  * This function is called by the Block Layer just before the
3749  * system hibernates.
3750  *
3751  * @dd Pointer to the driver data structure.
3752  *
3753  * return value
3754  *      0       Suspend was successful
3755  *      -EFAULT Suspend was not successful
3756  */
3757 static int mtip_hw_suspend(struct driver_data *dd)
3758 {
3759         /*
3760          * Send standby immediate (E0h) to the drive
3761          * so that it saves its state.
3762          */
3763         if (mtip_standby_immediate(dd->port) != 0) {
3764                 dev_err(&dd->pdev->dev,
3765                         "Failed standby-immediate command\n");
3766                 return -EFAULT;
3767         }
3768
3769         /* Disable interrupts on the HBA.*/
3770         writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3771                         dd->mmio + HOST_CTL);
3772         mtip_deinit_port(dd->port);
3773
3774         return 0;
3775 }
3776
3777 /*
3778  * Resume function
3779  *
3780  * This function is called by the Block Layer as the
3781  * system resumes.
3782  *
3783  * @dd Pointer to the driver data structure.
3784  *
3785  * return value
3786  *      0       Resume was successful
3787  *      -EFAULT Resume was not successful
3788  */
3789 static int mtip_hw_resume(struct driver_data *dd)
3790 {
3791         /* Perform any needed hardware setup steps */
3792         hba_setup(dd);
3793
3794         /* Reset the HBA */
3795         if (mtip_hba_reset(dd) != 0) {
3796                 dev_err(&dd->pdev->dev,
3797                         "Unable to reset the HBA\n");
3798                 return -EFAULT;
3799         }
3800
3801         /*
3802          * Enable the port, DMA engine, and FIS reception specific
3803          * h/w in controller.
3804          */
3805         mtip_init_port(dd->port);
3806         mtip_start_port(dd->port);
3807
3808         /* Enable interrupts on the HBA.*/
3809         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
3810                         dd->mmio + HOST_CTL);
3811
3812         return 0;
3813 }
3814
3815 /*
3816  * Helper function for reusing disk name
3817  * upon hot insertion.
3818  */
3819 static int rssd_disk_name_format(char *prefix,
3820                                  int index,
3821                                  char *buf,
3822                                  int buflen)
3823 {
3824         const int base = 'z' - 'a' + 1;
3825         char *begin = buf + strlen(prefix);
3826         char *end = buf + buflen;
3827         char *p;
3828         int unit;
3829
3830         p = end - 1;
3831         *p = '\0';
3832         unit = base;
3833         do {
3834                 if (p == begin)
3835                         return -EINVAL;
3836                 *--p = 'a' + (index % unit);
3837                 index = (index / unit) - 1;
3838         } while (index >= 0);
3839
3840         memmove(begin, p, end - p);
3841         memcpy(buf, prefix, strlen(prefix));
3842
3843         return 0;
3844 }
3845
3846 /*
3847  * Block layer IOCTL handler.
3848  *
3849  * @dev Pointer to the block_device structure.
3850  * @mode ignored
3851  * @cmd IOCTL command passed from the user application.
3852  * @arg Argument passed from the user application.
3853  *
3854  * return value
3855  *      0        IOCTL completed successfully.
3856  *      -ENOTTY  IOCTL not supported or invalid driver data
3857  *                 structure pointer.
3858  */
3859 static int mtip_block_ioctl(struct block_device *dev,
3860                             fmode_t mode,
3861                             unsigned cmd,
3862                             unsigned long arg)
3863 {
3864         struct driver_data *dd = dev->bd_disk->private_data;
3865
3866         if (!capable(CAP_SYS_ADMIN))
3867                 return -EACCES;
3868
3869         if (!dd)
3870                 return -ENOTTY;
3871
3872         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)))
3873                 return -ENOTTY;
3874
3875         switch (cmd) {
3876         case BLKFLSBUF:
3877                 return -ENOTTY;
3878         default:
3879                 return mtip_hw_ioctl(dd, cmd, arg);
3880         }
3881 }
3882
3883 #ifdef CONFIG_COMPAT
3884 /*
3885  * Block layer compat IOCTL handler.
3886  *
3887  * @dev Pointer to the block_device structure.
3888  * @mode ignored
3889  * @cmd IOCTL command passed from the user application.
3890  * @arg Argument passed from the user application.
3891  *
3892  * return value
3893  *      0        IOCTL completed successfully.
3894  *      -ENOTTY  IOCTL not supported or invalid driver data
3895  *                 structure pointer.
