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