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(&c->flags);
 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 static int nfs_wait_bit_uninterruptible(void *word)
 140 {
 141         io_schedule();
 142         return 0;
 143 }
 144
 145 /*
 146  * nfs_page_group_lock - lock the head of the page group
 147  * @req - request in group that is to be locked
 148  *
 149  * this lock must be held if modifying the page group list
 150  */
 151 void
 152 nfs_page_group_lock(struct nfs_page *req)
 153 {
 154         struct nfs_page *head = req->wb_head;
 155
 156         WARN_ON_ONCE(head != head->wb_head);
 157
 158         wait_on_bit_lock(&head->wb_flags, PG_HEADLOCK,
 159                         nfs_wait_bit_uninterruptible,
 160                         TASK_UNINTERRUPTIBLE);
 161 }
 162
 163 /*
 164  * nfs_page_group_unlock - unlock the head of the page group
 165  * @req - request in group that is to be unlocked
 166  */
 167 void
 168 nfs_page_group_unlock(struct nfs_page *req)
 169 {
 170         struct nfs_page *head = req->wb_head;
 171
 172         WARN_ON_ONCE(head != head->wb_head);
 173
 174         smp_mb__before_atomic();
 175         clear_bit(PG_HEADLOCK, &head->wb_flags);
 176         smp_mb__after_atomic();
 177         wake_up_bit(&head->wb_flags, PG_HEADLOCK);
 178 }
 179
 180 /*
 181  * nfs_page_group_sync_on_bit_locked
 182  *
 183  * must be called with page group lock held
 184  */
 185 static bool
 186 nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
 187 {
 188         struct nfs_page *head = req->wb_head;
 189         struct nfs_page *tmp;
 190
 191         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags));
 192         WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags));
 193
 194         tmp = req->wb_this_page;
 195         while (tmp != req) {
 196                 if (!test_bit(bit, &tmp->wb_flags))
 197                         return false;
 198                 tmp = tmp->wb_this_page;
 199         }
 200
 201         /* true! reset all bits */
 202         tmp = req;
 203         do {
 204                 clear_bit(bit, &tmp->wb_flags);
 205                 tmp = tmp->wb_this_page;
 206         } while (tmp != req);
 207
 208         return true;
 209 }
 210
 211 /*
 212  * nfs_page_group_sync_on_bit - set bit on current request, but only
 213  *   return true if the bit is set for all requests in page group
 214  * @req - request in page group
 215  * @bit - PG_* bit that is used to sync page group
 216  */
 217 bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit)
 218 {
 219         bool ret;
 220
 221         nfs_page_group_lock(req);
 222         ret = nfs_page_group_sync_on_bit_locked(req, bit);
 223         nfs_page_group_unlock(req);
 224
 225         return ret;
 226 }
 227
 228 /*
 229  * nfs_page_group_init - Initialize the page group linkage for @req
 230  * @req - a new nfs request
 231  * @prev - the previous request in page group, or NULL if @req is the first
 232  *         or only request in the group (the head).
 233  */
 234 static inline void
 235 nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev)
 236 {
 237         WARN_ON_ONCE(prev == req);
 238
 239         if (!prev) {
 240                 /* a head request */
 241                 req->wb_head = req;
 242                 req->wb_this_page = req;
 243         } else {
 244                 /* a subrequest */
 245                 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head);
 246                 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags));
 247                 req->wb_head = prev->wb_head;
 248                 req->wb_this_page = prev->wb_this_page;
 249                 prev->wb_this_page = req;
 250
 251                 /* All subrequests take a ref on the head request until
 252                  * nfs_page_group_destroy is called */
 253                 kref_get(&req->wb_head->wb_kref);
 254
 255                 /* grab extra ref if head request has extra ref from
 256                  * the write/commit path to handle handoff between write
 257                  * and commit lists */
 258                 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) {
 259                         set_bit(PG_INODE_REF, &req->wb_flags);
 260                         kref_get(&req->wb_kref);
 261                 }
 262         }
 263 }
 264
 265 /*
 266  * nfs_page_group_destroy - sync the destruction of page groups
 267  * @req - request that no longer needs the page group
 268  *
 269  * releases the page group reference from each member once all
 270  * members have called this function.
