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, 2017, 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/buffer_head.h>
  38 #include <linux/errno.h>
  39 #include <linux/fs.h>
  40 #include <linux/kernel.h>
  41 #include <linux/mm.h>
  42 #include <linux/mpage.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/string.h>
  45 #include <linux/unistd.h>
  46 #include <linux/writeback.h>
  47
  48 #ifdef HAVE_MIGRATE_H
  49 #include <linux/migrate.h>
  50 #elif defined(HAVE_MIGRATE_MODE_H)
  51 #include <linux/migrate_mode.h>
  52 #endif
  53
  54 #define DEBUG_SUBSYSTEM S_LLITE
  55
  56 #include "llite_internal.h"
  57 #include <lustre_compat.h>
  58
  59 /**
  60  * Implements Linux VM address_space::invalidatepage() method. This method is
  61  * called when the page is truncate from a file, either as a result of
  62  * explicit truncate, or when inode is removed from memory (as a result of
  63  * final iput(), umount, or memory pressure induced icache shrinking).
  64  *
  65  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  66  * aligned truncate). Lustre leaves partially truncated page in the cache,
  67  * relying on struct inode::i_size to limit further accesses.
  68  */
  69 static void ll_invalidatepage(struct page *vmpage,
  70 #ifdef HAVE_INVALIDATE_RANGE
  71                                 unsigned int offset, unsigned int length
  72 #else
  73                                 unsigned long offset
  74 #endif
  75                              )
  76 {
  77         struct inode     *inode;
  78         struct lu_env    *env;
  79         struct cl_page   *page;
  80         struct cl_object *obj;
  81
  82         LASSERT(PageLocked(vmpage));
  83         LASSERT(!PageWriteback(vmpage));
  84
  85         /*
  86          * It is safe to not check anything in invalidatepage/releasepage
  87          * below because they are run with page locked and all our io is
  88          * happening with locked page too
  89          */
  90 #ifdef HAVE_INVALIDATE_RANGE
  91         if (offset == 0 && length == PAGE_SIZE) {
  92 #else
  93         if (offset == 0) {
  94 #endif
  95                 /* See the comment in ll_releasepage() */
  96                 env = cl_env_percpu_get();
  97                 LASSERT(!IS_ERR(env));
  98
  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
 110                 cl_env_percpu_put(env);
 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         struct cl_object        *obj;
 123         struct cl_page          *page;
 124         struct address_space    *mapping;
 125         int result = 0;
 126
 127         LASSERT(PageLocked(vmpage));
 128         if (PageWriteback(vmpage) || PageDirty(vmpage))
 129                 return 0;
 130
 131         mapping = vmpage->mapping;
 132         if (mapping == NULL)
 133                 return 1;
 134
 135         obj = ll_i2info(mapping->host)->lli_clob;
 136         if (obj == NULL)
 137                 return 1;
 138
 139         /* 1 for caller, 1 for cl_page and 1 for page cache */
 140         if (page_count(vmpage) > 3)
 141                 return 0;
 142
 143         page = cl_vmpage_page(vmpage, obj);
 144         if (page == NULL)
 145                 return 1;
 146
 147         env = cl_env_percpu_get();
 148         LASSERT(!IS_ERR(env));
 149
 150         if (!cl_page_in_use(page)) {
 151                 result = 1;
 152                 cl_page_delete(env, page);
 153         }
 154
 155         /* To use percpu env array, the call path can not be rescheduled;
 156          * otherwise percpu array will be messed if ll_releaspage() called
 157          * again on the same CPU.
 158          *
 159          * If this page holds the last refc of cl_object, the following
 160          * call path may cause reschedule:
 161          *   cl_page_put -> cl_page_free -> cl_object_put ->
 162          *     lu_object_put -> lu_object_free -> lov_delete_raid0.
 163          *
 164          * However, the kernel can't get rid of this inode until all pages have
 165          * been cleaned up. Now that we hold page lock here, it's pretty safe
 166          * that we won't get into object delete path.
