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