fs/nfs/pagelist.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/pagelist.c
   4  *
   5  * A set of helper functions for managing NFS read and write requests.
   6  * The main purpose of these routines is to provide support for the
   7  * coalescing of several requests into a single RPC call.
   8  *
   9  * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
  10  *
  11  */
  12
  13 #include <linux/slab.h>
  14 #include <linux/file.h>
  15 #include <linux/sched.h>
  16 #include <linux/sunrpc/clnt.h>
  17 #include <linux/nfs.h>
  18 #include <linux/nfs3.h>
  19 #include <linux/nfs4.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_page.h>
  22 #include <linux/nfs_mount.h>
  23 #include <linux/export.h>
  24
  25 #include "internal.h"
  26 #include "pnfs.h"
  27 #include "nfstrace.h"
  28
  29 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  30
  31 static struct kmem_cache *nfs_page_cachep;
  32 static const struct rpc_call_ops nfs_pgio_common_ops;
  33
  34 static struct nfs_pgio_mirror *
  35 nfs_pgio_get_mirror(struct nfs_pageio_descriptor *desc, u32 idx)
  36 {
  37         if (desc->pg_ops->pg_get_mirror)
  38                 return desc->pg_ops->pg_get_mirror(desc, idx);
  39         return &desc->pg_mirrors[0];
  40 }
  41
  42 struct nfs_pgio_mirror *
  43 nfs_pgio_current_mirror(struct nfs_pageio_descriptor *desc)
  44 {
  45         return nfs_pgio_get_mirror(desc, desc->pg_mirror_idx);
  46 }
  47 EXPORT_SYMBOL_GPL(nfs_pgio_current_mirror);
  48
  49 static u32
  50 nfs_pgio_set_current_mirror(struct nfs_pageio_descriptor *desc, u32 idx)
  51 {
  52         if (desc->pg_ops->pg_set_mirror)
  53                 return desc->pg_ops->pg_set_mirror(desc, idx);
  54         return desc->pg_mirror_idx;
  55 }
  56
  57 void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
  58                        struct nfs_pgio_header *hdr,
  59                        void (*release)(struct nfs_pgio_header *hdr))
  60 {
  61         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
  62
  63
  64         hdr->req = nfs_list_entry(mirror->pg_list.next);
  65         hdr->inode = desc->pg_inode;
  66         hdr->cred = nfs_req_openctx(hdr->req)->cred;
  67         hdr->io_start = req_offset(hdr->req);
  68         hdr->good_bytes = mirror->pg_count;
  69         hdr->io_completion = desc->pg_io_completion;
  70         hdr->dreq = desc->pg_dreq;
  71         hdr->release = release;
  72         hdr->completion_ops = desc->pg_completion_ops;
  73         if (hdr->completion_ops->init_hdr)
  74                 hdr->completion_ops->init_hdr(hdr);
  75
  76         hdr->pgio_mirror_idx = desc->pg_mirror_idx;
  77 }
  78 EXPORT_SYMBOL_GPL(nfs_pgheader_init);
  79
  80 void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
  81 {
  82         unsigned int new = pos - hdr->io_start;
  83
  84         trace_nfs_pgio_error(hdr, error, pos);
  85         if (hdr->good_bytes > new) {
  86                 hdr->good_bytes = new;
  87                 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
  88                 if (!test_and_set_bit(NFS_IOHDR_ERROR, &hdr->flags))
  89                         hdr->error = error;
  90         }
  91 }
  92
  93 static inline struct nfs_page *
  94 nfs_page_alloc(void)
  95 {
  96         struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_KERNEL);
  97         if (p)
  98                 INIT_LIST_HEAD(&p->wb_list);
  99         return p;
 100 }
 101
 102 static inline void
 103 nfs_page_free(struct nfs_page *p)
 104 {
 105         kmem_cache_free(nfs_page_cachep, p);
 106 }
 107
 108 /**
 109  * nfs_iocounter_wait - wait for i/o to complete
 110  * @l_ctx: nfs_lock_context with io_counter to use
 111  *
 112  * returns -ERESTARTSYS if interrupted by a fatal signal.
 113  * Otherwise returns 0 once the io_count hits 0.
 114  */
 115 int
 116 nfs_iocounter_wait(struct nfs_lock_context *l_ctx)
 117 {
 118         return wait_var_event_killable(&l_ctx->io_count,
 119                                        !atomic_read(&l_ctx->io_count));
 120 }
 121
 122 /**
 123  * nfs_async_iocounter_wait - wait on a rpc_waitqueue for I/O
 124  * to complete
 125  * @task: the rpc_task that should wait
 126  * @l_ctx: nfs_lock_context with io_counter to check
 127  *
 128  * Returns true if there is outstanding I/O to wait on and the
 129  * task has been put to sleep.
 130  */
 131 bool
 132 nfs_async_iocounter_wait(struct rpc_task *task, struct nfs_lock_context *l_ctx)
 133 {
 134         struct inode *inode = d_inode(l_ctx->open_context->dentry);
 135         bool ret = false;
 136
 137         if (atomic_read(&l_ctx->io_count) > 0) {
 138                 rpc_sleep_on(&NFS_SERVER(inode)->uoc_rpcwaitq, task, NULL);
 139                 ret = true;
 140         }
 141
 142         if (atomic_read(&l_ctx->io_count) == 0) {
 143                 rpc_wake_up_queued_task(&NFS_SERVER(inode)->uoc_rpcwaitq, task);
 144                 ret = false;
 145         }
 146
 147         return ret;
 148 }
 149 EXPORT_SYMBOL_GPL(nfs_async_iocounter_wait);
 150
 151 /*
 152  * nfs_page_lock_head_request - page lock the head of the page group
 153  * @req: any member of the page group
 154  */
 155 struct nfs_page *
 156 nfs_page_group_lock_head(struct nfs_page *req)
 157 {
 158         struct nfs_page *head = req->wb_head;
 159
 160         while (!nfs_lock_request(head)) {
 161                 int ret = nfs_wait_on_request(head);
 162                 if (ret < 0)
 163                         return ERR_PTR(ret);
 164         }
 165         if (head != req)
 166                 kref_get(&head->wb_kref);
 167         return head;
 168 }
 169
 170 /*
 171  * nfs_unroll_locks -  unlock all newly locked reqs and wait on @req
 172  * @head: head request of page group, must be holding head lock
 173  * @req: request that couldn't lock and needs to wait on the req bit lock
 174  *
 175  * This is a helper function for nfs_lock_and_join_requests
 176  * returns 0 on success, < 0 on error.
