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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2014, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/lustre/llite/rw26.c
  33  *
  34  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  35  */
  36
  37 #include <linux/kernel.h>
  38 #include <linux/mm.h>
  39 #include <linux/string.h>
  40 #include <linux/stat.h>
  41 #include <linux/errno.h>
  42 #include <linux/unistd.h>
  43 #include <asm/uaccess.h>
  44
  45 #ifdef HAVE_MIGRATE_H
  46 #include <linux/migrate.h>
  47 #elif defined(HAVE_MIGRATE_MODE_H)
  48 #include <linux/migrate_mode.h>
  49 #endif
  50 #include <linux/fs.h>
  51 #include <linux/buffer_head.h>
  52 #include <linux/mpage.h>
  53 #include <linux/writeback.h>
  54 #include <linux/stat.h>
  55 #include <asm/uaccess.h>
  56 #include <linux/mm.h>
  57 #include <linux/pagemap.h>
  58
  59 #define DEBUG_SUBSYSTEM S_LLITE
  60
  61 #include "llite_internal.h"
  62 #include <lustre_compat.h>
  63
  64 /**
  65  * Implements Linux VM address_space::invalidatepage() method. This method is
  66  * called when the page is truncate from a file, either as a result of
  67  * explicit truncate, or when inode is removed from memory (as a result of
  68  * final iput(), umount, or memory pressure induced icache shrinking).
  69  *
  70  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  71  * aligned truncate). Lustre leaves partially truncated page in the cache,
  72  * relying on struct inode::i_size to limit further accesses.
  73  */
  74 static void ll_invalidatepage(struct page *vmpage,
  75 #ifdef HAVE_INVALIDATE_RANGE
  76                                 unsigned int offset, unsigned int length
  77 #else
  78                                 unsigned long offset
  79 #endif
  80                              )
  81 {
  82         struct inode     *inode;
  83         struct lu_env    *env;
  84         struct cl_page   *page;
  85         struct cl_object *obj;
  86
  87         LASSERT(PageLocked(vmpage));
  88         LASSERT(!PageWriteback(vmpage));
  89
  90         /*
  91          * It is safe to not check anything in invalidatepage/releasepage
  92          * below because they are run with page locked and all our io is
  93          * happening with locked page too
  94          */
  95 #ifdef HAVE_INVALIDATE_RANGE
  96         if (offset == 0 && length == PAGE_SIZE) {
  97 #else
  98         if (offset == 0) {
  99 #endif
 100                 /* See the comment in ll_releasepage() */
 101                 env = cl_env_percpu_get();
 102                 LASSERT(!IS_ERR(env));
 103
 104                 inode = vmpage->mapping->host;
 105                 obj = ll_i2info(inode)->lli_clob;
 106                 if (obj != NULL) {
 107                         page = cl_vmpage_page(vmpage, obj);
 108                         if (page != NULL) {
 109                                 cl_page_delete(env, page);
 110                                 cl_page_put(env, page);
 111                         }
 112                 } else
 113                         LASSERT(vmpage->private == 0);
 114
 115                 cl_env_percpu_put(env);
 116         }
 117 }
 118
 119 #ifdef HAVE_RELEASEPAGE_WITH_INT
 120 #define RELEASEPAGE_ARG_TYPE int
 121 #else
 122 #define RELEASEPAGE_ARG_TYPE gfp_t
 123 #endif
 124 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 125 {
 126         struct lu_env           *env;
 127         struct cl_object        *obj;
 128         struct cl_page          *page;
 129         struct address_space    *mapping;
 130         int result = 0;
 131
 132         LASSERT(PageLocked(vmpage));
 133         if (PageWriteback(vmpage) || PageDirty(vmpage))
 134                 return 0;
 135
 136         mapping = vmpage->mapping;
 137         if (mapping == NULL)
 138                 return 1;
 139
 140         obj = ll_i2info(mapping->host)->lli_clob;
 141         if (obj == NULL)
 142                 return 1;
 143
 144         /* 1 for caller, 1 for cl_page and 1 for page cache */
 145         if (page_count(vmpage) > 3)
 146                 return 0;
 147
 148         page = cl_vmpage_page(vmpage, obj);
 149         if (page == NULL)
 150                 return 1;
 151
 152         env = cl_env_percpu_get();
 153         LASSERT(!IS_ERR(env));
 154
 155         if (!cl_page_in_use(page)) {
 156                 result = 1;
 157                 cl_page_delete(env, page);
 158         }
 159
 160         /* To use percpu env array, the call path can not be rescheduled;
 161          * otherwise percpu array will be messed if ll_releaspage() called
 162          * again on the same CPU.