 167          */
 168         LASSERT(cl_object_refc(obj) > 1);
 169         cl_page_put(env, page);
 170
 171         cl_env_percpu_put(env);
 172         return result;
 173 }
 174
 175 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 176
 177 static ssize_t
 178 ll_direct_IO_seg(const struct lu_env *env, struct cl_io *io, int rw,
 179                  struct inode *inode, size_t size, loff_t file_offset,
 180                  struct page **pages, int page_count)
 181 {
 182         struct cl_page *clp;
 183         struct cl_2queue *queue;
 184         struct cl_object *obj = io->ci_obj;
 185         int i;
 186         ssize_t rc = 0;
 187         size_t page_size = cl_page_size(obj);
 188         size_t orig_size = size;
 189         bool do_io;
 190         int io_pages = 0;
 191
 192         ENTRY;
 193         queue = &io->ci_queue;
 194         cl_2queue_init(queue);
 195         for (i = 0; i < page_count; i++) {
 196                 LASSERT(!(file_offset & (page_size - 1)));
 197                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 198                                    pages[i], CPT_TRANSIENT);
 199                 if (IS_ERR(clp)) {
 200                         rc = PTR_ERR(clp);
 201                         break;
 202                 }
 203
 204                 rc = cl_page_own(env, io, clp);
 205                 if (rc) {
 206                         LASSERT(clp->cp_state == CPS_FREEING);
 207                         cl_page_put(env, clp);
 208                         break;
 209                 }
 210
 211                 do_io = true;
 212
 213                 /* check the page type: if the page is a host page, then do
 214                  * write directly
 215                  */
 216                 if (clp->cp_type == CPT_CACHEABLE) {
 217                         struct page *vmpage = cl_page_vmpage(clp);
 218                         struct page *src_page;
 219                         struct page *dst_page;
 220                         void *src;
 221                         void *dst;
 222
 223                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 224                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 225
 226                         src = ll_kmap_atomic(src_page, KM_USER0);
 227                         dst = ll_kmap_atomic(dst_page, KM_USER1);
 228                         memcpy(dst, src, min(page_size, size));
 229                         ll_kunmap_atomic(dst, KM_USER1);
 230                         ll_kunmap_atomic(src, KM_USER0);
 231
 232                         /* make sure page will be added to the transfer by
 233                          * cl_io_submit()->...->vvp_page_prep_write().
 234                          */
 235                         if (rw == WRITE)
 236                                 set_page_dirty(vmpage);
 237
 238                         if (rw == READ) {
 239                                 /* do not issue the page for read, since it
 240                                  * may reread a ra page which has NOT uptodate
 241                                  * bit set.
 242                                  */
 243                                 cl_page_disown(env, io, clp);
 244                                 do_io = false;
 245                         }
 246                 }
 247
 248                 if (likely(do_io)) {
 249                         cl_2queue_add(queue, clp);
 250
 251                         /*
 252                          * Set page clip to tell transfer formation engine
 253                          * that page has to be sent even if it is beyond KMS.