3896  */
3897 static int mtip_block_compat_ioctl(struct block_device *dev,
3898                             fmode_t mode,
3899                             unsigned cmd,
3900                             unsigned long arg)
3901 {
3902         struct driver_data *dd = dev->bd_disk->private_data;
3903
3904         if (!capable(CAP_SYS_ADMIN))
3905                 return -EACCES;
3906
3907         if (!dd)
3908                 return -ENOTTY;
3909
3910         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)))
3911                 return -ENOTTY;
3912
3913         switch (cmd) {
3914         case BLKFLSBUF:
3915                 return -ENOTTY;
3916         case HDIO_DRIVE_TASKFILE: {
3917                 struct mtip_compat_ide_task_request_s __user *compat_req_task;
3918                 ide_task_request_t req_task;
3919                 int compat_tasksize, outtotal, ret;
3920
3921                 compat_tasksize =
3922                         sizeof(struct mtip_compat_ide_task_request_s);
3923
3924                 compat_req_task =
3925                         (struct mtip_compat_ide_task_request_s __user *) arg;
3926
3927                 if (copy_from_user(&req_task, (void __user *) arg,
3928                         compat_tasksize - (2 * sizeof(compat_long_t))))
3929                         return -EFAULT;
3930
3931                 if (get_user(req_task.out_size, &compat_req_task->out_size))
3932                         return -EFAULT;
3933
3934                 if (get_user(req_task.in_size, &compat_req_task->in_size))
3935                         return -EFAULT;
3936
3937                 outtotal = sizeof(struct mtip_compat_ide_task_request_s);
3938
3939                 ret = exec_drive_taskfile(dd, (void __user *) arg,
3940                                                 &req_task, outtotal);
3941
3942                 if (copy_to_user((void __user *) arg, &req_task,
3943                                 compat_tasksize -
3944                                 (2 * sizeof(compat_long_t))))
3945                         return -EFAULT;
3946
3947                 if (put_user(req_task.out_size, &compat_req_task->out_size))
3948                         return -EFAULT;
3949
3950                 if (put_user(req_task.in_size, &compat_req_task->in_size))
3951                         return -EFAULT;
3952
3953                 return ret;
3954         }
3955         default:
3956                 return mtip_hw_ioctl(dd, cmd, arg);
3957         }
3958 }
3959 #endif
3960
3961 /*
3962  * Obtain the geometry of the device.
3963  *
3964  * You may think that this function is obsolete, but some applications,
3965  * fdisk for example still used CHS values. This function describes the
3966  * device as having 224 heads and 56 sectors per cylinder. These values are
3967  * chosen so that each cylinder is aligned on a 4KB boundary. Since a
3968  * partition is described in terms of a start and end cylinder this means
3969  * that each partition is also 4KB aligned. Non-aligned partitions adversely
3970  * affects performance.
3971  *
3972  * @dev Pointer to the block_device strucutre.
3973  * @geo Pointer to a hd_geometry structure.
3974  *
3975  * return value
3976  *      0       Operation completed successfully.
3977  *      -ENOTTY An error occurred while reading the drive capacity.
3978  */
3979 static int mtip_block_getgeo(struct block_device *dev,
3980                                 struct hd_geometry *geo)
3981 {
3982         struct driver_data *dd = dev->bd_disk->private_data;
3983         sector_t capacity;
3984
3985         if (!dd)
3986                 return -ENOTTY;
3987
3988         if (!(mtip_hw_get_capacity(dd, &capacity))) {
3989                 dev_warn(&dd->pdev->dev,
3990                         "Could not get drive capacity.\n");
3991                 return -ENOTTY;
3992         }
3993
3994         geo->heads = 224;
3995         geo->sectors = 56;
3996         sector_div(capacity, (geo->heads * geo->sectors));
3997         geo->cylinders = capacity;
3998         return 0;
3999 }
4000
4001 /*
4002  * Block device operation function.
4003  *
4004  * This structure contains pointers to the functions required by the block
4005  * layer.
4006  */
4007 static const struct block_device_operations mtip_block_ops = {
4008         .ioctl          = mtip_block_ioctl,
4009 #ifdef CONFIG_COMPAT
4010         .compat_ioctl   = mtip_block_compat_ioctl,
4011 #endif
4012         .getgeo         = mtip_block_getgeo,
4013         .owner          = THIS_MODULE
4014 };
4015
4016 /*
4017  * Block layer make request function.
4018  *
4019  * This function is called by the kernel to process a BIO for
4020  * the P320 device.
4021  *
4022  * @queue Pointer to the request queue. Unused other than to obtain
4023  *              the driver data structure.
4024  * @bio   Pointer to the BIO.
4025  *
4026  */
4027 static void mtip_make_request(struct request_queue *queue, struct bio *bio)
4028 {
4029         struct driver_data *dd = queue->queuedata;
4030         struct scatterlist *sg;
4031         struct bio_vec bvec;
4032         struct bvec_iter iter;
4033         int nents = 0;
4034         int tag = 0, unaligned = 0;
4035
4036         if (unlikely(dd->dd_flag & MTIP_DDF_STOP_IO)) {
4037                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
4038                                                         &dd->dd_flag))) {
4039                         bio_endio(bio, -ENXIO);
4040                         return;
4041                 }
4042                 if (unlikely(test_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag))) {
4043                         bio_endio(bio, -ENODATA);
4044                         return;
4045                 }
4046                 if (unlikely(test_bit(MTIP_DDF_WRITE_PROTECT_BIT,
4047                                                         &dd->dd_flag) &&
4048                                 bio_data_dir(bio))) {
4049                         bio_endio(bio, -ENODATA);
4050                         return;
4051                 }
4052                 if (unlikely(test_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag))) {
4053                         bio_endio(bio, -ENODATA);
4054                         return;
4055                 }
4056                 if (test_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag)) {
4057                         bio_endio(bio, -ENXIO);
4058                         return;
4059                 }
