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 #ifdef HAVE_MIGRATE_H
  50 #include <linux/migrate.h>
  51 #elif defined(HAVE_MIGRATE_MODE_H)
  52 #include <linux/migrate_mode.h>
  53 #endif
  54 #include <linux/fs.h>
  55 #include <linux/buffer_head.h>
  56 #include <linux/mpage.h>
  57 #include <linux/writeback.h>
  58 #include <linux/stat.h>
  59 #include <asm/uaccess.h>
  60 #include <linux/mm.h>
  61 #include <linux/pagemap.h>
  62
  63 #define DEBUG_SUBSYSTEM S_LLITE
  64
  65 #include "llite_internal.h"
  66 #include <linux/lustre_compat25.h>
  67
  68 /**
  69  * Implements Linux VM address_space::invalidatepage() method. This method is
  70  * called when the page is truncate from a file, either as a result of
  71  * explicit truncate, or when inode is removed from memory (as a result of
  72  * final iput(), umount, or memory pressure induced icache shrinking).
  73  *
  74  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  75  * aligned truncate). Lustre leaves partially truncated page in the cache,
  76  * relying on struct inode::i_size to limit further accesses.
  77  */
  78 static void ll_invalidatepage(struct page *vmpage,
  79 #ifdef HAVE_INVALIDATE_RANGE
  80                                 unsigned int offset, unsigned int length
  81 #else
  82                                 unsigned long offset
  83 #endif
  84                              )
  85 {
  86         struct inode     *inode;
  87         struct lu_env    *env;
  88         struct cl_page   *page;
  89         struct cl_object *obj;
  90
  91         int refcheck;
  92
  93         LASSERT(PageLocked(vmpage));
  94         LASSERT(!PageWriteback(vmpage));
  95
  96         /*
  97          * It is safe to not check anything in invalidatepage/releasepage
  98          * below because they are run with page locked and all our io is
  99          * happening with locked page too
 100          */
 101 #ifdef HAVE_INVALIDATE_RANGE
 102         if (offset == 0 && length == PAGE_CACHE_SIZE) {
 103 #else
 104         if (offset == 0) {
 105 #endif
 106                 env = cl_env_get(&refcheck);
 107                 if (!IS_ERR(env)) {
 108                         inode = vmpage->mapping->host;
 109                         obj = ll_i2info(inode)->lli_clob;
 110                         if (obj != NULL) {
 111                                 page = cl_vmpage_page(vmpage, obj);
 112                                 if (page != NULL) {
 113                                         cl_page_delete(env, page);
 114                                         cl_page_put(env, page);
 115                                 }
 116                         } else
 117                                 LASSERT(vmpage->private == 0);
 118                         cl_env_put(env, &refcheck);
 119                 }
 120         }
 121 }
 122
 123 #ifdef HAVE_RELEASEPAGE_WITH_INT
 124 #define RELEASEPAGE_ARG_TYPE int
 125 #else
 126 #define RELEASEPAGE_ARG_TYPE gfp_t
 127 #endif
 128 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 129 {
 130         struct lu_env           *env;
 131         void                    *cookie;
 132         struct cl_object        *obj;
 133         struct cl_page          *page;
 134         struct address_space    *mapping;
 135         int result = 0;
 136
 137         LASSERT(PageLocked(vmpage));
 138         if (PageWriteback(vmpage) || PageDirty(vmpage))
 139                 return 0;
 140
 141         mapping = vmpage->mapping;
 142         if (mapping == NULL)
 143                 return 1;
 144
 145         obj = ll_i2info(mapping->host)->lli_clob;
 146         if (obj == NULL)
 147                 return 1;
 148
 149         /* 1 for caller, 1 for cl_page and 1 for page cache */
 150         if (page_count(vmpage) > 3)
 151                 return 0;
 152
 153         page = cl_vmpage_page(vmpage, obj);
 154         if (page == NULL)
 155                 return 1;
 156
 157         cookie = cl_env_reenter();
 158         env = cl_env_percpu_get();
 159         LASSERT(!IS_ERR(env));
 160
 161         if (!cl_page_in_use(page)) {
 162                 result = 1;
 163                 cl_page_delete(env, page);
 164         }
 165
 166         /* To use percpu env array, the call path can not be rescheduled;
 167          * otherwise percpu array will be messed if ll_releaspage() called
 168          * again on the same CPU.