 254                          */
 255                         cl_page_clip(env, clp, 0, min(size, page_size));
 256
 257                         ++io_pages;
 258                 }
 259
 260                 /* drop the reference count for cl_page_find */
 261                 cl_page_put(env, clp);
 262                 size -= page_size;
 263                 file_offset += page_size;
 264         }
 265
 266         if (rc == 0 && io_pages) {
 267                 rc = cl_io_submit_sync(env, io,
 268                                        rw == READ ? CRT_READ : CRT_WRITE,
 269                                        queue, 0);
 270         }
 271         if (rc == 0)
 272                 rc = orig_size;
 273
 274         cl_2queue_discard(env, io, queue);
 275         cl_2queue_disown(env, io, queue);
 276         cl_2queue_fini(env, queue);
 277         RETURN(rc);
 278 }
 279
 280 /*  ll_free_user_pages - tear down page struct array
 281  *  @pages: array of page struct pointers underlying target buffer */
 282 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 283 {
 284         int i;
 285
 286         for (i = 0; i < npages; i++) {
 287                 if (pages[i] == NULL)
 288                         break;
 289                 if (do_dirty)
 290                         set_page_dirty_lock(pages[i]);
 291                 put_page(pages[i]);
 292         }
 293
 294 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
 295         kvfree(pages);
 296 #else
 297         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 298 #endif
 299 }
 300
 301 #ifdef KMALLOC_MAX_SIZE
 302 #define MAX_MALLOC KMALLOC_MAX_SIZE
 303 #else
 304 #define MAX_MALLOC (128 * 1024)
 305 #endif
 306
 307 /* This is the maximum size of a single O_DIRECT request, based on the
 308  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 309  * representing PAGE_SIZE worth of user data, into a single buffer, and
 310  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 311  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 312 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_SIZE) & \
 313                       ~(DT_MAX_BRW_SIZE - 1))
 314
 315 #ifndef HAVE_IOV_ITER_RW
 316 # define iov_iter_rw(iter)      rw
 317 #endif
 318
 319 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
 320 static ssize_t
 321 ll_direct_IO(
 322 # ifndef HAVE_IOV_ITER_RW
 323              int rw,
 324 # endif
 325              struct kiocb *iocb, struct iov_iter *iter
 326 # ifndef HAVE_DIRECTIO_2ARGS
 327              , loff_t file_offset
 328 # endif
 329              )
 330 {
 331 #ifdef HAVE_DIRECTIO_2ARGS
 332         loff_t file_offset = iocb->ki_pos;
 333 #endif
 334         struct ll_cl_context *lcc;
 335         const struct lu_env *env;
 336         struct cl_io *io;
 337         struct file *file = iocb->ki_filp;
 338         struct inode *inode = file->f_mapping->host;
 339         ssize_t count = iov_iter_count(iter);
 340         ssize_t tot_bytes = 0, result = 0;
 341         size_t size = MAX_DIO_SIZE;
 342
 343         /* Check EOF by ourselves */
 344         if (iov_iter_rw(iter) == READ && file_offset >= i_size_read(inode))
 345                 return 0;
 346         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 347         if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
 348                 return -EINVAL;
 349
 350         CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
 351                "offset=%lld=%llx, pages %zd (max %lu)\n",
 352                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 353                file_offset, file_offset, count >> PAGE_SHIFT,
 354                MAX_DIO_SIZE >> PAGE_SHIFT);
 355
 356         /* Check that all user buffers are aligned as well */
 357         if (iov_iter_alignment(iter) & ~PAGE_MASK)
 358                 return -EINVAL;
 359
 360         lcc = ll_cl_find(file);
 361         if (lcc == NULL)
 362                 RETURN(-EIO);
 363
 364         env = lcc->lcc_env;
 365         LASSERT(!IS_ERR(env));
 366         io = lcc->lcc_io;
 367         LASSERT(io != NULL);
 368
 369         /* 0. Need locking between buffered and direct access. and race with
 370          *    size changing by concurrent truncates and writes.
 371          * 1. Need inode mutex to operate transient pages.
 372          */
 373         if (iov_iter_rw(iter) == READ)
 374                 inode_lock(inode);
 375
 376         while (iov_iter_count(iter)) {
 377                 struct page **pages;
 378                 size_t offs;
 379
 380                 count = min_t(size_t, iov_iter_count(iter), size);
 381                 if (iov_iter_rw(iter) == READ) {
 382                         if (file_offset >= i_size_read(inode))
 383                                 break;
 384
 385                         if (file_offset + count > i_size_read(inode))
 386                                 count = i_size_read(inode) - file_offset;
 387                 }
 388
 389                 result = iov_iter_get_pages_alloc(iter, &pages, count, &offs);
 390                 if (likely(result > 0)) {
 391                         int n = DIV_ROUND_UP(result + offs, PAGE_SIZE);
 392
 393                         result = ll_direct_IO_seg(env, io, iov_iter_rw(iter),
 394                                                   inode, result, file_offset,
 395                                                   pages, n);
 396                         ll_free_user_pages(pages, n,
 397                                            iov_iter_rw(iter) == READ);
 398
 399                 }
 400                 if (unlikely(result <= 0)) {
 401                         /* If we can't allocate a large enough buffer
 402                          * for the request, shrink it to a smaller
 403                          * PAGE_SIZE multiple and try again.