 271  */
 272 static void
 273 nfs_page_group_destroy(struct kref *kref)
 274 {
 275         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 276         struct nfs_page *tmp, *next;
 277
 278         /* subrequests must release the ref on the head request */
 279         if (req->wb_head != req)
 280                 nfs_release_request(req->wb_head);
 281
 282         if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN))
 283                 return;
 284
 285         tmp = req;
 286         do {
 287                 next = tmp->wb_this_page;
 288                 /* unlink and free */
 289                 tmp->wb_this_page = tmp;
 290                 tmp->wb_head = tmp;
 291                 nfs_free_request(tmp);
 292                 tmp = next;
 293         } while (tmp != req);
 294 }
 295
 296 /**
 297  * nfs_create_request - Create an NFS read/write request.
 298  * @ctx: open context to use
 299  * @page: page to write
 300  * @last: last nfs request created for this page group or NULL if head
 301  * @offset: starting offset within the page for the write
 302  * @count: number of bytes to read/write
 303  *
 304  * The page must be locked by the caller. This makes sure we never
 305  * create two different requests for the same page.
 306  * User should ensure it is safe to sleep in this function.
 307  */
 308 struct nfs_page *
 309 nfs_create_request(struct nfs_open_context *ctx, struct page *page,
 310                    struct nfs_page *last, unsigned int offset,
 311                    unsigned int count)
 312 {
 313         struct nfs_page         *req;
 314         struct nfs_lock_context *l_ctx;
 315
 316         if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
 317                 return ERR_PTR(-EBADF);
 318         /* try to allocate the request struct */
 319         req = nfs_page_alloc();
 320         if (req == NULL)
 321                 return ERR_PTR(-ENOMEM);
 322
 323         /* get lock context early so we can deal with alloc failures */
 324         l_ctx = nfs_get_lock_context(ctx);
 325         if (IS_ERR(l_ctx)) {
 326                 nfs_page_free(req);
 327                 return ERR_CAST(l_ctx);
 328         }
 329         req->wb_lock_context = l_ctx;
 330         nfs_iocounter_inc(&l_ctx->io_count);
 331
 332         /* Initialize the request struct. Initially, we assume a
 333          * long write-back delay. This will be adjusted in
 334          * update_nfs_request below if the region is not locked. */
 335         req->wb_page    = page;
 336         req->wb_index   = page_file_index(page);
 337         page_cache_get(page);
 338         req->wb_offset  = offset;
 339         req->wb_pgbase  = offset;
 340         req->wb_bytes   = count;
 341         req->wb_context = get_nfs_open_context(ctx);
 342         kref_init(&req->wb_kref);
 343         nfs_page_group_init(req, last);
 344         return req;
 345 }
 346
 347 /**
 348  * nfs_unlock_request - Unlock request and wake up sleepers.
 349  * @req:
 350  */
 351 void nfs_unlock_request(struct nfs_page *req)
 352 {
 353         if (!NFS_WBACK_BUSY(req)) {
 354                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 355                 BUG();
 356         }
 357         smp_mb__before_atomic();
 358         clear_bit(PG_BUSY, &req->wb_flags);
 359         smp_mb__after_atomic();
 360         wake_up_bit(&req->wb_flags, PG_BUSY);
 361 }
 362
 363 /**
 364  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
 365  * @req:
 366  */
 367 void nfs_unlock_and_release_request(struct nfs_page *req)
 368 {
 369         nfs_unlock_request(req);
 370         nfs_release_request(req);
 371 }
 372
 373 /*
 374  * nfs_clear_request - Free up all resources allocated to the request
 375  * @req:
 376  *
 377  * Release page and open context resources associated with a read/write
 378  * request after it has completed.
