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, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/lustre/llite/rw26.c
  37  *
  38  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  39  */
  40
  41 #include <linux/kernel.h>
  42 #include <linux/mm.h>
  43 #include <linux/string.h>
  44 #include <linux/stat.h>
  45 #include <linux/errno.h>
  46 #include <linux/unistd.h>
  47 #include <asm/uaccess.h>
  48
  49 #ifdef HAVE_MIGRATE_H
  50 #include <linux/migrate.h>
  51 #elif defined(HAVE_MIGRATE_MODE_H)
  52 #include <linux/migrate_mode.h>
  53 #endif
  54 #include <linux/fs.h>
  55 #include <linux/buffer_head.h>
  56 #include <linux/mpage.h>
  57 #include <linux/writeback.h>
  58 #include <linux/stat.h>
  59 #include <asm/uaccess.h>
  60 #include <linux/mm.h>
  61 #include <linux/pagemap.h>
  62
  63 #define DEBUG_SUBSYSTEM S_LLITE
  64
  65 #include "llite_internal.h"
  66 #include <lustre_compat.h>
  67
  68 /**
  69  * Implements Linux VM address_space::invalidatepage() method. This method is
  70  * called when the page is truncate from a file, either as a result of
  71  * explicit truncate, or when inode is removed from memory (as a result of
  72  * final iput(), umount, or memory pressure induced icache shrinking).
  73  *
  74  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  75  * aligned truncate). Lustre leaves partially truncated page in the cache,
  76  * relying on struct inode::i_size to limit further accesses.
  77  */
  78 static void ll_invalidatepage(struct page *vmpage,
  79 #ifdef HAVE_INVALIDATE_RANGE
  80                                 unsigned int offset, unsigned int length
  81 #else
  82                                 unsigned long offset
  83 #endif
  84                              )
  85 {
  86         struct inode     *inode;
  87         struct lu_env    *env;
  88         struct cl_page   *page;
  89         struct cl_object *obj;
  90
  91         __u16 refcheck;
  92
  93         LASSERT(PageLocked(vmpage));
  94         LASSERT(!PageWriteback(vmpage));
  95
  96         /*
  97          * It is safe to not check anything in invalidatepage/releasepage
  98          * below because they are run with page locked and all our io is
  99          * happening with locked page too
 100          */
 101 #ifdef HAVE_INVALIDATE_RANGE
 102         if (offset == 0 && length == PAGE_CACHE_SIZE) {
 103 #else
 104         if (offset == 0) {
 105 #endif
 106                 env = cl_env_get(&refcheck);
 107                 if (!IS_ERR(env)) {
 108                         inode = vmpage->mapping->host;
 109                         obj = ll_i2info(inode)->lli_clob;
 110                         if (obj != NULL) {
 111                                 page = cl_vmpage_page(vmpage, obj);
 112                                 if (page != NULL) {
 113                                         cl_page_delete(env, page);
 114                                         cl_page_put(env, page);
 115                                 }
 116                         } else
 117                                 LASSERT(vmpage->private == 0);
 118                         cl_env_put(env, &refcheck);
 119                 }
 120         }
 121 }
 122
 123 #ifdef HAVE_RELEASEPAGE_WITH_INT
 124 #define RELEASEPAGE_ARG_TYPE int
 125 #else
 126 #define RELEASEPAGE_ARG_TYPE gfp_t
 127 #endif
 128 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 129 {
 130         struct lu_env           *env;
 131         void                    *cookie;
 132         struct cl_object        *obj;
 133         struct cl_page          *page;
 134         struct address_space    *mapping;
 135         int result = 0;
 136
 137         LASSERT(PageLocked(vmpage));
 138         if (PageWriteback(vmpage) || PageDirty(vmpage))
 139                 return 0;
 140
 141         mapping = vmpage->mapping;
 142         if (mapping == NULL)
 143                 return 1;
 144
 145         obj = ll_i2info(mapping->host)->lli_clob;
 146         if (obj == NULL)
 147                 return 1;
 148
 149         /* 1 for caller, 1 for cl_page and 1 for page cache */
 150         if (page_count(vmpage) > 3)
 151                 return 0;
 152
 153         page = cl_vmpage_page(vmpage, obj);
 154         if (page == NULL)
 155                 return 1;
 156
 157         cookie = cl_env_reenter();
 158         env = cl_env_percpu_get();
 159         LASSERT(!IS_ERR(env));
 160
 161         if (!cl_page_in_use(page)) {
 162                 result = 1;
 163                 cl_page_delete(env, page);
 164         }
 165
 166         /* To use percpu env array, the call path can not be rescheduled;
 167          * otherwise percpu array will be messed if ll_releaspage() called
 168          * again on the same CPU.
