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 && !ret);
 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  * @nonblock - if true don't block waiting for lock
 143  *
 144  * this lock must be held if modifying the page group list
 145  *
 146  * return 0 on success, < 0 on error: -EDELAY if nonblocking or the
 147  * result from wait_on_bit_lock
 148  *
 149  * NOTE: calling with nonblock=false should always have set the
 150  *       lock bit (see fs/buffer.c and other uses of wait_on_bit_lock
 151  *       with TASK_UNINTERRUPTIBLE), so there is no need to check the result.
 152  */
 153 int
 154 nfs_page_group_lock(struct nfs_page *req, bool nonblock)
 155 {
 156         struct nfs_page *head = req->wb_head;
 157
 158         WARN_ON_ONCE(head != head->wb_head);
 159
 160         if (!test_and_set_bit(PG_HEADLOCK, &head->wb_flags))
 161                 return 0;
 162
 163         if (!nonblock)
 164                 return wait_on_bit_lock(&head->wb_flags, PG_HEADLOCK,
 165                                 TASK_UNINTERRUPTIBLE);
 166
 167         return -EAGAIN;
 168 }
 169
 170 /*
 171  * nfs_page_group_lock_wait - wait for the lock to clear, but don't grab it
 172  * @req - a request in the group
 173  *
 174  * This is a blocking call to wait for the group lock to be cleared.
 175  */
 176 void
 177 nfs_page_group_lock_wait(struct nfs_page *req)
 178 {
 179         struct nfs_page *head = req->wb_head;
 180
 181         WARN_ON_ONCE(head != head->wb_head);
 182
 183         wait_on_bit(&head->wb_flags, PG_HEADLOCK,
 184                 TASK_UNINTERRUPTIBLE);
 185 }
 186
 187 /*
 188  * nfs_page_group_unlock - unlock the head of the page group
 189  * @req - request in group that is to be unlocked
 190  */
 191 void
 192 nfs_page_group_unlock(struct nfs_page *req)
 193 {
 194         struct nfs_page *head = req->wb_head;
 195
 196         WARN_ON_ONCE(head != head->wb_head);
 197
 198         smp_mb__before_atomic();
 199         clear_bit(PG_HEADLOCK, &head->wb_flags);
 200         smp_mb__after_atomic();
 201         wake_up_bit(&head->wb_flags, PG_HEADLOCK);
 202 }
 203
 204 /*
 205  * nfs_page_group_sync_on_bit_locked
 206  *
 207  * must be called with page group lock held
 208  */
 209 static bool
 210 nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
 211 {
 212         struct nfs_page *head = req->wb_head;
 213         struct nfs_page *tmp;
 214
 215         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags));
 216         WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags));
 217
 218         tmp = req->wb_this_page;
 219         while (tmp != req) {
 220                 if (!test_bit(bit, &tmp->wb_flags))
 221                         return false;
 222                 tmp = tmp->wb_this_page;
 223         }
 224
 225         /* true! reset all bits */
 226         tmp = req;
 227         do {
 228                 clear_bit(bit, &tmp->wb_flags);
 229                 tmp = tmp->wb_this_page;
 230         } while (tmp != req);
 231
 232         return true;
 233 }
 234
 235 /*
 236  * nfs_page_group_sync_on_bit - set bit on current request, but only
 237  *   return true if the bit is set for all requests in page group
 238  * @req - request in page group
 239  * @bit - PG_* bit that is used to sync page group
 240  */
 241 bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit)
 242 {
 243         bool ret;
 244
 245         nfs_page_group_lock(req, false);
 246         ret = nfs_page_group_sync_on_bit_locked(req, bit);
 247         nfs_page_group_unlock(req);
 248
 249         return ret;
 250 }
 251
 252 /*
 253  * nfs_page_group_init - Initialize the page group linkage for @req
 254  * @req - a new nfs request
 255  * @prev - the previous request in page group, or NULL if @req is the first
 256  *         or only request in the group (the head).