 379  */
 380 static void nfs_clear_request(struct nfs_page *req)
 381 {
 382         struct page *page = req->wb_page;
 383         struct nfs_open_context *ctx = req->wb_context;
 384         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 385
 386         if (page != NULL) {
 387                 page_cache_release(page);
 388                 req->wb_page = NULL;
 389         }
 390         if (l_ctx != NULL) {
 391                 nfs_iocounter_dec(&l_ctx->io_count);
 392                 nfs_put_lock_context(l_ctx);
 393                 req->wb_lock_context = NULL;
 394         }
 395         if (ctx != NULL) {
 396                 put_nfs_open_context(ctx);
 397                 req->wb_context = NULL;
 398         }
 399 }
 400
 401 /**
 402  * nfs_release_request - Release the count on an NFS read/write request
 403  * @req: request to release
 404  *
 405  * Note: Should never be called with the spinlock held!
 406  */
 407 void nfs_free_request(struct nfs_page *req)
 408 {
 409         WARN_ON_ONCE(req->wb_this_page != req);
 410
 411         /* extra debug: make sure no sync bits are still set */
 412         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
 413         WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
 414         WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
 415         WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
 416         WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
 417
 418         /* Release struct file and open context */
 419         nfs_clear_request(req);
 420         nfs_page_free(req);
 421 }
 422
 423 void nfs_release_request(struct nfs_page *req)
 424 {
 425         kref_put(&req->wb_kref, nfs_page_group_destroy);
 426 }
 427
 428 /**
 429  * nfs_wait_on_request - Wait for a request to complete.
 430  * @req: request to wait upon.
 431  *
 432  * Interruptible by fatal signals only.
 433  * The user is responsible for holding a count on the request.
 434  */
 435 int
 436 nfs_wait_on_request(struct nfs_page *req)
 437 {
 438         return wait_on_bit(&req->wb_flags, PG_BUSY,
 439                         nfs_wait_bit_uninterruptible,
 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 static inline struct nfs_rw_header *NFS_RW_HEADER(struct nfs_pgio_header *hdr)
 466 {
 467         return container_of(hdr, struct nfs_rw_header, header);
 468 }
 469
 470 /**
 471  * nfs_rw_header_alloc - Allocate a header for a read or write
 472  * @ops: Read or write function vector
 473  */
 474 struct nfs_rw_header *nfs_rw_header_alloc(const struct nfs_rw_ops *ops)
 475 {
 476         struct nfs_rw_header *header = ops->rw_alloc_header();
 477
 478         if (header) {
 479                 struct nfs_pgio_header *hdr = &header->header;
 480
 481                 INIT_LIST_HEAD(&hdr->pages);
 482                 spin_lock_init(&hdr->lock);
 483                 atomic_set(&hdr->refcnt, 0);
 484                 hdr->rw_ops = ops;
 485         }
 486         return header;
 487 }
 488 EXPORT_SYMBOL_GPL(nfs_rw_header_alloc);
 489
 490 /*
 491  * nfs_rw_header_free - Free a read or write header
 492  * @hdr: The header to free
 493  */
 494 void nfs_rw_header_free(struct nfs_pgio_header *hdr)
 495 {
 496         hdr->rw_ops->rw_free_header(NFS_RW_HEADER(hdr));
 497 }
 498 EXPORT_SYMBOL_GPL(nfs_rw_header_free);
 499
 500 /**
 501  * nfs_pgio_data_alloc - Allocate pageio data
 502  * @hdr: The header making a request
 503  * @pagecount: Number of pages to create
 504  */
 505 static struct nfs_pgio_data *nfs_pgio_data_alloc(struct nfs_pgio_header *hdr,
 506                                                  unsigned int pagecount)
 507 {
 508         struct nfs_pgio_data *data, *prealloc;
 509
 510         prealloc = &NFS_RW_HEADER(hdr)->rpc_data;
 511         if (prealloc->header == NULL)
 512                 data = prealloc;
 513         else
 514                 data = kzalloc(sizeof(*data), GFP_KERNEL);
 515         if (!data)
 516                 goto out;
 517
 518         if (nfs_pgarray_set(&data->pages, pagecount)) {
 519                 data->header = hdr;
 520                 atomic_inc(&hdr->refcnt);
 521         } else {
 522                 if (data != prealloc)
 523                         kfree(data);
 524                 data = NULL;
 525         }
 526 out:
 527         return data;
 528 }
 529
 530 /**
 531  * nfs_pgio_data_release - Properly free pageio data
 532  * @data: The data to release
 533  */
 534 void nfs_pgio_data_release(struct nfs_pgio_data *data)
 535 {
 536         struct nfs_pgio_header *hdr = data->header;
 537         struct nfs_rw_header *pageio_header = NFS_RW_HEADER(hdr);
 538
 539         put_nfs_open_context(data->args.context);
 540         if (data->pages.pagevec != data->pages.page_array)
 541                 kfree(data->pages.pagevec);
 542         if (data == &pageio_header->rpc_data) {
 543                 data->header = NULL;
 544                 data = NULL;
 545         }
 546         if (atomic_dec_and_test(&hdr->refcnt))
 547                 hdr->completion_ops->completion(hdr);
 548         /* Note: we only free the rpc_task after callbacks are done.