 169          *
 170          * If this page holds the last refc of cl_object, the following
 171          * call path may cause reschedule:
 172          *   cl_page_put -> cl_page_free -> cl_object_put ->
 173          *     lu_object_put -> lu_object_free -> lov_delete_raid0.
 174          *
 175          * However, the kernel can't get rid of this inode until all pages have
 176          * been cleaned up. Now that we hold page lock here, it's pretty safe
 177          * that we won't get into object delete path.
 178          */
 179         LASSERT(cl_object_refc(obj) > 1);
 180         cl_page_put(env, page);
 181
 182         cl_env_percpu_put(env);
 183         cl_env_reexit(cookie);
 184         return result;
 185 }
 186
 187 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 188
 189 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 190                            int rw, struct inode *inode,
 191                            struct ll_dio_pages *pv)
 192 {
 193         struct cl_page    *clp;
 194         struct cl_2queue  *queue;
 195         struct cl_object  *obj = io->ci_obj;
 196         int i;
 197         ssize_t rc = 0;
 198         loff_t file_offset  = pv->ldp_start_offset;
 199         size_t size         = pv->ldp_size;
 200         int page_count      = pv->ldp_nr;
 201         struct page **pages = pv->ldp_pages;
 202         size_t page_size    = cl_page_size(obj);
 203         bool do_io;
 204         int  io_pages       = 0;
 205         ENTRY;
 206
 207         queue = &io->ci_queue;
 208         cl_2queue_init(queue);
 209         for (i = 0; i < page_count; i++) {
 210                 if (pv->ldp_offsets)
 211                     file_offset = pv->ldp_offsets[i];
 212
 213                 LASSERT(!(file_offset & (page_size - 1)));
 214                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 215                                    pv->ldp_pages[i], CPT_TRANSIENT);
 216                 if (IS_ERR(clp)) {
 217                         rc = PTR_ERR(clp);
 218                         break;
 219                 }
 220
 221                 rc = cl_page_own(env, io, clp);
 222                 if (rc) {
 223                         LASSERT(clp->cp_state == CPS_FREEING);
 224                         cl_page_put(env, clp);
 225                         break;
 226                 }
 227
 228                 do_io = true;
 229
 230                 /* check the page type: if the page is a host page, then do
 231                  * write directly */
 232                 if (clp->cp_type == CPT_CACHEABLE) {
 233                         struct page *vmpage = cl_page_vmpage(clp);
 234                         struct page *src_page;
 235                         struct page *dst_page;
 236                         void       *src;
 237                         void       *dst;
 238
 239                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 240                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 241
 242                         src = ll_kmap_atomic(src_page, KM_USER0);
 243                         dst = ll_kmap_atomic(dst_page, KM_USER1);
 244                         memcpy(dst, src, min(page_size, size));
 245                         ll_kunmap_atomic(dst, KM_USER1);
 246                         ll_kunmap_atomic(src, KM_USER0);
 247
 248                         /* make sure page will be added to the transfer by
 249                          * cl_io_submit()->...->vvp_page_prep_write(). */
 250                         if (rw == WRITE)
 251                                 set_page_dirty(vmpage);
 252
 253                         if (rw == READ) {
 254                                 /* do not issue the page for read, since it
 255                                  * may reread a ra page which has NOT uptodate
 256                                  * bit set. */
 257                                 cl_page_disown(env, io, clp);
 258                                 do_io = false;
 259                         }
 260                 }
 261
 262                 if (likely(do_io)) {
 263                         cl_2queue_add(queue, clp);
 264
 265                         /*
 266                          * Set page clip to tell transfer formation engine
 267                          * that page has to be sent even if it is beyond KMS.