 257  */
 258 static inline void
 259 nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev)
 260 {
 261         struct inode *inode;
 262         WARN_ON_ONCE(prev == req);
 263
 264         if (!prev) {
 265                 /* a head request */
 266                 req->wb_head = req;
 267                 req->wb_this_page = req;
 268         } else {
 269                 /* a subrequest */
 270                 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head);
 271                 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags));
 272                 req->wb_head = prev->wb_head;
 273                 req->wb_this_page = prev->wb_this_page;
 274                 prev->wb_this_page = req;
 275
 276                 /* All subrequests take a ref on the head request until
 277                  * nfs_page_group_destroy is called */
 278                 kref_get(&req->wb_head->wb_kref);
 279
 280                 /* grab extra ref and bump the request count if head request
 281                  * has extra ref from the write/commit path to handle handoff
 282                  * between write and commit lists. */
 283                 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) {
 284                         inode = page_file_mapping(req->wb_page)->host;
 285                         set_bit(PG_INODE_REF, &req->wb_flags);
 286                         kref_get(&req->wb_kref);
 287                         spin_lock(&inode->i_lock);
 288                         NFS_I(inode)->nrequests++;
 289                         spin_unlock(&inode->i_lock);
 290                 }
 291         }
 292 }
 293
 294 /*
 295  * nfs_page_group_destroy - sync the destruction of page groups
 296  * @req - request that no longer needs the page group
 297  *
 298  * releases the page group reference from each member once all
 299  * members have called this function.
 300  */
 301 static void
 302 nfs_page_group_destroy(struct kref *kref)
 303 {
 304         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 305         struct nfs_page *tmp, *next;
 306
 307         /* subrequests must release the ref on the head request */
 308         if (req->wb_head != req)
 309                 nfs_release_request(req->wb_head);
 310
 311         if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN))
 312                 return;
 313
 314         tmp = req;
 315         do {
 316                 next = tmp->wb_this_page;
 317                 /* unlink and free */
 318                 tmp->wb_this_page = tmp;
 319                 tmp->wb_head = tmp;
 320                 nfs_free_request(tmp);
 321                 tmp = next;
 322         } while (tmp != req);
 323 }
 324
 325 /**
 326  * nfs_create_request - Create an NFS read/write request.
 327  * @ctx: open context to use
 328  * @page: page to write
 329  * @last: last nfs request created for this page group or NULL if head
 330  * @offset: starting offset within the page for the write
 331  * @count: number of bytes to read/write
 332  *
 333  * The page must be locked by the caller. This makes sure we never
 334  * create two different requests for the same page.
 335  * User should ensure it is safe to sleep in this function.
 336  */
 337 struct nfs_page *
 338 nfs_create_request(struct nfs_open_context *ctx, struct page *page,
 339                    struct nfs_page *last, unsigned int offset,
 340                    unsigned int count)
 341 {
 342         struct nfs_page         *req;
 343         struct nfs_lock_context *l_ctx;
 344
 345         if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
 346                 return ERR_PTR(-EBADF);
 347         /* try to allocate the request struct */
 348         req = nfs_page_alloc();
 349         if (req == NULL)
 350                 return ERR_PTR(-ENOMEM);
 351
 352         /* get lock context early so we can deal with alloc failures */
 353         l_ctx = nfs_get_lock_context(ctx);
 354         if (IS_ERR(l_ctx)) {
 355                 nfs_page_free(req);
 356                 return ERR_CAST(l_ctx);
 357         }
 358         req->wb_lock_context = l_ctx;
 359         nfs_iocounter_inc(&l_ctx->io_count);
 360
 361         /* Initialize the request struct. Initially, we assume a
 362          * long write-back delay. This will be adjusted in
 363          * update_nfs_request below if the region is not locked. */
 364         req->wb_page    = page;
 365         req->wb_index   = page_file_index(page);
 366         page_cache_get(page);
 367         req->wb_offset  = offset;
 368         req->wb_pgbase  = offset;
 369         req->wb_bytes   = count;
 370         req->wb_context = get_nfs_open_context(ctx);
 371         kref_init(&req->wb_kref);
 372         nfs_page_group_init(req, last);
 373         return req;
 374 }
 375
 376 /**
 377  * nfs_unlock_request - Unlock request and wake up sleepers.
