fs/nfs/pagelist.c

   1 /*
   2  * linux/fs/nfs/pagelist.c
   3  *
   4  * A set of helper functions for managing NFS read and write requests.
   5  * The main purpose of these routines is to provide support for the
   6  * coalescing of several requests into a single RPC call.
   7  *
   8  * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
   9  *
  10  */
  11
  12 #include <linux/slab.h>
  13 #include <linux/file.h>
  14 #include <linux/sched.h>
  15 #include <linux/sunrpc/clnt.h>
  16 #include <linux/nfs.h>
  17 #include <linux/nfs3.h>
  18 #include <linux/nfs4.h>
  19 #include <linux/nfs_page.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_mount.h>
  22 #include <linux/export.h>
  23
  24 #include "internal.h"
  25 #include "pnfs.h"
  26
  27 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  28
  29 static struct kmem_cache *nfs_page_cachep;
  30 static const struct rpc_call_ops nfs_pgio_common_ops;
  31
  32 static bool nfs_pgarray_set(struct nfs_page_array *p, unsigned int pagecount)
  33 {
  34         p->npages = pagecount;
  35         if (pagecount <= ARRAY_SIZE(p->page_array))
  36                 p->pagevec = p->page_array;
  37         else {
  38                 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL);
  39                 if (!p->pagevec)
  40                         p->npages = 0;
  41         }
  42         return p->pagevec != NULL;
  43 }
  44
  45 void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
  46                        struct nfs_pgio_header *hdr,
  47                        void (*release)(struct nfs_pgio_header *hdr))
  48 {
  49         hdr->req = nfs_list_entry(desc->pg_list.next);
  50         hdr->inode = desc->pg_inode;
  51         hdr->cred = hdr->req->wb_context->cred;
  52         hdr->io_start = req_offset(hdr->req);
  53         hdr->good_bytes = desc->pg_count;
  54         hdr->dreq = desc->pg_dreq;
  55         hdr->layout_private = desc->pg_layout_private;
  56         hdr->release = release;
  57         hdr->completion_ops = desc->pg_completion_ops;
  58         if (hdr->completion_ops->init_hdr)
  59                 hdr->completion_ops->init_hdr(hdr);
  60 }
  61 EXPORT_SYMBOL_GPL(nfs_pgheader_init);
  62
  63 void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
  64 {
  65         spin_lock(&hdr->lock);
  66         if (pos < hdr->io_start + hdr->good_bytes) {
  67                 set_bit(NFS_IOHDR_ERROR, &hdr->flags);
  68                 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
  69                 hdr->good_bytes = pos - hdr->io_start;
  70                 hdr->error = error;
  71         }
  72         spin_unlock(&hdr->lock);
  73 }
  74
  75 static inline struct nfs_page *
  76 nfs_page_alloc(void)
  77 {
  78         struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_NOIO);
  79         if (p)
  80                 INIT_LIST_HEAD(&p->wb_list);
  81         return p;
  82 }
  83
  84 static inline void
  85 nfs_page_free(struct nfs_page *p)
  86 {
  87         kmem_cache_free(nfs_page_cachep, p);
  88 }
  89
  90 static void
  91 nfs_iocounter_inc(struct nfs_io_counter *c)
  92 {
  93         atomic_inc(&c->io_count);
  94 }
  95
  96 static void
  97 nfs_iocounter_dec(struct nfs_io_counter *c)
  98 {
  99         if (atomic_dec_and_test(&c->io_count)) {
 100                 clear_bit(NFS_IO_INPROGRESS, &c->flags);
 101                 smp_mb__after_atomic();
 102                 wake_up_bit(&c->flags, NFS_IO_INPROGRESS);
 103         }
 104 }
 105
 106 static int
 107 __nfs_iocounter_wait(struct nfs_io_counter *c)
 108 {
 109         wait_queue_head_t *wq = bit_waitqueue(&c->flags, NFS_IO_INPROGRESS);
 110         DEFINE_WAIT_BIT(q, &c->flags, NFS_IO_INPROGRESS);
 111         int ret = 0;
 112
 113         do {
 114                 prepare_to_wait(wq, &q.wait, TASK_KILLABLE);
 115                 set_bit(NFS_IO_INPROGRESS, &c->flags);
 116                 if (atomic_read(&c->io_count) == 0)
 117                         break;
 118                 ret = nfs_wait_bit_killable(&q.key);
 119         } while (atomic_read(&c->io_count) != 0);
 120         finish_wait(wq, &q.wait);
 121         return ret;
 122 }
 123
 124 /**
 125  * nfs_iocounter_wait - wait for i/o to complete
 126  * @c: nfs_io_counter to use
 127  *
 128  * returns -ERESTARTSYS if interrupted by a fatal signal.
