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