 378  * @req:
 379  */
 380 void nfs_unlock_request(struct nfs_page *req)
 381 {
 382         if (!NFS_WBACK_BUSY(req)) {
 383                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 384                 BUG();
 385         }
 386         smp_mb__before_atomic();
 387         clear_bit(PG_BUSY, &req->wb_flags);
 388         smp_mb__after_atomic();
 389         wake_up_bit(&req->wb_flags, PG_BUSY);
 390 }
 391
 392 /**
 393  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
 394  * @req:
 395  */
 396 void nfs_unlock_and_release_request(struct nfs_page *req)
 397 {
 398         nfs_unlock_request(req);
 399         nfs_release_request(req);
 400 }
 401
 402 /*
 403  * nfs_clear_request - Free up all resources allocated to the request
 404  * @req:
 405  *
 406  * Release page and open context resources associated with a read/write
 407  * request after it has completed.
 408  */
 409 static void nfs_clear_request(struct nfs_page *req)
 410 {
 411         struct page *page = req->wb_page;
 412         struct nfs_open_context *ctx = req->wb_context;
 413         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 414
 415         if (page != NULL) {
 416                 page_cache_release(page);
 417                 req->wb_page = NULL;
 418         }
 419         if (l_ctx != NULL) {
 420                 nfs_iocounter_dec(&l_ctx->io_count);
 421                 nfs_put_lock_context(l_ctx);
 422                 req->wb_lock_context = NULL;
 423         }
 424         if (ctx != NULL) {
 425                 put_nfs_open_context(ctx);
 426                 req->wb_context = NULL;
 427         }
 428 }
 429
 430 /**
 431  * nfs_release_request - Release the count on an NFS read/write request
 432  * @req: request to release
 433  *
 434  * Note: Should never be called with the spinlock held!
 435  */
 436 void nfs_free_request(struct nfs_page *req)
 437 {
 438         WARN_ON_ONCE(req->wb_this_page != req);
 439
 440         /* extra debug: make sure no sync bits are still set */
 441         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
 442         WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
 443         WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
 444         WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
 445         WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
 446
 447         /* Release struct file and open context */
 448         nfs_clear_request(req);
 449         nfs_page_free(req);
 450 }
 451
 452 void nfs_release_request(struct nfs_page *req)
 453 {
 454         kref_put(&req->wb_kref, nfs_page_group_destroy);
 455 }
 456
 457 /**
 458  * nfs_wait_on_request - Wait for a request to complete.
 459  * @req: request to wait upon.
 460  *
 461  * Interruptible by fatal signals only.
 462  * The user is responsible for holding a count on the request.
 463  */
 464 int
 465 nfs_wait_on_request(struct nfs_page *req)
 466 {
 467         return wait_on_bit_io(&req->wb_flags, PG_BUSY,
 468                               TASK_UNINTERRUPTIBLE);
 469 }
 470
 471 /*
 472  * nfs_generic_pg_test - determine if requests can be coalesced
 473  * @desc: pointer to descriptor
 474  * @prev: previous request in desc, or NULL
 475  * @req: this request
 476  *
 477  * Returns zero if @req can be coalesced into @desc, otherwise it returns
 478  * the size of the request.
 479  */
 480 size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc,
 481                            struct nfs_page *prev, struct nfs_page *req)
 482 {
 483         if (desc->pg_count > desc->pg_bsize) {
 484                 /* should never happen */
 485                 WARN_ON_ONCE(1);
 486                 return 0;
 487         }
 488
 489         /*
 490          * Limit the request size so that we can still allocate a page array
 491          * for it without upsetting the slab allocator.
 492          */
 493         if (((desc->pg_count + req->wb_bytes) >> PAGE_SHIFT) *
 494                         sizeof(struct page) > PAGE_SIZE)
 495                 return 0;
 496
 497         return min(desc->pg_bsize - desc->pg_count, (size_t)req->wb_bytes);
 498 }
 499 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
 500
 501 struct nfs_pgio_header *nfs_pgio_header_alloc(const struct nfs_rw_ops *ops)
 502 {
 503         struct nfs_pgio_header *hdr = ops->rw_alloc_header();
 504
 505         if (hdr) {
 506                 INIT_LIST_HEAD(&hdr->pages);
 507                 spin_lock_init(&hdr->lock);
 508                 hdr->rw_ops = ops;
 509         }
 510         return hdr;
 511 }
 512 EXPORT_SYMBOL_GPL(nfs_pgio_header_alloc);
 513
 514 /*
 515  * nfs_pgio_header_free - Free a read or write header
 516  * @hdr: The header to free
 517  */
 518 void nfs_pgio_header_free(struct nfs_pgio_header *hdr)
 519 {
 520         hdr->rw_ops->rw_free_header(hdr);
 521 }
 522 EXPORT_SYMBOL_GPL(nfs_pgio_header_free);
 523
 524 /**
 525  * nfs_pgio_data_destroy - make @hdr suitable for reuse
 526  *
 527  * Frees memory and releases refs from nfs_generic_pgio, so that it may
 528  * be called again.