 129  * Otherwise returns 0 once the io_count hits 0.
 130  */
 131 int
 132 nfs_iocounter_wait(struct nfs_io_counter *c)
 133 {
 134         if (atomic_read(&c->io_count) == 0)
 135                 return 0;
 136         return __nfs_iocounter_wait(c);
 137 }
 138
 139 /*
 140  * nfs_page_group_lock - lock the head of the page group
 141  * @req - request in group that is to be locked
 142  *
 143  * this lock must be held if modifying the page group list
 144  */
 145 void
 146 nfs_page_group_lock(struct nfs_page *req)
 147 {
 148         struct nfs_page *head = req->wb_head;
 149
 150         WARN_ON_ONCE(head != head->wb_head);
 151
 152         wait_on_bit_lock(&head->wb_flags, PG_HEADLOCK,
 153                         TASK_UNINTERRUPTIBLE);
 154 }
 155
 156 /*
 157  * nfs_page_group_unlock - unlock the head of the page group
 158  * @req - request in group that is to be unlocked
 159  */
 160 void
 161 nfs_page_group_unlock(struct nfs_page *req)
 162 {
 163         struct nfs_page *head = req->wb_head;
 164
 165         WARN_ON_ONCE(head != head->wb_head);
 166
 167         smp_mb__before_atomic();
 168         clear_bit(PG_HEADLOCK, &head->wb_flags);
 169         smp_mb__after_atomic();
 170         wake_up_bit(&head->wb_flags, PG_HEADLOCK);
 171 }
 172
 173 /*
 174  * nfs_page_group_sync_on_bit_locked
 175  *
 176  * must be called with page group lock held
 177  */
 178 static bool
 179 nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
 180 {
 181         struct nfs_page *head = req->wb_head;
 182         struct nfs_page *tmp;
 183
 184         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags));
 185         WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags));
 186
 187         tmp = req->wb_this_page;
 188         while (tmp != req) {
 189                 if (!test_bit(bit, &tmp->wb_flags))
 190                         return false;
 191                 tmp = tmp->wb_this_page;
 192         }
 193
 194         /* true! reset all bits */
 195         tmp = req;
 196         do {
 197                 clear_bit(bit, &tmp->wb_flags);
 198                 tmp = tmp->wb_this_page;
 199         } while (tmp != req);
 200
 201         return true;
 202 }
 203
 204 /*
 205  * nfs_page_group_sync_on_bit - set bit on current request, but only
 206  *   return true if the bit is set for all requests in page group
 207  * @req - request in page group
 208  * @bit - PG_* bit that is used to sync page group
 209  */
 210 bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit)
 211 {
 212         bool ret;
 213
 214         nfs_page_group_lock(req);
 215         ret = nfs_page_group_sync_on_bit_locked(req, bit);
 216         nfs_page_group_unlock(req);
 217
 218         return ret;
 219 }
 220
 221 /*
 222  * nfs_page_group_init - Initialize the page group linkage for @req
 223  * @req - a new nfs request
 224  * @prev - the previous request in page group, or NULL if @req is the first
 225  *         or only request in the group (the head).
 226  */
 227 static inline void
 228 nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev)
 229 {
 230         WARN_ON_ONCE(prev == req);
 231
 232         if (!prev) {
 233                 /* a head request */
 234                 req->wb_head = req;
 235                 req->wb_this_page = req;
 236         } else {
 237                 /* a subrequest */
 238                 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head);
 239                 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags));
 240                 req->wb_head = prev->wb_head;
 241                 req->wb_this_page = prev->wb_this_page;
 242                 prev->wb_this_page = req;
 243
 244                 /* All subrequests take a ref on the head request until
 245                  * nfs_page_group_destroy is called */
 246                 kref_get(&req->wb_head->wb_kref);
 247
 248                 /* grab extra ref if head request has extra ref from
 249                  * the write/commit path to handle handoff between write
 250                  * and commit lists */
 251                 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) {
 252                         set_bit(PG_INODE_REF, &req->wb_flags);
 253                         kref_get(&req->wb_kref);
 254                 }
 255         }
 256 }
 257
 258 /*
 259  * nfs_page_group_destroy - sync the destruction of page groups
 260  * @req - request that no longer needs the page group
 261  *
 262  * releases the page group reference from each member once all
 263  * members have called this function.
 264  */
 265 static void
 266 nfs_page_group_destroy(struct kref *kref)
 267 {
 268         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 269         struct nfs_page *tmp, *next;
 270
 271         /* subrequests must release the ref on the head request */
 272         if (req->wb_head != req)
 273                 nfs_release_request(req->wb_head);
 274
 275         if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN))
 276                 return;
 277
 278         tmp = req;
 279         do {
 280                 next = tmp->wb_this_page;
 281                 /* unlink and free */
 282                 tmp->wb_this_page = tmp;
 283                 tmp->wb_head = tmp;
 284                 nfs_free_request(tmp);
 285                 tmp = next;
 286         } while (tmp != req);
 287 }
 288
 289 /**
 290  * nfs_create_request - Create an NFS read/write request.
