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 <lustre_lite.h>
  66 #include "llite_internal.h"
  67 #include <linux/lustre_compat25.h>
  68
  69 /**
  70  * Implements Linux VM address_space::invalidatepage() method. This method is
  71  * called when the page is truncate from a file, either as a result of
  72  * explicit truncate, or when inode is removed from memory (as a result of
  73  * final iput(), umount, or memory pressure induced icache shrinking).
  74  *
  75  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  76  * aligned truncate). Lustre leaves partially truncated page in the cache,
  77  * relying on struct inode::i_size to limit further accesses.
  78  */
  79 static void ll_invalidatepage(struct page *vmpage,
  80 #ifdef HAVE_INVALIDATE_RANGE
  81                                 unsigned int offset, unsigned int length
  82 #else
  83                                 unsigned long offset
  84 #endif
  85                              )
  86 {
  87         struct inode     *inode;
  88         struct lu_env    *env;
  89         struct cl_page   *page;
  90         struct cl_object *obj;
  91
  92         int refcheck;
  93
  94         LASSERT(PageLocked(vmpage));
  95         LASSERT(!PageWriteback(vmpage));
  96
  97         /*
  98          * It is safe to not check anything in invalidatepage/releasepage
  99          * below because they are run with page locked and all our io is
 100          * happening with locked page too
 101          */
 102 #ifdef HAVE_INVALIDATE_RANGE
 103         if (offset == 0 && length == PAGE_CACHE_SIZE) {
 104 #else
 105         if (offset == 0) {
 106 #endif
 107                 env = cl_env_get(&refcheck);
 108                 if (!IS_ERR(env)) {
 109                         inode = vmpage->mapping->host;
 110                         obj = ll_i2info(inode)->lli_clob;
 111                         if (obj != NULL) {
 112                                 page = cl_vmpage_page(vmpage, obj);
 113                                 if (page != NULL) {
 114                                         cl_page_delete(env, page);
 115                                         cl_page_put(env, page);
 116                                 }
 117                         } else
 118                                 LASSERT(vmpage->private == 0);
 119                         cl_env_put(env, &refcheck);
 120                 }
 121         }
 122 }
 123
 124 #ifdef HAVE_RELEASEPAGE_WITH_INT
 125 #define RELEASEPAGE_ARG_TYPE int
 126 #else
 127 #define RELEASEPAGE_ARG_TYPE gfp_t
 128 #endif
 129 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 130 {
 131         struct lu_env           *env;
 132         void                    *cookie;
 133         struct cl_object        *obj;
 134         struct cl_page          *page;
 135         struct address_space    *mapping;
 136         int result = 0;
 137
 138         LASSERT(PageLocked(vmpage));
 139         if (PageWriteback(vmpage) || PageDirty(vmpage))
 140                 return 0;
 141
 142         mapping = vmpage->mapping;
 143         if (mapping == NULL)
 144                 return 1;
 145
 146         obj = ll_i2info(mapping->host)->lli_clob;
 147         if (obj == NULL)
 148                 return 1;
 149
 150         /* 1 for caller, 1 for cl_page and 1 for page cache */
 151         if (page_count(vmpage) > 3)
 152                 return 0;
 153
 154         page = cl_vmpage_page(vmpage, obj);
 155         if (page == NULL)
 156                 return 1;
 157
 158         cookie = cl_env_reenter();
 159         env = cl_env_percpu_get();
 160         LASSERT(!IS_ERR(env));
 161
 162         if (!cl_page_in_use(page)) {
 163                 result = 1;
 164                 cl_page_delete(env, page);
 165         }
 166
 167         /* To use percpu env array, the call path can not be rescheduled;
 168          * otherwise percpu array will be messed if ll_releaspage() called
 169          * again on the same CPU.
 170          *
 171          * If this page holds the last refc of cl_object, the following
 172          * call path may cause reschedule:
 173          *   cl_page_put -> cl_page_free -> cl_object_put ->
 174          *     lu_object_put -> lu_object_free -> lov_delete_raid0 ->
 175          *     cl_locks_prune.