 268                          */
 269                         cl_page_clip(env, clp, 0, min(size, page_size));
 270
 271                         ++io_pages;
 272                 }
 273
 274                 /* drop the reference count for cl_page_find */
 275                 cl_page_put(env, clp);
 276                 size -= page_size;
 277                 file_offset += page_size;
 278         }
 279
 280         if (rc == 0 && io_pages) {
 281                 rc = cl_io_submit_sync(env, io,
 282                                        rw == READ ? CRT_READ : CRT_WRITE,
 283                                        queue, 0);
 284         }
 285         if (rc == 0)
 286                 rc = pv->ldp_size;
 287
 288         cl_2queue_discard(env, io, queue);
 289         cl_2queue_disown(env, io, queue);
 290         cl_2queue_fini(env, queue);
 291         RETURN(rc);
 292 }
 293 EXPORT_SYMBOL(ll_direct_rw_pages);
 294
 295 static ssize_t
 296 ll_direct_IO_seg(const struct lu_env *env, struct cl_io *io, int rw,
 297                  struct inode *inode, size_t size, loff_t file_offset,
 298                  struct page **pages, int page_count)
 299 {
 300         struct ll_dio_pages pvec = { .ldp_pages         = pages,
 301                                      .ldp_nr            = page_count,
 302                                      .ldp_size          = size,
 303                                      .ldp_offsets       = NULL,
 304                                      .ldp_start_offset  = file_offset
 305                                    };
 306
 307         return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 308 }
 309
 310 /*  ll_free_user_pages - tear down page struct array
 311  *  @pages: array of page struct pointers underlying target buffer */
 312 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 313 {
 314         int i;
 315
 316         for (i = 0; i < npages; i++) {
 317                 if (pages[i] == NULL)
 318                         break;
 319                 if (do_dirty)
 320                         set_page_dirty_lock(pages[i]);
 321                 page_cache_release(pages[i]);
 322         }
 323
 324 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
 325         kvfree(pages);
 326 #else
 327         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 328 #endif
 329 }
 330
 331 #ifdef KMALLOC_MAX_SIZE
 332 #define MAX_MALLOC KMALLOC_MAX_SIZE
 333 #else
 334 #define MAX_MALLOC (128 * 1024)
 335 #endif
 336
 337 /* This is the maximum size of a single O_DIRECT request, based on the
 338  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 339  * representing PAGE_SIZE worth of user data, into a single buffer, and
 340  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 341  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 342 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
 343                       ~(DT_MAX_BRW_SIZE - 1))
 344
 345 #ifndef HAVE_IOV_ITER_RW
 346 # define iov_iter_rw(iter)      rw
 347 #endif
 348
 349 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
 350 static ssize_t
 351 ll_direct_IO(
 352 # ifndef HAVE_IOV_ITER_RW
 353              int rw,
 354 # endif
 355              struct kiocb *iocb, struct iov_iter *iter,
 356              loff_t file_offset)
 357 {
 358         struct lu_env *env;
 359         struct cl_io *io;
 360         struct file *file = iocb->ki_filp;
 361         struct inode *inode = file->f_mapping->host;
 362         ssize_t count = iov_iter_count(iter);
 363         ssize_t tot_bytes = 0, result = 0;
 364         size_t size = MAX_DIO_SIZE;
 365         __u16 refcheck;
 366
 367         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 368         if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
 369                 return -EINVAL;
 370
 371         CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
 372                "offset=%lld=%llx, pages %zd (max %lu)\n",
 373                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 374                file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
 375                MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
 376
 377         /* Check that all user buffers are aligned as well */
 378         if (iov_iter_alignment(iter) & ~PAGE_MASK)
 379                 return -EINVAL;
 380
 381         env = cl_env_get(&refcheck);
 382         LASSERT(!IS_ERR(env));
 383         io = vvp_env_io(env)->vui_cl.cis_io;
 384         LASSERT(io != NULL);
 385
 386         /* 0. Need locking between buffered and direct access. and race with
 387          *    size changing by concurrent truncates and writes.