4060         }
4061
4062         if (unlikely(bio->bi_rw & REQ_DISCARD)) {
4063                 bio_endio(bio, mtip_send_trim(dd, bio->bi_iter.bi_sector,
4064                                                 bio_sectors(bio)));
4065                 return;
4066         }
4067
4068         if (unlikely(!bio_has_data(bio))) {
4069                 blk_queue_flush(queue, 0);
4070                 bio_endio(bio, 0);
4071                 return;
4072         }
4073
4074         if (bio_data_dir(bio) == WRITE && bio_sectors(bio) <= 64 &&
4075                                                         dd->unal_qdepth) {
4076                 if (bio->bi_iter.bi_sector % 8 != 0)
4077                         /* Unaligned on 4k boundaries */
4078                         unaligned = 1;
4079                 else if (bio_sectors(bio) % 8 != 0) /* Aligned but not 4k/8k */
4080                         unaligned = 1;
4081         }
4082
4083         sg = mtip_hw_get_scatterlist(dd, &tag, unaligned);
4084         if (likely(sg != NULL)) {
4085                 blk_queue_bounce(queue, &bio);
4086
4087                 if (unlikely((bio)->bi_vcnt > MTIP_MAX_SG)) {
4088                         dev_warn(&dd->pdev->dev,
4089                                 "Maximum number of SGL entries exceeded\n");
4090                         bio_io_error(bio);
4091                         mtip_hw_release_scatterlist(dd, tag, unaligned);
4092                         return;
4093                 }
4094
4095                 /* Create the scatter list for this bio. */
4096                 bio_for_each_segment(bvec, bio, iter) {
4097                         sg_set_page(&sg[nents],
4098                                         bvec.bv_page,
4099                                         bvec.bv_len,
4100                                         bvec.bv_offset);
4101                         nents++;
4102                 }
4103
4104                 /* Issue the read/write. */
4105                 mtip_hw_submit_io(dd,
4106                                 bio->bi_iter.bi_sector,
4107                                 bio_sectors(bio),
4108                                 nents,
4109                                 tag,
4110                                 bio_endio,
4111                                 bio,
4112                                 bio_data_dir(bio),
4113                                 unaligned);
4114         } else
4115                 bio_io_error(bio);
4116 }
4117
4118 /*
4119  * Block layer initialization function.
4120  *
4121  * This function is called once by the PCI layer for each P320
4122  * device that is connected to the system.
4123  *
4124  * @dd Pointer to the driver data structure.
4125  *
4126  * return value
4127  *      0 on success else an error code.
4128  */
4129 static int mtip_block_initialize(struct driver_data *dd)
4130 {
4131         int rv = 0, wait_for_rebuild = 0;
4132         sector_t capacity;
4133         unsigned int index = 0;
4134         struct kobject *kobj;
4135         unsigned char thd_name[16];
4136
4137         if (dd->disk)
4138                 goto skip_create_disk; /* hw init done, before rebuild */
4139
4140         /* Initialize the protocol layer. */
4141         wait_for_rebuild = mtip_hw_init(dd);
4142         if (wait_for_rebuild < 0) {
4143                 dev_err(&dd->pdev->dev,
4144                         "Protocol layer initialization failed\n");
4145                 rv = -EINVAL;
4146                 goto protocol_init_error;
4147         }
4148
4149         dd->disk = alloc_disk_node(MTIP_MAX_MINORS, dd->numa_node);
4150         if (dd->disk  == NULL) {
4151                 dev_err(&dd->pdev->dev,
4152                         "Unable to allocate gendisk structure\n");
4153                 rv = -EINVAL;
4154                 goto alloc_disk_error;
4155         }
4156
4157         /* Generate the disk name, implemented same as in sd.c */
4158         do {
4159                 if (!ida_pre_get(&rssd_index_ida, GFP_KERNEL))
4160                         goto ida_get_error;
4161
4162                 spin_lock(&rssd_index_lock);
4163                 rv = ida_get_new(&rssd_index_ida, &index);
4164                 spin_unlock(&rssd_index_lock);
4165         } while (rv == -EAGAIN);
4166
4167         if (rv)
4168                 goto ida_get_error;
4169
4170         rv = rssd_disk_name_format("rssd",
4171                                 index,
4172                                 dd->disk->disk_name,
4173                                 DISK_NAME_LEN);
4174         if (rv)
4175                 goto disk_index_error;
4176
4177         dd->disk->driverfs_dev  = &dd->pdev->dev;
4178         dd->disk->major         = dd->major;
4179         dd->disk->first_minor   = dd->instance * MTIP_MAX_MINORS;
4180         dd->disk->fops          = &mtip_block_ops;
4181         dd->disk->private_data  = dd;
4182         dd->index               = index;
4183
4184         mtip_hw_debugfs_init(dd);
4185
4186         /*
4187          * if rebuild pending, start the service thread, and delay the block
4188          * queue creation and add_disk()
4189          */
4190         if (wait_for_rebuild == MTIP_FTL_REBUILD_MAGIC)
4191                 goto start_service_thread;
4192
4193 skip_create_disk:
4194         /* Allocate the request queue. */
4195         dd->queue = blk_alloc_queue_node(GFP_KERNEL, dd->numa_node);
4196         if (dd->queue == NULL) {
4197                 dev_err(&dd->pdev->dev,
4198                         "Unable to allocate request queue\n");
4199                 rv = -ENOMEM;
4200                 goto block_queue_alloc_init_error;
4201         }
4202
4203         /* Attach our request function to the request queue. */
4204         blk_queue_make_request(dd->queue, mtip_make_request);
4205
4206         dd->disk->queue         = dd->queue;
4207         dd->queue->queuedata    = dd;
4208
4209         /* Set device limits. */
4210         set_bit(QUEUE_FLAG_NONROT, &dd->queue->queue_flags);
4211         blk_queue_max_segments(dd->queue, MTIP_MAX_SG);
4212         blk_queue_physical_block_size(dd->queue, 4096);
4213         blk_queue_max_hw_sectors(dd->queue, 0xffff);
4214         blk_queue_max_segment_size(dd->queue, 0x400000);
4215         blk_queue_io_min(dd->queue, 4096);
4216
4217         /*
4218          * write back cache is not supported in the device. FUA depends on
4219          * write back cache support, hence setting flush support to zero.