 176          *
 177          * However, the kernel can't get rid of this inode until all pages have
 178          * been cleaned up. Now that we hold page lock here, it's pretty safe
 179          * that we won't get into object delete path.
 180          */
 181         LASSERT(cl_object_refc(obj) > 1);
 182         cl_page_put(env, page);
 183
 184         cl_env_percpu_put(env);
 185         cl_env_reexit(cookie);
 186         return result;
 187 }
 188
 189 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 190
 191 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 192                                     size_t size, struct page ***pages,
 193                                     int *max_pages)
 194 {
 195         int result = -ENOMEM;
 196
 197         /* set an arbitrary limit to prevent arithmetic overflow */
 198         if (size > MAX_DIRECTIO_SIZE) {
 199                 *pages = NULL;
 200                 return -EFBIG;
 201         }
 202
 203         *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >>
 204                      PAGE_CACHE_SHIFT;
 205         *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
 206
 207         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 208         if (*pages) {
 209                 down_read(&current->mm->mmap_sem);
 210                 result = get_user_pages(current, current->mm, user_addr,
 211                                         *max_pages, (rw == READ), 0, *pages,
 212                                         NULL);
 213                 up_read(&current->mm->mmap_sem);
 214                 if (unlikely(result <= 0))
 215                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 216         }
 217
 218         return result;
 219 }
 220
 221 /*  ll_free_user_pages - tear down page struct array
 222  *  @pages: array of page struct pointers underlying target buffer */
 223 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 224 {
 225         int i;
 226
 227         for (i = 0; i < npages; i++) {
 228                 if (pages[i] == NULL)
 229                         break;
 230                 if (do_dirty)
 231                         set_page_dirty_lock(pages[i]);
 232                 page_cache_release(pages[i]);
 233         }
 234
 235         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 236 }
 237
 238 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 239                            int rw, struct inode *inode,
 240                            struct ll_dio_pages *pv)
 241 {
 242         struct cl_page    *clp;
 243         struct cl_2queue  *queue;
 244         struct cl_object  *obj = io->ci_obj;
 245         int i;
 246         ssize_t rc = 0;
 247         loff_t file_offset  = pv->ldp_start_offset;
 248         long size           = pv->ldp_size;
 249         int page_count      = pv->ldp_nr;
 250         struct page **pages = pv->ldp_pages;
 251         long page_size      = cl_page_size(obj);
 252         bool do_io;
 253         int  io_pages       = 0;
 254         ENTRY;
 255
 256         queue = &io->ci_queue;
 257         cl_2queue_init(queue);
 258         for (i = 0; i < page_count; i++) {
 259                 if (pv->ldp_offsets)
 260                     file_offset = pv->ldp_offsets[i];
 261
 262                 LASSERT(!(file_offset & (page_size - 1)));
 263                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 264                                    pv->ldp_pages[i], CPT_TRANSIENT);
 265                 if (IS_ERR(clp)) {
 266                         rc = PTR_ERR(clp);
 267                         break;
 268                 }
 269
 270                 rc = cl_page_own(env, io, clp);
 271                 if (rc) {
 272                         LASSERT(clp->cp_state == CPS_FREEING);
 273                         cl_page_put(env, clp);
 274                         break;
 275                 }
 276
 277                 do_io = true;
 278
 279                 /* check the page type: if the page is a host page, then do
 280                  * write directly */
 281                 if (clp->cp_type == CPT_CACHEABLE) {
 282                         struct page *vmpage = cl_page_vmpage(clp);
 283                         struct page *src_page;
 284                         struct page *dst_page;
 285                         void       *src;
 286                         void       *dst;
 287
 288                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 289                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 290
 291                         src = ll_kmap_atomic(src_page, KM_USER0);
 292                         dst = ll_kmap_atomic(dst_page, KM_USER1);
 293                         memcpy(dst, src, min(page_size, size));
 294                         ll_kunmap_atomic(dst, KM_USER1);
 295                         ll_kunmap_atomic(src, KM_USER0);
 296
 297                         /* make sure page will be added to the transfer by
 298                          * cl_io_submit()->...->vvp_page_prep_write(). */
 299                         if (rw == WRITE)
 300                                 set_page_dirty(vmpage);
 301
 302                         if (rw == READ) {
 303                                 /* do not issue the page for read, since it
 304                                  * may reread a ra page which has NOT uptodate
 305                                  * bit set. */
 306                                 cl_page_disown(env, io, clp);
 307                                 do_io = false;
 308                         }
 309                 }
 310
 311                 if (likely(do_io)) {
 312                         cl_2queue_add(queue, clp);
 313
 314                         /*
 315                          * Set page clip to tell transfer formation engine
 316                          * that page has to be sent even if it is beyond KMS.