 163          *
 164          * If this page holds the last refc of cl_object, the following
 165          * call path may cause reschedule:
 166          *   cl_page_put -> cl_page_free -> cl_object_put ->
 167          *     lu_object_put -> lu_object_free -> lov_delete_raid0.
 168          *
 169          * However, the kernel can't get rid of this inode until all pages have
 170          * been cleaned up. Now that we hold page lock here, it's pretty safe
 171          * that we won't get into object delete path.
 172          */
 173         LASSERT(cl_object_refc(obj) > 1);
 174         cl_page_put(env, page);
 175
 176         cl_env_percpu_put(env);
 177         return result;
 178 }
 179
 180 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 181
 182 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 183                            int rw, struct inode *inode,
 184                            struct ll_dio_pages *pv)
 185 {
 186         struct cl_page    *clp;
 187         struct cl_2queue  *queue;
 188         struct cl_object  *obj = io->ci_obj;
 189         int i;
 190         ssize_t rc = 0;
 191         loff_t file_offset  = pv->ldp_start_offset;
 192         size_t size         = pv->ldp_size;
 193         int page_count      = pv->ldp_nr;
 194         struct page **pages = pv->ldp_pages;
 195         size_t page_size    = cl_page_size(obj);
 196         bool do_io;
 197         int  io_pages       = 0;
 198         ENTRY;
 199
 200         queue = &io->ci_queue;
 201         cl_2queue_init(queue);
 202         for (i = 0; i < page_count; i++) {
 203                 if (pv->ldp_offsets)
 204                     file_offset = pv->ldp_offsets[i];
 205
 206                 LASSERT(!(file_offset & (page_size - 1)));
 207                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 208                                    pv->ldp_pages[i], CPT_TRANSIENT);
 209                 if (IS_ERR(clp)) {
 210                         rc = PTR_ERR(clp);
 211                         break;
 212                 }
 213
 214                 rc = cl_page_own(env, io, clp);
 215                 if (rc) {
 216                         LASSERT(clp->cp_state == CPS_FREEING);
 217                         cl_page_put(env, clp);
 218                         break;
 219                 }
 220
 221                 do_io = true;
 222
 223                 /* check the page type: if the page is a host page, then do
 224                  * write directly */
 225                 if (clp->cp_type == CPT_CACHEABLE) {
 226                         struct page *vmpage = cl_page_vmpage(clp);
 227                         struct page *src_page;
 228                         struct page *dst_page;
 229                         void       *src;
 230                         void       *dst;
 231
 232                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 233                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 234
 235                         src = ll_kmap_atomic(src_page, KM_USER0);
 236                         dst = ll_kmap_atomic(dst_page, KM_USER1);
 237                         memcpy(dst, src, min(page_size, size));
 238                         ll_kunmap_atomic(dst, KM_USER1);
 239                         ll_kunmap_atomic(src, KM_USER0);
 240
 241                         /* make sure page will be added to the transfer by
 242                          * cl_io_submit()->...->vvp_page_prep_write(). */
 243                         if (rw == WRITE)
 244                                 set_page_dirty(vmpage);
 245
 246                         if (rw == READ) {
 247                                 /* do not issue the page for read, since it
 248                                  * may reread a ra page which has NOT uptodate
 249                                  * bit set. */
 250                                 cl_page_disown(env, io, clp);
 251                                 do_io = false;
 252                         }
 253                 }
 254
 255                 if (likely(do_io)) {
 256                         cl_2queue_add(queue, clp);
 257
 258                         /*
 259                          * Set page clip to tell transfer formation engine
 260                          * that page has to be sent even if it is beyond KMS.
