Pileus Git - ~andy/linux/blob - drivers/staging/lustre/lustre/llite/rw26.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/lustre/llite/rw26.c
  37  *
  38  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  39  */
  40
  41 #include <linux/kernel.h>
  42 #include <linux/mm.h>
  43 #include <linux/string.h>
  44 #include <linux/stat.h>
  45 #include <linux/errno.h>
  46 #include <linux/unistd.h>
  47 #include <asm/uaccess.h>
  48
  49 #include <linux/migrate.h>
  50 #include <linux/fs.h>
  51 #include <linux/buffer_head.h>
  52 #include <linux/mpage.h>
  53 #include <linux/writeback.h>
  54 #include <linux/pagemap.h>
  55
  56 #define DEBUG_SUBSYSTEM S_LLITE
  57
  58 #include <lustre_lite.h>
  59 #include "llite_internal.h"
  60 #include <linux/lustre_compat25.h>
  61
  62 /**
  63  * Implements Linux VM address_space::invalidatepage() method. This method is
  64  * called when the page is truncate from a file, either as a result of
  65  * explicit truncate, or when inode is removed from memory (as a result of
  66  * final iput(), umount, or memory pressure induced icache shrinking).
  67  *
  68  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  69  * aligned truncate). Lustre leaves partially truncated page in the cache,
  70  * relying on struct inode::i_size to limit further accesses.
  71  */
  72 static void ll_invalidatepage(struct page *vmpage, unsigned int offset,
  73                               unsigned int length)
  74 {
  75         struct inode     *inode;
  76         struct lu_env    *env;
  77         struct cl_page   *page;
  78         struct cl_object *obj;
  79
  80         int refcheck;
  81
  82         LASSERT(PageLocked(vmpage));
  83         LASSERT(!PageWriteback(vmpage));
  84
  85         /*
  86          * It is safe to not check anything in invalidatepage/releasepage
  87          * below because they are run with page locked and all our io is
  88          * happening with locked page too
  89          */
  90         if (offset == 0 && length == PAGE_CACHE_SIZE) {
  91                 env = cl_env_get(&refcheck);
  92                 if (!IS_ERR(env)) {
  93                         inode = vmpage->mapping->host;
  94                         obj = ll_i2info(inode)->lli_clob;
  95                         if (obj != NULL) {
  96                                 page = cl_vmpage_page(vmpage, obj);
  97                                 if (page != NULL) {
  98                                         lu_ref_add(&page->cp_reference,
  99                                                    "delete", vmpage);
 100                                         cl_page_delete(env, page);
 101                                         lu_ref_del(&page->cp_reference,
 102                                                    "delete", vmpage);
 103                                         cl_page_put(env, page);
 104                                 }
 105                         } else
 106                                 LASSERT(vmpage->private == 0);
 107                         cl_env_put(env, &refcheck);
 108                 }
 109         }
 110 }
 111
 112 #ifdef HAVE_RELEASEPAGE_WITH_INT
 113 #define RELEASEPAGE_ARG_TYPE int
 114 #else
 115 #define RELEASEPAGE_ARG_TYPE gfp_t
 116 #endif
 117 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 118 {
 119         struct cl_env_nest nest;
 120         struct lu_env     *env;
 121         struct cl_object  *obj;
 122         struct cl_page    *page;
 123         struct address_space *mapping;
 124         int result;
 125
 126         LASSERT(PageLocked(vmpage));
 127         if (PageWriteback(vmpage) || PageDirty(vmpage))
 128                 return 0;
 129
 130         mapping = vmpage->mapping;
 131         if (mapping == NULL)
 132                 return 1;
 133
 134         obj = ll_i2info(mapping->host)->lli_clob;
 135         if (obj == NULL)
 136                 return 1;
 137
 138         /* 1 for page allocator, 1 for cl_page and 1 for page cache */
 139         if (page_count(vmpage) > 3)
 140                 return 0;
 141
 142         /* TODO: determine what gfp should be used by @gfp_mask. */
 143         env = cl_env_nested_get(&nest);
 144         if (IS_ERR(env))
 145                 /* If we can't allocate an env we won't call cl_page_put()
 146                  * later on which further means it's impossible to drop
 147                  * page refcount by cl_page, so ask kernel to not free
 148                  * this page. */
 149                 return 0;
 150
 151         page = cl_vmpage_page(vmpage, obj);
 152         result = page == NULL;
 153         if (page != NULL) {
 154                 if (!cl_page_in_use(page)) {
 155                         result = 1;
 156                         cl_page_delete(env, page);
 157                 }
 158                 cl_page_put(env, page);
 159         }
 160         cl_env_nested_put(&nest, env);
 161         return result;
 162 }
 163
 164 static int ll_set_page_dirty(struct page *vmpage)
 165 {
 166 #if 0
 167         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 168         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 169         struct vvp_page   *cpg;
 170
 171         /*
 172          * XXX should page method be called here?