 291  * @ctx: open context to use
 292  * @page: page to write
 293  * @last: last nfs request created for this page group or NULL if head
 294  * @offset: starting offset within the page for the write
 295  * @count: number of bytes to read/write
 296  *
 297  * The page must be locked by the caller. This makes sure we never
 298  * create two different requests for the same page.
 299  * User should ensure it is safe to sleep in this function.
 300  */
 301 struct nfs_page *
 302 nfs_create_request(struct nfs_open_context *ctx, struct page *page,
 303                    struct nfs_page *last, unsigned int offset,
 304                    unsigned int count)
 305 {
 306         struct nfs_page         *req;
 307         struct nfs_lock_context *l_ctx;
 308
 309         if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
 310                 return ERR_PTR(-EBADF);
 311         /* try to allocate the request struct */
 312         req = nfs_page_alloc();
 313         if (req == NULL)
 314                 return ERR_PTR(-ENOMEM);
 315
 316         /* get lock context early so we can deal with alloc failures */
 317         l_ctx = nfs_get_lock_context(ctx);
 318         if (IS_ERR(l_ctx)) {
 319                 nfs_page_free(req);
 320                 return ERR_CAST(l_ctx);
 321         }
 322         req->wb_lock_context = l_ctx;
 323         nfs_iocounter_inc(&l_ctx->io_count);
 324
 325         /* Initialize the request struct. Initially, we assume a
 326          * long write-back delay. This will be adjusted in
 327          * update_nfs_request below if the region is not locked. */
 328         req->wb_page    = page;
 329         req->wb_index   = page_file_index(page);
 330         page_cache_get(page);
 331         req->wb_offset  = offset;
 332         req->wb_pgbase  = offset;
 333         req->wb_bytes   = count;
 334         req->wb_context = get_nfs_open_context(ctx);
 335         kref_init(&req->wb_kref);
 336         nfs_page_group_init(req, last);
 337         return req;
 338 }
 339
 340 /**
 341  * nfs_unlock_request - Unlock request and wake up sleepers.
 342  * @req:
 343  */
 344 void nfs_unlock_request(struct nfs_page *req)
 345 {
 346         if (!NFS_WBACK_BUSY(req)) {
 347                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 348                 BUG();
 349         }
 350         smp_mb__before_atomic();
 351         clear_bit(PG_BUSY, &req->wb_flags);
 352         smp_mb__after_atomic();
 353         wake_up_bit(&req->wb_flags, PG_BUSY);
 354 }
 355
 356 /**
 357  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
 358  * @req:
 359  */
 360 void nfs_unlock_and_release_request(struct nfs_page *req)
 361 {
 362         nfs_unlock_request(req);
 363         nfs_release_request(req);
 364 }
 365
 366 /*
 367  * nfs_clear_request - Free up all resources allocated to the request
 368  * @req:
 369  *
 370  * Release page and open context resources associated with a read/write
 371  * request after it has completed.
 372  */
 373 static void nfs_clear_request(struct nfs_page *req)
 374 {
 375         struct page *page = req->wb_page;
 376         struct nfs_open_context *ctx = req->wb_context;
 377         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 378
 379         if (page != NULL) {
 380                 page_cache_release(page);
 381                 req->wb_page = NULL;
 382         }
 383         if (l_ctx != NULL) {
 384                 nfs_iocounter_dec(&l_ctx->io_count);
 385                 nfs_put_lock_context(l_ctx);
 386                 req->wb_lock_context = NULL;
 387         }
 388         if (ctx != NULL) {
 389                 put_nfs_open_context(ctx);
 390                 req->wb_context = NULL;
 391         }
 392 }
 393
 394 /**
 395  * nfs_release_request - Release the count on an NFS read/write request
 396  * @req: request to release
 397  *
 398  * Note: Should never be called with the spinlock held!
 399  */
 400 void nfs_free_request(struct nfs_page *req)
 401 {
 402         WARN_ON_ONCE(req->wb_this_page != req);
 403
 404         /* extra debug: make sure no sync bits are still set */
 405         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
 406         WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
 407         WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
 408         WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
 409         WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
 410
 411         /* Release struct file and open context */
 412         nfs_clear_request(req);
 413         nfs_page_free(req);
 414 }
 415
 416 void nfs_release_request(struct nfs_page *req)
 417 {
 418         kref_put(&req->wb_kref, nfs_page_group_destroy);
 419 }
 420
 421 /**
 422  * nfs_wait_on_request - Wait for a request to complete.