 317                          */
 318                         cl_page_clip(env, clp, 0, min(size, page_size));
 319
 320                         ++io_pages;
 321                 }
 322
 323                 /* drop the reference count for cl_page_find */
 324                 cl_page_put(env, clp);
 325                 size -= page_size;
 326                 file_offset += page_size;
 327         }
 328
 329         if (rc == 0 && io_pages) {
 330                 rc = cl_io_submit_sync(env, io,
 331                                        rw == READ ? CRT_READ : CRT_WRITE,
 332                                        queue, 0);
 333         }
 334         if (rc == 0)
 335                 rc = pv->ldp_size;
 336
 337         cl_2queue_discard(env, io, queue);
 338         cl_2queue_disown(env, io, queue);
 339         cl_2queue_fini(env, queue);
 340         RETURN(rc);
 341 }
 342 EXPORT_SYMBOL(ll_direct_rw_pages);
 343
 344 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 345                                    int rw, struct inode *inode,
 346                                    struct address_space *mapping,
 347                                    size_t size, loff_t file_offset,
 348                                    struct page **pages, int page_count)
 349 {
 350     struct ll_dio_pages pvec = { .ldp_pages        = pages,
 351                                  .ldp_nr           = page_count,
 352                                  .ldp_size         = size,
 353                                  .ldp_offsets      = NULL,
 354                                  .ldp_start_offset = file_offset
 355                                };
 356
 357     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 358 }
 359
 360 #ifdef KMALLOC_MAX_SIZE
 361 #define MAX_MALLOC KMALLOC_MAX_SIZE
 362 #else
 363 #define MAX_MALLOC (128 * 1024)
 364 #endif
 365
 366 /* This is the maximum size of a single O_DIRECT request, based on the
 367  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 368  * representing PAGE_SIZE worth of user data, into a single buffer, and
 369  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 370  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 371 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
 372                       ~(DT_MAX_BRW_SIZE - 1))
 373 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 374                                const struct iovec *iov, loff_t file_offset,
 375                                unsigned long nr_segs)
 376 {
 377         struct lu_env *env;
 378         struct cl_io *io;
 379         struct file *file = iocb->ki_filp;
 380         struct inode *inode = file->f_mapping->host;
 381         struct ccc_object *obj = cl_inode2ccc(inode);
 382         long count = iov_length(iov, nr_segs);
 383         long tot_bytes = 0, result = 0;
 384         struct ll_inode_info *lli = ll_i2info(inode);
 385         unsigned long seg = 0;
 386         long size = MAX_DIO_SIZE;
 387         int refcheck;
 388         ENTRY;
 389
 390         if (!lli->lli_has_smd)
 391                 RETURN(-EBADF);
 392
 393         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 394         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 395                 RETURN(-EINVAL);
 396
 397         CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%lu (max %lu), "
 398                "offset=%lld=%llx, pages %lu (max %lu)\n",
 399                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 400                file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
 401                MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
 402
 403         /* Check that all user buffers are aligned as well */
 404         for (seg = 0; seg < nr_segs; seg++) {
 405                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 406                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 407                         RETURN(-EINVAL);
 408         }
 409
 410         env = cl_env_get(&refcheck);
 411         LASSERT(!IS_ERR(env));
 412         io = ccc_env_io(env)->cui_cl.cis_io;
 413         LASSERT(io != NULL);
 414
 415         /* 0. Need locking between buffered and direct access. and race with
 416          *    size changing by concurrent truncates and writes.