 529  *
 530  * @hdr: A header that has had nfs_generic_pgio called
 531  */
 532 void nfs_pgio_data_destroy(struct nfs_pgio_header *hdr)
 533 {
 534         if (hdr->args.context)
 535                 put_nfs_open_context(hdr->args.context);
 536         if (hdr->page_array.pagevec != hdr->page_array.page_array)
 537                 kfree(hdr->page_array.pagevec);
 538 }
 539 EXPORT_SYMBOL_GPL(nfs_pgio_data_destroy);
 540
 541 /**
 542  * nfs_pgio_rpcsetup - Set up arguments for a pageio call
 543  * @hdr: The pageio hdr
 544  * @count: Number of bytes to read
 545  * @offset: Initial offset
 546  * @how: How to commit data (writes only)
 547  * @cinfo: Commit information for the call (writes only)
 548  */
 549 static void nfs_pgio_rpcsetup(struct nfs_pgio_header *hdr,
 550                               unsigned int count, unsigned int offset,
 551                               int how, struct nfs_commit_info *cinfo)
 552 {
 553         struct nfs_page *req = hdr->req;
 554
 555         /* Set up the RPC argument and reply structs
 556          * NB: take care not to mess about with hdr->commit et al. */
 557
 558         hdr->args.fh     = NFS_FH(hdr->inode);
 559         hdr->args.offset = req_offset(req) + offset;
 560         /* pnfs_set_layoutcommit needs this */
 561         hdr->mds_offset = hdr->args.offset;
 562         hdr->args.pgbase = req->wb_pgbase + offset;
 563         hdr->args.pages  = hdr->page_array.pagevec;
 564         hdr->args.count  = count;
 565         hdr->args.context = get_nfs_open_context(req->wb_context);
 566         hdr->args.lock_context = req->wb_lock_context;
 567         hdr->args.stable  = NFS_UNSTABLE;
 568         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 569         case 0:
 570                 break;
 571         case FLUSH_COND_STABLE:
 572                 if (nfs_reqs_to_commit(cinfo))
 573                         break;
 574         default:
 575                 hdr->args.stable = NFS_FILE_SYNC;
 576         }
 577
 578         hdr->res.fattr   = &hdr->fattr;
 579         hdr->res.count   = count;
 580         hdr->res.eof     = 0;
 581         hdr->res.verf    = &hdr->verf;
 582         nfs_fattr_init(&hdr->fattr);
 583 }
 584
 585 /**
 586  * nfs_pgio_prepare - Prepare pageio hdr to go over the wire
 587  * @task: The current task
 588  * @calldata: pageio header to prepare
 589  */
 590 static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
 591 {
 592         struct nfs_pgio_header *hdr = calldata;
 593         int err;
 594         err = NFS_PROTO(hdr->inode)->pgio_rpc_prepare(task, hdr);
 595         if (err)
 596                 rpc_exit(task, err);
 597 }
 598
 599 int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_header *hdr,
 600                       const struct rpc_call_ops *call_ops, int how, int flags)
 601 {
 602         struct rpc_task *task;
 603         struct rpc_message msg = {
 604                 .rpc_argp = &hdr->args,
 605                 .rpc_resp = &hdr->res,
 606                 .rpc_cred = hdr->cred,
 607         };
 608         struct rpc_task_setup task_setup_data = {
 609                 .rpc_client = clnt,
 610                 .task = &hdr->task,
 611                 .rpc_message = &msg,
 612                 .callback_ops = call_ops,
 613                 .callback_data = hdr,
 614                 .workqueue = nfsiod_workqueue,
 615                 .flags = RPC_TASK_ASYNC | flags,
 616         };
 617         int ret = 0;
 618
 619         hdr->rw_ops->rw_initiate(hdr, &msg, &task_setup_data, how);
 620
 621         dprintk("NFS: %5u initiated pgio call "
 622                 "(req %s/%llu, %u bytes @ offset %llu)\n",
 623                 hdr->task.tk_pid,
 624                 hdr->inode->i_sb->s_id,
 625                 (unsigned long long)NFS_FILEID(hdr->inode),
 626                 hdr->args.count,