 173          */
 174         LASSERT(&obj->co_cl == page->cp_obj);
 175         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 176         /*
 177          * XXX cannot do much here, because page is possibly not locked:
 178          * sys_munmap()->...
 179          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 180          */
 181         vvp_write_pending(obj, cpg);
 182 #endif
 183         return __set_page_dirty_nobuffers(vmpage);
 184 }
 185
 186 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 187
 188 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 189                                     size_t size, struct page ***pages,
 190                                     int *max_pages)
 191 {
 192         int result = -ENOMEM;
 193
 194         /* set an arbitrary limit to prevent arithmetic overflow */
 195         if (size > MAX_DIRECTIO_SIZE) {
 196                 *pages = NULL;
 197                 return -EFBIG;
 198         }
 199
 200         *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 201         *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
 202
 203         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 204         if (*pages) {
 205                 result = get_user_pages_fast(user_addr, *max_pages,
 206                                              (rw == READ), *pages);
 207                 if (unlikely(result <= 0))
 208                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 209         }
 210
 211         return result;
 212 }
 213
 214 /*  ll_free_user_pages - tear down page struct array
 215  *  @pages: array of page struct pointers underlying target buffer */
 216 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 217 {
 218         int i;
 219
 220         for (i = 0; i < npages; i++) {
 221                 if (pages[i] == NULL)
 222                         break;
 223                 if (do_dirty)
 224                         set_page_dirty_lock(pages[i]);
 225                 page_cache_release(pages[i]);
 226         }
 227
 228         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 229 }
 230
 231 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 232                            int rw, struct inode *inode,
 233                            struct ll_dio_pages *pv)
 234 {
 235         struct cl_page    *clp;
 236         struct cl_2queue  *queue;
 237         struct cl_object  *obj = io->ci_obj;
 238         int i;
 239         ssize_t rc = 0;
 240         loff_t file_offset  = pv->ldp_start_offset;
 241         long size          = pv->ldp_size;
 242         int page_count      = pv->ldp_nr;
 243         struct page **pages = pv->ldp_pages;
 244         long page_size      = cl_page_size(obj);
 245         bool do_io;
 246         int  io_pages       = 0;
 247
 248         queue = &io->ci_queue;
 249         cl_2queue_init(queue);
 250         for (i = 0; i < page_count; i++) {
 251                 if (pv->ldp_offsets)
 252                     file_offset = pv->ldp_offsets[i];
 253
 254                 LASSERT(!(file_offset & (page_size - 1)));
 255                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 256                                    pv->ldp_pages[i], CPT_TRANSIENT);
 257                 if (IS_ERR(clp)) {
 258                         rc = PTR_ERR(clp);
 259                         break;
 260                 }
 261
 262                 rc = cl_page_own(env, io, clp);
 263                 if (rc) {
 264                         LASSERT(clp->cp_state == CPS_FREEING);
 265                         cl_page_put(env, clp);
 266                         break;
 267                 }
 268
 269                 do_io = true;
 270
 271                 /* check the page type: if the page is a host page, then do
 272                  * write directly */
 273                 if (clp->cp_type == CPT_CACHEABLE) {
 274                         struct page *vmpage = cl_page_vmpage(env, clp);
 275                         struct page *src_page;
 276                         struct page *dst_page;
 277                         void       *src;
 278                         void       *dst;
 279
 280                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 281                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 282
 283                         src = kmap_atomic(src_page);
 284                         dst = kmap_atomic(dst_page);
 285                         memcpy(dst, src, min(page_size, size));
 286                         kunmap_atomic(dst);
 287                         kunmap_atomic(src);
 288
 289                         /* make sure page will be added to the transfer by
 290                          * cl_io_submit()->...->vvp_page_prep_write(). */
 291                         if (rw == WRITE)
 292                                 set_page_dirty(vmpage);
 293
 294                         if (rw == READ) {
 295                                 /* do not issue the page for read, since it
 296                                  * may reread a ra page which has NOT uptodate
 297                                  * bit set. */
 298                                 cl_page_disown(env, io, clp);
 299                                 do_io = false;
 300                         }
 301                 }
 302
 303                 if (likely(do_io)) {
 304                         cl_2queue_add(queue, clp);
 305
 306                         /*
 307                          * Set page clip to tell transfer formation engine
 308                          * that page has to be sent even if it is beyond KMS.