4220          */
4221         blk_queue_flush(dd->queue, 0);
4222
4223         /* Signal trim support */
4224         if (dd->trim_supp == true) {
4225                 set_bit(QUEUE_FLAG_DISCARD, &dd->queue->queue_flags);
4226                 dd->queue->limits.discard_granularity = 4096;
4227                 blk_queue_max_discard_sectors(dd->queue,
4228                         MTIP_MAX_TRIM_ENTRY_LEN * MTIP_MAX_TRIM_ENTRIES);
4229                 dd->queue->limits.discard_zeroes_data = 0;
4230         }
4231
4232         /* Set the capacity of the device in 512 byte sectors. */
4233         if (!(mtip_hw_get_capacity(dd, &capacity))) {
4234                 dev_warn(&dd->pdev->dev,
4235                         "Could not read drive capacity\n");
4236                 rv = -EIO;
4237                 goto read_capacity_error;
4238         }
4239         set_capacity(dd->disk, capacity);
4240
4241         /* Enable the block device and add it to /dev */
4242         add_disk(dd->disk);
4243
4244         dd->bdev = bdget_disk(dd->disk, 0);
4245         /*
4246          * Now that the disk is active, initialize any sysfs attributes
4247          * managed by the protocol layer.
4248          */
4249         kobj = kobject_get(&disk_to_dev(dd->disk)->kobj);
4250         if (kobj) {
4251                 mtip_hw_sysfs_init(dd, kobj);
4252                 kobject_put(kobj);
4253         }
4254
4255         if (dd->mtip_svc_handler) {
4256                 set_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag);
4257                 return rv; /* service thread created for handling rebuild */
4258         }
4259
4260 start_service_thread:
4261         sprintf(thd_name, "mtip_svc_thd_%02d", index);
4262         dd->mtip_svc_handler = kthread_create_on_node(mtip_service_thread,
4263                                                 dd, dd->numa_node, "%s",
4264                                                 thd_name);
4265
4266         if (IS_ERR(dd->mtip_svc_handler)) {
4267                 dev_err(&dd->pdev->dev, "service thread failed to start\n");
4268                 dd->mtip_svc_handler = NULL;
4269                 rv = -EFAULT;
4270                 goto kthread_run_error;
4271         }
4272         wake_up_process(dd->mtip_svc_handler);
4273         if (wait_for_rebuild == MTIP_FTL_REBUILD_MAGIC)
4274                 rv = wait_for_rebuild;
4275
4276         return rv;
4277
4278 kthread_run_error:
4279         bdput(dd->bdev);
4280         dd->bdev = NULL;
4281
4282         /* Delete our gendisk. This also removes the device from /dev */
4283         del_gendisk(dd->disk);
4284
4285 read_capacity_error:
4286         blk_cleanup_queue(dd->queue);
4287
4288 block_queue_alloc_init_error:
4289         mtip_hw_debugfs_exit(dd);
4290 disk_index_error:
4291         spin_lock(&rssd_index_lock);
4292         ida_remove(&rssd_index_ida, index);
4293         spin_unlock(&rssd_index_lock);
4294
4295 ida_get_error:
4296         put_disk(dd->disk);
4297
4298 alloc_disk_error:
4299         mtip_hw_exit(dd); /* De-initialize the protocol layer. */
4300
4301 protocol_init_error:
4302         return rv;
4303 }
4304
4305 /*
4306  * Block layer deinitialization function.
4307  *
4308  * Called by the PCI layer as each P320 device is removed.
4309  *
4310  * @dd Pointer to the driver data structure.
4311  *
4312  * return value
4313  *      0
4314  */
4315 static int mtip_block_remove(struct driver_data *dd)
4316 {
4317         struct kobject *kobj;
4318
4319         if (!dd->sr) {
4320                 mtip_hw_debugfs_exit(dd);
4321
4322                 if (dd->mtip_svc_handler) {
4323                         set_bit(MTIP_PF_SVC_THD_STOP_BIT, &dd->port->flags);
4324                         wake_up_interruptible(&dd->port->svc_wait);
4325                         kthread_stop(dd->mtip_svc_handler);
4326                 }
4327
4328                 /* Clean up the sysfs attributes, if created */
4329                 if (test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag)) {
4330                         kobj = kobject_get(&disk_to_dev(dd->disk)->kobj);
4331                         if (kobj) {
4332                                 mtip_hw_sysfs_exit(dd, kobj);
4333                                 kobject_put(kobj);
4334                         }
4335                 }
4336                 /*
4337                  * Delete our gendisk structure. This also removes the device
4338                  * from /dev
4339                  */
4340                 if (dd->bdev) {
4341                         bdput(dd->bdev);
4342                         dd->bdev = NULL;
4343                 }
4344                 if (dd->disk) {
4345                         if (dd->disk->queue) {
4346                                 del_gendisk(dd->disk);
4347                                 blk_cleanup_queue(dd->queue);
4348                                 dd->queue = NULL;
4349                         } else
4350                                 put_disk(dd->disk);
4351                 }
4352                 dd->disk  = NULL;
4353
4354                 spin_lock(&rssd_index_lock);
4355                 ida_remove(&rssd_index_ida, dd->index);
4356                 spin_unlock(&rssd_index_lock);
4357         } else {
4358                 dev_info(&dd->pdev->dev, "device %s surprise removal\n",
4359                                                 dd->disk->disk_name);
4360         }
4361
4362         /* De-initialize the protocol layer. */
4363         mtip_hw_exit(dd);
4364
4365         return 0;
4366 }
4367
4368 /*
4369  * Function called by the PCI layer when just before the
4370  * machine shuts down.