 177  */
 178 static void
 179 nfs_unroll_locks(struct nfs_page *head, struct nfs_page *req)
 180 {
 181         struct nfs_page *tmp;
 182
 183         /* relinquish all the locks successfully grabbed this run */
 184         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
 185                 if (!kref_read(&tmp->wb_kref))
 186                         continue;
 187                 nfs_unlock_and_release_request(tmp);
 188         }
 189 }
 190
 191 /*
 192  * nfs_page_group_lock_subreq -  try to lock a subrequest
 193  * @head: head request of page group
 194  * @subreq: request to lock
 195  *
 196  * This is a helper function for nfs_lock_and_join_requests which
 197  * must be called with the head request and page group both locked.
 198  * On error, it returns with the page group unlocked.
 199  */
 200 static int
 201 nfs_page_group_lock_subreq(struct nfs_page *head, struct nfs_page *subreq)
 202 {
 203         int ret;
 204
 205         if (!kref_get_unless_zero(&subreq->wb_kref))
 206                 return 0;
 207         while (!nfs_lock_request(subreq)) {
 208                 nfs_page_group_unlock(head);
 209                 ret = nfs_wait_on_request(subreq);
 210                 if (!ret)
 211                         ret = nfs_page_group_lock(head);
 212                 if (ret < 0) {
 213                         nfs_unroll_locks(head, subreq);
 214                         nfs_release_request(subreq);
 215                         return ret;
 216                 }
 217         }
 218         return 0;
 219 }
 220
 221 /*
 222  * nfs_page_group_lock_subrequests -  try to lock the subrequests
 223  * @head: head request of page group
 224  *
 225  * This is a helper function for nfs_lock_and_join_requests which
 226  * must be called with the head request locked.
 227  */
 228 int nfs_page_group_lock_subrequests(struct nfs_page *head)
 229 {
 230         struct nfs_page *subreq;
 231         int ret;
 232
 233         ret = nfs_page_group_lock(head);
 234         if (ret < 0)
 235                 return ret;
 236         /* lock each request in the page group */
 237         for (subreq = head->wb_this_page; subreq != head;
 238                         subreq = subreq->wb_this_page) {
 239                 ret = nfs_page_group_lock_subreq(head, subreq);
 240                 if (ret < 0)
 241                         return ret;
 242         }
 243         nfs_page_group_unlock(head);
 244         return 0;
 245 }
 246
 247 /*
 248  * nfs_page_set_headlock - set the request PG_HEADLOCK
 249  * @req: request that is to be locked
 250  *
 251  * this lock must be held when modifying req->wb_head
 252  *
 253  * return 0 on success, < 0 on error
 254  */
 255 int
 256 nfs_page_set_headlock(struct nfs_page *req)
 257 {
 258         if (!test_and_set_bit(PG_HEADLOCK, &req->wb_flags))
 259                 return 0;
 260
 261         set_bit(PG_CONTENDED1, &req->wb_flags);
 262         smp_mb__after_atomic();
 263         return wait_on_bit_lock(&req->wb_flags, PG_HEADLOCK,
 264                                 TASK_UNINTERRUPTIBLE);
 265 }
 266
 267 /*
 268  * nfs_page_clear_headlock - clear the request PG_HEADLOCK
 269  * @req: request that is to be locked
 270  */
 271 void
 272 nfs_page_clear_headlock(struct nfs_page *req)
 273 {
 274         smp_mb__before_atomic();
 275         clear_bit(PG_HEADLOCK, &req->wb_flags);
 276         smp_mb__after_atomic();
 277         if (!test_bit(PG_CONTENDED1, &req->wb_flags))
 278                 return;
 279         wake_up_bit(&req->wb_flags, PG_HEADLOCK);
 280 }
 281
 282 /*
 283  * nfs_page_group_lock - lock the head of the page group
 284  * @req: request in group that is to be locked
 285  *
 286  * this lock must be held when traversing or modifying the page
 287  * group list
 288  *
 289  * return 0 on success, < 0 on error
 290  */
 291 int
 292 nfs_page_group_lock(struct nfs_page *req)
 293 {
 294         int ret;
 295
 296         ret = nfs_page_set_headlock(req);
 297         if (ret || req->wb_head == req)
 298                 return ret;
 299         return nfs_page_set_headlock(req->wb_head);
 300 }
 301
 302 /*
 303  * nfs_page_group_unlock - unlock the head of the page group
 304  * @req: request in group that is to be unlocked
 305  */
 306 void
 307 nfs_page_group_unlock(struct nfs_page *req)
 308 {
 309         if (req != req->wb_head)
 310                 nfs_page_clear_headlock(req->wb_head);
 311         nfs_page_clear_headlock(req);
 312 }
 313
 314 /*
 315  * nfs_page_group_sync_on_bit_locked
 316  *
 317  * must be called with page group lock held
 318  */
 319 static bool
 320 nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
 321 {
 322         struct nfs_page *head = req->wb_head;
 323         struct nfs_page *tmp;
 324
 325         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags));
 326         WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags));
 327
 328         tmp = req->wb_this_page;
 329         while (tmp != req) {
 330                 if (!test_bit(bit, &tmp->wb_flags))
 331                         return false;
 332                 tmp = tmp->wb_this_page;
 333         }
 334
 335         /* true! reset all bits */
 336         tmp = req;
 337         do {
 338                 clear_bit(bit, &tmp->wb_flags);
 339                 tmp = tmp->wb_this_page;
 340         } while (tmp != req);
 341
 342         return true;
 343 }
 344
 345 /*
 346  * nfs_page_group_sync_on_bit - set bit on current request, but only
 347  *   return true if the bit is set for all requests in page group
 348  * @req - request in page group
 349  * @bit - PG_* bit that is used to sync page group
 350  */
 351 bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit)
 352 {
 353         bool ret;
 354
 355         nfs_page_group_lock(req);
 356         ret = nfs_page_group_sync_on_bit_locked(req, bit);
 357         nfs_page_group_unlock(req);
 358
 359         return ret;
 360 }
 361
 362 /*
 363  * nfs_page_group_init - Initialize the page group linkage for @req
 364  * @req - a new nfs request
 365  * @prev - the previous request in page group, or NULL if @req is the first
 366  *         or only request in the group (the head).