 627                 (unsigned long long)hdr->args.offset);
 628
 629         task = rpc_run_task(&task_setup_data);
 630         if (IS_ERR(task)) {
 631                 ret = PTR_ERR(task);
 632                 goto out;
 633         }
 634         if (how & FLUSH_SYNC) {
 635                 ret = rpc_wait_for_completion_task(task);
 636                 if (ret == 0)
 637                         ret = task->tk_status;
 638         }
 639         rpc_put_task(task);
 640 out:
 641         return ret;
 642 }
 643 EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
 644
 645 /**
 646  * nfs_pgio_error - Clean up from a pageio error
 647  * @desc: IO descriptor
 648  * @hdr: pageio header
 649  */
 650 static int nfs_pgio_error(struct nfs_pageio_descriptor *desc,
 651                           struct nfs_pgio_header *hdr)
 652 {
 653         set_bit(NFS_IOHDR_REDO, &hdr->flags);
 654         nfs_pgio_data_destroy(hdr);
 655         hdr->completion_ops->completion(hdr);
 656         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 657         return -ENOMEM;
 658 }
 659
 660 /**
 661  * nfs_pgio_release - Release pageio data
 662  * @calldata: The pageio header to release
 663  */
 664 static void nfs_pgio_release(void *calldata)
 665 {
 666         struct nfs_pgio_header *hdr = calldata;
 667         if (hdr->rw_ops->rw_release)
 668                 hdr->rw_ops->rw_release(hdr);
 669         nfs_pgio_data_destroy(hdr);
 670         hdr->completion_ops->completion(hdr);
 671 }
 672
 673 /**
 674  * nfs_pageio_init - initialise a page io descriptor
 675  * @desc: pointer to descriptor
 676  * @inode: pointer to inode
 677  * @doio: pointer to io function
 678  * @bsize: io block size
 679  * @io_flags: extra parameters for the io function
 680  */
 681 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 682                      struct inode *inode,
 683                      const struct nfs_pageio_ops *pg_ops,
 684                      const struct nfs_pgio_completion_ops *compl_ops,
 685                      const struct nfs_rw_ops *rw_ops,
 686                      size_t bsize,
 687                      int io_flags)
 688 {
 689         INIT_LIST_HEAD(&desc->pg_list);
 690         desc->pg_bytes_written = 0;
 691         desc->pg_count = 0;
 692         desc->pg_bsize = bsize;
 693         desc->pg_base = 0;
 694         desc->pg_moreio = 0;
 695         desc->pg_recoalesce = 0;
 696         desc->pg_inode = inode;
 697         desc->pg_ops = pg_ops;
 698         desc->pg_completion_ops = compl_ops;
 699         desc->pg_rw_ops = rw_ops;
 700         desc->pg_ioflags = io_flags;
 701         desc->pg_error = 0;
 702         desc->pg_lseg = NULL;
 703         desc->pg_dreq = NULL;
 704         desc->pg_layout_private = NULL;
 705 }
 706 EXPORT_SYMBOL_GPL(nfs_pageio_init);
 707
 708 /**
 709  * nfs_pgio_result - Basic pageio error handling
 710  * @task: The task that ran
 711  * @calldata: Pageio header to check
 712  */
 713 static void nfs_pgio_result(struct rpc_task *task, void *calldata)
 714 {
 715         struct nfs_pgio_header *hdr = calldata;
 716         struct inode *inode = hdr->inode;
 717
 718         dprintk("NFS: %s: %5u, (status %d)\n", __func__,
 719                 task->tk_pid, task->tk_status);
 720
 721         if (hdr->rw_ops->rw_done(task, hdr, inode) != 0)
 722                 return;
 723         if (task->tk_status < 0)
 724                 nfs_set_pgio_error(hdr, task->tk_status, hdr->args.offset);
 725         else
 726                 hdr->rw_ops->rw_result(task, hdr);
 727 }
 728
 729 /*
 730  * Create an RPC task for the given read or write request and kick it.
 731  * The page must have been locked by the caller.
 732  *
 733  * It may happen that the page we're passed is not marked dirty.
 734  * This is the case if nfs_updatepage detects a conflicting request
 735  * that has been written but not committed.