 261                          */
 262                         cl_page_clip(env, clp, 0, min(size, page_size));
 263
 264                         ++io_pages;
 265                 }
 266
 267                 /* drop the reference count for cl_page_find */
 268                 cl_page_put(env, clp);
 269                 size -= page_size;
 270                 file_offset += page_size;
 271         }
 272
 273         if (rc == 0 && io_pages) {
 274                 rc = cl_io_submit_sync(env, io,
 275                                        rw == READ ? CRT_READ : CRT_WRITE,
 276                                        queue, 0);
 277         }
 278         if (rc == 0)
 279                 rc = pv->ldp_size;
 280
 281         cl_2queue_discard(env, io, queue);
 282         cl_2queue_disown(env, io, queue);
 283         cl_2queue_fini(env, queue);
 284         RETURN(rc);
 285 }
 286 EXPORT_SYMBOL(ll_direct_rw_pages);
 287
 288 static ssize_t
 289 ll_direct_IO_seg(const struct lu_env *env, struct cl_io *io, int rw,
 290                  struct inode *inode, size_t size, loff_t file_offset,
 291                  struct page **pages, int page_count)
 292 {
 293         struct ll_dio_pages pvec = { .ldp_pages         = pages,
 294                                      .ldp_nr            = page_count,
 295                                      .ldp_size          = size,
 296                                      .ldp_offsets       = NULL,
 297                                      .ldp_start_offset  = file_offset
 298                                    };
 299
 300         return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 301 }
 302
 303 /*  ll_free_user_pages - tear down page struct array
 304  *  @pages: array of page struct pointers underlying target buffer */
 305 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 306 {
 307         int i;
 308
 309         for (i = 0; i < npages; i++) {
 310                 if (pages[i] == NULL)
 311                         break;
 312                 if (do_dirty)
 313                         set_page_dirty_lock(pages[i]);
 314                 put_page(pages[i]);
 315         }
 316
 317 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
 318         kvfree(pages);
 319 #else
 320         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 321 #endif
 322 }
 323
 324 #ifdef KMALLOC_MAX_SIZE
 325 #define MAX_MALLOC KMALLOC_MAX_SIZE
 326 #else
 327 #define MAX_MALLOC (128 * 1024)
 328 #endif
 329
 330 /* This is the maximum size of a single O_DIRECT request, based on the
 331  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 332  * representing PAGE_SIZE worth of user data, into a single buffer, and
 333  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 334  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 335 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_SIZE) & \
 336                       ~(DT_MAX_BRW_SIZE - 1))
 337
 338 #ifndef HAVE_IOV_ITER_RW
 339 # define iov_iter_rw(iter)      rw
 340 #endif
 341
 342 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
 343 static ssize_t
 344 ll_direct_IO(
 345 # ifndef HAVE_IOV_ITER_RW
 346              int rw,
 347 # endif
 348              struct kiocb *iocb, struct iov_iter *iter,
 349              loff_t file_offset)
 350 {
 351         struct ll_cl_context *lcc;
 352         const struct lu_env *env;
 353         struct cl_io *io;
 354         struct file *file = iocb->ki_filp;
 355         struct inode *inode = file->f_mapping->host;
 356         ssize_t count = iov_iter_count(iter);
 357         ssize_t tot_bytes = 0, result = 0;
 358         size_t size = MAX_DIO_SIZE;
 359
 360         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 361         if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
 362                 return -EINVAL;
 363
 364         CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
 365                "offset=%lld=%llx, pages %zd (max %lu)\n",
 366                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 367                file_offset, file_offset, count >> PAGE_SHIFT,
 368                MAX_DIO_SIZE >> PAGE_SHIFT);
 369
 370         /* Check that all user buffers are aligned as well */
 371         if (iov_iter_alignment(iter) & ~PAGE_MASK)
 372                 return -EINVAL;
 373
 374         lcc = ll_cl_find(file);
 375         if (lcc == NULL)
 376                 RETURN(-EIO);
 377
 378         env = lcc->lcc_env;
 379         LASSERT(!IS_ERR(env));
 380         io = lcc->lcc_io;
 381         LASSERT(io != NULL);
 382
 383         /* 0. Need locking between buffered and direct access. and race with
 384          *    size changing by concurrent truncates and writes.