 367  */
 368 static inline void
 369 nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev)
 370 {
 371         struct inode *inode;
 372         WARN_ON_ONCE(prev == req);
 373
 374         if (!prev) {
 375                 /* a head request */
 376                 req->wb_head = req;
 377                 req->wb_this_page = req;
 378         } else {
 379                 /* a subrequest */
 380                 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head);
 381                 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags));
 382                 req->wb_head = prev->wb_head;
 383                 req->wb_this_page = prev->wb_this_page;
 384                 prev->wb_this_page = req;
 385
 386                 /* All subrequests take a ref on the head request until
 387                  * nfs_page_group_destroy is called */
 388                 kref_get(&req->wb_head->wb_kref);
 389
 390                 /* grab extra ref and bump the request count if head request
 391                  * has extra ref from the write/commit path to handle handoff
 392                  * between write and commit lists. */
 393                 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) {
 394                         inode = page_file_mapping(req->wb_page)->host;
 395                         set_bit(PG_INODE_REF, &req->wb_flags);
 396                         kref_get(&req->wb_kref);
 397                         atomic_long_inc(&NFS_I(inode)->nrequests);
 398                 }
 399         }
 400 }
 401
 402 /*
 403  * nfs_page_group_destroy - sync the destruction of page groups
 404  * @req - request that no longer needs the page group
 405  *
 406  * releases the page group reference from each member once all
 407  * members have called this function.
 408  */
 409 static void
 410 nfs_page_group_destroy(struct kref *kref)
 411 {
 412         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 413         struct nfs_page *head = req->wb_head;
 414         struct nfs_page *tmp, *next;
 415
 416         if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN))
 417                 goto out;
 418
 419         tmp = req;
 420         do {
 421                 next = tmp->wb_this_page;
 422                 /* unlink and free */
 423                 tmp->wb_this_page = tmp;
 424                 tmp->wb_head = tmp;
 425                 nfs_free_request(tmp);
 426                 tmp = next;
 427         } while (tmp != req);
 428 out:
 429         /* subrequests must release the ref on the head request */
 430         if (head != req)
 431                 nfs_release_request(head);
 432 }
 433
 434 static struct nfs_page *
 435 __nfs_create_request(struct nfs_lock_context *l_ctx, struct page *page,
 436                    unsigned int pgbase, unsigned int offset,
 437                    unsigned int count)
 438 {
 439         struct nfs_page         *req;
 440         struct nfs_open_context *ctx = l_ctx->open_context;
 441
 442         if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
 443                 return ERR_PTR(-EBADF);
 444         /* try to allocate the request struct */
 445         req = nfs_page_alloc();
 446         if (req == NULL)
 447                 return ERR_PTR(-ENOMEM);
 448
 449         req->wb_lock_context = l_ctx;
 450         refcount_inc(&l_ctx->count);
 451         atomic_inc(&l_ctx->io_count);
 452
 453         /* Initialize the request struct. Initially, we assume a
 454          * long write-back delay. This will be adjusted in
 455          * update_nfs_request below if the region is not locked. */
 456         req->wb_page    = page;
 457         if (page) {
 458                 req->wb_index = page_index(page);
 459                 get_page(page);
 460         }
 461         req->wb_offset  = offset;
 462         req->wb_pgbase  = pgbase;
 463         req->wb_bytes   = count;
 464         kref_init(&req->wb_kref);
 465         req->wb_nio = 0;
 466         return req;
 467 }
 468
 469 /**
 470  * nfs_create_request - Create an NFS read/write request.
 471  * @ctx: open context to use
 472  * @page: page to write
 473  * @offset: starting offset within the page for the write
 474  * @count: number of bytes to read/write
 475  *
 476  * The page must be locked by the caller. This makes sure we never
 477  * create two different requests for the same page.
 478  * User should ensure it is safe to sleep in this function.
 479  */
 480 struct nfs_page *
 481 nfs_create_request(struct nfs_open_context *ctx, struct page *page,
 482                    unsigned int offset, unsigned int count)
 483 {
 484         struct nfs_lock_context *l_ctx = nfs_get_lock_context(ctx);
 485         struct nfs_page *ret;
 486
 487         if (IS_ERR(l_ctx))
 488                 return ERR_CAST(l_ctx);
 489         ret = __nfs_create_request(l_ctx, page, offset, offset, count);
 490         if (!IS_ERR(ret))
 491                 nfs_page_group_init(ret, NULL);
 492         nfs_put_lock_context(l_ctx);
 493         return ret;
 494 }
 495
 496 static struct nfs_page *
 497 nfs_create_subreq(struct nfs_page *req,
 498                   unsigned int pgbase,
 499                   unsigned int offset,
 500                   unsigned int count)
 501 {
 502         struct nfs_page *last;
 503         struct nfs_page *ret;
 504
 505         ret = __nfs_create_request(req->wb_lock_context, req->wb_page,
 506                         pgbase, offset, count);
 507         if (!IS_ERR(ret)) {
 508                 /* find the last request */
 509                 for (last = req->wb_head;
 510                      last->wb_this_page != req->wb_head;
 511                      last = last->wb_this_page)
 512                         ;
 513
 514                 nfs_lock_request(ret);
 515                 ret->wb_index = req->wb_index;
 516                 nfs_page_group_init(ret, last);
 517                 ret->wb_nio = req->wb_nio;
 518         }
 519         return ret;
 520 }
 521
 522 /**
 523  * nfs_unlock_request - Unlock request and wake up sleepers.
 524  * @req: pointer to request
 525  */
 526 void nfs_unlock_request(struct nfs_page *req)
 527 {
 528         if (!NFS_WBACK_BUSY(req)) {
 529                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 530                 BUG();
 531         }
 532         smp_mb__before_atomic();
 533         clear_bit(PG_BUSY, &req->wb_flags);
 534         smp_mb__after_atomic();
 535         if (!test_bit(PG_CONTENDED2, &req->wb_flags))
 536                 return;
 537         wake_up_bit(&req->wb_flags, PG_BUSY);
 538 }
 539
 540 /**
 541  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
 542  * @req: pointer to request
 543  */
 544 void nfs_unlock_and_release_request(struct nfs_page *req)
 545 {
 546         nfs_unlock_request(req);
 547         nfs_release_request(req);
 548 }
 549
 550 /*
 551  * nfs_clear_request - Free up all resources allocated to the request
 552  * @req:
 553  *
 554  * Release page and open context resources associated with a read/write
 555  * request after it has completed.
 556  */
 557 static void nfs_clear_request(struct nfs_page *req)
 558 {
 559         struct page *page = req->wb_page;
 560         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 561         struct nfs_open_context *ctx;
 562
 563         if (page != NULL) {
 564                 put_page(page);
 565                 req->wb_page = NULL;
 566         }
 567         if (l_ctx != NULL) {
 568                 if (atomic_dec_and_test(&l_ctx->io_count)) {
 569                         wake_up_var(&l_ctx->io_count);
 570                         ctx = l_ctx->open_context;
 571                         if (test_bit(NFS_CONTEXT_UNLOCK, &ctx->flags))
 572                                 rpc_wake_up(&NFS_SERVER(d_inode(ctx->dentry))->uoc_rpcwaitq);
 573                 }
 574                 nfs_put_lock_context(l_ctx);
 575                 req->wb_lock_context = NULL;
 576         }
 577 }
 578
 579 /**
 580  * nfs_release_request - Release the count on an NFS read/write request
 581  * @req: request to release
 582  *
 583  * Note: Should never be called with the spinlock held!