 736  */
 737 int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
 738                      struct nfs_pgio_header *hdr)
 739 {
 740         struct nfs_page         *req;
 741         struct page             **pages,
 742                                 *last_page;
 743         struct list_head *head = &desc->pg_list;
 744         struct nfs_commit_info cinfo;
 745         unsigned int pagecount, pageused;
 746
 747         pagecount = nfs_page_array_len(desc->pg_base, desc->pg_count);
 748         if (!nfs_pgarray_set(&hdr->page_array, pagecount))
 749                 return nfs_pgio_error(desc, hdr);
 750
 751         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
 752         pages = hdr->page_array.pagevec;
 753         last_page = NULL;
 754         pageused = 0;
 755         while (!list_empty(head)) {
 756                 req = nfs_list_entry(head->next);
 757                 nfs_list_remove_request(req);
 758                 nfs_list_add_request(req, &hdr->pages);
 759
 760                 if (!last_page || last_page != req->wb_page) {
 761                         pageused++;
 762                         if (pageused > pagecount)
 763                                 break;
 764                         *pages++ = last_page = req->wb_page;
 765                 }
 766         }
 767         if (WARN_ON_ONCE(pageused != pagecount))
 768                 return nfs_pgio_error(desc, hdr);
 769
 770         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 771             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
 772                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 773
 774         /* Set up the argument struct */
 775         nfs_pgio_rpcsetup(hdr, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
 776         desc->pg_rpc_callops = &nfs_pgio_common_ops;
 777         return 0;
 778 }
 779 EXPORT_SYMBOL_GPL(nfs_generic_pgio);
 780
 781 static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
 782 {
 783         struct nfs_pgio_header *hdr;
 784         int ret;
 785
 786         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
 787         if (!hdr) {
 788                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 789                 return -ENOMEM;
 790         }
 791         nfs_pgheader_init(desc, hdr, nfs_pgio_header_free);
 792         ret = nfs_generic_pgio(desc, hdr);
 793         if (ret == 0)
 794                 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
 795                                         hdr, desc->pg_rpc_callops,
 796                                         desc->pg_ioflags, 0);
 797         return ret;
 798 }
 799
 800 static bool nfs_match_open_context(const struct nfs_open_context *ctx1,
 801                 const struct nfs_open_context *ctx2)
 802 {
 803         return ctx1->cred == ctx2->cred && ctx1->state == ctx2->state;
 804 }
 805
 806 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
 807                 const struct nfs_lock_context *l2)
 808 {
 809         return l1->lockowner.l_owner == l2->lockowner.l_owner
 810                 && l1->lockowner.l_pid == l2->lockowner.l_pid;
 811 }
 812
 813 /**
 814  * nfs_can_coalesce_requests - test two requests for compatibility
 815  * @prev: pointer to nfs_page
 816  * @req: pointer to nfs_page
 817  *
 818  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
 819  * page data area they describe is contiguous, and that their RPC
 820  * credentials, NFSv4 open state, and lockowners are the same.
 821  *
 822  * Return 'true' if this is the case, else return 'false'.
 823  */
 824 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
 825                                       struct nfs_page *req,
 826                                       struct nfs_pageio_descriptor *pgio)
 827 {
 828         size_t size;
 829
 830         if (prev) {
 831                 if (!nfs_match_open_context(req->wb_context, prev->wb_context))
 832                         return false;
 833                 if (req->wb_context->dentry->d_inode->i_flock != NULL &&
 834                     !nfs_match_lock_context(req->wb_lock_context,
 835                                             prev->wb_lock_context))
 836                         return false;
 837                 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
 838                         return false;
 839                 if (req->wb_page == prev->wb_page) {
 840                         if (req->wb_pgbase != prev->wb_pgbase + prev->wb_bytes)
 841                                 return false;
 842                 } else {
 843                         if (req->wb_pgbase != 0 ||
 844                             prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
 845                                 return false;
 846                 }
 847         }
 848         size = pgio->pg_ops->pg_test(pgio, prev, req);
 849         WARN_ON_ONCE(size > req->wb_bytes);
 850         if (size && size < req->wb_bytes)
 851                 req->wb_bytes = size;
 852         return size > 0;
 853 }
 854
 855 /**
 856  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
 857  * @desc: destination io descriptor
 858  * @req: request
 859  *
 860  * Returns true if the request 'req' was successfully coalesced into the
 861  * existing list of pages 'desc'.