4371  *
4372  * If a protocol layer shutdown function is present it will be called
4373  * by this function.
4374  *
4375  * @dd Pointer to the driver data structure.
4376  *
4377  * return value
4378  *      0
4379  */
4380 static int mtip_block_shutdown(struct driver_data *dd)
4381 {
4382         /* Delete our gendisk structure, and cleanup the blk queue. */
4383         if (dd->disk) {
4384                 dev_info(&dd->pdev->dev,
4385                         "Shutting down %s ...\n", dd->disk->disk_name);
4386
4387                 if (dd->disk->queue) {
4388                         del_gendisk(dd->disk);
4389                         blk_cleanup_queue(dd->queue);
4390                 } else
4391                         put_disk(dd->disk);
4392                 dd->disk  = NULL;
4393                 dd->queue = NULL;
4394         }
4395
4396         spin_lock(&rssd_index_lock);
4397         ida_remove(&rssd_index_ida, dd->index);
4398         spin_unlock(&rssd_index_lock);
4399
4400         mtip_hw_shutdown(dd);
4401         return 0;
4402 }
4403
4404 static int mtip_block_suspend(struct driver_data *dd)
4405 {
4406         dev_info(&dd->pdev->dev,
4407                 "Suspending %s ...\n", dd->disk->disk_name);
4408         mtip_hw_suspend(dd);
4409         return 0;
4410 }
4411
4412 static int mtip_block_resume(struct driver_data *dd)
4413 {
4414         dev_info(&dd->pdev->dev, "Resuming %s ...\n",
4415                 dd->disk->disk_name);
4416         mtip_hw_resume(dd);
4417         return 0;
4418 }
4419
4420 static void drop_cpu(int cpu)
4421 {
4422         cpu_use[cpu]--;
4423 }
4424
4425 static int get_least_used_cpu_on_node(int node)
4426 {
4427         int cpu, least_used_cpu, least_cnt;
4428         const struct cpumask *node_mask;
4429
4430         node_mask = cpumask_of_node(node);
4431         least_used_cpu = cpumask_first(node_mask);
4432         least_cnt = cpu_use[least_used_cpu];
4433         cpu = least_used_cpu;
4434
4435         for_each_cpu(cpu, node_mask) {
4436                 if (cpu_use[cpu] < least_cnt) {
4437                         least_used_cpu = cpu;
4438                         least_cnt = cpu_use[cpu];
4439                 }
4440         }
4441         cpu_use[least_used_cpu]++;
4442         return least_used_cpu;
4443 }
4444
4445 /* Helper for selecting a node in round robin mode */
4446 static inline int mtip_get_next_rr_node(void)
4447 {
4448         static int next_node = -1;
4449
4450         if (next_node == -1) {
4451                 next_node = first_online_node;
4452                 return next_node;
4453         }
4454
4455         next_node = next_online_node(next_node);
4456         if (next_node == MAX_NUMNODES)
4457                 next_node = first_online_node;
4458         return next_node;
4459 }
4460
4461 static DEFINE_HANDLER(0);
4462 static DEFINE_HANDLER(1);
4463 static DEFINE_HANDLER(2);
4464 static DEFINE_HANDLER(3);
4465 static DEFINE_HANDLER(4);
4466 static DEFINE_HANDLER(5);
4467 static DEFINE_HANDLER(6);
4468 static DEFINE_HANDLER(7);
4469
4470 /*
4471  * Called for each supported PCI device detected.
4472  *
4473  * This function allocates the private data structure, enables the
4474  * PCI device and then calls the block layer initialization function.
4475  *
4476  * return value
4477  *      0 on success else an error code.