 584  */
 585 void nfs_free_request(struct nfs_page *req)
 586 {
 587         WARN_ON_ONCE(req->wb_this_page != req);
 588
 589         /* extra debug: make sure no sync bits are still set */
 590         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
 591         WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
 592         WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
 593         WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
 594         WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
 595
 596         /* Release struct file and open context */
 597         nfs_clear_request(req);
 598         nfs_page_free(req);
 599 }
 600
 601 void nfs_release_request(struct nfs_page *req)
 602 {
 603         kref_put(&req->wb_kref, nfs_page_group_destroy);
 604 }
 605 EXPORT_SYMBOL_GPL(nfs_release_request);
 606
 607 /**
 608  * nfs_wait_on_request - Wait for a request to complete.
 609  * @req: request to wait upon.
 610  *
 611  * Interruptible by fatal signals only.
 612  * The user is responsible for holding a count on the request.
 613  */
 614 int
 615 nfs_wait_on_request(struct nfs_page *req)
 616 {
 617         if (!test_bit(PG_BUSY, &req->wb_flags))
 618                 return 0;
 619         set_bit(PG_CONTENDED2, &req->wb_flags);
 620         smp_mb__after_atomic();
 621         return wait_on_bit_io(&req->wb_flags, PG_BUSY,
 622                               TASK_UNINTERRUPTIBLE);
 623 }
 624 EXPORT_SYMBOL_GPL(nfs_wait_on_request);
 625
 626 /*
 627  * nfs_generic_pg_test - determine if requests can be coalesced
 628  * @desc: pointer to descriptor
 629  * @prev: previous request in desc, or NULL
 630  * @req: this request
 631  *
 632  * Returns zero if @req cannot be coalesced into @desc, otherwise it returns
 633  * the size of the request.
 634  */
 635 size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc,
 636                            struct nfs_page *prev, struct nfs_page *req)
 637 {
 638         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
 639
 640
 641         if (mirror->pg_count > mirror->pg_bsize) {
 642                 /* should never happen */
 643                 WARN_ON_ONCE(1);
 644                 return 0;
 645         }
 646
 647         /*
 648          * Limit the request size so that we can still allocate a page array
 649          * for it without upsetting the slab allocator.
 650          */
 651         if (((mirror->pg_count + req->wb_bytes) >> PAGE_SHIFT) *
 652                         sizeof(struct page *) > PAGE_SIZE)
 653                 return 0;
 654
 655         return min(mirror->pg_bsize - mirror->pg_count, (size_t)req->wb_bytes);
 656 }
 657 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
 658
 659 struct nfs_pgio_header *nfs_pgio_header_alloc(const struct nfs_rw_ops *ops)
 660 {
 661         struct nfs_pgio_header *hdr = ops->rw_alloc_header();
 662
 663         if (hdr) {
 664                 INIT_LIST_HEAD(&hdr->pages);
 665                 hdr->rw_ops = ops;
 666         }
 667         return hdr;
 668 }
 669 EXPORT_SYMBOL_GPL(nfs_pgio_header_alloc);
 670
 671 /**
 672  * nfs_pgio_data_destroy - make @hdr suitable for reuse
 673  *
 674  * Frees memory and releases refs from nfs_generic_pgio, so that it may
 675  * be called again.
 676  *
 677  * @hdr: A header that has had nfs_generic_pgio called
 678  */
 679 static void nfs_pgio_data_destroy(struct nfs_pgio_header *hdr)
 680 {
 681         if (hdr->args.context)
 682                 put_nfs_open_context(hdr->args.context);
 683         if (hdr->page_array.pagevec != hdr->page_array.page_array)
 684                 kfree(hdr->page_array.pagevec);
 685 }
 686
 687 /*
 688  * nfs_pgio_header_free - Free a read or write header
 689  * @hdr: The header to free
 690  */
 691 void nfs_pgio_header_free(struct nfs_pgio_header *hdr)
 692 {
 693         nfs_pgio_data_destroy(hdr);
 694         hdr->rw_ops->rw_free_header(hdr);
 695 }
 696 EXPORT_SYMBOL_GPL(nfs_pgio_header_free);
 697
 698 /**
 699  * nfs_pgio_rpcsetup - Set up arguments for a pageio call
 700  * @hdr: The pageio hdr
 701  * @count: Number of bytes to read
 702  * @how: How to commit data (writes only)
 703  * @cinfo: Commit information for the call (writes only)
 704  */
 705 static void nfs_pgio_rpcsetup(struct nfs_pgio_header *hdr,
 706                               unsigned int count,
 707                               int how, struct nfs_commit_info *cinfo)
 708 {
 709         struct nfs_page *req = hdr->req;
 710
 711         /* Set up the RPC argument and reply structs
 712          * NB: take care not to mess about with hdr->commit et al. */
 713
 714         hdr->args.fh     = NFS_FH(hdr->inode);
 715         hdr->args.offset = req_offset(req);
 716         /* pnfs_set_layoutcommit needs this */
 717         hdr->mds_offset = hdr->args.offset;
 718         hdr->args.pgbase = req->wb_pgbase;
 719         hdr->args.pages  = hdr->page_array.pagevec;
 720         hdr->args.count  = count;
 721         hdr->args.context = get_nfs_open_context(nfs_req_openctx(req));
 722         hdr->args.lock_context = req->wb_lock_context;
 723         hdr->args.stable  = NFS_UNSTABLE;
 724         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 725         case 0:
 726                 break;
 727         case FLUSH_COND_STABLE:
 728                 if (nfs_reqs_to_commit(cinfo))
 729                         break;
 730                 fallthrough;
 731         default:
 732                 hdr->args.stable = NFS_FILE_SYNC;
 733         }
 734
 735         hdr->res.fattr   = &hdr->fattr;
 736         hdr->res.count   = 0;
 737         hdr->res.eof     = 0;
