]> Pileus Git - ~andy/linux/blob - drivers/infiniband/ulp/iser/iser_initiator.c
IB/iser: Post initial receive buffers before sending the final login request
[~andy/linux] / drivers / infiniband / ulp / iser / iser_initiator.c
1 /*
2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 #include <linux/kernel.h>
33 #include <linux/slab.h>
34 #include <linux/mm.h>
35 #include <linux/scatterlist.h>
36 #include <linux/kfifo.h>
37 #include <scsi/scsi_cmnd.h>
38 #include <scsi/scsi_host.h>
39
40 #include "iscsi_iser.h"
41
42 /* Register user buffer memory and initialize passive rdma
43  *  dto descriptor. Total data size is stored in
44  *  iser_task->data[ISER_DIR_IN].data_len
45  */
46 static int iser_prepare_read_cmd(struct iscsi_task *task,
47                                  unsigned int edtl)
48
49 {
50         struct iscsi_iser_task *iser_task = task->dd_data;
51         struct iser_regd_buf *regd_buf;
52         int err;
53         struct iser_hdr *hdr = &iser_task->desc.iser_header;
54         struct iser_data_buf *buf_in = &iser_task->data[ISER_DIR_IN];
55
56         err = iser_dma_map_task_data(iser_task,
57                                      buf_in,
58                                      ISER_DIR_IN,
59                                      DMA_FROM_DEVICE);
60         if (err)
61                 return err;
62
63         if (edtl > iser_task->data[ISER_DIR_IN].data_len) {
64                 iser_err("Total data length: %ld, less than EDTL: "
65                          "%d, in READ cmd BHS itt: %d, conn: 0x%p\n",
66                          iser_task->data[ISER_DIR_IN].data_len, edtl,
67                          task->itt, iser_task->iser_conn);
68                 return -EINVAL;
69         }
70
71         err = iser_reg_rdma_mem(iser_task,ISER_DIR_IN);
72         if (err) {
73                 iser_err("Failed to set up Data-IN RDMA\n");
74                 return err;
75         }
76         regd_buf = &iser_task->rdma_regd[ISER_DIR_IN];
77
78         hdr->flags    |= ISER_RSV;
79         hdr->read_stag = cpu_to_be32(regd_buf->reg.rkey);
80         hdr->read_va   = cpu_to_be64(regd_buf->reg.va);
81
82         iser_dbg("Cmd itt:%d READ tags RKEY:%#.4X VA:%#llX\n",
83                  task->itt, regd_buf->reg.rkey,
84                  (unsigned long long)regd_buf->reg.va);
85
86         return 0;
87 }
88
89 /* Register user buffer memory and initialize passive rdma
90  *  dto descriptor. Total data size is stored in
91  *  task->data[ISER_DIR_OUT].data_len
92  */
93 static int
94 iser_prepare_write_cmd(struct iscsi_task *task,
95                        unsigned int imm_sz,
96                        unsigned int unsol_sz,
97                        unsigned int edtl)
98 {
99         struct iscsi_iser_task *iser_task = task->dd_data;
100         struct iser_regd_buf *regd_buf;
101         int err;
102         struct iser_hdr *hdr = &iser_task->desc.iser_header;
103         struct iser_data_buf *buf_out = &iser_task->data[ISER_DIR_OUT];
104         struct ib_sge *tx_dsg = &iser_task->desc.tx_sg[1];
105
106         err = iser_dma_map_task_data(iser_task,
107                                      buf_out,
108                                      ISER_DIR_OUT,
109                                      DMA_TO_DEVICE);
110         if (err)
111                 return err;
112
113         if (edtl > iser_task->data[ISER_DIR_OUT].data_len) {
114                 iser_err("Total data length: %ld, less than EDTL: %d, "
115                          "in WRITE cmd BHS itt: %d, conn: 0x%p\n",
116                          iser_task->data[ISER_DIR_OUT].data_len,
117                          edtl, task->itt, task->conn);
118                 return -EINVAL;
119         }
120
121         err = iser_reg_rdma_mem(iser_task,ISER_DIR_OUT);
122         if (err != 0) {
123                 iser_err("Failed to register write cmd RDMA mem\n");
124                 return err;
125         }
126
127         regd_buf = &iser_task->rdma_regd[ISER_DIR_OUT];
128
129         if (unsol_sz < edtl) {
130                 hdr->flags     |= ISER_WSV;
131                 hdr->write_stag = cpu_to_be32(regd_buf->reg.rkey);
132                 hdr->write_va   = cpu_to_be64(regd_buf->reg.va + unsol_sz);
133
134                 iser_dbg("Cmd itt:%d, WRITE tags, RKEY:%#.4X "
135                          "VA:%#llX + unsol:%d\n",