 388          * 1. Need inode mutex to operate transient pages.
 389          */
 390         if (iov_iter_rw(iter) == READ)
 391                 mutex_lock(&inode->i_mutex);
 392
 393         while (iov_iter_count(iter)) {
 394                 struct page **pages;
 395                 size_t offs;
 396
 397                 count = min_t(size_t, iov_iter_count(iter), size);
 398                 if (iov_iter_rw(iter) == READ) {
 399                         if (file_offset >= i_size_read(inode))
 400                                 break;
 401
 402                         if (file_offset + count > i_size_read(inode))
 403                                 count = i_size_read(inode) - file_offset;
 404                 }
 405
 406                 result = iov_iter_get_pages_alloc(iter, &pages, count, &offs);
 407                 if (likely(result > 0)) {
 408                         int n = DIV_ROUND_UP(result + offs, PAGE_SIZE);
 409
 410                         result = ll_direct_IO_seg(env, io, iov_iter_rw(iter),
 411                                                   inode, result, file_offset,
 412                                                   pages, n);
 413                         ll_free_user_pages(pages, n,
 414                                            iov_iter_rw(iter) == READ);
 415
 416                 }
 417                 if (unlikely(result <= 0)) {
 418                         /* If we can't allocate a large enough buffer
 419                          * for the request, shrink it to a smaller
 420                          * PAGE_SIZE multiple and try again.
 421                          * We should always be able to kmalloc for a
 422                          * page worth of page pointers = 4MB on i386. */
 423                         if (result == -ENOMEM &&
 424                             size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
 425                                     PAGE_CACHE_SIZE) {
 426                                 size = ((((size / 2) - 1) |
 427                                         ~PAGE_MASK) + 1) & PAGE_MASK;
 428                                 CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
 429                                        size);
 430                                 continue;
 431                         }
 432
 433                         GOTO(out, result);
 434                 }
 435
 436                 iov_iter_advance(iter, result);
 437                 tot_bytes += result;
 438                 file_offset += result;
 439         }
 440 out:
 441         if (iov_iter_rw(iter) == READ)
 442                 mutex_unlock(&inode->i_mutex);
 443
 444         if (tot_bytes > 0) {
 445                 struct vvp_io *vio = vvp_env_io(env);
 446
 447                 /* no commit async for direct IO */
 448                 vio->u.write.vui_written += tot_bytes;
 449         }
 450
 451         cl_env_put(env, &refcheck);
 452         return tot_bytes ? : result;
 453 }
 454 #else /* !HAVE_DIRECTIO_ITER && !HAVE_IOV_ITER_RW */
 455
 456 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 457                                     size_t size, struct page ***pages,
 458                                     int *max_pages)
 459 {
 460         int result = -ENOMEM;
 461
 462         /* set an arbitrary limit to prevent arithmetic overflow */
 463         if (size > MAX_DIRECTIO_SIZE) {
 464                 *pages = NULL;
 465                 return -EFBIG;
 466         }
 467
 468         *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >>
 469                       PAGE_CACHE_SHIFT;
 470         *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
 471