 738         hdr->res.verf    = &hdr->verf;
 739         nfs_fattr_init(&hdr->fattr);
 740 }
 741
 742 /**
 743  * nfs_pgio_prepare - Prepare pageio hdr to go over the wire
 744  * @task: The current task
 745  * @calldata: pageio header to prepare
 746  */
 747 static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
 748 {
 749         struct nfs_pgio_header *hdr = calldata;
 750         int err;
 751         err = NFS_PROTO(hdr->inode)->pgio_rpc_prepare(task, hdr);
 752         if (err)
 753                 rpc_exit(task, err);
 754 }
 755
 756 int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_header *hdr,
 757                       const struct cred *cred, const struct nfs_rpc_ops *rpc_ops,
 758                       const struct rpc_call_ops *call_ops, int how, int flags)
 759 {
 760         struct rpc_task *task;
 761         struct rpc_message msg = {
 762                 .rpc_argp = &hdr->args,
 763                 .rpc_resp = &hdr->res,
 764                 .rpc_cred = cred,
 765         };
 766         struct rpc_task_setup task_setup_data = {
 767                 .rpc_client = clnt,
 768                 .task = &hdr->task,
 769                 .rpc_message = &msg,
 770                 .callback_ops = call_ops,
 771                 .callback_data = hdr,
 772                 .workqueue = nfsiod_workqueue,
 773                 .flags = RPC_TASK_ASYNC | flags,
 774         };
 775
 776         hdr->rw_ops->rw_initiate(hdr, &msg, rpc_ops, &task_setup_data, how);
 777
 778         dprintk("NFS: initiated pgio call "
 779                 "(req %s/%llu, %u bytes @ offset %llu)\n",
 780                 hdr->inode->i_sb->s_id,
 781                 (unsigned long long)NFS_FILEID(hdr->inode),
 782                 hdr->args.count,
 783                 (unsigned long long)hdr->args.offset);
 784
 785         task = rpc_run_task(&task_setup_data);
 786         if (IS_ERR(task))
 787                 return PTR_ERR(task);
 788         rpc_put_task(task);
 789         return 0;
 790 }
 791 EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
 792
 793 /**
 794  * nfs_pgio_error - Clean up from a pageio error
 795  * @hdr: pageio header
 796  */
 797 static void nfs_pgio_error(struct nfs_pgio_header *hdr)
 798 {
 799         set_bit(NFS_IOHDR_REDO, &hdr->flags);
 800         hdr->completion_ops->completion(hdr);
 801 }
 802
 803 /**
 804  * nfs_pgio_release - Release pageio data
 805  * @calldata: The pageio header to release
 806  */
 807 static void nfs_pgio_release(void *calldata)
 808 {
 809         struct nfs_pgio_header *hdr = calldata;
 810         hdr->completion_ops->completion(hdr);
 811 }
 812
 813 static void nfs_pageio_mirror_init(struct nfs_pgio_mirror *mirror,
 814                                    unsigned int bsize)
 815 {
 816         INIT_LIST_HEAD(&mirror->pg_list);
 817         mirror->pg_bytes_written = 0;
 818         mirror->pg_count = 0;
 819         mirror->pg_bsize = bsize;
 820         mirror->pg_base = 0;
 821         mirror->pg_recoalesce = 0;
 822 }
 823
 824 /**
 825  * nfs_pageio_init - initialise a page io descriptor
 826  * @desc: pointer to descriptor
 827  * @inode: pointer to inode
 828  * @pg_ops: pointer to pageio operations
 829  * @compl_ops: pointer to pageio completion operations
 830  * @rw_ops: pointer to nfs read/write operations
 831  * @bsize: io block size
 832  * @io_flags: extra parameters for the io function
 833  */
 834 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 835                      struct inode *inode,
 836                      const struct nfs_pageio_ops *pg_ops,
 837                      const struct nfs_pgio_completion_ops *compl_ops,
 838                      const struct nfs_rw_ops *rw_ops,
 839                      size_t bsize,
 840                      int io_flags)
 841 {
 842         desc->pg_moreio = 0;
 843         desc->pg_inode = inode;
 844         desc->pg_ops = pg_ops;
 845         desc->pg_completion_ops = compl_ops;
 846         desc->pg_rw_ops = rw_ops;
 847         desc->pg_ioflags = io_flags;
 848         desc->pg_error = 0;
 849         desc->pg_lseg = NULL;
 850         desc->pg_io_completion = NULL;
 851         desc->pg_dreq = NULL;
 852         desc->pg_bsize = bsize;
 853
 854         desc->pg_mirror_count = 1;
 855         desc->pg_mirror_idx = 0;
 856
 857         desc->pg_mirrors_dynamic = NULL;
 858         desc->pg_mirrors = desc->pg_mirrors_static;
 859         nfs_pageio_mirror_init(&desc->pg_mirrors[0], bsize);
 860         desc->pg_maxretrans = 0;
 861 }
 862
 863 /**
 864  * nfs_pgio_result - Basic pageio error handling
 865  * @task: The task that ran
 866  * @calldata: Pageio header to check
 867  */
 868 static void nfs_pgio_result(struct rpc_task *task, void *calldata)
 869 {
 870         struct nfs_pgio_header *hdr = calldata;
 871         struct inode *inode = hdr->inode;
 872
 873         dprintk("NFS: %s: %5u, (status %d)\n", __func__,
 874                 task->tk_pid, task->tk_status);
 875
 876         if (hdr->rw_ops->rw_done(task, hdr, inode) != 0)
 877                 return;
 878         if (task->tk_status < 0)
 879                 nfs_set_pgio_error(hdr, task->tk_status, hdr->args.offset);
 880         else
 881                 hdr->rw_ops->rw_result(task, hdr);
 882 }
 883
 884 /*
 885  * Create an RPC task for the given read or write request and kick it.
 886  * The page must have been locked by the caller.
 887  *
 888  * It may happen that the page we're passed is not marked dirty.
 889  * This is the case if nfs_updatepage detects a conflicting request
 890  * that has been written but not committed.