136                          task->itt, regd_buf->reg.rkey,
137                          (unsigned long long)regd_buf->reg.va, unsol_sz);
138         }
139
140         if (imm_sz > 0) {
141                 iser_dbg("Cmd itt:%d, WRITE, adding imm.data sz: %d\n",
142                          task->itt, imm_sz);
143                 tx_dsg->addr   = regd_buf->reg.va;
144                 tx_dsg->length = imm_sz;
145                 tx_dsg->lkey   = regd_buf->reg.lkey;
146                 iser_task->desc.num_sge = 2;
147         }
148
149         return 0;
150 }
151
152 /* creates a new tx descriptor and adds header regd buffer */
153 static void iser_create_send_desc(struct iser_conn      *ib_conn,
154                                   struct iser_tx_desc   *tx_desc)
155 {
156         struct iser_device *device = ib_conn->device;
157
158         ib_dma_sync_single_for_cpu(device->ib_device,
159                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
160
161         memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
162         tx_desc->iser_header.flags = ISER_VER;
163
164         tx_desc->num_sge = 1;
165
166         if (tx_desc->tx_sg[0].lkey != device->mr->lkey) {
167                 tx_desc->tx_sg[0].lkey = device->mr->lkey;
168                 iser_dbg("sdesc %p lkey mismatch, fixing\n", tx_desc);
169         }
170 }
171
172
173 int iser_alloc_rx_descriptors(struct iser_conn *ib_conn)
174 {
175         int i, j;
176         u64 dma_addr;
177         struct iser_rx_desc *rx_desc;
178         struct ib_sge       *rx_sg;
179         struct iser_device  *device = ib_conn->device;
180
181         ib_conn->rx_descs = kmalloc(ISER_QP_MAX_RECV_DTOS *
182                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
183         if (!ib_conn->rx_descs)
184                 goto rx_desc_alloc_fail;
185
186         rx_desc = ib_conn->rx_descs;
187
188         for (i = 0; i < ISER_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
189                 dma_addr = ib_dma_map_single(device->ib_device, (void *)rx_desc,
190                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
191                 if (ib_dma_mapping_error(device->ib_device, dma_addr))
192                         goto rx_desc_dma_map_failed;
193
194                 rx_desc->dma_addr = dma_addr;
195
196                 rx_sg = &rx_desc->rx_sg;
197                 rx_sg->addr   = rx_desc->dma_addr;
198                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
199                 rx_sg->lkey   = device->mr->lkey;
200         }
201
202         ib_conn->rx_desc_head = 0;
203         return 0;
204
205 rx_desc_dma_map_failed:
206         rx_desc = ib_conn->rx_descs;
207         for (j = 0; j < i; j++, rx_desc++)
208                 ib_dma_unmap_single(device->ib_device, rx_desc->dma_addr,
209                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
210         kfree(ib_conn->rx_descs);
211         ib_conn->rx_descs = NULL;
212 rx_desc_alloc_fail:
213         iser_err("failed allocating rx descriptors / data buffers\n");
214         return -ENOMEM;
215 }
216
217 void iser_free_rx_descriptors(struct iser_conn *ib_conn)
218 {
219         int i;
220         struct iser_rx_desc *rx_desc;
221         struct iser_device *device = ib_conn->device;
222
223         if (!ib_conn->rx_descs)
224                 return;
225
226         rx_desc = ib_conn->rx_descs;
227         for (i = 0; i < ISER_QP_MAX_RECV_DTOS; i++, rx_desc++)
228                 ib_dma_unmap_single(device->ib_device, rx_desc->dma_addr,
229                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
230         kfree(ib_conn->rx_descs);
231 }
232
233 static int iser_post_rx_bufs(struct iscsi_conn *conn, struct iscsi_hdr *req)
234 {
235         struct iscsi_iser_conn *iser_conn = conn->dd_data;
236
237         iser_dbg("req op %x flags %x\n", req->opcode, req->flags);
238         /* check if this is the last login - going to full feature phase */
239         if ((req->flags & ISCSI_FULL_FEATURE_PHASE) != ISCSI_FULL_FEATURE_PHASE)
240                 return 0;
241
242         /*
243          * Check that there is one posted recv buffer (for the last login
244          * response) and no posted send buffers left - they must have been
245          * consumed during previous login phases.