 404                          * We should always be able to kmalloc for a
 405                          * page worth of page pointers = 4MB on i386. */
 406                         if (result == -ENOMEM &&
 407                             size > (PAGE_SIZE / sizeof(*pages)) *
 408                                     PAGE_SIZE) {
 409                                 size = ((((size / 2) - 1) |
 410                                         ~PAGE_MASK) + 1) & PAGE_MASK;
 411                                 CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
 412                                        size);
 413                                 continue;
 414                         }
 415
 416                         GOTO(out, result);
 417                 }
 418
 419                 iov_iter_advance(iter, result);
 420                 tot_bytes += result;
 421                 file_offset += result;
 422         }
 423 out:
 424         if (iov_iter_rw(iter) == READ)
 425                 inode_unlock(inode);
 426
 427         if (tot_bytes > 0) {
 428                 struct vvp_io *vio = vvp_env_io(env);
 429
 430                 /* no commit async for direct IO */
 431                 vio->u.write.vui_written += tot_bytes;
 432         }
 433
 434         return tot_bytes ? : result;
 435 }
 436 #else /* !HAVE_DIRECTIO_ITER && !HAVE_IOV_ITER_RW */
 437
 438 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 439                                     size_t size, struct page ***pages,
 440                                     int *max_pages)
 441 {
 442         int result = -ENOMEM;
 443
 444         /* set an arbitrary limit to prevent arithmetic overflow */
 445         if (size > MAX_DIRECTIO_SIZE) {
 446                 *pages = NULL;
 447                 return -EFBIG;
 448         }
 449
 450         *max_pages = (user_addr + size + PAGE_SIZE - 1) >>
 451                       PAGE_SHIFT;
 452         *max_pages -= user_addr >> PAGE_SHIFT;
 453
 454         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 455         if (*pages) {
 456                 down_read(&current->mm->mmap_sem);
 457                 result = get_user_pages(current, current->mm, user_addr,
 458                                         *max_pages, (rw == READ), 0, *pages,
 459                                         NULL);
 460                 up_read(&current->mm->mmap_sem);
 461                 if (unlikely(result <= 0))
 462                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 463         }
 464
 465         return result;
 466 }
 467
 468 static ssize_t
 469 ll_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
 470              loff_t file_offset, unsigned long nr_segs)
 471 {
 472         struct ll_cl_context *lcc;
 473         const struct lu_env *env;
 474         struct cl_io *io;
 475         struct file *file = iocb->ki_filp;
 476         struct inode *inode = file->f_mapping->host;
 477         ssize_t count = iov_length(iov, nr_segs);
 478         ssize_t tot_bytes = 0, result = 0;
 479         unsigned long seg = 0;
 480         size_t size = MAX_DIO_SIZE;
 481         ENTRY;
 482
 483         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 484         if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
 485                 RETURN(-EINVAL);
 486
 487         CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
 488                "offset=%lld=%llx, pages %zd (max %lu)\n",
 489                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 490                file_offset, file_offset, count >> PAGE_SHIFT,
 491                MAX_DIO_SIZE >> PAGE_SHIFT);
 492
 493         /* Check that all user buffers are aligned as well */
 494         for (seg = 0; seg < nr_segs; seg++) {
 495                 if (((unsigned long)iov[seg].iov_base & ~PAGE_MASK) ||
 496                     (iov[seg].iov_len & ~PAGE_MASK))
 497                         RETURN(-EINVAL);
 498         }
 499
 500         lcc = ll_cl_find(file);
 501         if (lcc == NULL)
 502                 RETURN(-EIO);
 503
 504         env = lcc->lcc_env;
 505         LASSERT(!IS_ERR(env));
 506         io = lcc->lcc_io;
 507         LASSERT(io != NULL);
 508
 509         for (seg = 0; seg < nr_segs; seg++) {
 510                 size_t iov_left = iov[seg].iov_len;
 511                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 512