 472         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 473         if (*pages) {
 474                 down_read(&current->mm->mmap_sem);
 475                 result = get_user_pages(current, current->mm, user_addr,
 476                                         *max_pages, (rw == READ), 0, *pages,
 477                                         NULL);
 478                 up_read(&current->mm->mmap_sem);
 479                 if (unlikely(result <= 0))
 480                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 481         }
 482
 483         return result;
 484 }
 485
 486 static ssize_t
 487 ll_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
 488              loff_t file_offset, unsigned long nr_segs)
 489 {
 490         struct lu_env *env;
 491         struct cl_io *io;
 492         struct file *file = iocb->ki_filp;
 493         struct inode *inode = file->f_mapping->host;
 494         ssize_t count = iov_length(iov, nr_segs);
 495         ssize_t tot_bytes = 0, result = 0;
 496         unsigned long seg = 0;
 497         size_t size = MAX_DIO_SIZE;
 498         __u16 refcheck;
 499         ENTRY;
 500
 501         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 502         if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
 503                 RETURN(-EINVAL);
 504
 505         CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
 506                "offset=%lld=%llx, pages %zd (max %lu)\n",
 507                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 508                file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
 509                MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
 510
 511         /* Check that all user buffers are aligned as well */
 512         for (seg = 0; seg < nr_segs; seg++) {
 513                 if (((unsigned long)iov[seg].iov_base & ~PAGE_MASK) ||
 514                     (iov[seg].iov_len & ~PAGE_MASK))
 515                         RETURN(-EINVAL);
 516         }
 517
 518         env = cl_env_get(&refcheck);
 519         LASSERT(!IS_ERR(env));
 520         io = vvp_env_io(env)->vui_cl.cis_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_CACHE_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_CACHE_SIZE / sizeof(*pages)) *
 562                                            PAGE_CACHE_SIZE) {
 563                                         size = ((((size / 2) - 1) |
 564                                                  ~PAGE_CACHE_MASK) + 1) &
 565                                                 PAGE_CACHE_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         cl_env_put(env, &refcheck);
 589         RETURN(tot_bytes ? tot_bytes : result);
 590 }
 591 #endif /* HAVE_DIRECTIO_ITER || HAVE_IOV_ITER_RW */
 592
 593 /**
 594  * Prepare partially written-to page for a write.
 595  */
 596 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
 597                                    struct cl_page *pg)
 598 {
 599         struct cl_attr *attr   = vvp_env_thread_attr(env);
 600         struct cl_object *obj  = io->ci_obj;
 601         struct vvp_page *vpg   = cl_object_page_slice(obj, pg);
 602         loff_t          offset = cl_offset(obj, vvp_index(vpg));
 603         int             result;
 604
 605         cl_object_attr_lock(obj);
 606         result = cl_object_attr_get(env, obj, attr);
 607         cl_object_attr_unlock(obj);
 608         if (result == 0) {
 609                 /*
 610                  * If are writing to a new page, no need to read old data.
 611                  * The extent locking will have updated the KMS, and for our
 612                  * purposes here we can treat it like i_size.