246          */
247         WARN_ON(iser_conn->ib_conn->post_recv_buf_count != 1);
248         WARN_ON(atomic_read(&iser_conn->ib_conn->post_send_buf_count) != 0);
249
250         iser_dbg("Initially post: %d\n", ISER_MIN_POSTED_RX);
251         /* Initial post receive buffers */
252         if (iser_post_recvm(iser_conn->ib_conn, ISER_MIN_POSTED_RX))
253                 return -ENOMEM;
254
255         return 0;
256 }
257
258 /**
259  * iser_send_command - send command PDU
260  */
261 int iser_send_command(struct iscsi_conn *conn,
262                       struct iscsi_task *task)
263 {
264         struct iscsi_iser_conn *iser_conn = conn->dd_data;
265         struct iscsi_iser_task *iser_task = task->dd_data;
266         unsigned long edtl;
267         int err;
268         struct iser_data_buf *data_buf;
269         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)task->hdr;
270         struct scsi_cmnd *sc  =  task->sc;
271         struct iser_tx_desc *tx_desc = &iser_task->desc;
272
273         edtl = ntohl(hdr->data_length);
274
275         /* build the tx desc regd header and add it to the tx desc dto */
276         tx_desc->type = ISCSI_TX_SCSI_COMMAND;
277         iser_create_send_desc(iser_conn->ib_conn, tx_desc);
278
279         if (hdr->flags & ISCSI_FLAG_CMD_READ)
280                 data_buf = &iser_task->data[ISER_DIR_IN];
281         else
282                 data_buf = &iser_task->data[ISER_DIR_OUT];
283
284         if (scsi_sg_count(sc)) { /* using a scatter list */
285                 data_buf->buf  = scsi_sglist(sc);
286                 data_buf->size = scsi_sg_count(sc);
287         }
288
289         data_buf->data_len = scsi_bufflen(sc);
290
291         if (hdr->flags & ISCSI_FLAG_CMD_READ) {
292                 err = iser_prepare_read_cmd(task, edtl);
293                 if (err)
294                         goto send_command_error;
295         }
296         if (hdr->flags & ISCSI_FLAG_CMD_WRITE) {
297                 err = iser_prepare_write_cmd(task,
298                                              task->imm_count,
299                                              task->imm_count +
300                                              task->unsol_r2t.data_length,
301                                              edtl);
302                 if (err)
303                         goto send_command_error;
304         }
305
306         iser_task->status = ISER_TASK_STATUS_STARTED;
307
308         err = iser_post_send(iser_conn->ib_conn, tx_desc);
309         if (!err)
310                 return 0;
311
312 send_command_error:
313         iser_err("conn %p failed task->itt %d err %d\n",conn, task->itt, err);
314         return err;
315 }
316
317 /**
318  * iser_send_data_out - send data out PDU
319  */
320 int iser_send_data_out(struct iscsi_conn *conn,
321                        struct iscsi_task *task,
322                        struct iscsi_data *hdr)
323 {
324         struct iscsi_iser_conn *iser_conn = conn->dd_data;
325         struct iscsi_iser_task *iser_task = task->dd_data;
326         struct iser_tx_desc *tx_desc = NULL;
327         struct iser_regd_buf *regd_buf;
328         unsigned long buf_offset;
329         unsigned long data_seg_len;
330         uint32_t itt;
331         int err = 0;
332         struct ib_sge *tx_dsg;
333
334         itt = (__force uint32_t)hdr->itt;
335         data_seg_len = ntoh24(hdr->dlength);
336         buf_offset   = ntohl(hdr->offset);
337
338         iser_dbg("%s itt %d dseg_len %d offset %d\n",
339                  __func__,(int)itt,(int)data_seg_len,(int)buf_offset);
340
341         tx_desc = kmem_cache_zalloc(ig.desc_cache, GFP_ATOMIC);
342         if (tx_desc == NULL) {
343                 iser_err("Failed to alloc desc for post dataout\n");
344                 return -ENOMEM;
345         }
346
347         tx_desc->type = ISCSI_TX_DATAOUT;
348         tx_desc->iser_header.flags = ISER_VER;
349         memcpy(&tx_desc->iscsi_header, hdr, sizeof(struct iscsi_hdr));
350
351         /* build the tx desc */
352         iser_initialize_task_headers(task, tx_desc);
353
354         regd_buf = &iser_task->rdma_regd[ISER_DIR_OUT];
355         tx_dsg = &tx_desc->tx_sg[1];
356         tx_dsg->addr    = regd_buf->reg.va + buf_offset;
357         tx_dsg->length  = data_seg_len;