 385          * 1. Need inode mutex to operate transient pages.
 386          */
 387         if (iov_iter_rw(iter) == READ)
 388                 inode_lock(inode);
 389
 390         while (iov_iter_count(iter)) {
 391                 struct page **pages;
 392                 size_t offs;
 393
 394                 count = min_t(size_t, iov_iter_count(iter), size);
 395                 if (iov_iter_rw(iter) == READ) {
 396                         if (file_offset >= i_size_read(inode))
 397                                 break;
 398
 399                         if (file_offset + count > i_size_read(inode))
 400                                 count = i_size_read(inode) - file_offset;
 401                 }
 402
 403                 result = iov_iter_get_pages_alloc(iter, &pages, count, &offs);
 404                 if (likely(result > 0)) {
 405                         int n = DIV_ROUND_UP(result + offs, PAGE_SIZE);
 406
 407                         result = ll_direct_IO_seg(env, io, iov_iter_rw(iter),
 408                                                   inode, result, file_offset,
 409                                                   pages, n);
 410                         ll_free_user_pages(pages, n,
 411                                            iov_iter_rw(iter) == READ);
 412
 413                 }
 414                 if (unlikely(result <= 0)) {
 415                         /* If we can't allocate a large enough buffer
 416                          * for the request, shrink it to a smaller
 417                          * PAGE_SIZE multiple and try again.
 418                          * We should always be able to kmalloc for a
 419                          * page worth of page pointers = 4MB on i386. */
 420                         if (result == -ENOMEM &&
 421                             size > (PAGE_SIZE / sizeof(*pages)) *
 422                                     PAGE_SIZE) {
 423                                 size = ((((size / 2) - 1) |
 424                                         ~PAGE_MASK) + 1) & PAGE_MASK;
 425                                 CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
 426                                        size);
 427                                 continue;
 428                         }
 429
 430                         GOTO(out, result);
 431                 }
 432
 433                 iov_iter_advance(iter, result);
 434                 tot_bytes += result;
 435                 file_offset += result;
 436         }
 437 out:
 438         if (iov_iter_rw(iter) == READ)
 439                 inode_unlock(inode);
 440
 441         if (tot_bytes > 0) {
 442                 struct vvp_io *vio = vvp_env_io(env);
 443
 444                 /* no commit async for direct IO */
 445                 vio->u.write.vui_written += tot_bytes;
 446         }
 447
 448         return tot_bytes ? : result;
 449 }
 450 #else /* !HAVE_DIRECTIO_ITER && !HAVE_IOV_ITER_RW */
 451
 452 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 453                                     size_t size, struct page ***pages,
 454                                     int *max_pages)
 455 {
 456         int result = -ENOMEM;
 457
 458         /* set an arbitrary limit to prevent arithmetic overflow */
 459         if (size > MAX_DIRECTIO_SIZE) {
 460                 *pages = NULL;
 461                 return -EFBIG;
 462         }
 463
 464         *max_pages = (user_addr + size + PAGE_SIZE - 1) >>
 465                       PAGE_SHIFT;
 466         *max_pages -= user_addr >> PAGE_SHIFT;
 467
 468         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 469         if (*pages) {
 470                 down_read(&current->mm->mmap_sem);
 471                 result = get_user_pages(current, current->mm, user_addr,
 472                                         *max_pages, (rw == READ), 0, *pages,
 473                                         NULL);
 474                 up_read(&current->mm->mmap_sem);
 475                 if (unlikely(result <= 0))
 476                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 477         }
 478
 479         return result;
 480 }
 481
 482 static ssize_t
 483 ll_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
 484              loff_t file_offset, unsigned long nr_segs)
 485 {
 486         struct ll_cl_context *lcc;
 487         const struct lu_env *env;
 488         struct cl_io *io;
 489         struct file *file = iocb->ki_filp;
 490         struct inode *inode = file->f_mapping->host;
 491         ssize_t count = iov_length(iov, nr_segs);
 492         ssize_t tot_bytes = 0, result = 0;
 493         unsigned long seg = 0;
 494         size_t size = MAX_DIO_SIZE;
 495         ENTRY;
 496
 497         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 498         if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
 499                 RETURN(-EINVAL);
 500
 501         CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
 502                "offset=%lld=%llx, pages %zd (max %lu)\n",