 423  * @req: request to wait upon.
 424  *
 425  * Interruptible by fatal signals only.
 426  * The user is responsible for holding a count on the request.
 427  */
 428 int
 429 nfs_wait_on_request(struct nfs_page *req)
 430 {
 431         return wait_on_bit_io(&req->wb_flags, PG_BUSY,
 432                               TASK_UNINTERRUPTIBLE);
 433 }
 434
 435 /*
 436  * nfs_generic_pg_test - determine if requests can be coalesced
 437  * @desc: pointer to descriptor
 438  * @prev: previous request in desc, or NULL
 439  * @req: this request
 440  *
 441  * Returns zero if @req can be coalesced into @desc, otherwise it returns
 442  * the size of the request.
 443  */
 444 size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc,
 445                            struct nfs_page *prev, struct nfs_page *req)
 446 {
 447         if (desc->pg_count > desc->pg_bsize) {
 448                 /* should never happen */
 449                 WARN_ON_ONCE(1);
 450                 return 0;
 451         }
 452
 453         return min(desc->pg_bsize - desc->pg_count, (size_t)req->wb_bytes);
 454 }
 455 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
 456
 457 static inline struct nfs_rw_header *NFS_RW_HEADER(struct nfs_pgio_header *hdr)
 458 {
 459         return container_of(hdr, struct nfs_rw_header, header);
 460 }
 461
 462 /**
 463  * nfs_rw_header_alloc - Allocate a header for a read or write
 464  * @ops: Read or write function vector
 465  */
 466 struct nfs_rw_header *nfs_rw_header_alloc(const struct nfs_rw_ops *ops)
 467 {
 468         struct nfs_rw_header *header = ops->rw_alloc_header();
 469
 470         if (header) {
 471                 struct nfs_pgio_header *hdr = &header->header;
 472
 473                 INIT_LIST_HEAD(&hdr->pages);
 474                 spin_lock_init(&hdr->lock);
 475                 atomic_set(&hdr->refcnt, 0);
 476                 hdr->rw_ops = ops;
 477         }
 478         return header;
 479 }
 480 EXPORT_SYMBOL_GPL(nfs_rw_header_alloc);
 481
 482 /*
 483  * nfs_rw_header_free - Free a read or write header
 484  * @hdr: The header to free
 485  */
 486 void nfs_rw_header_free(struct nfs_pgio_header *hdr)
 487 {
 488         hdr->rw_ops->rw_free_header(NFS_RW_HEADER(hdr));
 489 }
 490 EXPORT_SYMBOL_GPL(nfs_rw_header_free);
 491
 492 /**
 493  * nfs_pgio_data_alloc - Allocate pageio data
 494  * @hdr: The header making a request
 495  * @pagecount: Number of pages to create
 496  */
 497 static struct nfs_pgio_data *nfs_pgio_data_alloc(struct nfs_pgio_header *hdr,
 498                                                  unsigned int pagecount)
 499 {
 500         struct nfs_pgio_data *data, *prealloc;
 501
 502         prealloc = &NFS_RW_HEADER(hdr)->rpc_data;
 503         if (prealloc->header == NULL)
 504                 data = prealloc;
 505         else
 506                 data = kzalloc(sizeof(*data), GFP_KERNEL);
 507         if (!data)
 508                 goto out;
 509
 510         if (nfs_pgarray_set(&data->pages, pagecount)) {
 511                 data->header = hdr;
 512                 atomic_inc(&hdr->refcnt);
 513         } else {
 514                 if (data != prealloc)
 515                         kfree(data);
 516                 data = NULL;
 517         }
 518 out:
 519         return data;
 520 }
 521
 522 /**
 523  * nfs_pgio_data_release - Properly free pageio data
 524  * @data: The data to release
 525  */
 526 void nfs_pgio_data_release(struct nfs_pgio_data *data)
 527 {
 528         struct nfs_pgio_header *hdr = data->header;
 529         struct nfs_rw_header *pageio_header = NFS_RW_HEADER(hdr);
 530
 531         put_nfs_open_context(data->args.context);
 532         if (data->pages.pagevec != data->pages.page_array)
 533                 kfree(data->pages.pagevec);
 534         if (data == &pageio_header->rpc_data) {
 535                 data->header = NULL;
 536                 data = NULL;
 537         }
 538         if (atomic_dec_and_test(&hdr->refcnt))
 539                 hdr->completion_ops->completion(hdr);
 540         /* Note: we only free the rpc_task after callbacks are done.