 417          * 1. Need inode mutex to operate transient pages.
 418          */
 419         if (rw == READ)
 420                 mutex_lock(&inode->i_mutex);
 421
 422         LASSERT(obj->cob_transient_pages == 0);
 423         for (seg = 0; seg < nr_segs; seg++) {
 424                 long iov_left = iov[seg].iov_len;
 425                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 426
 427                 if (rw == READ) {
 428                         if (file_offset >= i_size_read(inode))
 429                                 break;
 430                         if (file_offset + iov_left > i_size_read(inode))
 431                                 iov_left = i_size_read(inode) - file_offset;
 432                 }
 433
 434                 while (iov_left > 0) {
 435                         struct page **pages;
 436                         int page_count, max_pages = 0;
 437                         long bytes;
 438
 439                         bytes = min(size, iov_left);
 440                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 441                                                        &pages, &max_pages);
 442                         if (likely(page_count > 0)) {
 443                                 if (unlikely(page_count <  max_pages))
 444                                         bytes = page_count << PAGE_CACHE_SHIFT;
 445                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
 446                                                              file->f_mapping,
 447                                                              bytes, file_offset,
 448                                                              pages, page_count);
 449                                 ll_free_user_pages(pages, max_pages, rw==READ);
 450                         } else if (page_count == 0) {
 451                                 GOTO(out, result = -EFAULT);
 452                         } else {
 453                                 result = page_count;
 454                         }
 455                         if (unlikely(result <= 0)) {
 456                                 /* If we can't allocate a large enough buffer
 457                                  * for the request, shrink it to a smaller
 458                                  * PAGE_SIZE multiple and try again.
 459                                  * We should always be able to kmalloc for a
 460                                  * page worth of page pointers = 4MB on i386. */
 461                                 if (result == -ENOMEM &&
 462                                     size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
 463                                            PAGE_CACHE_SIZE) {
 464                                         size = ((((size / 2) - 1) |
 465                                                  ~CFS_PAGE_MASK) + 1) &
 466                                                 CFS_PAGE_MASK;
 467                                         CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
 468                                                size);
 469                                         continue;
 470                                 }
 471
 472                                 GOTO(out, result);
 473                         }
 474
 475                         tot_bytes += result;
 476                         file_offset += result;
 477                         iov_left -= result;
 478                         user_addr += result;
 479                 }
 480         }
 481 out:
 482         LASSERT(obj->cob_transient_pages == 0);
 483         if (rw == READ)
 484                 mutex_unlock(&inode->i_mutex);
 485
 486         if (tot_bytes > 0) {
 487                 struct ccc_io *cio = ccc_env_io(env);
 488
 489                 /* no commit async for direct IO */
 490                 cio->u.write.cui_written += tot_bytes;
 491         }
 492
 493         cl_env_put(env, &refcheck);
 494         RETURN(tot_bytes ? tot_bytes : result);
 495 }
 496
 497 /**
 498  * Prepare partially written-to page for a write.
 499  */
 500 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
 501                                    struct cl_page *pg)
 502 {
 503         struct cl_attr *attr   = ccc_env_thread_attr(env);
 504         struct cl_object *obj  = io->ci_obj;
 505         struct ccc_page *cp    = cl_object_page_slice(obj, pg);
 506         loff_t          offset = cl_offset(obj, ccc_index(cp));
 507         int             result;
 508
 509         cl_object_attr_lock(obj);
 510         result = cl_object_attr_get(env, obj, attr);
 511         cl_object_attr_unlock(obj);
 512         if (result == 0) {
 513                 /*
 514                  * If are writing to a new page, no need to read old data.
 515                  * The extent locking will have updated the KMS, and for our
 516                  * purposes here we can treat it like i_size.