 503                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 504                file_offset, file_offset, count >> PAGE_SHIFT,
 505                MAX_DIO_SIZE >> PAGE_SHIFT);
 506
 507         /* Check that all user buffers are aligned as well */
 508         for (seg = 0; seg < nr_segs; seg++) {
 509                 if (((unsigned long)iov[seg].iov_base & ~PAGE_MASK) ||
 510                     (iov[seg].iov_len & ~PAGE_MASK))
 511                         RETURN(-EINVAL);
 512         }
 513
 514         lcc = ll_cl_find(file);
 515         if (lcc == NULL)
 516                 RETURN(-EIO);
 517
 518         env = lcc->lcc_env;
 519         LASSERT(!IS_ERR(env));
 520         io = lcc->lcc_io;
 521         LASSERT(io != NULL);
 522
 523         for (seg = 0; seg < nr_segs; seg++) {
 524                 size_t iov_left = iov[seg].iov_len;
 525                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 526
 527                 if (rw == READ) {
 528                         if (file_offset >= i_size_read(inode))
 529                                 break;
 530                         if (file_offset + iov_left > i_size_read(inode))
 531                                 iov_left = i_size_read(inode) - file_offset;
 532                 }
 533
 534                 while (iov_left > 0) {
 535                         struct page **pages;
 536                         int page_count, max_pages = 0;
 537                         size_t bytes;
 538
 539                         bytes = min(size, iov_left);
 540                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 541                                                        &pages, &max_pages);
 542                         if (likely(page_count > 0)) {
 543                                 if (unlikely(page_count <  max_pages))
 544                                         bytes = page_count << PAGE_SHIFT;
 545                                 result = ll_direct_IO_seg(env, io, rw, inode,
 546                                                           bytes, file_offset,
 547                                                           pages, page_count);
 548                                 ll_free_user_pages(pages, max_pages, rw==READ);
 549                         } else if (page_count == 0) {
 550                                 GOTO(out, result = -EFAULT);
 551                         } else {
 552                                 result = page_count;
 553                         }
 554                         if (unlikely(result <= 0)) {
 555                                 /* If we can't allocate a large enough buffer
 556                                  * for the request, shrink it to a smaller
 557                                  * PAGE_SIZE multiple and try again.
 558                                  * We should always be able to kmalloc for a
 559                                  * page worth of page pointers = 4MB on i386. */
 560                                 if (result == -ENOMEM &&
 561                                     size > (PAGE_SIZE / sizeof(*pages)) *
 562                                            PAGE_SIZE) {
 563                                         size = ((((size / 2) - 1) |
 564                                                  ~PAGE_MASK) + 1) &
 565                                                 PAGE_MASK;
 566                                         CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
 567                                                size);
 568                                         continue;
 569                                 }
 570
 571                                 GOTO(out, result);
 572                         }
 573
 574                         tot_bytes += result;
 575                         file_offset += result;
 576                         iov_left -= result;
 577                         user_addr += result;
 578                 }
 579         }
 580 out:
 581         if (tot_bytes > 0) {
 582                 struct vvp_io *vio = vvp_env_io(env);
 583
 584                 /* no commit async for direct IO */
 585                 vio->u.write.vui_written += tot_bytes;
 586         }
 587
 588         RETURN(tot_bytes ? tot_bytes : result);
 589 }
 590 #endif /* HAVE_DIRECTIO_ITER || HAVE_IOV_ITER_RW */
 591
 592 /**
 593  * Prepare partially written-to page for a write.
 594  */
 595 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
 596                                    struct cl_page *pg)
 597 {
 598         struct cl_attr *attr   = vvp_env_thread_attr(env);
 599         struct cl_object *obj  = io->ci_obj;
 600         struct vvp_page *vpg   = cl_object_page_slice(obj, pg);
 601         loff_t          offset = cl_offset(obj, vvp_index(vpg));
 602         int             result;
 603
 604         cl_object_attr_lock(obj);
 605         result = cl_object_attr_get(env, obj, attr);
 606         cl_object_attr_unlock(obj);
 607         if (result == 0) {
 608                 /*
 609                  * If are writing to a new page, no need to read old data.