 891  */
 892 int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
 893                      struct nfs_pgio_header *hdr)
 894 {
 895         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
 896
 897         struct nfs_page         *req;
 898         struct page             **pages,
 899                                 *last_page;
 900         struct list_head *head = &mirror->pg_list;
 901         struct nfs_commit_info cinfo;
 902         struct nfs_page_array *pg_array = &hdr->page_array;
 903         unsigned int pagecount, pageused;
 904         gfp_t gfp_flags = GFP_KERNEL;
 905
 906         pagecount = nfs_page_array_len(mirror->pg_base, mirror->pg_count);
 907         pg_array->npages = pagecount;
 908
 909         if (pagecount <= ARRAY_SIZE(pg_array->page_array))
 910                 pg_array->pagevec = pg_array->page_array;
 911         else {
 912                 pg_array->pagevec = kcalloc(pagecount, sizeof(struct page *), gfp_flags);
 913                 if (!pg_array->pagevec) {
 914                         pg_array->npages = 0;
 915                         nfs_pgio_error(hdr);
 916                         desc->pg_error = -ENOMEM;
 917                         return desc->pg_error;
 918                 }
 919         }
 920
 921         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
 922         pages = hdr->page_array.pagevec;
 923         last_page = NULL;
 924         pageused = 0;
 925         while (!list_empty(head)) {
 926                 req = nfs_list_entry(head->next);
 927                 nfs_list_move_request(req, &hdr->pages);
 928
 929                 if (!last_page || last_page != req->wb_page) {
 930                         pageused++;
 931                         if (pageused > pagecount)
 932                                 break;
 933                         *pages++ = last_page = req->wb_page;
 934                 }
 935         }
 936         if (WARN_ON_ONCE(pageused != pagecount)) {
 937                 nfs_pgio_error(hdr);
 938                 desc->pg_error = -EINVAL;
 939                 return desc->pg_error;
 940         }
 941
 942         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 943             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
 944                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 945
 946         /* Set up the argument struct */
 947         nfs_pgio_rpcsetup(hdr, mirror->pg_count, desc->pg_ioflags, &cinfo);
 948         desc->pg_rpc_callops = &nfs_pgio_common_ops;
 949         return 0;
 950 }
 951 EXPORT_SYMBOL_GPL(nfs_generic_pgio);
 952
 953 static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
 954 {
 955         struct nfs_pgio_header *hdr;
 956         int ret;
 957
 958         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
 959         if (!hdr) {
 960                 desc->pg_error = -ENOMEM;
 961                 return desc->pg_error;
 962         }
 963         nfs_pgheader_init(desc, hdr, nfs_pgio_header_free);
 964         ret = nfs_generic_pgio(desc, hdr);
 965         if (ret == 0)
 966                 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
 967                                         hdr,
 968                                         hdr->cred,
 969                                         NFS_PROTO(hdr->inode),
 970                                         desc->pg_rpc_callops,
 971                                         desc->pg_ioflags,
 972                                         RPC_TASK_CRED_NOREF);
 973         return ret;
 974 }
 975
 976 static struct nfs_pgio_mirror *
 977 nfs_pageio_alloc_mirrors(struct nfs_pageio_descriptor *desc,
 978                 unsigned int mirror_count)
 979 {
 980         struct nfs_pgio_mirror *ret;
 981         unsigned int i;
 982
 983         kfree(desc->pg_mirrors_dynamic);
 984         desc->pg_mirrors_dynamic = NULL;
 985         if (mirror_count == 1)
 986                 return desc->pg_mirrors_static;
 987         ret = kmalloc_array(mirror_count, sizeof(*ret), GFP_KERNEL);
 988         if (ret != NULL) {
 989                 for (i = 0; i < mirror_count; i++)
 990                         nfs_pageio_mirror_init(&ret[i], desc->pg_bsize);
 991                 desc->pg_mirrors_dynamic = ret;
 992         }
 993         return ret;
 994 }
 995
 996 /*
 997  * nfs_pageio_setup_mirroring - determine if mirroring is to be used
 998  *                              by calling the pg_get_mirror_count op
 999  */
1000 static void nfs_pageio_setup_mirroring(struct nfs_pageio_descriptor *pgio,
1001                                        struct nfs_page *req)
1002 {
1003         unsigned int mirror_count = 1;
1004
1005         if (pgio->pg_ops->pg_get_mirror_count)
1006                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
1007         if (mirror_count == pgio->pg_mirror_count || pgio->pg_error < 0)
1008                 return;
1009
1010         if (!mirror_count || mirror_count > NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX) {
1011                 pgio->pg_error = -EINVAL;
1012                 return;
1013         }
1014
1015         pgio->pg_mirrors = nfs_pageio_alloc_mirrors(pgio, mirror_count);
1016         if (pgio->pg_mirrors == NULL) {
1017                 pgio->pg_error = -ENOMEM;
1018                 pgio->pg_mirrors = pgio->pg_mirrors_static;
1019                 mirror_count = 1;
1020         }
1021         pgio->pg_mirror_count = mirror_count;
1022 }
1023
1024 static void nfs_pageio_cleanup_mirroring(struct nfs_pageio_descriptor *pgio)
1025 {
1026         pgio->pg_mirror_count = 1;
1027         pgio->pg_mirror_idx = 0;
1028         pgio->pg_mirrors = pgio->pg_mirrors_static;
1029         kfree(pgio->pg_mirrors_dynamic);
1030         pgio->pg_mirrors_dynamic = NULL;
1031 }
1032
1033 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
1034                 const struct nfs_lock_context *l2)
1035 {
1036         return l1->lockowner == l2->lockowner;
1037 }
1038
1039 /**
1040  * nfs_coalesce_size - test two requests for compatibility
1041  * @prev: pointer to nfs_page
1042  * @req: pointer to nfs_page
1043  * @pgio: pointer to nfs_pagio_descriptor
1044  *
1045  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
1046  * page data area they describe is contiguous, and that their RPC
1047  * credentials, NFSv4 open state, and lockowners are the same.
1048  *
1049  * Returns size of the request that can be coalesced
1050  */
1051 static unsigned int nfs_coalesce_size(struct nfs_page *prev,
1052                                       struct nfs_page *req,
1053                                       struct nfs_pageio_descriptor *pgio)
1054 {
1055         struct file_lock_context *flctx;
1056
1057         if (prev) {
1058                 if (!nfs_match_open_context(nfs_req_openctx(req), nfs_req_openctx(prev)))
1059                         return 0;
1060                 flctx = d_inode(nfs_req_openctx(req)->dentry)->i_flctx;
1061                 if (flctx != NULL &&
1062                     !(list_empty_careful(&flctx->flc_posix) &&
1063                       list_empty_careful(&flctx->flc_flock)) &&
1064                     !nfs_match_lock_context(req->wb_lock_context,
1065                                             prev->wb_lock_context))
1066                         return 0;
1067                 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
1068                         return 0;
1069                 if (req->wb_page == prev->wb_page) {
1070                         if (req->wb_pgbase != prev->wb_pgbase + prev->wb_bytes)
1071                                 return 0;
1072                 } else {
1073                         if (req->wb_pgbase != 0 ||
1074                             prev->wb_pgbase + prev->wb_bytes != PAGE_SIZE)
1075                                 return 0;
1076                 }
1077         }
1078         return pgio->pg_ops->pg_test(pgio, prev, req);
1079 }
1080
1081 /**
1082  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
1083  * @desc: destination io descriptor
1084  * @req: request
1085  *
1086  * If the request 'req' was successfully coalesced into the existing list
1087  * of pages 'desc', it returns the size of req.