 541          * See the comment in rpc_free_task() for why
 542          */
 543         kfree(data);
 544 }
 545 EXPORT_SYMBOL_GPL(nfs_pgio_data_release);
 546
 547 /**
 548  * nfs_pgio_rpcsetup - Set up arguments for a pageio call
 549  * @data: The pageio data
 550  * @count: Number of bytes to read
 551  * @offset: Initial offset
 552  * @how: How to commit data (writes only)
 553  * @cinfo: Commit information for the call (writes only)
 554  */
 555 static void nfs_pgio_rpcsetup(struct nfs_pgio_data *data,
 556                               unsigned int count, unsigned int offset,
 557                               int how, struct nfs_commit_info *cinfo)
 558 {
 559         struct nfs_page *req = data->header->req;
 560
 561         /* Set up the RPC argument and reply structs
 562          * NB: take care not to mess about with data->commit et al. */
 563
 564         data->args.fh     = NFS_FH(data->header->inode);
 565         data->args.offset = req_offset(req) + offset;
 566         /* pnfs_set_layoutcommit needs this */
 567         data->mds_offset = data->args.offset;
 568         data->args.pgbase = req->wb_pgbase + offset;
 569         data->args.pages  = data->pages.pagevec;
 570         data->args.count  = count;
 571         data->args.context = get_nfs_open_context(req->wb_context);
 572         data->args.lock_context = req->wb_lock_context;
 573         data->args.stable  = NFS_UNSTABLE;
 574         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 575         case 0:
 576                 break;
 577         case FLUSH_COND_STABLE:
 578                 if (nfs_reqs_to_commit(cinfo))
 579                         break;
 580         default:
 581                 data->args.stable = NFS_FILE_SYNC;
 582         }
 583
 584         data->res.fattr   = &data->fattr;
 585         data->res.count   = count;
 586         data->res.eof     = 0;
 587         data->res.verf    = &data->verf;
 588         nfs_fattr_init(&data->fattr);
 589 }
 590
 591 /**
 592  * nfs_pgio_prepare - Prepare pageio data to go over the wire
 593  * @task: The current task
 594  * @calldata: pageio data to prepare
 595  */
 596 static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
 597 {
 598         struct nfs_pgio_data *data = calldata;
 599         int err;
 600         err = NFS_PROTO(data->header->inode)->pgio_rpc_prepare(task, data);
 601         if (err)
 602                 rpc_exit(task, err);
 603 }
 604
 605 int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_data *data,
 606                       const struct rpc_call_ops *call_ops, int how, int flags)
 607 {
 608         struct rpc_task *task;
 609         struct rpc_message msg = {
 610                 .rpc_argp = &data->args,
 611                 .rpc_resp = &data->res,
 612                 .rpc_cred = data->header->cred,
 613         };
 614         struct rpc_task_setup task_setup_data = {
 615                 .rpc_client = clnt,
 616                 .task = &data->task,
 617                 .rpc_message = &msg,
 618                 .callback_ops = call_ops,
 619                 .callback_data = data,
 620                 .workqueue = nfsiod_workqueue,
 621                 .flags = RPC_TASK_ASYNC | flags,
 622         };
 623         int ret = 0;
 624
 625         data->header->rw_ops->rw_initiate(data, &msg, &task_setup_data, how);
 626
 627         dprintk("NFS: %5u initiated pgio call "
 628                 "(req %s/%llu, %u bytes @ offset %llu)\n",
 629                 data->task.tk_pid,
 630                 data->header->inode->i_sb->s_id,
 631                 (unsigned long long)NFS_FILEID(data->header->inode),
 632                 data->args.count,
 633                 (unsigned long long)data->args.offset);
 634
 635         task = rpc_run_task(&task_setup_data);
 636         if (IS_ERR(task)) {
 637                 ret = PTR_ERR(task);
 638                 goto out;
 639         }
 640         if (how & FLUSH_SYNC) {
 641                 ret = rpc_wait_for_completion_task(task);
 642                 if (ret == 0)
 643                         ret = task->tk_status;
 644         }
 645         rpc_put_task(task);
 646 out:
 647         return ret;
 648 }
 649 EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
 650
 651 /**
 652  * nfs_pgio_error - Clean up from a pageio error
 653  * @desc: IO descriptor
 654  * @hdr: pageio header
 655  */
 656 static int nfs_pgio_error(struct nfs_pageio_descriptor *desc,
 657                           struct nfs_pgio_header *hdr)
 658 {
 659         set_bit(NFS_IOHDR_REDO, &hdr->flags);
 660         nfs_pgio_data_release(hdr->data);
 661         hdr->data = NULL;
 662         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 663         return -ENOMEM;
 664 }
 665
 666 /**
 667  * nfs_pgio_release - Release pageio data
 668  * @calldata: The pageio data to release
 669  */
 670 static void nfs_pgio_release(void *calldata)
 671 {
 672         struct nfs_pgio_data *data = calldata;
 673         if (data->header->rw_ops->rw_release)
 674                 data->header->rw_ops->rw_release(data);
 675         nfs_pgio_data_release(data);
 676 }
 677
 678 /**
 679  * nfs_pageio_init - initialise a page io descriptor
 680  * @desc: pointer to descriptor
 681  * @inode: pointer to inode
 682  * @doio: pointer to io function
 683  * @bsize: io block size
 684  * @io_flags: extra parameters for the io function
 685  */