 613                  */
 614                 if (attr->cat_kms <= offset) {
 615                         char *kaddr = ll_kmap_atomic(vpg->vpg_page, KM_USER0);
 616
 617                         memset(kaddr, 0, cl_page_size(obj));
 618                         ll_kunmap_atomic(kaddr, KM_USER0);
 619                 } else if (vpg->vpg_defer_uptodate)
 620                         vpg->vpg_ra_used = 1;
 621                 else
 622                         result = ll_page_sync_io(env, io, pg, CRT_READ);
 623         }
 624         return result;
 625 }
 626
 627 static int ll_write_begin(struct file *file, struct address_space *mapping,
 628                           loff_t pos, unsigned len, unsigned flags,
 629                           struct page **pagep, void **fsdata)
 630 {
 631         struct ll_cl_context *lcc;
 632         const struct lu_env  *env;
 633         struct cl_io   *io;
 634         struct cl_page *page;
 635
 636         struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
 637         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 638         struct page *vmpage = NULL;
 639         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 640         unsigned to = from + len;
 641         int result = 0;
 642         ENTRY;
 643
 644         CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
 645
 646         lcc = ll_cl_find(file);
 647         if (lcc == NULL) {
 648                 io = NULL;
 649                 GOTO(out, result = -EIO);
 650         }
 651
 652         env = lcc->lcc_env;
 653         io  = lcc->lcc_io;
 654
 655         /* To avoid deadlock, try to lock page first. */
 656         vmpage = grab_cache_page_nowait(mapping, index);
 657
 658         if (unlikely(vmpage == NULL ||
 659                      PageDirty(vmpage) || PageWriteback(vmpage))) {
 660                 struct vvp_io *vio = vvp_env_io(env);
 661                 struct cl_page_list *plist = &vio->u.write.vui_queue;
 662
 663                 /* if the page is already in dirty cache, we have to commit
 664                  * the pages right now; otherwise, it may cause deadlock
 665                  * because it holds page lock of a dirty page and request for
 666                  * more grants. It's okay for the dirty page to be the first
 667                  * one in commit page list, though. */
 668                 if (vmpage != NULL && plist->pl_nr > 0) {
 669                         unlock_page(vmpage);
 670                         page_cache_release(vmpage);
 671                         vmpage = NULL;
 672                 }
 673
 674                 /* commit pages and then wait for page lock */
 675                 result = vvp_io_write_commit(env, io);
 676                 if (result < 0)
 677                         GOTO(out, result);
 678
 679                 if (vmpage == NULL) {
 680                         vmpage = grab_cache_page_write_begin(mapping, index,
 681                                                              flags);
 682                         if (vmpage == NULL)
 683                                 GOTO(out, result = -ENOMEM);
 684                 }
 685         }
 686
 687         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 688         if (IS_ERR(page))
 689                 GOTO(out, result = PTR_ERR(page));
 690
 691         lcc->lcc_page = page;
 692         lu_ref_add(&page->cp_reference, "cl_io", io);
 693
 694         cl_page_assume(env, io, page);
 695         if (!PageUptodate(vmpage)) {
 696                 /*
 697                  * We're completely overwriting an existing page,
 698                  * so _don't_ set it up to date until commit_write
 699                  */
 700                 if (from == 0 && to == PAGE_SIZE) {
 701                         CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
 702                         POISON_PAGE(vmpage, 0x11);
 703                 } else {
 704                         /* TODO: can be optimized at OSC layer to check if it
 705                          * is a lockless IO. In that case, it's not necessary
 706                          * to read the data. */
 707                         result = ll_prepare_partial_page(env, io, page);
 708                         if (result == 0)
 709                                 SetPageUptodate(vmpage);
 710                 }
 711         }
 712         if (result < 0)
 713                 cl_page_unassume(env, io, page);
 714         EXIT;
 715 out:
 716         if (result < 0) {
 717                 if (vmpage != NULL) {
 718                         unlock_page(vmpage);
 719                         page_cache_release(vmpage);
 720                 }