4478  */
4479 static int mtip_pci_probe(struct pci_dev *pdev,
4480                         const struct pci_device_id *ent)
4481 {
4482         int rv = 0;
4483         struct driver_data *dd = NULL;
4484         char cpu_list[256];
4485         const struct cpumask *node_mask;
4486         int cpu, i = 0, j = 0;
4487         int my_node = NUMA_NO_NODE;
4488         unsigned long flags;
4489
4490         /* Allocate memory for this devices private data. */
4491         my_node = pcibus_to_node(pdev->bus);
4492         if (my_node != NUMA_NO_NODE) {
4493                 if (!node_online(my_node))
4494                         my_node = mtip_get_next_rr_node();
4495         } else {
4496                 dev_info(&pdev->dev, "Kernel not reporting proximity, choosing a node\n");
4497                 my_node = mtip_get_next_rr_node();
4498         }
4499         dev_info(&pdev->dev, "NUMA node %d (closest: %d,%d, probe on %d:%d)\n",
4500                 my_node, pcibus_to_node(pdev->bus), dev_to_node(&pdev->dev),
4501                 cpu_to_node(smp_processor_id()), smp_processor_id());
4502
4503         dd = kzalloc_node(sizeof(struct driver_data), GFP_KERNEL, my_node);
4504         if (dd == NULL) {
4505                 dev_err(&pdev->dev,
4506                         "Unable to allocate memory for driver data\n");
4507                 return -ENOMEM;
4508         }
4509
4510         /* Attach the private data to this PCI device.  */
4511         pci_set_drvdata(pdev, dd);
4512
4513         rv = pcim_enable_device(pdev);
4514         if (rv < 0) {
4515                 dev_err(&pdev->dev, "Unable to enable device\n");
4516                 goto iomap_err;
4517         }
4518
4519         /* Map BAR5 to memory. */
4520         rv = pcim_iomap_regions(pdev, 1 << MTIP_ABAR, MTIP_DRV_NAME);
4521         if (rv < 0) {
4522                 dev_err(&pdev->dev, "Unable to map regions\n");
4523                 goto iomap_err;
4524         }
4525
4526         if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(64))) {
4527                 rv = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
4528
4529                 if (rv) {
4530                         rv = pci_set_consistent_dma_mask(pdev,
4531                                                 DMA_BIT_MASK(32));
4532                         if (rv) {
4533                                 dev_warn(&pdev->dev,
4534                                         "64-bit DMA enable failed\n");
4535                                 goto setmask_err;
4536                         }
4537                 }
4538         }
4539
4540         /* Copy the info we may need later into the private data structure. */
4541         dd->major       = mtip_major;
4542         dd->instance    = instance;
4543         dd->pdev        = pdev;
4544         dd->numa_node   = my_node;
4545
4546         INIT_LIST_HEAD(&dd->online_list);
4547         INIT_LIST_HEAD(&dd->remove_list);
4548
4549         memset(dd->workq_name, 0, 32);
4550         snprintf(dd->workq_name, 31, "mtipq%d", dd->instance);
4551
4552         dd->isr_workq = create_workqueue(dd->workq_name);
4553         if (!dd->isr_workq) {
4554                 dev_warn(&pdev->dev, "Can't create wq %d\n", dd->instance);
4555                 rv = -ENOMEM;
4556                 goto block_initialize_err;
4557         }
4558
4559         memset(cpu_list, 0, sizeof(cpu_list));
4560
4561         node_mask = cpumask_of_node(dd->numa_node);
4562         if (!cpumask_empty(node_mask)) {
4563                 for_each_cpu(cpu, node_mask)
4564                 {
4565                         snprintf(&cpu_list[j], 256 - j, "%d ", cpu);
4566                         j = strlen(cpu_list);
4567                 }
4568
4569                 dev_info(&pdev->dev, "Node %d on package %d has %d cpu(s): %s\n",
4570                         dd->numa_node,
4571                         topology_physical_package_id(cpumask_first(node_mask)),
4572                         nr_cpus_node(dd->numa_node),
4573                         cpu_list);
4574         } else
4575                 dev_dbg(&pdev->dev, "mtip32xx: node_mask empty\n");
4576
4577         dd->isr_binding = get_least_used_cpu_on_node(dd->numa_node);
4578         dev_info(&pdev->dev, "Initial IRQ binding node:cpu %d:%d\n",
4579                 cpu_to_node(dd->isr_binding), dd->isr_binding);
4580
4581         /* first worker context always runs in ISR */
4582         dd->work[0].cpu_binding = dd->isr_binding;
4583         dd->work[1].cpu_binding = get_least_used_cpu_on_node(dd->numa_node);
4584         dd->work[2].cpu_binding = get_least_used_cpu_on_node(dd->numa_node);
4585         dd->work[3].cpu_binding = dd->work[0].cpu_binding;
4586         dd->work[4].cpu_binding = dd->work[1].cpu_binding;
4587         dd->work[5].cpu_binding = dd->work[2].cpu_binding;
4588         dd->work[6].cpu_binding = dd->work[2].cpu_binding;
4589         dd->work[7].cpu_binding = dd->work[1].cpu_binding;
4590
4591         /* Log the bindings */
4592         for_each_present_cpu(cpu) {
4593                 memset(cpu_list, 0, sizeof(cpu_list));
4594                 for (i = 0, j = 0; i < MTIP_MAX_SLOT_GROUPS; i++) {
4595                         if (dd->work[i].cpu_binding == cpu) {
4596                                 snprintf(&cpu_list[j], 256 - j, "%d ", i);
4597                                 j = strlen(cpu_list);
4598                         }
4599                 }
4600                 if (j)
4601                         dev_info(&pdev->dev, "CPU %d: WQs %s\n", cpu, cpu_list);
4602         }
4603
4604         INIT_WORK(&dd->work[0].work, mtip_workq_sdbf0);
4605         INIT_WORK(&dd->work[1].work, mtip_workq_sdbf1);
4606         INIT_WORK(&dd->work[2].work, mtip_workq_sdbf2);
4607         INIT_WORK(&dd->work[3].work, mtip_workq_sdbf3);
4608         INIT_WORK(&dd->work[4].work, mtip_workq_sdbf4);
4609         INIT_WORK(&dd->work[5].work, mtip_workq_sdbf5);
4610         INIT_WORK(&dd->work[6].work, mtip_workq_sdbf6);
4611         INIT_WORK(&dd->work[7].work, mtip_workq_sdbf7);
4612
4613         pci_set_master(pdev);
4614         rv = pci_enable_msi(pdev);
4615         if (rv) {
4616                 dev_warn(&pdev->dev,
4617                         "Unable to enable MSI interrupt.\n");
4618                 goto block_initialize_err;
4619         }
4620
4621         /* Initialize the block layer. */
4622         rv = mtip_block_initialize(dd);
4623         if (rv < 0) {
4624                 dev_err(&pdev->dev,
4625                         "Unable to initialize block layer\n");
4626                 goto block_initialize_err;
4627         }
4628
4629         /*
4630          * Increment the instance count so that each device has a unique
4631          * instance number.