 549          * See the comment in rpc_free_task() for why
 550          */
 551         kfree(data);
 552 }
 553 EXPORT_SYMBOL_GPL(nfs_pgio_data_release);
 554
 555 /**
 556  * nfs_pgio_rpcsetup - Set up arguments for a pageio call
 557  * @data: The pageio data
 558  * @count: Number of bytes to read
 559  * @offset: Initial offset
 560  * @how: How to commit data (writes only)
 561  * @cinfo: Commit information for the call (writes only)
 562  */
 563 static void nfs_pgio_rpcsetup(struct nfs_pgio_data *data,
 564                               unsigned int count, unsigned int offset,
 565                               int how, struct nfs_commit_info *cinfo)
 566 {
 567         struct nfs_page *req = data->header->req;
 568
 569         /* Set up the RPC argument and reply structs
 570          * NB: take care not to mess about with data->commit et al. */
 571
 572         data->args.fh     = NFS_FH(data->header->inode);
 573         data->args.offset = req_offset(req) + offset;
 574         /* pnfs_set_layoutcommit needs this */
 575         data->mds_offset = data->args.offset;
 576         data->args.pgbase = req->wb_pgbase + offset;
 577         data->args.pages  = data->pages.pagevec;
 578         data->args.count  = count;
 579         data->args.context = get_nfs_open_context(req->wb_context);
 580         data->args.lock_context = req->wb_lock_context;
 581         data->args.stable  = NFS_UNSTABLE;
 582         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 583         case 0:
 584                 break;
 585         case FLUSH_COND_STABLE:
 586                 if (nfs_reqs_to_commit(cinfo))
 587                         break;
 588         default:
 589                 data->args.stable = NFS_FILE_SYNC;
 590         }
 591
 592         data->res.fattr   = &data->fattr;
 593         data->res.count   = count;
 594         data->res.eof     = 0;
 595         data->res.verf    = &data->verf;
 596         nfs_fattr_init(&data->fattr);
 597 }
 598
 599 /**
 600  * nfs_pgio_prepare - Prepare pageio data to go over the wire
 601  * @task: The current task
 602  * @calldata: pageio data to prepare
 603  */
 604 static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
 605 {
 606         struct nfs_pgio_data *data = calldata;
 607         int err;
 608         err = NFS_PROTO(data->header->inode)->pgio_rpc_prepare(task, data);
 609         if (err)
 610                 rpc_exit(task, err);
 611 }
 612
 613 int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_data *data,
 614                       const struct rpc_call_ops *call_ops, int how, int flags)
 615 {
 616         struct rpc_task *task;
 617         struct rpc_message msg = {
 618                 .rpc_argp = &data->args,
 619                 .rpc_resp = &data->res,
 620                 .rpc_cred = data->header->cred,
 621         };
 622         struct rpc_task_setup task_setup_data = {
 623                 .rpc_client = clnt,
 624                 .task = &data->task,
 625                 .rpc_message = &msg,
 626                 .callback_ops = call_ops,
 627                 .callback_data = data,
 628                 .workqueue = nfsiod_workqueue,
 629                 .flags = RPC_TASK_ASYNC | flags,
 630         };
 631         int ret = 0;
 632
 633         data->header->rw_ops->rw_initiate(data, &msg, &task_setup_data, how);
 634
 635         dprintk("NFS: %5u initiated pgio call "
 636                 "(req %s/%llu, %u bytes @ offset %llu)\n",
 637                 data->task.tk_pid,
 638                 data->header->inode->i_sb->s_id,
 639                 (unsigned long long)NFS_FILEID(data->header->inode),
 640                 data->args.count,