 862  */
 863 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
 864                                      struct nfs_page *req)
 865 {
 866         struct nfs_page *prev = NULL;
 867         if (desc->pg_count != 0) {
 868                 prev = nfs_list_entry(desc->pg_list.prev);
 869         } else {
 870                 if (desc->pg_ops->pg_init)
 871                         desc->pg_ops->pg_init(desc, req);
 872                 desc->pg_base = req->wb_pgbase;
 873         }
 874         if (!nfs_can_coalesce_requests(prev, req, desc))
 875                 return 0;
 876         nfs_list_remove_request(req);
 877         nfs_list_add_request(req, &desc->pg_list);
 878         desc->pg_count += req->wb_bytes;
 879         return 1;
 880 }
 881
 882 /*
 883  * Helper for nfs_pageio_add_request and nfs_pageio_complete
 884  */
 885 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
 886 {
 887         if (!list_empty(&desc->pg_list)) {
 888                 int error = desc->pg_ops->pg_doio(desc);
 889                 if (error < 0)
 890                         desc->pg_error = error;
 891                 else
 892                         desc->pg_bytes_written += desc->pg_count;
 893         }
 894         if (list_empty(&desc->pg_list)) {
 895                 desc->pg_count = 0;
 896                 desc->pg_base = 0;
 897         }
 898 }
 899
 900 /**
 901  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
 902  * @desc: destination io descriptor
 903  * @req: request
 904  *
 905  * This may split a request into subrequests which are all part of the
 906  * same page group.
 907  *
 908  * Returns true if the request 'req' was successfully coalesced into the
 909  * existing list of pages 'desc'.
 910  */
 911 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 912                            struct nfs_page *req)
 913 {
 914         struct nfs_page *subreq;
 915         unsigned int bytes_left = 0;
 916         unsigned int offset, pgbase;
 917
 918         nfs_page_group_lock(req, false);
 919
 920         subreq = req;
 921         bytes_left = subreq->wb_bytes;
 922         offset = subreq->wb_offset;
 923         pgbase = subreq->wb_pgbase;
 924
 925         do {
 926                 if (!nfs_pageio_do_add_request(desc, subreq)) {
 927                         /* make sure pg_test call(s) did nothing */
 928                         WARN_ON_ONCE(subreq->wb_bytes != bytes_left);
 929                         WARN_ON_ONCE(subreq->wb_offset != offset);
 930                         WARN_ON_ONCE(subreq->wb_pgbase != pgbase);
 931
 932                         nfs_page_group_unlock(req);
 933                         desc->pg_moreio = 1;
 934                         nfs_pageio_doio(desc);
 935                         if (desc->pg_error < 0)
 936                                 return 0;
 937                         if (desc->pg_recoalesce)
 938                                 return 0;
 939                         /* retry add_request for this subreq */
 940                         nfs_page_group_lock(req, false);
 941                         continue;
 942                 }
 943
 944                 /* check for buggy pg_test call(s) */
 945                 WARN_ON_ONCE(subreq->wb_bytes + subreq->wb_pgbase > PAGE_SIZE);
 946                 WARN_ON_ONCE(subreq->wb_bytes > bytes_left);
 947                 WARN_ON_ONCE(subreq->wb_bytes == 0);
 948
 949                 bytes_left -= subreq->wb_bytes;
 950                 offset += subreq->wb_bytes;
 951                 pgbase += subreq->wb_bytes;
 952
 953                 if (bytes_left) {
 954                         subreq = nfs_create_request(req->wb_context,
 955                                         req->wb_page,
 956                                         subreq, pgbase, bytes_left);
 957                         if (IS_ERR(subreq))
 958                                 goto err_ptr;
 959                         nfs_lock_request(subreq);
 960                         subreq->wb_offset  = offset;
 961                         subreq->wb_index = req->wb_index;
 962                 }
 963         } while (bytes_left > 0);
 964
 965         nfs_page_group_unlock(req);
 966         return 1;
 967 err_ptr:
 968         desc->pg_error = PTR_ERR(subreq);
 969         nfs_page_group_unlock(req);
 970         return 0;
 971 }
 972
 973 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
 974 {
 975         LIST_HEAD(head);
 976
 977         do {
 978                 list_splice_init(&desc->pg_list, &head);
 979                 desc->pg_bytes_written -= desc->pg_count;