 686 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 687                      struct inode *inode,
 688                      const struct nfs_pageio_ops *pg_ops,
 689                      const struct nfs_pgio_completion_ops *compl_ops,
 690                      const struct nfs_rw_ops *rw_ops,
 691                      size_t bsize,
 692                      int io_flags)
 693 {
 694         INIT_LIST_HEAD(&desc->pg_list);
 695         desc->pg_bytes_written = 0;
 696         desc->pg_count = 0;
 697         desc->pg_bsize = bsize;
 698         desc->pg_base = 0;
 699         desc->pg_moreio = 0;
 700         desc->pg_recoalesce = 0;
 701         desc->pg_inode = inode;
 702         desc->pg_ops = pg_ops;
 703         desc->pg_completion_ops = compl_ops;
 704         desc->pg_rw_ops = rw_ops;
 705         desc->pg_ioflags = io_flags;
 706         desc->pg_error = 0;
 707         desc->pg_lseg = NULL;
 708         desc->pg_dreq = NULL;
 709         desc->pg_layout_private = NULL;
 710 }
 711 EXPORT_SYMBOL_GPL(nfs_pageio_init);
 712
 713 /**
 714  * nfs_pgio_result - Basic pageio error handling
 715  * @task: The task that ran
 716  * @calldata: Pageio data to check
 717  */
 718 static void nfs_pgio_result(struct rpc_task *task, void *calldata)
 719 {
 720         struct nfs_pgio_data *data = calldata;
 721         struct inode *inode = data->header->inode;
 722
 723         dprintk("NFS: %s: %5u, (status %d)\n", __func__,
 724                 task->tk_pid, task->tk_status);
 725
 726         if (data->header->rw_ops->rw_done(task, data, inode) != 0)
 727                 return;
 728         if (task->tk_status < 0)
 729                 nfs_set_pgio_error(data->header, task->tk_status, data->args.offset);
 730         else
 731                 data->header->rw_ops->rw_result(task, data);
 732 }
 733
 734 /*
 735  * Create an RPC task for the given read or write request and kick it.
 736  * The page must have been locked by the caller.
 737  *
 738  * It may happen that the page we're passed is not marked dirty.
 739  * This is the case if nfs_updatepage detects a conflicting request
 740  * that has been written but not committed.
 741  */
 742 int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
 743                      struct nfs_pgio_header *hdr)
 744 {
 745         struct nfs_page         *req;
 746         struct page             **pages;
 747         struct nfs_pgio_data    *data;
 748         struct list_head *head = &desc->pg_list;
 749         struct nfs_commit_info cinfo;
 750
 751         data = nfs_pgio_data_alloc(hdr, nfs_page_array_len(desc->pg_base,
 752                                                            desc->pg_count));
 753         if (!data)
 754                 return nfs_pgio_error(desc, hdr);
 755
 756         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
 757         pages = data->pages.pagevec;
 758         while (!list_empty(head)) {
 759                 req = nfs_list_entry(head->next);
 760                 nfs_list_remove_request(req);
 761                 nfs_list_add_request(req, &hdr->pages);
 762                 *pages++ = req->wb_page;
 763         }
 764
 765         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 766             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
 767                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 768
 769         /* Set up the argument struct */
 770         nfs_pgio_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
 771         hdr->data = data;
 772         desc->pg_rpc_callops = &nfs_pgio_common_ops;
 773         return 0;
 774 }
 775 EXPORT_SYMBOL_GPL(nfs_generic_pgio);
 776
 777 static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
 778 {
 779         struct nfs_rw_header *rw_hdr;
 780         struct nfs_pgio_header *hdr;
 781         int ret;
 782
 783         rw_hdr = nfs_rw_header_alloc(desc->pg_rw_ops);
 784         if (!rw_hdr) {
 785                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 786                 return -ENOMEM;
 787         }
 788         hdr = &rw_hdr->header;
 789         nfs_pgheader_init(desc, hdr, nfs_rw_header_free);
 790         atomic_inc(&hdr->refcnt);
 791         ret = nfs_generic_pgio(desc, hdr);
 792         if (ret == 0)
 793                 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
 794                                         hdr->data, desc->pg_rpc_callops,
 795                                         desc->pg_ioflags, 0);
 796         if (atomic_dec_and_test(&hdr->refcnt))
 797                 hdr->completion_ops->completion(hdr);
 798         return ret;
 799 }
 800
 801 static bool nfs_match_open_context(const struct nfs_open_context *ctx1,
 802                 const struct nfs_open_context *ctx2)
 803 {
 804         return ctx1->cred == ctx2->cred && ctx1->state == ctx2->state;
 805 }
 806
 807 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
 808                 const struct nfs_lock_context *l2)
 809 {
 810         return l1->lockowner.l_owner == l2->lockowner.l_owner
 811                 && l1->lockowner.l_pid == l2->lockowner.l_pid;
 812 }
 813
 814 /**
 815  * nfs_can_coalesce_requests - test two requests for compatibility
 816  * @prev: pointer to nfs_page
 817  * @req: pointer to nfs_page
 818  *
 819  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
 820  * page data area they describe is contiguous, and that their RPC
 821  * credentials, NFSv4 open state, and lockowners are the same.