 309                          */
 310                         cl_page_clip(env, clp, 0, min(size, page_size));
 311
 312                         ++io_pages;
 313                 }
 314
 315                 /* drop the reference count for cl_page_find */
 316                 cl_page_put(env, clp);
 317                 size -= page_size;
 318                 file_offset += page_size;
 319         }
 320
 321         if (rc == 0 && io_pages) {
 322                 rc = cl_io_submit_sync(env, io,
 323                                        rw == READ ? CRT_READ : CRT_WRITE,
 324                                        queue, 0);
 325         }
 326         if (rc == 0)
 327                 rc = pv->ldp_size;
 328
 329         cl_2queue_discard(env, io, queue);
 330         cl_2queue_disown(env, io, queue);
 331         cl_2queue_fini(env, queue);
 332         return rc;
 333 }
 334 EXPORT_SYMBOL(ll_direct_rw_pages);
 335
 336 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 337                                    int rw, struct inode *inode,
 338                                    struct address_space *mapping,
 339                                    size_t size, loff_t file_offset,
 340                                    struct page **pages, int page_count)
 341 {
 342     struct ll_dio_pages pvec = { .ldp_pages     = pages,
 343                                  .ldp_nr           = page_count,
 344                                  .ldp_size       = size,
 345                                  .ldp_offsets      = NULL,
 346                                  .ldp_start_offset = file_offset
 347                                };
 348
 349     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 350 }
 351
 352 #ifdef KMALLOC_MAX_SIZE
 353 #define MAX_MALLOC KMALLOC_MAX_SIZE
 354 #else
 355 #define MAX_MALLOC (128 * 1024)
 356 #endif
 357
 358 /* This is the maximum size of a single O_DIRECT request, based on the
 359  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 360  * representing PAGE_SIZE worth of user data, into a single buffer, and
 361  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 362  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 363 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
 364                       ~(DT_MAX_BRW_SIZE - 1))
 365 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 366                                const struct iovec *iov, loff_t file_offset,
 367                                unsigned long nr_segs)
 368 {
 369         struct lu_env *env;
 370         struct cl_io *io;
 371         struct file *file = iocb->ki_filp;
 372         struct inode *inode = file->f_mapping->host;
 373         struct ccc_object *obj = cl_inode2ccc(inode);
 374         long count = iov_length(iov, nr_segs);
 375         long tot_bytes = 0, result = 0;
 376         struct ll_inode_info *lli = ll_i2info(inode);
 377         unsigned long seg = 0;
 378         long size = MAX_DIO_SIZE;
 379         int refcheck;
 380
 381         if (!lli->lli_has_smd)
 382                 return -EBADF;
 383
 384         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 385         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 386                 return -EINVAL;
 387
 388         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
 389                "offset=%lld=%llx, pages %lu (max %lu)\n",
 390                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 391                file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
 392                MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
 393
 394         /* Check that all user buffers are aligned as well */
 395         for (seg = 0; seg < nr_segs; seg++) {
 396                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 397                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 398                         return -EINVAL;
 399         }
 400
 401         env = cl_env_get(&refcheck);
 402         LASSERT(!IS_ERR(env));
 403         io = ccc_env_io(env)->cui_cl.cis_io;
 404         LASSERT(io != NULL);
 405
 406         /* 0. Need locking between buffered and direct access. and race with
 407          *    size changing by concurrent truncates and writes.
 408          * 1. Need inode mutex to operate transient pages.
 409          */
 410         if (rw == READ)
 411                 mutex_lock(&inode->i_mutex);
 412
 413         LASSERT(obj->cob_transient_pages == 0);
 414         for (seg = 0; seg < nr_segs; seg++) {
 415                 long iov_left = iov[seg].iov_len;
 416                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 417
 418                 if (rw == READ) {
 419                         if (file_offset >= i_size_read(inode))
 420                                 break;
 421                         if (file_offset + iov_left > i_size_read(inode))
 422                                 iov_left = i_size_read(inode) - file_offset;
 423                 }
 424
 425                 while (iov_left > 0) {
 426                         struct page **pages;
 427                         int page_count, max_pages = 0;
 428                         long bytes;
 429
 430                         bytes = min(size, iov_left);
 431                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 432                                                        &pages, &max_pages);
 433                         if (likely(page_count > 0)) {
 434                                 if (unlikely(page_count <  max_pages))
 435                                         bytes = page_count << PAGE_CACHE_SHIFT;
 436                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
 437                                                              file->f_mapping,
 438                                                              bytes, file_offset,
 439                                                              pages, page_count);
 440                                 ll_free_user_pages(pages, max_pages, rw==READ);
 441                         } else if (page_count == 0) {
 442                                 GOTO(out, result = -EFAULT);
 443                         } else {
 444                                 result = page_count;
 445                         }
 446                         if (unlikely(result <= 0)) {
 447                                 /* If we can't allocate a large enough buffer
 448                                  * for the request, shrink it to a smaller
 449                                  * PAGE_SIZE multiple and try again.