 610                  * The extent locking will have updated the KMS, and for our
 611                  * purposes here we can treat it like i_size.
 612                  */
 613                 if (attr->cat_kms <= offset) {
 614                         char *kaddr = ll_kmap_atomic(vpg->vpg_page, KM_USER0);
 615
 616                         memset(kaddr, 0, cl_page_size(obj));
 617                         ll_kunmap_atomic(kaddr, KM_USER0);
 618                 } else if (vpg->vpg_defer_uptodate)
 619                         vpg->vpg_ra_used = 1;
 620                 else
 621                         result = ll_page_sync_io(env, io, pg, CRT_READ);
 622         }
 623         return result;
 624 }
 625
 626 static int ll_write_begin(struct file *file, struct address_space *mapping,
 627                           loff_t pos, unsigned len, unsigned flags,
 628                           struct page **pagep, void **fsdata)
 629 {
 630         struct ll_cl_context *lcc;
 631         const struct lu_env  *env = NULL;
 632         struct cl_io   *io;
 633         struct cl_page *page = NULL;
 634
 635         struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
 636         pgoff_t index = pos >> PAGE_SHIFT;
 637         struct page *vmpage = NULL;
 638         unsigned from = pos & (PAGE_SIZE - 1);
 639         unsigned to = from + len;
 640         int result = 0;
 641         ENTRY;
 642
 643         CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
 644
 645         lcc = ll_cl_find(file);
 646         if (lcc == NULL) {
 647                 io = NULL;
 648                 GOTO(out, result = -EIO);
 649         }
 650
 651         env = lcc->lcc_env;
 652         io  = lcc->lcc_io;
 653
 654         /* To avoid deadlock, try to lock page first. */
 655         vmpage = grab_cache_page_nowait(mapping, index);
 656
 657         if (unlikely(vmpage == NULL ||
 658                      PageDirty(vmpage) || PageWriteback(vmpage))) {
 659                 struct vvp_io *vio = vvp_env_io(env);
 660                 struct cl_page_list *plist = &vio->u.write.vui_queue;
 661
 662                 /* if the page is already in dirty cache, we have to commit
 663                  * the pages right now; otherwise, it may cause deadlock
 664                  * because it holds page lock of a dirty page and request for
 665                  * more grants. It's okay for the dirty page to be the first
 666                  * one in commit page list, though. */
 667                 if (vmpage != NULL && plist->pl_nr > 0) {
 668                         unlock_page(vmpage);
 669                         put_page(vmpage);
 670                         vmpage = NULL;
 671                 }
 672
 673                 /* commit pages and then wait for page lock */
 674                 result = vvp_io_write_commit(env, io);
 675                 if (result < 0)
 676                         GOTO(out, result);
 677
 678                 if (vmpage == NULL) {
 679                         vmpage = grab_cache_page_write_begin(mapping, index,
 680                                                              flags);
 681                         if (vmpage == NULL)
 682                                 GOTO(out, result = -ENOMEM);
 683                 }
 684         }
 685
 686         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 687         if (IS_ERR(page))
 688                 GOTO(out, result = PTR_ERR(page));
 689
 690         lcc->lcc_page = page;
 691         lu_ref_add(&page->cp_reference, "cl_io", io);
 692
 693         cl_page_assume(env, io, page);
 694         if (!PageUptodate(vmpage)) {
 695                 /*
 696                  * We're completely overwriting an existing page,
 697                  * so _don't_ set it up to date until commit_write
 698                  */
 699                 if (from == 0 && to == PAGE_SIZE) {
 700                         CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
 701                         POISON_PAGE(vmpage, 0x11);
 702                 } else {
 703                         /* TODO: can be optimized at OSC layer to check if it
 704                          * is a lockless IO. In that case, it's not necessary
 705                          * to read the data. */
 706                         result = ll_prepare_partial_page(env, io, page);
 707                         if (result == 0)