 980                 desc->pg_count = 0;
 981                 desc->pg_base = 0;
 982                 desc->pg_recoalesce = 0;
 983                 desc->pg_moreio = 0;
 984
 985                 while (!list_empty(&head)) {
 986                         struct nfs_page *req;
 987
 988                         req = list_first_entry(&head, struct nfs_page, wb_list);
 989                         nfs_list_remove_request(req);
 990                         if (__nfs_pageio_add_request(desc, req))
 991                                 continue;
 992                         if (desc->pg_error < 0)
 993                                 return 0;
 994                         break;
 995                 }
 996         } while (desc->pg_recoalesce);
 997         return 1;
 998 }
 999
1000 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1001                 struct nfs_page *req)
1002 {
1003         int ret;
1004
1005         do {
1006                 ret = __nfs_pageio_add_request(desc, req);
1007                 if (ret)
1008                         break;
1009                 if (desc->pg_error < 0)
1010                         break;
1011                 ret = nfs_do_recoalesce(desc);
1012         } while (ret);
1013         return ret;
1014 }
1015
1016 /*
1017  * nfs_pageio_resend - Transfer requests to new descriptor and resend
1018  * @hdr - the pgio header to move request from
1019  * @desc - the pageio descriptor to add requests to
1020  *
1021  * Try to move each request (nfs_page) from @hdr to @desc then attempt
1022  * to send them.
1023  *
1024  * Returns 0 on success and < 0 on error.
1025  */
1026 int nfs_pageio_resend(struct nfs_pageio_descriptor *desc,
1027                       struct nfs_pgio_header *hdr)
1028 {
1029         LIST_HEAD(failed);
1030
1031         desc->pg_dreq = hdr->dreq;
1032         while (!list_empty(&hdr->pages)) {
1033                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1034
1035                 nfs_list_remove_request(req);
1036                 if (!nfs_pageio_add_request(desc, req))
1037                         nfs_list_add_request(req, &failed);
1038         }
1039         nfs_pageio_complete(desc);
1040         if (!list_empty(&failed)) {
1041                 list_move(&failed, &hdr->pages);
1042                 return -EIO;
1043         }
1044         return 0;
1045 }
1046 EXPORT_SYMBOL_GPL(nfs_pageio_resend);
1047
1048 /**
1049  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
1050  * @desc: pointer to io descriptor
1051  */
1052 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1053 {
1054         for (;;) {
1055                 nfs_pageio_doio(desc);
1056                 if (!desc->pg_recoalesce)
1057                         break;
1058                 if (!nfs_do_recoalesce(desc))
1059                         break;
1060         }
1061 }
1062
1063 /**
1064  * nfs_pageio_cond_complete - Conditional I/O completion
1065  * @desc: pointer to io descriptor
1066  * @index: page index
1067  *
1068  * It is important to ensure that processes don't try to take locks
1069  * on non-contiguous ranges of pages as that might deadlock. This
1070  * function should be called before attempting to wait on a locked
1071  * nfs_page. It will complete the I/O if the page index 'index'
1072  * is not contiguous with the existing list of pages in 'desc'.
1073  */
1074 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1075 {
1076         if (!list_empty(&desc->pg_list)) {
1077                 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
1078                 if (index != prev->wb_index + 1)
1079                         nfs_pageio_complete(desc);
1080         }
1081 }
1082
1083 int __init nfs_init_nfspagecache(void)
1084 {
1085         nfs_page_cachep = kmem_cache_create("nfs_page",
1086                                             sizeof(struct nfs_page),
1087                                             0, SLAB_HWCACHE_ALIGN,
1088                                             NULL);
1089         if (nfs_page_cachep == NULL)
1090                 return -ENOMEM;
1091
1092         return 0;
1093 }
1094
1095 void nfs_destroy_nfspagecache(void)
1096 {
1097         kmem_cache_destroy(nfs_page_cachep);
1098 }
1099
1100 static const struct rpc_call_ops nfs_pgio_common_ops = {
1101         .rpc_call_prepare = nfs_pgio_prepare,
1102         .rpc_call_done = nfs_pgio_result,
1103         .rpc_release = nfs_pgio_release,
1104 };
1105
1106 const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1107         .pg_test = nfs_generic_pg_test,
1108         .pg_doio = nfs_generic_pg_pgios,
1109 };