 169          *
 170          * If this page holds the last refc of cl_object, the following
 171          * call path may cause reschedule:
 172          *   cl_page_put -> cl_page_free -> cl_object_put ->
 173          *     lu_object_put -> lu_object_free -> lov_delete_raid0.
 174          *
 175          * However, the kernel can't get rid of this inode until all pages have
 176          * been cleaned up. Now that we hold page lock here, it's pretty safe
 177          * that we won't get into object delete path.
 178          */
 179         LASSERT(cl_object_refc(obj) > 1);
 180         cl_page_put(env, page);
 181
 182         cl_env_percpu_put(env);
 183         cl_env_reexit(cookie);
 184         return result;
 185 }
 186
 187 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 188
 189 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 190                                     size_t size, struct page ***pages,
 191                                     int *max_pages)
 192 {
 193         int result = -ENOMEM;
 194
 195         /* set an arbitrary limit to prevent arithmetic overflow */
 196         if (size > MAX_DIRECTIO_SIZE) {
 197                 *pages = NULL;
 198                 return -EFBIG;
 199         }
 200
 201         *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >>
 202                      PAGE_CACHE_SHIFT;
 203         *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
 204
 205         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 206         if (*pages) {
 207                 down_read(&current->mm->mmap_sem);
 208                 result = get_user_pages(current, current->mm, user_addr,
 209                                         *max_pages, (rw == READ), 0, *pages,
 210                                         NULL);
 211                 up_read(&current->mm->mmap_sem);
 212                 if (unlikely(result <= 0))
 213                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 214         }
 215
 216         return result;
 217 }
 218
 219 /*  ll_free_user_pages - tear down page struct array
 220  *  @pages: array of page struct pointers underlying target buffer */
 221 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 222 {
 223         int i;
 224
 225         for (i = 0; i < npages; i++) {
 226                 if (pages[i] == NULL)
 227                         break;
 228                 if (do_dirty)
 229                         set_page_dirty_lock(pages[i]);
 230                 page_cache_release(pages[i]);
 231         }
 232
 233         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 234 }
 235
 236 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 237                            int rw, struct inode *inode,
 238                            struct ll_dio_pages *pv)
 239 {
 240         struct cl_page    *clp;
 241         struct cl_2queue  *queue;
 242         struct cl_object  *obj = io->ci_obj;
 243         int i;
 244         ssize_t rc = 0;
 245         loff_t file_offset  = pv->ldp_start_offset;
 246         long size           = pv->ldp_size;
 247         int page_count      = pv->ldp_nr;
 248         struct page **pages = pv->ldp_pages;
 249         long page_size      = cl_page_size(obj);
 250         bool do_io;
 251         int  io_pages       = 0;
 252         ENTRY;
 253
 254         queue = &io->ci_queue;
 255         cl_2queue_init(queue);
 256         for (i = 0; i < page_count; i++) {
 257                 if (pv->ldp_offsets)
 258                     file_offset = pv->ldp_offsets[i];
 259
 260                 LASSERT(!(file_offset & (page_size - 1)));
 261                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 262                                    pv->ldp_pages[i], CPT_TRANSIENT);
 263                 if (IS_ERR(clp)) {
 264                         rc = PTR_ERR(clp);
 265                         break;
 266                 }
 267
 268                 rc = cl_page_own(env, io, clp);
 269                 if (rc) {
 270                         LASSERT(clp->cp_state == CPS_FREEING);
 271                         cl_page_put(env, clp);
 272                         break;
 273                 }
 274
 275                 do_io = true;
 276
 277                 /* check the page type: if the page is a host page, then do
 278                  * write directly */
 279                 if (clp->cp_type == CPT_CACHEABLE) {
 280                         struct page *vmpage = cl_page_vmpage(clp);
 281                         struct page *src_page;
 282                         struct page *dst_page;
 283                         void       *src;
 284                         void       *dst;
 285
 286                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 287                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 288
 289                         src = ll_kmap_atomic(src_page, KM_USER0);
 290                         dst = ll_kmap_atomic(dst_page, KM_USER1);
 291                         memcpy(dst, src, min(page_size, size));
 292                         ll_kunmap_atomic(dst, KM_USER1);
 293                         ll_kunmap_atomic(src, KM_USER0);
 294
 295                         /* make sure page will be added to the transfer by
 296                          * cl_io_submit()->...->vvp_page_prep_write(). */
 297                         if (rw == WRITE)
 298                                 set_page_dirty(vmpage);
 299
 300                         if (rw == READ) {
 301                                 /* do not issue the page for read, since it
 302                                  * may reread a ra page which has NOT uptodate
 303                                  * bit set. */
 304                                 cl_page_disown(env, io, clp);
 305                                 do_io = false;
 306                         }
 307                 }
 308
 309                 if (likely(do_io)) {
 310                         cl_2queue_add(queue, clp);
 311
 312                         /*
 313                          * Set page clip to tell transfer formation engine
 314                          * that page has to be sent even if it is beyond KMS.