 517                  */
 518                 if (attr->cat_kms <= offset) {
 519                         char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
 520
 521                         memset(kaddr, 0, cl_page_size(obj));
 522                         ll_kunmap_atomic(kaddr, KM_USER0);
 523                 } else if (cp->cpg_defer_uptodate)
 524                         cp->cpg_ra_used = 1;
 525                 else
 526                         result = ll_page_sync_io(env, io, pg, CRT_READ);
 527         }
 528         return result;
 529 }
 530
 531 static int ll_write_begin(struct file *file, struct address_space *mapping,
 532                           loff_t pos, unsigned len, unsigned flags,
 533                           struct page **pagep, void **fsdata)
 534 {
 535         struct ll_cl_context *lcc;
 536         struct lu_env  *env;
 537         struct cl_io   *io;
 538         struct cl_page *page;
 539
 540         struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
 541         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 542         struct page *vmpage = NULL;
 543         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 544         unsigned to = from + len;
 545         int result = 0;
 546         ENTRY;
 547
 548         CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
 549
 550         lcc = ll_cl_init(file, NULL);
 551         if (IS_ERR(lcc))
 552                 GOTO(out, result = PTR_ERR(lcc));
 553
 554         env = lcc->lcc_env;
 555         io  = lcc->lcc_io;
 556
 557         /* To avoid deadlock, try to lock page first. */
 558         vmpage = grab_cache_page_nowait(mapping, index);
 559         if (unlikely(vmpage == NULL ||
 560                      PageDirty(vmpage) || PageWriteback(vmpage))) {
 561                 struct ccc_io *cio = ccc_env_io(env);
 562                 struct cl_page_list *plist = &cio->u.write.cui_queue;
 563
 564                 /* if the page is already in dirty cache, we have to commit
 565                  * the pages right now; otherwise, it may cause deadlock
 566                  * because it holds page lock of a dirty page and request for
 567                  * more grants. It's okay for the dirty page to be the first
 568                  * one in commit page list, though. */
 569                 if (vmpage != NULL && plist->pl_nr > 0) {
 570                         unlock_page(vmpage);
 571                         page_cache_release(vmpage);
 572                         vmpage = NULL;
 573                 }
 574
 575                 /* commit pages and then wait for page lock */
 576                 result = vvp_io_write_commit(env, io);
 577                 if (result < 0)
 578                         GOTO(out, result);
 579
 580                 if (vmpage == NULL) {
 581                         vmpage = grab_cache_page_write_begin(mapping, index,
 582                                                              flags);
 583                         if (vmpage == NULL)
 584                                 GOTO(out, result = -ENOMEM);
 585                 }
 586         }
 587
 588         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 589         if (IS_ERR(page))
 590                 GOTO(out, result = PTR_ERR(page));
 591
 592         lcc->lcc_page = page;
 593         lu_ref_add(&page->cp_reference, "cl_io", io);
 594
 595         cl_page_assume(env, io, page);
 596         if (!PageUptodate(vmpage)) {
 597                 /*
 598                  * We're completely overwriting an existing page,
 599                  * so _don't_ set it up to date until commit_write
 600                  */
 601                 if (from == 0 && to == PAGE_SIZE) {
 602                         CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
 603                         POISON_PAGE(vmpage, 0x11);
 604                 } else {
 605                         /* TODO: can be optimized at OSC layer to check if it
 606                          * is a lockless IO. In that case, it's not necessary
 607                          * to read the data. */
 608                         result = ll_prepare_partial_page(env, io, page);
 609                         if (result == 0)
 610                                 SetPageUptodate(vmpage);
 611                 }
 612         }
 613         if (result < 0)
 614                 cl_page_unassume(env, io, page);
 615         EXIT;
 616 out:
 617         if (result < 0) {
 618                 if (vmpage != NULL) {
 619                         unlock_page(vmpage);