 513                 if (rw == READ) {
 514                         if (file_offset >= i_size_read(inode))
 515                                 break;
 516                         if (file_offset + iov_left > i_size_read(inode))
 517                                 iov_left = i_size_read(inode) - file_offset;
 518                 }
 519
 520                 while (iov_left > 0) {
 521                         struct page **pages;
 522                         int page_count, max_pages = 0;
 523                         size_t bytes;
 524
 525                         bytes = min(size, iov_left);
 526                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 527                                                        &pages, &max_pages);
 528                         if (likely(page_count > 0)) {
 529                                 if (unlikely(page_count <  max_pages))
 530                                         bytes = page_count << PAGE_SHIFT;
 531                                 result = ll_direct_IO_seg(env, io, rw, inode,
 532                                                           bytes, file_offset,
 533                                                           pages, page_count);
 534                                 ll_free_user_pages(pages, max_pages, rw==READ);
 535                         } else if (page_count == 0) {
 536                                 GOTO(out, result = -EFAULT);
 537                         } else {
 538                                 result = page_count;
 539                         }
 540                         if (unlikely(result <= 0)) {
 541                                 /* If we can't allocate a large enough buffer
 542                                  * for the request, shrink it to a smaller
 543                                  * PAGE_SIZE multiple and try again.
 544                                  * We should always be able to kmalloc for a
 545                                  * page worth of page pointers = 4MB on i386. */
 546                                 if (result == -ENOMEM &&
 547                                     size > (PAGE_SIZE / sizeof(*pages)) *
 548                                            PAGE_SIZE) {
 549                                         size = ((((size / 2) - 1) |
 550                                                  ~PAGE_MASK) + 1) &
 551                                                 PAGE_MASK;
 552                                         CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
 553                                                size);
 554                                         continue;
 555                                 }
 556
 557                                 GOTO(out, result);
 558                         }
 559
 560                         tot_bytes += result;
 561                         file_offset += result;
 562                         iov_left -= result;
 563                         user_addr += result;
 564                 }
 565         }
 566 out:
 567         if (tot_bytes > 0) {
 568                 struct vvp_io *vio = vvp_env_io(env);
 569
 570                 /* no commit async for direct IO */
 571                 vio->u.write.vui_written += tot_bytes;
 572         }
 573
 574         RETURN(tot_bytes ? tot_bytes : result);
 575 }
 576 #endif /* HAVE_DIRECTIO_ITER || HAVE_IOV_ITER_RW */
 577
 578 /**
 579  * Prepare partially written-to page for a write.
 580  * @pg is owned when passed in and disowned when it returns non-zero result to
 581  * the caller.
 582  */
 583 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
 584                                    struct cl_page *pg, struct file *file)
 585 {
 586         struct cl_attr *attr   = vvp_env_thread_attr(env);
 587         struct cl_object *obj  = io->ci_obj;
 588         struct vvp_page *vpg   = cl_object_page_slice(obj, pg);
 589         loff_t          offset = cl_offset(obj, vvp_index(vpg));
 590         int             result;
 591         ENTRY;
 592
 593         cl_object_attr_lock(obj);
 594         result = cl_object_attr_get(env, obj, attr);
 595         cl_object_attr_unlock(obj);
 596         if (result) {
 597                 cl_page_disown(env, io, pg);
 598                 GOTO(out, result);
 599         }
 600
 601         /*
 602          * If are writing to a new page, no need to read old data.