1088  */
1089 static unsigned int
1090 nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
1091                 struct nfs_page *req)
1092 {
1093         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1094         struct nfs_page *prev = NULL;
1095         unsigned int size;
1096
1097         if (mirror->pg_count != 0) {
1098                 prev = nfs_list_entry(mirror->pg_list.prev);
1099         } else {
1100                 if (desc->pg_ops->pg_init)
1101                         desc->pg_ops->pg_init(desc, req);
1102                 if (desc->pg_error < 0)
1103                         return 0;
1104                 mirror->pg_base = req->wb_pgbase;
1105         }
1106
1107         if (desc->pg_maxretrans && req->wb_nio > desc->pg_maxretrans) {
1108                 if (NFS_SERVER(desc->pg_inode)->flags & NFS_MOUNT_SOFTERR)
1109                         desc->pg_error = -ETIMEDOUT;
1110                 else
1111                         desc->pg_error = -EIO;
1112                 return 0;
1113         }
1114
1115         size = nfs_coalesce_size(prev, req, desc);
1116         if (size < req->wb_bytes)
1117                 return size;
1118         nfs_list_move_request(req, &mirror->pg_list);
1119         mirror->pg_count += req->wb_bytes;
1120         return req->wb_bytes;
1121 }
1122
1123 /*
1124  * Helper for nfs_pageio_add_request and nfs_pageio_complete
1125  */
1126 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
1127 {
1128         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1129
1130
1131         if (!list_empty(&mirror->pg_list)) {
1132                 int error = desc->pg_ops->pg_doio(desc);
1133                 if (error < 0)
1134                         desc->pg_error = error;
1135                 else
1136                         mirror->pg_bytes_written += mirror->pg_count;
1137         }
1138         if (list_empty(&mirror->pg_list)) {
1139                 mirror->pg_count = 0;
1140                 mirror->pg_base = 0;
1141         }
1142 }
1143
1144 static void
1145 nfs_pageio_cleanup_request(struct nfs_pageio_descriptor *desc,
1146                 struct nfs_page *req)
1147 {
1148         LIST_HEAD(head);
1149
1150         nfs_list_move_request(req, &head);
1151         desc->pg_completion_ops->error_cleanup(&head, desc->pg_error);
1152 }
1153
1154 /**
1155  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
1156  * @desc: destination io descriptor
1157  * @req: request
1158  *
1159  * This may split a request into subrequests which are all part of the
1160  * same page group. If so, it will submit @req as the last one, to ensure
1161  * the pointer to @req is still valid in case of failure.
1162  *
1163  * Returns true if the request 'req' was successfully coalesced into the
1164  * existing list of pages 'desc'.
1165  */
1166 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1167                            struct nfs_page *req)
1168 {
1169         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1170         struct nfs_page *subreq;
1171         unsigned int size, subreq_size;
1172
1173         nfs_page_group_lock(req);
1174
1175         subreq = req;
1176         subreq_size = subreq->wb_bytes;
1177         for(;;) {
1178                 size = nfs_pageio_do_add_request(desc, subreq);
1179                 if (size == subreq_size) {
1180                         /* We successfully submitted a request */
1181                         if (subreq == req)
1182                                 break;
1183                         req->wb_pgbase += size;
1184                         req->wb_bytes -= size;
1185                         req->wb_offset += size;
1186                         subreq_size = req->wb_bytes;
1187                         subreq = req;
1188                         continue;
1189                 }
1190                 if (WARN_ON_ONCE(subreq != req)) {
1191                         nfs_page_group_unlock(req);
1192                         nfs_pageio_cleanup_request(desc, subreq);
1193                         subreq = req;
1194                         subreq_size = req->wb_bytes;
1195                         nfs_page_group_lock(req);
1196                 }
1197                 if (!size) {
1198                         /* Can't coalesce any more, so do I/O */
1199                         nfs_page_group_unlock(req);
1200                         desc->pg_moreio = 1;
1201                         nfs_pageio_doio(desc);
1202                         if (desc->pg_error < 0 || mirror->pg_recoalesce)
1203                                 return 0;
1204                         /* retry add_request for this subreq */
1205                         nfs_page_group_lock(req);
1206                         continue;
1207                 }
1208                 subreq = nfs_create_subreq(req, req->wb_pgbase,
1209                                 req->wb_offset, size);
1210                 if (IS_ERR(subreq))
1211                         goto err_ptr;
1212                 subreq_size = size;
1213         }
1214
1215         nfs_page_group_unlock(req);
1216         return 1;
1217 err_ptr:
1218         desc->pg_error = PTR_ERR(subreq);
1219         nfs_page_group_unlock(req);
1220         return 0;
1221 }
1222
1223 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
1224 {
1225         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1226         LIST_HEAD(head);
1227
1228         do {
1229                 list_splice_init(&mirror->pg_list, &head);
1230                 mirror->pg_bytes_written -= mirror->pg_count;
1231                 mirror->pg_count = 0;
1232                 mirror->pg_base = 0;
1233                 mirror->pg_recoalesce = 0;
1234
1235                 while (!list_empty(&head)) {
1236                         struct nfs_page *req;
1237
1238                         req = list_first_entry(&head, struct nfs_page, wb_list);
1239                         if (__nfs_pageio_add_request(desc, req))
1240                                 continue;
1241                         if (desc->pg_error < 0) {
1242                                 list_splice_tail(&head, &mirror->pg_list);
1243                                 mirror->pg_recoalesce = 1;
1244                                 return 0;
1245                         }
1246                         break;
1247                 }
1248         } while (mirror->pg_recoalesce);
1249         return 1;
1250 }
1251
1252 static int nfs_pageio_add_request_mirror(struct nfs_pageio_descriptor *desc,
1253                 struct nfs_page *req)
1254 {
1255         int ret;
1256
1257         do {
1258                 ret = __nfs_pageio_add_request(desc, req);
1259                 if (ret)
1260                         break;
1261                 if (desc->pg_error < 0)
1262                         break;
1263                 ret = nfs_do_recoalesce(desc);
1264         } while (ret);
1265
1266         return ret;
1267 }
1268
1269 static void nfs_pageio_error_cleanup(struct nfs_pageio_descriptor *desc)
1270 {
1271         u32 midx;
1272         struct nfs_pgio_mirror *mirror;
1273
1274         if (!desc->pg_error)