 721                 if (io)
 722                         io->ci_result = result;
 723         } else {
 724                 *pagep = vmpage;
 725                 *fsdata = lcc;
 726         }
 727         RETURN(result);
 728 }
 729
 730 static int ll_write_end(struct file *file, struct address_space *mapping,
 731                         loff_t pos, unsigned len, unsigned copied,
 732                         struct page *vmpage, void *fsdata)
 733 {
 734         struct ll_cl_context *lcc = fsdata;
 735         const struct lu_env *env;
 736         struct cl_io *io;
 737         struct vvp_io *vio;
 738         struct cl_page *page;
 739         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 740         bool unplug = false;
 741         int result = 0;
 742         ENTRY;
 743
 744         page_cache_release(vmpage);
 745
 746         LASSERT(lcc != NULL);
 747         env  = lcc->lcc_env;
 748         page = lcc->lcc_page;
 749         io   = lcc->lcc_io;
 750         vio  = vvp_env_io(env);
 751
 752         LASSERT(cl_page_is_owned(page, io));
 753         if (copied > 0) {
 754                 struct cl_page_list *plist = &vio->u.write.vui_queue;
 755
 756                 lcc->lcc_page = NULL; /* page will be queued */
 757
 758                 /* Add it into write queue */
 759                 cl_page_list_add(plist, page);
 760                 if (plist->pl_nr == 1) /* first page */
 761                         vio->u.write.vui_from = from;
 762                 else
 763                         LASSERT(from == 0);
 764                 vio->u.write.vui_to = from + copied;
 765
 766                 /* To address the deadlock in balance_dirty_pages() where
 767                  * this dirty page may be written back in the same thread. */
 768                 if (PageDirty(vmpage))
 769                         unplug = true;
 770
 771                 /* We may have one full RPC, commit it soon */
 772                 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
 773                         unplug = true;
 774
 775                 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
 776                               "queued page: %d.\n", plist->pl_nr);
 777         } else {
 778                 cl_page_disown(env, io, page);
 779
 780                 lcc->lcc_page = NULL;
 781                 lu_ref_del(&page->cp_reference, "cl_io", io);
 782                 cl_page_put(env, page);
 783
 784                 /* page list is not contiguous now, commit it now */
 785                 unplug = true;
 786         }
 787         if (unplug ||
 788             file->f_flags & O_SYNC || IS_SYNC(file->f_path.dentry->d_inode))
 789                 result = vvp_io_write_commit(env, io);
 790
 791         if (result < 0)
 792                 io->ci_result = result;
 793         RETURN(result >= 0 ? copied : result);
 794 }
 795
 796 #ifdef CONFIG_MIGRATION
 797 static int ll_migratepage(struct address_space *mapping,
 798                           struct page *newpage, struct page *page
 799 #ifdef HAVE_MIGRATEPAGE_4ARGS
 800                           , enum migrate_mode mode
 801 #endif
 802         )
 803 {
 804         /* Always fail page migration until we have a proper implementation */
 805         return -EIO;
 806 }
 807 #endif
 808
 809 #ifndef MS_HAS_NEW_AOPS
 810 const struct address_space_operations ll_aops = {
 811         .readpage       = ll_readpage,
 812         .direct_IO      = ll_direct_IO,
 813         .writepage      = ll_writepage,
 814         .writepages     = ll_writepages,
 815         .set_page_dirty = __set_page_dirty_nobuffers,
 816         .write_begin    = ll_write_begin,
 817         .write_end      = ll_write_end,
 818         .invalidatepage = ll_invalidatepage,
 819         .releasepage    = (void *)ll_releasepage,
 820 #ifdef CONFIG_MIGRATION
 821         .migratepage    = ll_migratepage,
 822 #endif
 823 };
 824 #else
 825 const struct address_space_operations_ext ll_aops = {
 826         .orig_aops.readpage             = ll_readpage,
 827         .orig_aops.direct_IO            = ll_direct_IO,
 828         .orig_aops.writepage            = ll_writepage,
 829         .orig_aops.writepages           = ll_writepages,
 830         .orig_aops.set_page_dirty       = __set_page_dirty_nobuffers,
 831         .orig_aops.invalidatepage       = ll_invalidatepage,
 832         .orig_aops.releasepage          = ll_releasepage,
 833 #ifdef CONFIG_MIGRATION
 834         .orig_aops.migratepage          = ll_migratepage,
 835 #endif
 836         .write_begin                    = ll_write_begin,
 837         .write_end                      = ll_write_end
 838 };
 839 #endif