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