4632          */
4633         instance++;
4634         if (rv != MTIP_FTL_REBUILD_MAGIC)
4635                 set_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag);
4636         else
4637                 rv = 0; /* device in rebuild state, return 0 from probe */
4638
4639         /* Add to online list even if in ftl rebuild */
4640         spin_lock_irqsave(&dev_lock, flags);
4641         list_add(&dd->online_list, &online_list);
4642         spin_unlock_irqrestore(&dev_lock, flags);
4643
4644         goto done;
4645
4646 block_initialize_err:
4647         pci_disable_msi(pdev);
4648         if (dd->isr_workq) {
4649                 flush_workqueue(dd->isr_workq);
4650                 destroy_workqueue(dd->isr_workq);
4651                 drop_cpu(dd->work[0].cpu_binding);
4652                 drop_cpu(dd->work[1].cpu_binding);
4653                 drop_cpu(dd->work[2].cpu_binding);
4654         }
4655 setmask_err:
4656         pcim_iounmap_regions(pdev, 1 << MTIP_ABAR);
4657
4658 iomap_err:
4659         kfree(dd);
4660         pci_set_drvdata(pdev, NULL);
4661         return rv;
4662 done:
4663         return rv;
4664 }
4665
4666 /*
4667  * Called for each probed device when the device is removed or the
4668  * driver is unloaded.
4669  *
4670  * return value
4671  *      None
4672  */
4673 static void mtip_pci_remove(struct pci_dev *pdev)
4674 {
4675         struct driver_data *dd = pci_get_drvdata(pdev);
4676         unsigned long flags, to;
4677
4678         set_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag);
4679
4680         spin_lock_irqsave(&dev_lock, flags);
4681         list_del_init(&dd->online_list);
4682         list_add(&dd->remove_list, &removing_list);
4683         spin_unlock_irqrestore(&dev_lock, flags);
4684
4685         mtip_check_surprise_removal(pdev);
4686         synchronize_irq(dd->pdev->irq);
4687
4688         /* Spin until workers are done */
4689         to = jiffies + msecs_to_jiffies(4000);
4690         do {
4691                 msleep(20);
4692         } while (atomic_read(&dd->irq_workers_active) != 0 &&
4693                 time_before(jiffies, to));
4694
4695         if (atomic_read(&dd->irq_workers_active) != 0) {
4696                 dev_warn(&dd->pdev->dev,
4697                         "Completion workers still active!\n");
4698         }
4699         /* Cleanup the outstanding commands */
4700         mtip_command_cleanup(dd);
4701
4702         /* Clean up the block layer. */
4703         mtip_block_remove(dd);
4704
4705         if (dd->isr_workq) {
4706                 flush_workqueue(dd->isr_workq);
4707                 destroy_workqueue(dd->isr_workq);
4708                 drop_cpu(dd->work[0].cpu_binding);
4709                 drop_cpu(dd->work[1].cpu_binding);
4710                 drop_cpu(dd->work[2].cpu_binding);
4711         }
4712
4713         pci_disable_msi(pdev);
4714
4715         spin_lock_irqsave(&dev_lock, flags);
4716         list_del_init(&dd->remove_list);
4717         spin_unlock_irqrestore(&dev_lock, flags);
4718
4719         if (!dd->sr)
4720                 kfree(dd);
4721         else
4722                 set_bit(MTIP_DDF_REMOVE_DONE_BIT, &dd->dd_flag);
4723
4724         pcim_iounmap_regions(pdev, 1 << MTIP_ABAR);
4725         pci_set_drvdata(pdev, NULL);
4726         pci_dev_put(pdev);
4727
4728 }
4729
4730 /*
4731  * Called for each probed device when the device is suspended.
4732  *
4733  * return value
4734  *      0  Success
4735  *      <0 Error
4736  */
4737 static int mtip_pci_suspend(struct pci_dev *pdev, pm_message_t mesg)
4738 {
4739         int rv = 0;
4740         struct driver_data *dd = pci_get_drvdata(pdev);
4741
4742         if (!dd) {
4743                 dev_err(&pdev->dev,
4744                         "Driver private datastructure is NULL\n");
4745                 return -EFAULT;
4746         }
4747
4748         set_bit(MTIP_DDF_RESUME_BIT, &dd->dd_flag);
4749
4750         /* Disable ports & interrupts then send standby immediate */
4751         rv = mtip_block_suspend(dd);
4752         if (rv < 0) {
4753                 dev_err(&pdev->dev,
4754                         "Failed to suspend controller\n");
4755                 return rv;
4756         }
4757
4758         /*
4759          * Save the pci config space to pdev structure &
4760          * disable the device
4761          */
4762         pci_save_state(pdev);
4763         pci_disable_device(pdev);
4764
4765         /* Move to Low power state*/
4766         pci_set_power_state(pdev, PCI_D3hot);
4767
4768         return rv;
4769 }
4770
4771 /*
4772  * Called for each probed device when the device is resumed.