 315                          */
 316                         cl_page_clip(env, clp, 0, min(size, page_size));
 317
 318                         ++io_pages;
 319                 }
 320
 321                 /* drop the reference count for cl_page_find */
 322                 cl_page_put(env, clp);
 323                 size -= page_size;
 324                 file_offset += page_size;
 325         }
 326
 327         if (rc == 0 && io_pages) {
 328                 rc = cl_io_submit_sync(env, io,
 329                                        rw == READ ? CRT_READ : CRT_WRITE,
 330                                        queue, 0);
 331         }
 332         if (rc == 0)
 333                 rc = pv->ldp_size;
 334
 335         cl_2queue_discard(env, io, queue);
 336         cl_2queue_disown(env, io, queue);
 337         cl_2queue_fini(env, queue);
 338         RETURN(rc);
 339 }
 340 EXPORT_SYMBOL(ll_direct_rw_pages);
 341
 342 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 343                                    int rw, struct inode *inode,
 344                                    struct address_space *mapping,
 345                                    size_t size, loff_t file_offset,
 346                                    struct page **pages, int page_count)
 347 {
 348     struct ll_dio_pages pvec = { .ldp_pages        = pages,
 349                                  .ldp_nr           = page_count,
 350                                  .ldp_size         = size,
 351                                  .ldp_offsets      = NULL,
 352                                  .ldp_start_offset = file_offset
 353                                };
 354
 355     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 356 }
 357
 358 #ifdef KMALLOC_MAX_SIZE
 359 #define MAX_MALLOC KMALLOC_MAX_SIZE
 360 #else
 361 #define MAX_MALLOC (128 * 1024)
 362 #endif
 363
 364 /* This is the maximum size of a single O_DIRECT request, based on the
 365  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 366  * representing PAGE_SIZE worth of user data, into a single buffer, and
 367  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 368  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 369 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
 370                       ~(DT_MAX_BRW_SIZE - 1))
 371 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 372                                const struct iovec *iov, loff_t file_offset,
 373                                unsigned long nr_segs)
 374 {
 375         struct lu_env *env;
 376         struct cl_io *io;
 377         struct file *file = iocb->ki_filp;
 378         struct inode *inode = file->f_mapping->host;
 379         struct ccc_object *obj = cl_inode2ccc(inode);
 380         long count = iov_length(iov, nr_segs);
 381         long tot_bytes = 0, result = 0;
 382         struct ll_inode_info *lli = ll_i2info(inode);
 383         unsigned long seg = 0;
 384         long size = MAX_DIO_SIZE;
 385         int refcheck;
 386         ENTRY;
 387
 388         if (!lli->lli_has_smd)
 389                 RETURN(-EBADF);
 390
 391         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 392         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 393                 RETURN(-EINVAL);
 394
 395         CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%lu (max %lu), "
 396                "offset=%lld=%llx, pages %lu (max %lu)\n",
 397                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 398                file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
 399                MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
 400
 401         /* Check that all user buffers are aligned as well */
 402         for (seg = 0; seg < nr_segs; seg++) {
 403                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 404                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 405                         RETURN(-EINVAL);
 406         }
 407
 408         env = cl_env_get(&refcheck);
 409         LASSERT(!IS_ERR(env));
 410         io = ccc_env_io(env)->cui_cl.cis_io;
 411         LASSERT(io != NULL);
 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
 475         if (tot_bytes > 0) {
 476                 struct ccc_io *cio = ccc_env_io(env);
 477
 478                 /* no commit async for direct IO */
 479                 cio->u.write.cui_written += tot_bytes;
 480         }
 481
 482         cl_env_put(env, &refcheck);
 483         RETURN(tot_bytes ? tot_bytes : result);
 484 }
 485
 486 /**
 487  * Prepare partially written-to page for a write.