 603          * The extent locking will have updated the KMS, and for our
 604          * purposes here we can treat it like i_size.
 605          */
 606         if (attr->cat_kms <= offset) {
 607                 char *kaddr = ll_kmap_atomic(vpg->vpg_page, KM_USER0);
 608
 609                 memset(kaddr, 0, cl_page_size(obj));
 610                 ll_kunmap_atomic(kaddr, KM_USER0);
 611                 GOTO(out, result = 0);
 612         }
 613
 614         if (vpg->vpg_defer_uptodate) {
 615                 vpg->vpg_ra_used = 1;
 616                 GOTO(out, result = 0);
 617         }
 618
 619         result = ll_io_read_page(env, io, pg, file);
 620         if (result)
 621                 GOTO(out, result);
 622
 623         /* ll_io_read_page() disowns the page */
 624         result = cl_page_own(env, io, pg);
 625         if (!result) {
 626                 if (!PageUptodate(cl_page_vmpage(pg))) {
 627                         cl_page_disown(env, io, pg);
 628                         result = -EIO;
 629                 }
 630         } else if (result == -ENOENT) {
 631                 /* page was truncated */
 632                 result = -EAGAIN;
 633         }
 634         EXIT;
 635
 636 out:
 637         return result;
 638 }
 639
 640 static int ll_tiny_write_begin(struct page *vmpage)
 641 {
 642         /* Page must be present, up to date, dirty, and not in writeback. */
 643         if (!vmpage || !PageUptodate(vmpage) || !PageDirty(vmpage) ||
 644             PageWriteback(vmpage))
 645                 return -ENODATA;
 646
 647         return 0;
 648 }
 649
 650 static int ll_write_begin(struct file *file, struct address_space *mapping,
 651                           loff_t pos, unsigned len, unsigned flags,
 652                           struct page **pagep, void **fsdata)
 653 {
 654         struct ll_cl_context *lcc = NULL;
 655         const struct lu_env  *env = NULL;
 656         struct cl_io   *io = NULL;
 657         struct cl_page *page = NULL;
 658
 659         struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
 660         pgoff_t index = pos >> PAGE_SHIFT;
 661         struct page *vmpage = NULL;
 662         unsigned from = pos & (PAGE_SIZE - 1);
 663         unsigned to = from + len;
 664         int result = 0;
 665         ENTRY;
 666
 667         CDEBUG(D_PAGE, "Writing %lu of %d to %d bytes\n", index, from, len);
 668
 669         lcc = ll_cl_find(file);
 670         if (lcc == NULL) {
 671                 vmpage = grab_cache_page_nowait(mapping, index);
 672                 result = ll_tiny_write_begin(vmpage);
 673                 GOTO(out, result);
 674         }
 675
 676         env = lcc->lcc_env;
 677         io  = lcc->lcc_io;
 678
 679         if (file->f_flags & O_DIRECT && io->ci_designated_mirror > 0) {
 680                 /* direct IO failed because it couldn't clean up cached pages,
 681                  * this causes a problem for mirror write because the cached
 682                  * page may belong to another mirror, which will result in
 683                  * problem submitting the I/O. */
 684                 GOTO(out, result = -EBUSY);
 685         }
 686
 687 again:
 688         /* To avoid deadlock, try to lock page first. */
 689         vmpage = grab_cache_page_nowait(mapping, index);
 690
 691         if (unlikely(vmpage == NULL ||
 692                      PageDirty(vmpage) || PageWriteback(vmpage))) {
 693                 struct vvp_io *vio = vvp_env_io(env);
 694                 struct cl_page_list *plist = &vio->u.write.vui_queue;
 695
 696                 /* if the page is already in dirty cache, we have to commit
 697                  * the pages right now; otherwise, it may cause deadlock
 698                  * because it holds page lock of a dirty page and request for
 699                  * more grants. It's okay for the dirty page to be the first
 700                  * one in commit page list, though. */
 701                 if (vmpage != NULL && plist->pl_nr > 0) {