 620                         page_cache_release(vmpage);
 621                 }
 622                 if (!IS_ERR(lcc))
 623                         ll_cl_fini(lcc);
 624         } else {
 625                 *pagep = vmpage;
 626                 *fsdata = lcc;
 627         }
 628         RETURN(result);
 629 }
 630
 631 static int ll_write_end(struct file *file, struct address_space *mapping,
 632                         loff_t pos, unsigned len, unsigned copied,
 633                         struct page *vmpage, void *fsdata)
 634 {
 635         struct ll_cl_context *lcc = fsdata;
 636         struct lu_env *env;
 637         struct cl_io *io;
 638         struct ccc_io *cio;
 639         struct cl_page *page;
 640         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 641         bool unplug = false;
 642         int result = 0;
 643         ENTRY;
 644
 645         page_cache_release(vmpage);
 646
 647         LASSERT(lcc != NULL);
 648         env  = lcc->lcc_env;
 649         page = lcc->lcc_page;
 650         io   = lcc->lcc_io;
 651         cio  = ccc_env_io(env);
 652
 653         LASSERT(cl_page_is_owned(page, io));
 654         if (copied > 0) {
 655                 struct cl_page_list *plist = &cio->u.write.cui_queue;
 656
 657                 lcc->lcc_page = NULL; /* page will be queued */
 658
 659                 /* Add it into write queue */
 660                 cl_page_list_add(plist, page);
 661                 if (plist->pl_nr == 1) /* first page */
 662                         cio->u.write.cui_from = from;
 663                 else
 664                         LASSERT(from == 0);
 665                 cio->u.write.cui_to = from + copied;
 666
 667                 /* To address the deadlock in balance_dirty_pages() where
 668                  * this dirty page may be written back in the same thread. */
 669                 if (PageDirty(vmpage))
 670                         unplug = true;
 671
 672                 /* We may have one full RPC, commit it soon */
 673                 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
 674                         unplug = true;
 675
 676                 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
 677                               "queued page: %d.\n", plist->pl_nr);
 678         } else {
 679                 cl_page_disown(env, io, page);
 680
 681                 /* page list is not contiguous now, commit it now */
 682                 unplug = true;
 683         }
 684
 685         if (unplug ||
 686             file->f_flags & O_SYNC || IS_SYNC(file->f_dentry->d_inode))
 687                 result = vvp_io_write_commit(env, io);
 688
 689         ll_cl_fini(lcc);
 690         RETURN(result >= 0 ? copied : result);
 691 }
 692
 693 #ifdef CONFIG_MIGRATION
 694 static int ll_migratepage(struct address_space *mapping,
 695                           struct page *newpage, struct page *page
 696 #ifdef HAVE_MIGRATEPAGE_4ARGS
 697                           , enum migrate_mode mode
 698 #endif
 699         )
 700 {
 701         /* Always fail page migration until we have a proper implementation */
 702         return -EIO;
 703 }
 704 #endif
 705
 706 #ifndef MS_HAS_NEW_AOPS
 707 const struct address_space_operations ll_aops = {
 708         .readpage       = ll_readpage,
 709         .direct_IO      = ll_direct_IO_26,
 710         .writepage      = ll_writepage,
 711         .writepages     = ll_writepages,
 712         .set_page_dirty = __set_page_dirty_nobuffers,
 713         .write_begin    = ll_write_begin,
 714         .write_end      = ll_write_end,
 715         .invalidatepage = ll_invalidatepage,
 716         .releasepage    = (void *)ll_releasepage,
 717 #ifdef CONFIG_MIGRATION
 718         .migratepage    = ll_migratepage,
 719 #endif
 720 };
 721 #else
 722 const struct address_space_operations_ext ll_aops = {
 723         .orig_aops.readpage             = ll_readpage,
 724         .orig_aops.direct_IO            = ll_direct_IO_26,
 725         .orig_aops.writepage            = ll_writepage,
 726         .orig_aops.writepages           = ll_writepages,
 727         .orig_aops.set_page_dirty       = __set_page_dirty_nobuffers,
 728         .orig_aops.invalidatepage       = ll_invalidatepage,
 729         .orig_aops.releasepage          = ll_releasepage,
 730 #ifdef CONFIG_MIGRATION
 731         .orig_aops.migratepage          = ll_migratepage,
 732 #endif
 733         .write_begin                    = ll_write_begin,
 734         .write_end                      = ll_write_end
 735 };
 736 #endif