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