 702                         unlock_page(vmpage);
 703                         put_page(vmpage);
 704                         vmpage = NULL;
 705                 }
 706
 707                 /* commit pages and then wait for page lock */
 708                 result = vvp_io_write_commit(env, io);
 709                 if (result < 0)
 710                         GOTO(out, result);
 711
 712                 if (vmpage == NULL) {
 713                         vmpage = grab_cache_page_write_begin(mapping, index,
 714                                                              flags);
 715                         if (vmpage == NULL)
 716                                 GOTO(out, result = -ENOMEM);
 717                 }
 718         }
 719
 720         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 721         if (IS_ERR(page))
 722                 GOTO(out, result = PTR_ERR(page));
 723
 724         lcc->lcc_page = page;
 725         lu_ref_add(&page->cp_reference, "cl_io", io);
 726
 727         cl_page_assume(env, io, page);
 728         if (!PageUptodate(vmpage)) {
 729                 /*
 730                  * We're completely overwriting an existing page,
 731                  * so _don't_ set it up to date until commit_write
 732                  */
 733                 if (from == 0 && to == PAGE_SIZE) {
 734                         CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
 735                         POISON_PAGE(vmpage, 0x11);
 736                 } else {
 737                         /* TODO: can be optimized at OSC layer to check if it
 738                          * is a lockless IO. In that case, it's not necessary
 739                          * to read the data. */
 740                         result = ll_prepare_partial_page(env, io, page, file);
 741                         if (result) {
 742                                 /* vmpage should have been unlocked */
 743                                 put_page(vmpage);
 744                                 vmpage = NULL;
 745
 746                                 if (result == -EAGAIN)
 747                                         goto again;
 748                                 GOTO(out, result);
 749                         }
 750                 }
 751         }
 752         EXIT;
 753 out:
 754         if (result < 0) {
 755                 if (vmpage != NULL) {
 756                         unlock_page(vmpage);
 757                         put_page(vmpage);
 758                 }
 759                 /* On tiny_write failure, page and io are always null. */
 760                 if (!IS_ERR_OR_NULL(page)) {
 761                         lu_ref_del(&page->cp_reference, "cl_io", io);
 762                         cl_page_put(env, page);
 763                 }
 764                 if (io)
 765                         io->ci_result = result;
 766         } else {
 767                 *pagep = vmpage;
 768                 *fsdata = lcc;
 769         }
 770         RETURN(result);
 771 }
 772
 773 static int ll_tiny_write_end(struct file *file, struct address_space *mapping,
 774                              loff_t pos, unsigned int len, unsigned int copied,
 775                              struct page *vmpage)
 776 {
 777         struct cl_page *clpage = (struct cl_page *) vmpage->private;
 778         loff_t kms = pos+copied;
 779         loff_t to = kms & (PAGE_SIZE-1) ? kms & (PAGE_SIZE-1) : PAGE_SIZE;
 780         __u16 refcheck;
 781         struct lu_env *env = cl_env_get(&refcheck);
 782         int rc = 0;
 783
 784         ENTRY;
 785
 786         if (IS_ERR(env)) {
 787                 rc = PTR_ERR(env);
 788                 goto out;
 789         }
 790
 791         /* This page is dirty in cache, so it should have a cl_page pointer
 792          * set in vmpage->private.
 793          */
 794         LASSERT(clpage != NULL);
 795
 796         if (copied == 0)
 797                 goto out_env;
 798
 799         /* Update the underlying size information in the OSC/LOV objects this
 800          * page is part of.