 822  *
 823  * Return 'true' if this is the case, else return 'false'.
 824  */
 825 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
 826                                       struct nfs_page *req,
 827                                       struct nfs_pageio_descriptor *pgio)
 828 {
 829         size_t size;
 830
 831         if (prev) {
 832                 if (!nfs_match_open_context(req->wb_context, prev->wb_context))
 833                         return false;
 834                 if (req->wb_context->dentry->d_inode->i_flock != NULL &&
 835                     !nfs_match_lock_context(req->wb_lock_context,
 836                                             prev->wb_lock_context))
 837                         return false;
 838                 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
 839                         return false;
 840         }
 841         size = pgio->pg_ops->pg_test(pgio, prev, req);
 842         WARN_ON_ONCE(size > req->wb_bytes);
 843         if (size && size < req->wb_bytes)
 844                 req->wb_bytes = size;
 845         return size > 0;
 846 }
 847
 848 /**
 849  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
 850  * @desc: destination io descriptor
 851  * @req: request
 852  *
 853  * Returns true if the request 'req' was successfully coalesced into the
 854  * existing list of pages 'desc'.
 855  */
 856 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
 857                                      struct nfs_page *req)
 858 {
 859         struct nfs_page *prev = NULL;
 860         if (desc->pg_count != 0) {
 861                 prev = nfs_list_entry(desc->pg_list.prev);
 862         } else {
 863                 if (desc->pg_ops->pg_init)
 864                         desc->pg_ops->pg_init(desc, req);
 865                 desc->pg_base = req->wb_pgbase;
 866         }
 867         if (!nfs_can_coalesce_requests(prev, req, desc))
 868                 return 0;
 869         nfs_list_remove_request(req);
 870         nfs_list_add_request(req, &desc->pg_list);
 871         desc->pg_count += req->wb_bytes;
 872         return 1;
 873 }
 874
 875 /*
 876  * Helper for nfs_pageio_add_request and nfs_pageio_complete
 877  */
 878 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
 879 {
 880         if (!list_empty(&desc->pg_list)) {
 881                 int error = desc->pg_ops->pg_doio(desc);
 882                 if (error < 0)
 883                         desc->pg_error = error;
 884                 else
 885                         desc->pg_bytes_written += desc->pg_count;
 886         }
 887         if (list_empty(&desc->pg_list)) {
 888                 desc->pg_count = 0;
 889                 desc->pg_base = 0;
 890         }
 891 }
 892
 893 /**
 894  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
 895  * @desc: destination io descriptor
 896  * @req: request
 897  *
 898  * This may split a request into subrequests which are all part of the
 899  * same page group.
 900  *
 901  * Returns true if the request 'req' was successfully coalesced into the
 902  * existing list of pages 'desc'.