 641                 (unsigned long long)data->args.offset);
 642
 643         task = rpc_run_task(&task_setup_data);
 644         if (IS_ERR(task)) {
 645                 ret = PTR_ERR(task);
 646                 goto out;
 647         }
 648         if (how & FLUSH_SYNC) {
 649                 ret = rpc_wait_for_completion_task(task);
 650                 if (ret == 0)
 651                         ret = task->tk_status;
 652         }
 653         rpc_put_task(task);
 654 out:
 655         return ret;
 656 }
 657 EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
 658
 659 /**
 660  * nfs_pgio_error - Clean up from a pageio error
 661  * @desc: IO descriptor
 662  * @hdr: pageio header
 663  */
 664 static int nfs_pgio_error(struct nfs_pageio_descriptor *desc,
 665                           struct nfs_pgio_header *hdr)
 666 {
 667         set_bit(NFS_IOHDR_REDO, &hdr->flags);
 668         nfs_pgio_data_release(hdr->data);
 669         hdr->data = NULL;
 670         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 671         return -ENOMEM;
 672 }
 673
 674 /**
 675  * nfs_pgio_release - Release pageio data
 676  * @calldata: The pageio data to release
 677  */
 678 static void nfs_pgio_release(void *calldata)
 679 {
 680         struct nfs_pgio_data *data = calldata;
 681         if (data->header->rw_ops->rw_release)
 682                 data->header->rw_ops->rw_release(data);
 683         nfs_pgio_data_release(data);
 684 }
 685
 686 /**
 687  * nfs_pageio_init - initialise a page io descriptor
 688  * @desc: pointer to descriptor
 689  * @inode: pointer to inode
 690  * @doio: pointer to io function
 691  * @bsize: io block size
 692  * @io_flags: extra parameters for the io function
 693  */
 694 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 695                      struct inode *inode,
 696                      const struct nfs_pageio_ops *pg_ops,
 697                      const struct nfs_pgio_completion_ops *compl_ops,
 698                      const struct nfs_rw_ops *rw_ops,
 699                      size_t bsize,
 700                      int io_flags)
 701 {
 702         INIT_LIST_HEAD(&desc->pg_list);
 703         desc->pg_bytes_written = 0;
 704         desc->pg_count = 0;
 705         desc->pg_bsize = bsize;
 706         desc->pg_base = 0;
 707         desc->pg_moreio = 0;
 708         desc->pg_recoalesce = 0;
 709         desc->pg_inode = inode;
 710         desc->pg_ops = pg_ops;
 711         desc->pg_completion_ops = compl_ops;
 712         desc->pg_rw_ops = rw_ops;
 713         desc->pg_ioflags = io_flags;
 714         desc->pg_error = 0;
 715         desc->pg_lseg = NULL;
 716         desc->pg_dreq = NULL;
 717         desc->pg_layout_private = NULL;
 718 }
 719 EXPORT_SYMBOL_GPL(nfs_pageio_init);
 720
 721 /**
 722  * nfs_pgio_result - Basic pageio error handling
 723  * @task: The task that ran
 724  * @calldata: Pageio data to check
 725  */
 726 static void nfs_pgio_result(struct rpc_task *task, void *calldata)
 727 {
 728         struct nfs_pgio_data *data = calldata;
 729         struct inode *inode = data->header->inode;
 730
 731         dprintk("NFS: %s: %5u, (status %d)\n", __func__,
 732                 task->tk_pid, task->tk_status);
 733
 734         if (data->header->rw_ops->rw_done(task, data, inode) != 0)
 735                 return;
 736         if (task->tk_status < 0)
 737                 nfs_set_pgio_error(data->header, task->tk_status, data->args.offset);
 738         else
 739                 data->header->rw_ops->rw_result(task, data);
 740 }
 741
 742 /*
 743  * Create an RPC task for the given read or write request and kick it.