 488  */
 489 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
 490                                    struct cl_page *pg)
 491 {
 492         struct cl_attr *attr   = ccc_env_thread_attr(env);
 493         struct cl_object *obj  = io->ci_obj;
 494         struct ccc_page *cp    = cl_object_page_slice(obj, pg);
 495         loff_t          offset = cl_offset(obj, ccc_index(cp));
 496         int             result;
 497
 498         cl_object_attr_lock(obj);
 499         result = cl_object_attr_get(env, obj, attr);
 500         cl_object_attr_unlock(obj);
 501         if (result == 0) {
 502                 /*
 503                  * If are writing to a new page, no need to read old data.
 504                  * The extent locking will have updated the KMS, and for our
 505                  * purposes here we can treat it like i_size.
 506                  */
 507                 if (attr->cat_kms <= offset) {
 508                         char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
 509
 510                         memset(kaddr, 0, cl_page_size(obj));
 511                         ll_kunmap_atomic(kaddr, KM_USER0);
 512                 } else if (cp->cpg_defer_uptodate)
 513                         cp->cpg_ra_used = 1;
 514                 else
 515                         result = ll_page_sync_io(env, io, pg, CRT_READ);
 516         }
 517         return result;
 518 }
 519
 520 static int ll_write_begin(struct file *file, struct address_space *mapping,
 521                           loff_t pos, unsigned len, unsigned flags,
 522                           struct page **pagep, void **fsdata)
 523 {
 524         struct ll_cl_context *lcc;
 525         const struct lu_env  *env;
 526         struct cl_io   *io;
 527         struct cl_page *page;
 528
 529         struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
 530         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 531         struct page *vmpage = NULL;
 532         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 533         unsigned to = from + len;
 534         int result = 0;
 535         ENTRY;
 536
 537         CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
 538
 539         lcc = ll_cl_find(file);
 540         if (lcc == NULL)
 541                 GOTO(out, result = -EIO);
 542
 543         env = lcc->lcc_env;
 544         io  = lcc->lcc_io;
 545
 546         /* To avoid deadlock, try to lock page first. */
 547         vmpage = grab_cache_page_nowait(mapping, index);
 548
 549         if (unlikely(vmpage == NULL ||
 550                      PageDirty(vmpage) || PageWriteback(vmpage))) {
 551                 struct ccc_io *cio = ccc_env_io(env);
 552                 struct cl_page_list *plist = &cio->u.write.cui_queue;
 553
 554                 /* if the page is already in dirty cache, we have to commit
 555                  * the pages right now; otherwise, it may cause deadlock
 556                  * because it holds page lock of a dirty page and request for
 557                  * more grants. It's okay for the dirty page to be the first
 558                  * one in commit page list, though. */
 559                 if (vmpage != NULL && plist->pl_nr > 0) {
 560                         unlock_page(vmpage);
 561                         page_cache_release(vmpage);
 562                         vmpage = NULL;
 563                 }
 564
 565                 /* commit pages and then wait for page lock */
 566                 result = vvp_io_write_commit(env, io);
 567                 if (result < 0)
 568                         GOTO(out, result);
 569
 570                 if (vmpage == NULL) {
 571                         vmpage = grab_cache_page_write_begin(mapping, index,
 572                                                              flags);
 573                         if (vmpage == NULL)
 574                                 GOTO(out, result = -ENOMEM);
 575                 }
 576         }
 577
 578         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 579         if (IS_ERR(page))
 580                 GOTO(out, result = PTR_ERR(page));
 581
 582         lcc->lcc_page = page;
 583         lu_ref_add(&page->cp_reference, "cl_io", io);
 584
 585         cl_page_assume(env, io, page);
 586         if (!PageUptodate(vmpage)) {
 587                 /*
 588                  * We're completely overwriting an existing page,
 589                  * so _don't_ set it up to date until commit_write
 590                  */
 591                 if (from == 0 && to == PAGE_SIZE) {
 592                         CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
 593                         POISON_PAGE(vmpage, 0x11);
 594                 } else {
 595                         /* TODO: can be optimized at OSC layer to check if it
 596                          * is a lockless IO. In that case, it's not necessary
 597                          * to read the data. */
 598                         result = ll_prepare_partial_page(env, io, page);