1275                 return;
1276
1277         for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1278                 mirror = nfs_pgio_get_mirror(desc, midx);
1279                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list,
1280                                 desc->pg_error);
1281         }
1282 }
1283
1284 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1285                            struct nfs_page *req)
1286 {
1287         u32 midx;
1288         unsigned int pgbase, offset, bytes;
1289         struct nfs_page *dupreq;
1290
1291         pgbase = req->wb_pgbase;
1292         offset = req->wb_offset;
1293         bytes = req->wb_bytes;
1294
1295         nfs_pageio_setup_mirroring(desc, req);
1296         if (desc->pg_error < 0)
1297                 goto out_failed;
1298
1299         /* Create the mirror instances first, and fire them off */
1300         for (midx = 1; midx < desc->pg_mirror_count; midx++) {
1301                 nfs_page_group_lock(req);
1302
1303                 dupreq = nfs_create_subreq(req,
1304                                 pgbase, offset, bytes);
1305
1306                 nfs_page_group_unlock(req);
1307                 if (IS_ERR(dupreq)) {
1308                         desc->pg_error = PTR_ERR(dupreq);
1309                         goto out_failed;
1310                 }
1311
1312                 nfs_pgio_set_current_mirror(desc, midx);
1313                 if (!nfs_pageio_add_request_mirror(desc, dupreq))
1314                         goto out_cleanup_subreq;
1315         }
1316
1317         nfs_pgio_set_current_mirror(desc, 0);
1318         if (!nfs_pageio_add_request_mirror(desc, req))
1319                 goto out_failed;
1320
1321         return 1;
1322
1323 out_cleanup_subreq:
1324         nfs_pageio_cleanup_request(desc, dupreq);
1325 out_failed:
1326         nfs_pageio_error_cleanup(desc);
1327         return 0;
1328 }
1329
1330 /*
1331  * nfs_pageio_complete_mirror - Complete I/O on the current mirror of an
1332  *                              nfs_pageio_descriptor
1333  * @desc: pointer to io descriptor
1334  * @mirror_idx: pointer to mirror index
1335  */
1336 static void nfs_pageio_complete_mirror(struct nfs_pageio_descriptor *desc,
1337                                        u32 mirror_idx)
1338 {
1339         struct nfs_pgio_mirror *mirror;
1340         u32 restore_idx;
1341
1342         restore_idx = nfs_pgio_set_current_mirror(desc, mirror_idx);
1343         mirror = nfs_pgio_current_mirror(desc);
1344
1345         for (;;) {
1346                 nfs_pageio_doio(desc);
1347                 if (desc->pg_error < 0 || !mirror->pg_recoalesce)
1348                         break;
1349                 if (!nfs_do_recoalesce(desc))
1350                         break;
1351         }
1352         nfs_pgio_set_current_mirror(desc, restore_idx);
1353 }
1354
1355 /*
1356  * nfs_pageio_resend - Transfer requests to new descriptor and resend
1357  * @hdr - the pgio header to move request from
1358  * @desc - the pageio descriptor to add requests to
1359  *
1360  * Try to move each request (nfs_page) from @hdr to @desc then attempt
1361  * to send them.
1362  *
1363  * Returns 0 on success and < 0 on error.
1364  */
1365 int nfs_pageio_resend(struct nfs_pageio_descriptor *desc,
1366                       struct nfs_pgio_header *hdr)
1367 {
1368         LIST_HEAD(pages);
1369
1370         desc->pg_io_completion = hdr->io_completion;
1371         desc->pg_dreq = hdr->dreq;
1372         list_splice_init(&hdr->pages, &pages);
1373         while (!list_empty(&pages)) {
1374                 struct nfs_page *req = nfs_list_entry(pages.next);
1375
1376                 if (!nfs_pageio_add_request(desc, req))
1377                         break;
1378         }
1379         nfs_pageio_complete(desc);
1380         if (!list_empty(&pages)) {
1381                 int err = desc->pg_error < 0 ? desc->pg_error : -EIO;
1382                 hdr->completion_ops->error_cleanup(&pages, err);
1383                 nfs_set_pgio_error(hdr, err, hdr->io_start);
1384                 return err;
1385         }
1386         return 0;
1387 }
1388 EXPORT_SYMBOL_GPL(nfs_pageio_resend);
1389
1390 /**
1391  * nfs_pageio_complete - Complete I/O then cleanup an nfs_pageio_descriptor
1392  * @desc: pointer to io descriptor
1393  */
1394 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1395 {
1396         u32 midx;
1397
1398         for (midx = 0; midx < desc->pg_mirror_count; midx++)
1399                 nfs_pageio_complete_mirror(desc, midx);
1400
1401         if (desc->pg_error < 0)
1402                 nfs_pageio_error_cleanup(desc);
1403         if (desc->pg_ops->pg_cleanup)
1404                 desc->pg_ops->pg_cleanup(desc);
1405         nfs_pageio_cleanup_mirroring(desc);
1406 }
1407
1408 /**
1409  * nfs_pageio_cond_complete - Conditional I/O completion
1410  * @desc: pointer to io descriptor
1411  * @index: page index
1412  *
1413  * It is important to ensure that processes don't try to take locks
1414  * on non-contiguous ranges of pages as that might deadlock. This
1415  * function should be called before attempting to wait on a locked
1416  * nfs_page. It will complete the I/O if the page index 'index'
1417  * is not contiguous with the existing list of pages in 'desc'.
1418  */
1419 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1420 {
1421         struct nfs_pgio_mirror *mirror;
1422         struct nfs_page *prev;
1423         u32 midx;
1424
1425         for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1426                 mirror = nfs_pgio_get_mirror(desc, midx);
1427                 if (!list_empty(&mirror->pg_list)) {
1428                         prev = nfs_list_entry(mirror->pg_list.prev);
1429                         if (index != prev->wb_index + 1) {
1430                                 nfs_pageio_complete(desc);
1431                                 break;
1432                         }
1433                 }
1434         }
1435 }
1436
1437 /*
1438  * nfs_pageio_stop_mirroring - stop using mirroring (set mirror count to 1)
1439  */
1440 void nfs_pageio_stop_mirroring(struct nfs_pageio_descriptor *pgio)
1441 {
1442         nfs_pageio_complete(pgio);
1443 }
1444
1445 int __init nfs_init_nfspagecache(void)
1446 {
1447         nfs_page_cachep = kmem_cache_create("nfs_page",
1448                                             sizeof(struct nfs_page),
1449                                             0, SLAB_HWCACHE_ALIGN,
1450                                             NULL);
1451         if (nfs_page_cachep == NULL)
1452                 return -ENOMEM;
1453
1454         return 0;
1455 }
1456
1457 void nfs_destroy_nfspagecache(void)
1458 {
1459         kmem_cache_destroy(nfs_page_cachep);
1460 }
1461
1462 static const struct rpc_call_ops nfs_pgio_common_ops = {
1463         .rpc_call_prepare = nfs_pgio_prepare,
1464         .rpc_call_done = nfs_pgio_result,
1465         .rpc_release = nfs_pgio_release,
1466 };
1467
1468 const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1469         .pg_test = nfs_generic_pg_test,
1470         .pg_doio = nfs_generic_pg_pgios,
1471 };