 744  * The page must have been locked by the caller.
 745  *
 746  * It may happen that the page we're passed is not marked dirty.
 747  * This is the case if nfs_updatepage detects a conflicting request
 748  * that has been written but not committed.
 749  */
 750 int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
 751                      struct nfs_pgio_header *hdr)
 752 {
 753         struct nfs_page         *req;
 754         struct page             **pages;
 755         struct nfs_pgio_data    *data;
 756         struct list_head *head = &desc->pg_list;
 757         struct nfs_commit_info cinfo;
 758
 759         data = nfs_pgio_data_alloc(hdr, nfs_page_array_len(desc->pg_base,
 760                                                            desc->pg_count));
 761         if (!data)
 762                 return nfs_pgio_error(desc, hdr);
 763
 764         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
 765         pages = data->pages.pagevec;
 766         while (!list_empty(head)) {
 767                 req = nfs_list_entry(head->next);
 768                 nfs_list_remove_request(req);
 769                 nfs_list_add_request(req, &hdr->pages);
 770                 *pages++ = req->wb_page;
 771         }
 772
 773         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 774             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
 775                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 776
 777         /* Set up the argument struct */
 778         nfs_pgio_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
 779         hdr->data = data;
 780         desc->pg_rpc_callops = &nfs_pgio_common_ops;
 781         return 0;
 782 }
 783 EXPORT_SYMBOL_GPL(nfs_generic_pgio);
 784
 785 static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
 786 {
 787         struct nfs_rw_header *rw_hdr;
 788         struct nfs_pgio_header *hdr;
 789         int ret;
 790
 791         rw_hdr = nfs_rw_header_alloc(desc->pg_rw_ops);
 792         if (!rw_hdr) {
 793                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 794                 return -ENOMEM;
 795         }
 796         hdr = &rw_hdr->header;
 797         nfs_pgheader_init(desc, hdr, nfs_rw_header_free);
 798         atomic_inc(&hdr->refcnt);
 799         ret = nfs_generic_pgio(desc, hdr);
 800         if (ret == 0)
 801                 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
 802                                         hdr->data, desc->pg_rpc_callops,
 803                                         desc->pg_ioflags, 0);
 804         if (atomic_dec_and_test(&hdr->refcnt))
 805                 hdr->completion_ops->completion(hdr);
 806         return ret;
 807 }
 808
 809 static bool nfs_match_open_context(const struct nfs_open_context *ctx1,
 810                 const struct nfs_open_context *ctx2)
 811 {
 812         return ctx1->cred == ctx2->cred && ctx1->state == ctx2->state;
 813 }
 814
 815 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
 816                 const struct nfs_lock_context *l2)
 817 {
 818         return l1->lockowner.l_owner == l2->lockowner.l_owner
 819                 && l1->lockowner.l_pid == l2->lockowner.l_pid;
 820 }
 821
 822 /**
 823  * nfs_can_coalesce_requests - test two requests for compatibility
 824  * @prev: pointer to nfs_page
 825  * @req: pointer to nfs_page
 826  *
 827  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
 828  * page data area they describe is contiguous, and that their RPC
 829  * credentials, NFSv4 open state, and lockowners are the same.
 830  *
 831  * Return 'true' if this is the case, else return 'false'.