 599                         if (result == 0)
 600                                 SetPageUptodate(vmpage);
 601                 }
 602         }
 603         if (result < 0)
 604                 cl_page_unassume(env, io, page);
 605         EXIT;
 606 out:
 607         if (result < 0) {
 608                 if (vmpage != NULL) {
 609                         unlock_page(vmpage);
 610                         page_cache_release(vmpage);
 611                 }
 612         } else {
 613                 *pagep = vmpage;
 614                 *fsdata = lcc;
 615         }
 616         RETURN(result);
 617 }
 618
 619 static int ll_write_end(struct file *file, struct address_space *mapping,
 620                         loff_t pos, unsigned len, unsigned copied,
 621                         struct page *vmpage, void *fsdata)
 622 {
 623         struct ll_cl_context *lcc = fsdata;
 624         const struct lu_env *env;
 625         struct cl_io *io;
 626         struct ccc_io *cio;
 627         struct cl_page *page;
 628         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 629         bool unplug = false;
 630         int result = 0;
 631         ENTRY;
 632
 633         page_cache_release(vmpage);
 634
 635         LASSERT(lcc != NULL);
 636         env  = lcc->lcc_env;
 637         page = lcc->lcc_page;
 638         io   = lcc->lcc_io;
 639         cio  = ccc_env_io(env);
 640
 641         LASSERT(cl_page_is_owned(page, io));
 642         if (copied > 0) {
 643                 struct cl_page_list *plist = &cio->u.write.cui_queue;
 644
 645                 lcc->lcc_page = NULL; /* page will be queued */
 646
 647                 /* Add it into write queue */
 648                 cl_page_list_add(plist, page);
 649                 if (plist->pl_nr == 1) /* first page */
 650                         cio->u.write.cui_from = from;
 651                 else
 652                         LASSERT(from == 0);
 653                 cio->u.write.cui_to = from + copied;
 654
 655                 /* To address the deadlock in balance_dirty_pages() where
 656                  * this dirty page may be written back in the same thread. */
 657                 if (PageDirty(vmpage))
 658                         unplug = true;
 659
 660                 /* We may have one full RPC, commit it soon */
 661                 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
 662                         unplug = true;
 663
 664                 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
 665                               "queued page: %d.\n", plist->pl_nr);
 666         } else {
 667                 cl_page_disown(env, io, page);
 668
 669                 lcc->lcc_page = NULL;
 670                 lu_ref_del(&page->cp_reference, "cl_io", io);
 671                 cl_page_put(env, page);
 672
 673                 /* page list is not contiguous now, commit it now */
 674                 unplug = true;
 675         }
 676         if (unplug ||
 677             file->f_flags & O_SYNC || IS_SYNC(file->f_dentry->d_inode))
 678                 result = vvp_io_write_commit(env, io);
 679
 680         RETURN(result >= 0 ? copied : result);
 681 }
 682
 683 #ifdef CONFIG_MIGRATION
 684 static int ll_migratepage(struct address_space *mapping,
 685                           struct page *newpage, struct page *page
 686 #ifdef HAVE_MIGRATEPAGE_4ARGS
 687                           , enum migrate_mode mode
 688 #endif
 689         )
 690 {
 691         /* Always fail page migration until we have a proper implementation */
 692         return -EIO;
 693 }
 694 #endif
 695
 696 #ifndef MS_HAS_NEW_AOPS
 697 const struct address_space_operations ll_aops = {
 698         .readpage       = ll_readpage,
 699         .direct_IO      = ll_direct_IO_26,
 700         .writepage      = ll_writepage,
 701         .writepages     = ll_writepages,
 702         .set_page_dirty = __set_page_dirty_nobuffers,
 703         .write_begin    = ll_write_begin,
 704         .write_end      = ll_write_end,
 705         .invalidatepage = ll_invalidatepage,
 706         .releasepage    = (void *)ll_releasepage,
 707 #ifdef CONFIG_MIGRATION
 708         .migratepage    = ll_migratepage,
 709 #endif
 710 };
 711 #else
 712 const struct address_space_operations_ext ll_aops = {
 713         .orig_aops.readpage             = ll_readpage,
 714         .orig_aops.direct_IO            = ll_direct_IO_26,
 715         .orig_aops.writepage            = ll_writepage,
 716         .orig_aops.writepages           = ll_writepages,
 717         .orig_aops.set_page_dirty       = __set_page_dirty_nobuffers,
 718         .orig_aops.invalidatepage       = ll_invalidatepage,
 719         .orig_aops.releasepage          = ll_releasepage,
 720 #ifdef CONFIG_MIGRATION
 721         .orig_aops.migratepage          = ll_migratepage,
 722 #endif
 723         .write_begin                    = ll_write_begin,
 724         .write_end                      = ll_write_end
 725 };
 726 #endif