 708                                 SetPageUptodate(vmpage);
 709                 }
 710         }
 711         if (result < 0)
 712                 cl_page_unassume(env, io, page);
 713         EXIT;
 714 out:
 715         if (result < 0) {
 716                 if (vmpage != NULL) {
 717                         unlock_page(vmpage);
 718                         put_page(vmpage);
 719                 }
 720                 if (!IS_ERR_OR_NULL(page)) {
 721                         lu_ref_del(&page->cp_reference, "cl_io", io);
 722                         cl_page_put(env, page);
 723                 }
 724                 if (io)
 725                         io->ci_result = result;
 726         } else {
 727                 *pagep = vmpage;
 728                 *fsdata = lcc;
 729         }
 730         RETURN(result);
 731 }
 732
 733 static int ll_write_end(struct file *file, struct address_space *mapping,
 734                         loff_t pos, unsigned len, unsigned copied,
 735                         struct page *vmpage, void *fsdata)
 736 {
 737         struct ll_cl_context *lcc = fsdata;
 738         const struct lu_env *env;
 739         struct cl_io *io;
 740         struct vvp_io *vio;
 741         struct cl_page *page;
 742         unsigned from = pos & (PAGE_SIZE - 1);
 743         bool unplug = false;
 744         int result = 0;
 745         ENTRY;
 746
 747         put_page(vmpage);
 748
 749         LASSERT(lcc != NULL);
 750         env  = lcc->lcc_env;
 751         page = lcc->lcc_page;
 752         io   = lcc->lcc_io;
 753         vio  = vvp_env_io(env);
 754
 755         LASSERT(cl_page_is_owned(page, io));
 756         if (copied > 0) {
 757                 struct cl_page_list *plist = &vio->u.write.vui_queue;
 758
 759                 lcc->lcc_page = NULL; /* page will be queued */
 760
 761                 /* Add it into write queue */
 762                 cl_page_list_add(plist, page);
 763                 if (plist->pl_nr == 1) /* first page */
 764                         vio->u.write.vui_from = from;
 765                 else
 766                         LASSERT(from == 0);
 767                 vio->u.write.vui_to = from + copied;
 768
 769                 /* To address the deadlock in balance_dirty_pages() where
 770                  * this dirty page may be written back in the same thread. */
 771                 if (PageDirty(vmpage))
 772                         unplug = true;
 773
 774                 /* We may have one full RPC, commit it soon */
 775                 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
 776                         unplug = true;
 777
 778                 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
 779                               "queued page: %d.\n", plist->pl_nr);
 780         } else {
 781                 cl_page_disown(env, io, page);
 782
 783                 lcc->lcc_page = NULL;
 784                 lu_ref_del(&page->cp_reference, "cl_io", io);
 785                 cl_page_put(env, page);
 786
 787                 /* page list is not contiguous now, commit it now */
 788                 unplug = true;
 789         }
 790         if (unplug ||
 791             file->f_flags & O_SYNC || IS_SYNC(file_inode(file)))
 792                 result = vvp_io_write_commit(env, io);
 793
 794         if (result < 0)
 795                 io->ci_result = result;
 796         RETURN(result >= 0 ? copied : result);
 797 }
 798
 799 #ifdef CONFIG_MIGRATION
 800 static int ll_migratepage(struct address_space *mapping,
 801                           struct page *newpage, struct page *page
 802 #ifdef HAVE_MIGRATEPAGE_4ARGS
 803                           , enum migrate_mode mode
 804 #endif
 805         )
 806 {
 807         /* Always fail page migration until we have a proper implementation */
 808         return -EIO;
 809 }
 810 #endif
 811
 812 const struct address_space_operations ll_aops = {
 813         .readpage       = ll_readpage,
 814         .direct_IO      = ll_direct_IO,
 815         .writepage      = ll_writepage,
 816         .writepages     = ll_writepages,
 817         .set_page_dirty = __set_page_dirty_nobuffers,
 818         .write_begin    = ll_write_begin,
 819         .write_end      = ll_write_end,
 820         .invalidatepage = ll_invalidatepage,
 821         .releasepage    = (void *)ll_releasepage,
 822 #ifdef CONFIG_MIGRATION
 823         .migratepage    = ll_migratepage,
 824 #endif
 825 };