358         tx_dsg->lkey    = regd_buf->reg.lkey;
359         tx_desc->num_sge = 2;
360
361         if (buf_offset + data_seg_len > iser_task->data[ISER_DIR_OUT].data_len) {
362                 iser_err("Offset:%ld & DSL:%ld in Data-Out "
363                          "inconsistent with total len:%ld, itt:%d\n",
364                          buf_offset, data_seg_len,
365                          iser_task->data[ISER_DIR_OUT].data_len, itt);
366                 err = -EINVAL;
367                 goto send_data_out_error;
368         }
369         iser_dbg("data-out itt: %d, offset: %ld, sz: %ld\n",
370                  itt, buf_offset, data_seg_len);
371
372
373         err = iser_post_send(iser_conn->ib_conn, tx_desc);
374         if (!err)
375                 return 0;
376
377 send_data_out_error:
378         kmem_cache_free(ig.desc_cache, tx_desc);
379         iser_err("conn %p failed err %d\n",conn, err);
380         return err;
381 }
382
383 int iser_send_control(struct iscsi_conn *conn,
384                       struct iscsi_task *task)
385 {
386         struct iscsi_iser_conn *iser_conn = conn->dd_data;
387         struct iscsi_iser_task *iser_task = task->dd_data;
388         struct iser_tx_desc *mdesc = &iser_task->desc;
389         unsigned long data_seg_len;
390         int err = 0;
391         struct iser_device *device;
392         struct iser_conn *ib_conn = iser_conn->ib_conn;
393
394         /* build the tx desc regd header and add it to the tx desc dto */
395         mdesc->type = ISCSI_TX_CONTROL;
396         iser_create_send_desc(iser_conn->ib_conn, mdesc);
397
398         device = iser_conn->ib_conn->device;
399
400         data_seg_len = ntoh24(task->hdr->dlength);
401
402         if (data_seg_len > 0) {
403                 struct ib_sge *tx_dsg = &mdesc->tx_sg[1];
404                 if (task != conn->login_task) {
405                         iser_err("data present on non login task!!!\n");
406                         goto send_control_error;
407                 }
408
409                 ib_dma_sync_single_for_cpu(device->ib_device,
410                         ib_conn->login_req_dma, task->data_count,
411                         DMA_TO_DEVICE);
412
413                 memcpy(iser_conn->ib_conn->login_req_buf, task->data,
414                                                         task->data_count);
415
416                 ib_dma_sync_single_for_device(device->ib_device,
417                         ib_conn->login_req_dma, task->data_count,
418                         DMA_TO_DEVICE);
419
420                 tx_dsg->addr    = iser_conn->ib_conn->login_req_dma;
421                 tx_dsg->length  = task->data_count;
422                 tx_dsg->lkey    = device->mr->lkey;
423                 mdesc->num_sge = 2;
424         }
425
426         if (task == conn->login_task) {
427                 err = iser_post_recvl(iser_conn->ib_conn);
428                 if (err)
429                         goto send_control_error;
430                 err = iser_post_rx_bufs(conn, task->hdr);
431                 if (err)
432                         goto send_control_error;
433         }
434
435         err = iser_post_send(iser_conn->ib_conn, mdesc);
436         if (!err)
437                 return 0;
438
439 send_control_error:
440         iser_err("conn %p failed err %d\n",conn, err);
441         return err;
442 }
443
444 /**
445  * iser_rcv_dto_completion - recv DTO completion
446  */
447 void iser_rcv_completion(struct iser_rx_desc *rx_desc,
448                          unsigned long rx_xfer_len,
449                          struct iser_conn *ib_conn)
450 {
451         struct iscsi_iser_conn *conn = ib_conn->iser_conn;
452         struct iscsi_hdr *hdr;
453         u64 rx_dma;
454         int rx_buflen, outstanding, count, err;
455
456         /* differentiate between login to all other PDUs */
457         if ((char *)rx_desc == ib_conn->login_resp_buf) {
458                 rx_dma = ib_conn->login_resp_dma;
459                 rx_buflen = ISER_RX_LOGIN_SIZE;
460         } else {
461                 rx_dma = rx_desc->dma_addr;
462                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
463         }
464
465         ib_dma_sync_single_for_cpu(ib_conn->device->ib_device, rx_dma,
466                         rx_buflen, DMA_FROM_DEVICE);