4773  *
4774  * return value
4775  *      0  Success
4776  *      <0 Error
4777  */
4778 static int mtip_pci_resume(struct pci_dev *pdev)
4779 {
4780         int rv = 0;
4781         struct driver_data *dd;
4782
4783         dd = pci_get_drvdata(pdev);
4784         if (!dd) {
4785                 dev_err(&pdev->dev,
4786                         "Driver private datastructure is NULL\n");
4787                 return -EFAULT;
4788         }
4789
4790         /* Move the device to active State */
4791         pci_set_power_state(pdev, PCI_D0);
4792
4793         /* Restore PCI configuration space */
4794         pci_restore_state(pdev);
4795
4796         /* Enable the PCI device*/
4797         rv = pcim_enable_device(pdev);
4798         if (rv < 0) {
4799                 dev_err(&pdev->dev,
4800                         "Failed to enable card during resume\n");
4801                 goto err;
4802         }
4803         pci_set_master(pdev);
4804
4805         /*
4806          * Calls hbaReset, initPort, & startPort function
4807          * then enables interrupts
4808          */
4809         rv = mtip_block_resume(dd);
4810         if (rv < 0)
4811                 dev_err(&pdev->dev, "Unable to resume\n");
4812
4813 err:
4814         clear_bit(MTIP_DDF_RESUME_BIT, &dd->dd_flag);
4815
4816         return rv;
4817 }
4818
4819 /*
4820  * Shutdown routine
4821  *
4822  * return value
4823  *      None
4824  */
4825 static void mtip_pci_shutdown(struct pci_dev *pdev)
4826 {
4827         struct driver_data *dd = pci_get_drvdata(pdev);
4828         if (dd)
4829                 mtip_block_shutdown(dd);
4830 }
4831
4832 /* Table of device ids supported by this driver. */
4833 static DEFINE_PCI_DEVICE_TABLE(mtip_pci_tbl) = {
4834         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320H_DEVICE_ID) },
4835         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320M_DEVICE_ID) },
4836         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320S_DEVICE_ID) },
4837         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P325M_DEVICE_ID) },
4838         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P420H_DEVICE_ID) },
4839         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P420M_DEVICE_ID) },
4840         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P425M_DEVICE_ID) },
4841         { 0 }
4842 };
4843
4844 /* Structure that describes the PCI driver functions. */
4845 static struct pci_driver mtip_pci_driver = {
4846         .name                   = MTIP_DRV_NAME,
4847         .id_table               = mtip_pci_tbl,
4848         .probe                  = mtip_pci_probe,
4849         .remove                 = mtip_pci_remove,
4850         .suspend                = mtip_pci_suspend,
4851         .resume                 = mtip_pci_resume,
4852         .shutdown               = mtip_pci_shutdown,
4853 };
4854
4855 MODULE_DEVICE_TABLE(pci, mtip_pci_tbl);
4856
4857 /*
4858  * Module initialization function.
4859  *
4860  * Called once when the module is loaded. This function allocates a major
4861  * block device number to the Cyclone devices and registers the PCI layer
4862  * of the driver.
4863  *
4864  * Return value
4865  *      0 on success else error code.
4866  */
4867 static int __init mtip_init(void)
4868 {
4869         int error;
4870
4871         pr_info(MTIP_DRV_NAME " Version " MTIP_DRV_VERSION "\n");
4872
4873         spin_lock_init(&dev_lock);
4874
4875         INIT_LIST_HEAD(&online_list);
4876         INIT_LIST_HEAD(&removing_list);
4877
4878         /* Allocate a major block device number to use with this driver. */
4879         error = register_blkdev(0, MTIP_DRV_NAME);
4880         if (error <= 0) {
4881                 pr_err("Unable to register block device (%d)\n",
4882                 error);
4883                 return -EBUSY;
4884         }
4885         mtip_major = error;
4886
4887         dfs_parent = debugfs_create_dir("rssd", NULL);
4888         if (IS_ERR_OR_NULL(dfs_parent)) {
4889                 pr_warn("Error creating debugfs parent\n");
4890                 dfs_parent = NULL;
4891         }
4892         if (dfs_parent) {
4893                 dfs_device_status = debugfs_create_file("device_status",
4894                                         S_IRUGO, dfs_parent, NULL,
4895                                         &mtip_device_status_fops);
4896                 if (IS_ERR_OR_NULL(dfs_device_status)) {
4897                         pr_err("Error creating device_status node\n");
4898                         dfs_device_status = NULL;
4899                 }
4900         }
4901
4902         /* Register our PCI operations. */
4903         error = pci_register_driver(&mtip_pci_driver);
4904         if (error) {
4905                 debugfs_remove(dfs_parent);
4906                 unregister_blkdev(mtip_major, MTIP_DRV_NAME);
4907         }
4908
4909         return error;
4910 }
4911
4912 /*
4913  * Module de-initialization function.
4914  *
4915  * Called once when the module is unloaded. This function deallocates
4916  * the major block device number allocated by mtip_init() and
4917  * unregisters the PCI layer of the driver.
4918  *
4919  * Return value
4920  *      none
4921  */
4922 static void __exit mtip_exit(void)
4923 {
4924         debugfs_remove_recursive(dfs_parent);
4925
4926         /* Release the allocated major block device number. */
4927         unregister_blkdev(mtip_major, MTIP_DRV_NAME);
4928
4929         /* Unregister the PCI driver. */
4930         pci_unregister_driver(&mtip_pci_driver);
4931 }
4932
4933 MODULE_AUTHOR("Micron Technology, Inc");
4934 MODULE_DESCRIPTION("Micron RealSSD PCIe Block Driver");
4935 MODULE_LICENSE("GPL");
4936 MODULE_VERSION(MTIP_DRV_VERSION);
4937
4938 module_init(mtip_init);
4939 module_exit(mtip_exit);