 832  */
 833 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
 834                                       struct nfs_page *req,
 835                                       struct nfs_pageio_descriptor *pgio)
 836 {
 837         size_t size;
 838
 839         if (prev) {
 840                 if (!nfs_match_open_context(req->wb_context, prev->wb_context))
 841                         return false;
 842                 if (req->wb_context->dentry->d_inode->i_flock != NULL &&
 843                     !nfs_match_lock_context(req->wb_lock_context,
 844                                             prev->wb_lock_context))
 845                         return false;
 846                 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
 847                         return false;
 848         }
 849         size = pgio->pg_ops->pg_test(pgio, prev, req);
 850         WARN_ON_ONCE(size > req->wb_bytes);
 851         if (size && size < req->wb_bytes)
 852                 req->wb_bytes = size;
 853         return size > 0;
 854 }
 855
 856 /**
 857  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
 858  * @desc: destination io descriptor
 859  * @req: request
 860  *
 861  * Returns true if the request 'req' was successfully coalesced into the
 862  * existing list of pages 'desc'.
 863  */
 864 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
 865                                      struct nfs_page *req)
 866 {
 867         struct nfs_page *prev = NULL;
 868         if (desc->pg_count != 0) {
 869                 prev = nfs_list_entry(desc->pg_list.prev);
 870         } else {
 871                 if (desc->pg_ops->pg_init)
 872                         desc->pg_ops->pg_init(desc, req);
 873                 desc->pg_base = req->wb_pgbase;
 874         }
 875         if (!nfs_can_coalesce_requests(prev, req, desc))
 876                 return 0;
 877         nfs_list_remove_request(req);
 878         nfs_list_add_request(req, &desc->pg_list);
 879         desc->pg_count += req->wb_bytes;
 880         return 1;
 881 }
 882
 883 /*
 884  * Helper for nfs_pageio_add_request and nfs_pageio_complete
 885  */
 886 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
 887 {
 888         if (!list_empty(&desc->pg_list)) {
 889                 int error = desc->pg_ops->pg_doio(desc);
 890                 if (error < 0)
 891                         desc->pg_error = error;
 892                 else
 893                         desc->pg_bytes_written += desc->pg_count;
 894         }
 895         if (list_empty(&desc->pg_list)) {
 896                 desc->pg_count = 0;
 897                 desc->pg_base = 0;
 898         }
 899 }
 900
 901 /**
 902  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
 903  * @desc: destination io descriptor
 904  * @req: request
 905  *
 906  * This may split a request into subrequests which are all part of the
 907  * same page group.
 908  *
 909  * Returns true if the request 'req' was successfully coalesced into the
 910  * existing list of pages 'desc'.
 911  */
 912 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 913                            struct nfs_page *req)
 914 {
 915         struct nfs_page *subreq;
 916         unsigned int bytes_left = 0;
 917         unsigned int offset, pgbase;
 918
 919         nfs_page_group_lock(req);
 920
 921         subreq = req;
 922         bytes_left = subreq->wb_bytes;
 923         offset = subreq->wb_offset;
 924         pgbase = subreq->wb_pgbase;
 925
 926         do {
 927                 if (!nfs_pageio_do_add_request(desc, subreq)) {
 928                         /* make sure pg_test call(s) did nothing */
 929                         WARN_ON_ONCE(subreq->wb_bytes != bytes_left);
 930                         WARN_ON_ONCE(subreq->wb_offset != offset);
 931                         WARN_ON_ONCE(subreq->wb_pgbase != pgbase);
 932
 933                         nfs_page_group_unlock(req);
 934                         desc->pg_moreio = 1;
 935                         nfs_pageio_doio(desc);
 936                         if (desc->pg_error < 0)
 937                                 return 0;
 938                         if (desc->pg_recoalesce)
 939                                 return 0;
 940                         /* retry add_request for this subreq */
 941                         nfs_page_group_lock(req);
 942                         continue;
 943                 }
 944
 945                 /* check for buggy pg_test call(s) */
 946                 WARN_ON_ONCE(subreq->wb_bytes + subreq->wb_pgbase > PAGE_SIZE);
 947                 WARN_ON_ONCE(subreq->wb_bytes > bytes_left);
 948                 WARN_ON_ONCE(subreq->wb_bytes == 0);
 949