467
468         hdr = &rx_desc->iscsi_header;
469
470         iser_dbg("op 0x%x itt 0x%x dlen %d\n", hdr->opcode,
471                         hdr->itt, (int)(rx_xfer_len - ISER_HEADERS_LEN));
472
473         iscsi_iser_recv(conn->iscsi_conn, hdr,
474                 rx_desc->data, rx_xfer_len - ISER_HEADERS_LEN);
475
476         ib_dma_sync_single_for_device(ib_conn->device->ib_device, rx_dma,
477                         rx_buflen, DMA_FROM_DEVICE);
478
479         /* decrementing conn->post_recv_buf_count only --after-- freeing the   *
480          * task eliminates the need to worry on tasks which are completed in   *
481          * parallel to the execution of iser_conn_term. So the code that waits *
482          * for the posted rx bufs refcount to become zero handles everything   */
483         conn->ib_conn->post_recv_buf_count--;
484
485         if (rx_dma == ib_conn->login_resp_dma)
486                 return;
487
488         outstanding = ib_conn->post_recv_buf_count;
489         if (outstanding + ISER_MIN_POSTED_RX <= ISER_QP_MAX_RECV_DTOS) {
490                 count = min(ISER_QP_MAX_RECV_DTOS - outstanding,
491                                                 ISER_MIN_POSTED_RX);
492                 err = iser_post_recvm(ib_conn, count);
493                 if (err)
494                         iser_err("posting %d rx bufs err %d\n", count, err);
495         }
496 }
497
498 void iser_snd_completion(struct iser_tx_desc *tx_desc,
499                         struct iser_conn *ib_conn)
500 {
501         struct iscsi_task *task;
502         struct iser_device *device = ib_conn->device;
503
504         if (tx_desc->type == ISCSI_TX_DATAOUT) {
505                 ib_dma_unmap_single(device->ib_device, tx_desc->dma_addr,
506                                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
507                 kmem_cache_free(ig.desc_cache, tx_desc);
508         }
509
510         atomic_dec(&ib_conn->post_send_buf_count);
511
512         if (tx_desc->type == ISCSI_TX_CONTROL) {
513                 /* this arithmetic is legal by libiscsi dd_data allocation */
514                 task = (void *) ((long)(void *)tx_desc -
515                                   sizeof(struct iscsi_task));
516                 if (task->hdr->itt == RESERVED_ITT)
517                         iscsi_put_task(task);
518         }
519 }
520
521 void iser_task_rdma_init(struct iscsi_iser_task *iser_task)
522
523 {
524         iser_task->status = ISER_TASK_STATUS_INIT;
525
526         iser_task->dir[ISER_DIR_IN] = 0;
527         iser_task->dir[ISER_DIR_OUT] = 0;
528
529         iser_task->data[ISER_DIR_IN].data_len  = 0;
530         iser_task->data[ISER_DIR_OUT].data_len = 0;
531
532         memset(&iser_task->rdma_regd[ISER_DIR_IN], 0,
533                sizeof(struct iser_regd_buf));
534         memset(&iser_task->rdma_regd[ISER_DIR_OUT], 0,
535                sizeof(struct iser_regd_buf));
536 }
537
538 void iser_task_rdma_finalize(struct iscsi_iser_task *iser_task)
539 {
540         int is_rdma_aligned = 1;
541         struct iser_regd_buf *regd;
542
543         /* if we were reading, copy back to unaligned sglist,
544          * anyway dma_unmap and free the copy
545          */
546         if (iser_task->data_copy[ISER_DIR_IN].copy_buf != NULL) {
547                 is_rdma_aligned = 0;
548                 iser_finalize_rdma_unaligned_sg(iser_task, ISER_DIR_IN);
549         }
550         if (iser_task->data_copy[ISER_DIR_OUT].copy_buf != NULL) {
551                 is_rdma_aligned = 0;
552                 iser_finalize_rdma_unaligned_sg(iser_task, ISER_DIR_OUT);
553         }
554
555         if (iser_task->dir[ISER_DIR_IN]) {
556                 regd = &iser_task->rdma_regd[ISER_DIR_IN];
557                 if (regd->reg.is_fmr)
558                         iser_unreg_mem(&regd->reg);
559         }
560
561         if (iser_task->dir[ISER_DIR_OUT]) {
562                 regd = &iser_task->rdma_regd[ISER_DIR_OUT];
563                 if (regd->reg.is_fmr)
564                         iser_unreg_mem(&regd->reg);
565         }
566
567        /* if the data was unaligned, it was already unmapped and then copied */
568        if (is_rdma_aligned)
569                 iser_dma_unmap_task_data(iser_task);
570 }