 801          */
 802         cl_page_touch(env, clpage, to);
 803
 804 out_env:
 805         cl_env_put(env, &refcheck);
 806
 807 out:
 808         /* Must return page unlocked. */
 809         unlock_page(vmpage);
 810
 811         RETURN(rc);
 812 }
 813
 814 static int ll_write_end(struct file *file, struct address_space *mapping,
 815                         loff_t pos, unsigned len, unsigned copied,
 816                         struct page *vmpage, void *fsdata)
 817 {
 818         struct ll_cl_context *lcc = fsdata;
 819         const struct lu_env *env;
 820         struct cl_io *io;
 821         struct vvp_io *vio;
 822         struct cl_page *page;
 823         unsigned from = pos & (PAGE_SIZE - 1);
 824         bool unplug = false;
 825         int result = 0;
 826         ENTRY;
 827
 828         put_page(vmpage);
 829
 830         CDEBUG(D_VFSTRACE, "pos %llu, len %u, copied %u\n", pos, len, copied);
 831
 832         if (lcc == NULL) {
 833                 result = ll_tiny_write_end(file, mapping, pos, len, copied,
 834                                            vmpage);
 835                 GOTO(out, result);
 836         }
 837
 838         LASSERT(lcc != NULL);
 839         env  = lcc->lcc_env;
 840         page = lcc->lcc_page;
 841         io   = lcc->lcc_io;
 842         vio  = vvp_env_io(env);
 843
 844         LASSERT(cl_page_is_owned(page, io));
 845         if (copied > 0) {
 846                 struct cl_page_list *plist = &vio->u.write.vui_queue;
 847
 848                 lcc->lcc_page = NULL; /* page will be queued */
 849
 850                 /* Add it into write queue */
 851                 cl_page_list_add(plist, page);
 852                 if (plist->pl_nr == 1) /* first page */
 853                         vio->u.write.vui_from = from;
 854                 else
 855                         LASSERT(from == 0);
 856                 vio->u.write.vui_to = from + copied;
 857
 858                 /* To address the deadlock in balance_dirty_pages() where
 859                  * this dirty page may be written back in the same thread. */
 860                 if (PageDirty(vmpage))
 861                         unplug = true;
 862
 863                 /* We may have one full RPC, commit it soon */
 864                 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
 865                         unplug = true;
 866
 867                 CL_PAGE_DEBUG(D_PAGE, env, page,
 868                               "queued page: %d.\n", plist->pl_nr);
 869         } else {
 870                 cl_page_disown(env, io, page);
 871
 872                 lcc->lcc_page = NULL;
 873                 lu_ref_del(&page->cp_reference, "cl_io", io);
 874                 cl_page_put(env, page);
 875
 876                 /* page list is not contiguous now, commit it now */
 877                 unplug = true;
 878         }
 879         if (unplug || io->u.ci_rw.rw_sync)
 880                 result = vvp_io_write_commit(env, io);
 881
 882         if (result < 0)
 883                 io->ci_result = result;
 884
 885
 886 out:
 887         RETURN(result >= 0 ? copied : result);
 888 }
 889
 890 #ifdef CONFIG_MIGRATION
 891 static int ll_migratepage(struct address_space *mapping,
 892                           struct page *newpage, struct page *page
 893 #ifdef HAVE_MIGRATEPAGE_4ARGS
 894                           , enum migrate_mode mode
 895 #endif
 896         )
 897 {
 898         /* Always fail page migration until we have a proper implementation */
 899         return -EIO;
 900 }
 901 #endif
 902
 903 const struct address_space_operations ll_aops = {
 904         .readpage       = ll_readpage,
 905         .direct_IO      = ll_direct_IO,
 906         .writepage      = ll_writepage,
 907         .writepages     = ll_writepages,
 908         .set_page_dirty = __set_page_dirty_nobuffers,
 909         .write_begin    = ll_write_begin,
 910         .write_end      = ll_write_end,
 911         .invalidatepage = ll_invalidatepage,
 912         .releasepage    = (void *)ll_releasepage,
 913 #ifdef CONFIG_MIGRATION
 914         .migratepage    = ll_migratepage,
 915 #endif
 916 };