 950                 bytes_left -= subreq->wb_bytes;
 951                 offset += subreq->wb_bytes;
 952                 pgbase += subreq->wb_bytes;
 953
 954                 if (bytes_left) {
 955                         subreq = nfs_create_request(req->wb_context,
 956                                         req->wb_page,
 957                                         subreq, pgbase, bytes_left);
 958                         if (IS_ERR(subreq))
 959                                 goto err_ptr;
 960                         nfs_lock_request(subreq);
 961                         subreq->wb_offset  = offset;
 962                         subreq->wb_index = req->wb_index;
 963                 }
 964         } while (bytes_left > 0);
 965
 966         nfs_page_group_unlock(req);
 967         return 1;
 968 err_ptr:
 969         desc->pg_error = PTR_ERR(subreq);
 970         nfs_page_group_unlock(req);
 971         return 0;
 972 }
 973
 974 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
 975 {
 976         LIST_HEAD(head);
 977
 978         do {
 979                 list_splice_init(&desc->pg_list, &head);
 980                 desc->pg_bytes_written -= desc->pg_count;
 981                 desc->pg_count = 0;
 982                 desc->pg_base = 0;
 983                 desc->pg_recoalesce = 0;
 984                 desc->pg_moreio = 0;
 985
 986                 while (!list_empty(&head)) {
 987                         struct nfs_page *req;
 988
 989                         req = list_first_entry(&head, struct nfs_page, wb_list);
 990                         nfs_list_remove_request(req);
 991                         if (__nfs_pageio_add_request(desc, req))
 992                                 continue;
 993                         if (desc->pg_error < 0)
 994                                 return 0;
 995                         break;
 996                 }
 997         } while (desc->pg_recoalesce);
 998         return 1;
 999 }
1000
1001 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1002                 struct nfs_page *req)
1003 {
1004         int ret;
1005
1006         do {
1007                 ret = __nfs_pageio_add_request(desc, req);
1008                 if (ret)
1009                         break;
1010                 if (desc->pg_error < 0)
1011                         break;
1012                 ret = nfs_do_recoalesce(desc);
1013         } while (ret);
1014         return ret;
1015 }
1016 EXPORT_SYMBOL_GPL(nfs_pageio_add_request);
1017
1018 /**
1019  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
1020  * @desc: pointer to io descriptor
1021  */
1022 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1023 {
1024         for (;;) {
1025                 nfs_pageio_doio(desc);
1026                 if (!desc->pg_recoalesce)
1027                         break;
1028                 if (!nfs_do_recoalesce(desc))
1029                         break;
1030         }
1031 }
1032 EXPORT_SYMBOL_GPL(nfs_pageio_complete);
1033
1034 /**
1035  * nfs_pageio_cond_complete - Conditional I/O completion
1036  * @desc: pointer to io descriptor
1037  * @index: page index
1038  *
1039  * It is important to ensure that processes don't try to take locks
1040  * on non-contiguous ranges of pages as that might deadlock. This
1041  * function should be called before attempting to wait on a locked
1042  * nfs_page. It will complete the I/O if the page index 'index'
1043  * is not contiguous with the existing list of pages in 'desc'.
1044  */
1045 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1046 {
1047         if (!list_empty(&desc->pg_list)) {
1048                 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
1049                 if (index != prev->wb_index + 1)
1050                         nfs_pageio_complete(desc);
1051         }
1052 }
1053
1054 int __init nfs_init_nfspagecache(void)
1055 {
1056         nfs_page_cachep = kmem_cache_create("nfs_page",
1057                                             sizeof(struct nfs_page),
1058                                             0, SLAB_HWCACHE_ALIGN,
1059                                             NULL);
1060         if (nfs_page_cachep == NULL)
1061                 return -ENOMEM;
1062
1063         return 0;
1064 }
1065
1066 void nfs_destroy_nfspagecache(void)
1067 {
1068         kmem_cache_destroy(nfs_page_cachep);
1069 }
1070
1071 static const struct rpc_call_ops nfs_pgio_common_ops = {
1072         .rpc_call_prepare = nfs_pgio_prepare,
1073         .rpc_call_done = nfs_pgio_result,
1074         .rpc_release = nfs_pgio_release,
1075 };
1076
1077 const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1078         .pg_test = nfs_generic_pg_test,
1079         .pg_doio = nfs_generic_pg_pgios,
1080 };