 450                                  * We should always be able to kmalloc for a
 451                                  * page worth of page pointers = 4MB on i386. */
 452                                 if (result == -ENOMEM &&
 453                                     size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
 454                                            PAGE_CACHE_SIZE) {
 455                                         size = ((((size / 2) - 1) |
 456                                                  ~CFS_PAGE_MASK) + 1) &
 457                                                 CFS_PAGE_MASK;
 458                                         CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
 459                                                size);
 460                                         continue;
 461                                 }
 462
 463                                 GOTO(out, result);
 464                         }
 465
 466                         tot_bytes += result;
 467                         file_offset += result;
 468                         iov_left -= result;
 469                         user_addr += result;
 470                 }
 471         }
 472 out:
 473         LASSERT(obj->cob_transient_pages == 0);
 474         if (rw == READ)
 475                 mutex_unlock(&inode->i_mutex);
 476
 477         if (tot_bytes > 0) {
 478                 if (rw == WRITE) {
 479                         struct lov_stripe_md *lsm;
 480
 481                         lsm = ccc_inode_lsm_get(inode);
 482                         LASSERT(lsm != NULL);
 483                         lov_stripe_lock(lsm);
 484                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 485                         lov_stripe_unlock(lsm);
 486                         ccc_inode_lsm_put(inode, lsm);
 487                 }
 488         }
 489
 490         cl_env_put(env, &refcheck);
 491         return tot_bytes ? : result;
 492 }
 493
 494 static int ll_write_begin(struct file *file, struct address_space *mapping,
 495                          loff_t pos, unsigned len, unsigned flags,
 496                          struct page **pagep, void **fsdata)
 497 {
 498         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 499         struct page *page;
 500         int rc;
 501         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 502
 503         page = grab_cache_page_write_begin(mapping, index, flags);
 504         if (!page)
 505                 return -ENOMEM;
 506
 507         *pagep = page;
 508
 509         rc = ll_prepare_write(file, page, from, from + len);
 510         if (rc) {
 511                 unlock_page(page);
 512                 page_cache_release(page);
 513         }
 514         return rc;
 515 }
 516
 517 static int ll_write_end(struct file *file, struct address_space *mapping,
 518                         loff_t pos, unsigned len, unsigned copied,
 519                         struct page *page, void *fsdata)
 520 {
 521         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 522         int rc;
 523
 524         rc = ll_commit_write(file, page, from, from + copied);
 525         unlock_page(page);
 526         page_cache_release(page);
 527
 528         return rc ?: copied;
 529 }
 530
 531 #ifdef CONFIG_MIGRATION
 532 int ll_migratepage(struct address_space *mapping,
 533                 struct page *newpage, struct page *page
 534                 , enum migrate_mode mode
 535                 )
 536 {
 537         /* Always fail page migration until we have a proper implementation */
 538         return -EIO;
 539 }
 540 #endif
 541
 542 #ifndef MS_HAS_NEW_AOPS
 543 struct address_space_operations ll_aops = {
 544         .readpage       = ll_readpage,
 545 //      .readpages      = ll_readpages,
 546         .direct_IO      = ll_direct_IO_26,
 547         .writepage      = ll_writepage,
 548         .writepages     = ll_writepages,
 549         .set_page_dirty = ll_set_page_dirty,
 550         .write_begin    = ll_write_begin,
 551         .write_end      = ll_write_end,
 552         .invalidatepage = ll_invalidatepage,
 553         .releasepage    = (void *)ll_releasepage,
 554 #ifdef CONFIG_MIGRATION
 555         .migratepage    = ll_migratepage,
 556 #endif
 557         .bmap      = NULL
 558 };
 559 #else
 560 struct address_space_operations_ext ll_aops = {
 561         .orig_aops.readpage       = ll_readpage,
 562 //      .orig_aops.readpages      = ll_readpages,
 563         .orig_aops.direct_IO      = ll_direct_IO_26,
 564         .orig_aops.writepage      = ll_writepage,
 565         .orig_aops.writepages     = ll_writepages,
 566         .orig_aops.set_page_dirty = ll_set_page_dirty,
 567         .orig_aops.prepare_write  = ll_prepare_write,
 568         .orig_aops.commit_write   = ll_commit_write,
 569         .orig_aops.invalidatepage = ll_invalidatepage,
 570         .orig_aops.releasepage    = ll_releasepage,
 571 #ifdef CONFIG_MIGRATION
 572         .orig_aops.migratepage    = ll_migratepage,
 573 #endif
 574         .orig_aops.bmap    = NULL,
 575         .write_begin    = ll_write_begin,
 576         .write_end      = ll_write_end
 577 };
 578 #endif