 903  */
 904 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 905                            struct nfs_page *req)
 906 {
 907         struct nfs_page *subreq;
 908         unsigned int bytes_left = 0;
 909         unsigned int offset, pgbase;
 910
 911         nfs_page_group_lock(req);
 912
 913         subreq = req;
 914         bytes_left = subreq->wb_bytes;
 915         offset = subreq->wb_offset;
 916         pgbase = subreq->wb_pgbase;
 917
 918         do {
 919                 if (!nfs_pageio_do_add_request(desc, subreq)) {
 920                         /* make sure pg_test call(s) did nothing */
 921                         WARN_ON_ONCE(subreq->wb_bytes != bytes_left);
 922                         WARN_ON_ONCE(subreq->wb_offset != offset);
 923                         WARN_ON_ONCE(subreq->wb_pgbase != pgbase);
 924
 925                         nfs_page_group_unlock(req);
 926                         desc->pg_moreio = 1;
 927                         nfs_pageio_doio(desc);
 928                         if (desc->pg_error < 0)
 929                                 return 0;
 930                         if (desc->pg_recoalesce)
 931                                 return 0;
 932                         /* retry add_request for this subreq */
 933                         nfs_page_group_lock(req);
 934                         continue;
 935                 }
 936
 937                 /* check for buggy pg_test call(s) */
 938                 WARN_ON_ONCE(subreq->wb_bytes + subreq->wb_pgbase > PAGE_SIZE);
 939                 WARN_ON_ONCE(subreq->wb_bytes > bytes_left);
 940                 WARN_ON_ONCE(subreq->wb_bytes == 0);
 941
 942                 bytes_left -= subreq->wb_bytes;
 943                 offset += subreq->wb_bytes;
 944                 pgbase += subreq->wb_bytes;
 945
 946                 if (bytes_left) {
 947                         subreq = nfs_create_request(req->wb_context,
 948                                         req->wb_page,
 949                                         subreq, pgbase, bytes_left);
 950                         if (IS_ERR(subreq))
 951                                 goto err_ptr;
 952                         nfs_lock_request(subreq);
 953                         subreq->wb_offset  = offset;
 954                         subreq->wb_index = req->wb_index;
 955                 }
 956         } while (bytes_left > 0);
 957
 958         nfs_page_group_unlock(req);
 959         return 1;
 960 err_ptr:
 961         desc->pg_error = PTR_ERR(subreq);
 962         nfs_page_group_unlock(req);
 963         return 0;
 964 }
 965
 966 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
 967 {
 968         LIST_HEAD(head);
 969
 970         do {
 971                 list_splice_init(&desc->pg_list, &head);
 972                 desc->pg_bytes_written -= desc->pg_count;
 973                 desc->pg_count = 0;
 974                 desc->pg_base = 0;
 975                 desc->pg_recoalesce = 0;
 976                 desc->pg_moreio = 0;
 977
 978                 while (!list_empty(&head)) {
 979                         struct nfs_page *req;
 980
 981                         req = list_first_entry(&head, struct nfs_page, wb_list);
 982                         nfs_list_remove_request(req);
 983                         if (__nfs_pageio_add_request(desc, req))
 984                                 continue;
 985                         if (desc->pg_error < 0)
 986                                 return 0;
 987                         break;
 988                 }
 989         } while (desc->pg_recoalesce);
 990         return 1;
 991 }
 992
 993 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 994                 struct nfs_page *req)
 995 {
 996         int ret;
 997
 998         do {
 999                 ret = __nfs_pageio_add_request(desc, req);
1000                 if (ret)
1001                         break;
1002                 if (desc->pg_error < 0)
1003                         break;
1004                 ret = nfs_do_recoalesce(desc);
1005         } while (ret);
1006         return ret;
1007 }
1008 EXPORT_SYMBOL_GPL(nfs_pageio_add_request);
1009
1010 /**
1011  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
1012  * @desc: pointer to io descriptor
1013  */
1014 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1015 {
1016         for (;;) {
1017                 nfs_pageio_doio(desc);
1018                 if (!desc->pg_recoalesce)
1019                         break;
1020                 if (!nfs_do_recoalesce(desc))
1021                         break;
1022         }
1023 }
1024 EXPORT_SYMBOL_GPL(nfs_pageio_complete);
1025
1026 /**
1027  * nfs_pageio_cond_complete - Conditional I/O completion
1028  * @desc: pointer to io descriptor
1029  * @index: page index
1030  *
1031  * It is important to ensure that processes don't try to take locks
1032  * on non-contiguous ranges of pages as that might deadlock. This
1033  * function should be called before attempting to wait on a locked
1034  * nfs_page. It will complete the I/O if the page index 'index'
1035  * is not contiguous with the existing list of pages in 'desc'.
1036  */
1037 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1038 {
1039         if (!list_empty(&desc->pg_list)) {
1040                 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
1041                 if (index != prev->wb_index + 1)
1042                         nfs_pageio_complete(desc);
1043         }
1044 }
1045
1046 int __init nfs_init_nfspagecache(void)
1047 {
1048         nfs_page_cachep = kmem_cache_create("nfs_page",
1049                                             sizeof(struct nfs_page),
1050                                             0, SLAB_HWCACHE_ALIGN,
1051                                             NULL);
1052         if (nfs_page_cachep == NULL)
1053                 return -ENOMEM;
1054
1055         return 0;
1056 }
1057
1058 void nfs_destroy_nfspagecache(void)
1059 {
1060         kmem_cache_destroy(nfs_page_cachep);
1061 }
1062
1063 static const struct rpc_call_ops nfs_pgio_common_ops = {
1064         .rpc_call_prepare = nfs_pgio_prepare,
1065         .rpc_call_done = nfs_pgio_result,
1066         .rpc_release = nfs_pgio_release,
1067 };
1068
1069 const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1070         .pg_test = nfs_generic_pg_test,
1071         .pg_doio = nfs_generic_pg_pgios,
1072 };