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 static struct kmem_cache *nfs_page_cachep;
  28
  29 static bool nfs_pgarray_set(struct nfs_page_array *p, unsigned int pagecount)
  30 {
  31         p->npages = pagecount;
  32         if (pagecount <= ARRAY_SIZE(p->page_array))
  33                 p->pagevec = p->page_array;
  34         else {
  35                 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL);
  36                 if (!p->pagevec)
  37                         p->npages = 0;
  38         }
  39         return p->pagevec != NULL;
  40 }
  41
  42 void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
  43                        struct nfs_pgio_header *hdr,
  44                        void (*release)(struct nfs_pgio_header *hdr))
  45 {
  46         hdr->req = nfs_list_entry(desc->pg_list.next);
  47         hdr->inode = desc->pg_inode;
  48         hdr->cred = hdr->req->wb_context->cred;
  49         hdr->io_start = req_offset(hdr->req);
  50         hdr->good_bytes = desc->pg_count;
  51         hdr->dreq = desc->pg_dreq;
  52         hdr->layout_private = desc->pg_layout_private;
  53         hdr->release = release;
  54         hdr->completion_ops = desc->pg_completion_ops;
  55         if (hdr->completion_ops->init_hdr)
  56                 hdr->completion_ops->init_hdr(hdr);
  57 }
  58 EXPORT_SYMBOL_GPL(nfs_pgheader_init);
  59
  60 void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
  61 {
  62         spin_lock(&hdr->lock);
  63         if (pos < hdr->io_start + hdr->good_bytes) {
  64                 set_bit(NFS_IOHDR_ERROR, &hdr->flags);
  65                 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
  66                 hdr->good_bytes = pos - hdr->io_start;
  67                 hdr->error = error;
  68         }
  69         spin_unlock(&hdr->lock);
  70 }
  71
  72 static inline struct nfs_page *
  73 nfs_page_alloc(void)
  74 {
  75         struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_NOIO);
  76         if (p)
  77                 INIT_LIST_HEAD(&p->wb_list);
  78         return p;
  79 }
  80
  81 static inline void
  82 nfs_page_free(struct nfs_page *p)
  83 {
  84         kmem_cache_free(nfs_page_cachep, p);
  85 }
  86
  87 static void
  88 nfs_iocounter_inc(struct nfs_io_counter *c)
  89 {
  90         atomic_inc(&c->io_count);
  91 }
  92
  93 static void
  94 nfs_iocounter_dec(struct nfs_io_counter *c)
  95 {
  96         if (atomic_dec_and_test(&c->io_count)) {
  97                 clear_bit(NFS_IO_INPROGRESS, &c->flags);
  98                 smp_mb__after_clear_bit();
  99                 wake_up_bit(&c->flags, NFS_IO_INPROGRESS);
 100         }
 101 }
 102
 103 static int
 104 __nfs_iocounter_wait(struct nfs_io_counter *c)
 105 {
 106         wait_queue_head_t *wq = bit_waitqueue(&c->flags, NFS_IO_INPROGRESS);
 107         DEFINE_WAIT_BIT(q, &c->flags, NFS_IO_INPROGRESS);
 108         int ret = 0;
 109
 110         do {
 111                 prepare_to_wait(wq, &q.wait, TASK_KILLABLE);
 112                 set_bit(NFS_IO_INPROGRESS, &c->flags);
 113                 if (atomic_read(&c->io_count) == 0)
 114                         break;
 115                 ret = nfs_wait_bit_killable(&c->flags);
 116         } while (atomic_read(&c->io_count) != 0);
 117         finish_wait(wq, &q.wait);
 118         return ret;
 119 }
 120
 121 /**
 122  * nfs_iocounter_wait - wait for i/o to complete
 123  * @c: nfs_io_counter to use
 124  *
 125  * returns -ERESTARTSYS if interrupted by a fatal signal.
 126  * Otherwise returns 0 once the io_count hits 0.
 127  */
 128 int
 129 nfs_iocounter_wait(struct nfs_io_counter *c)
 130 {
 131         if (atomic_read(&c->io_count) == 0)
 132                 return 0;
 133         return __nfs_iocounter_wait(c);
 134 }
 135
 136 /**
 137  * nfs_create_request - Create an NFS read/write request.
 138  * @ctx: open context to use
 139  * @inode: inode to which the request is attached
 140  * @page: page to write
 141  * @offset: starting offset within the page for the write
 142  * @count: number of bytes to read/write
 143  *
 144  * The page must be locked by the caller. This makes sure we never
 145  * create two different requests for the same page.
 146  * User should ensure it is safe to sleep in this function.
 147  */
 148 struct nfs_page *
 149 nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
 150                    struct page *page,
 151                    unsigned int offset, unsigned int count)
 152 {
 153         struct nfs_page         *req;
 154         struct nfs_lock_context *l_ctx;
 155
 156         if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
 157                 return ERR_PTR(-EBADF);
 158         /* try to allocate the request struct */
 159         req = nfs_page_alloc();
 160         if (req == NULL)
 161                 return ERR_PTR(-ENOMEM);
 162
 163         /* get lock context early so we can deal with alloc failures */
 164         l_ctx = nfs_get_lock_context(ctx);
 165         if (IS_ERR(l_ctx)) {
 166                 nfs_page_free(req);
 167                 return ERR_CAST(l_ctx);
 168         }
 169         req->wb_lock_context = l_ctx;
 170         nfs_iocounter_inc(&l_ctx->io_count);
 171
 172         /* Initialize the request struct. Initially, we assume a
 173          * long write-back delay. This will be adjusted in
 174          * update_nfs_request below if the region is not locked. */
 175         req->wb_page    = page;
 176         req->wb_index   = page_file_index(page);
 177         page_cache_get(page);
 178         req->wb_offset  = offset;
 179         req->wb_pgbase  = offset;
 180         req->wb_bytes   = count;
 181         req->wb_context = get_nfs_open_context(ctx);
 182         kref_init(&req->wb_kref);
 183         return req;
 184 }
 185
 186 /**
 187  * nfs_unlock_request - Unlock request and wake up sleepers.
 188  * @req:
 189  */
 190 void nfs_unlock_request(struct nfs_page *req)
 191 {
 192         if (!NFS_WBACK_BUSY(req)) {
 193                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 194                 BUG();
 195         }
 196         smp_mb__before_clear_bit();
 197         clear_bit(PG_BUSY, &req->wb_flags);
 198         smp_mb__after_clear_bit();
 199         wake_up_bit(&req->wb_flags, PG_BUSY);
 200 }
 201
 202 /**
 203  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
 204  * @req:
 205  */
 206 void nfs_unlock_and_release_request(struct nfs_page *req)
 207 {
 208         nfs_unlock_request(req);
 209         nfs_release_request(req);
 210 }
 211
 212 /*
 213  * nfs_clear_request - Free up all resources allocated to the request
 214  * @req:
 215  *
 216  * Release page and open context resources associated with a read/write
 217  * request after it has completed.
 218  */
 219 static void nfs_clear_request(struct nfs_page *req)
 220 {
 221         struct page *page = req->wb_page;
 222         struct nfs_open_context *ctx = req->wb_context;
 223         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 224
 225         if (page != NULL) {
 226                 page_cache_release(page);
 227                 req->wb_page = NULL;
 228         }
 229         if (l_ctx != NULL) {
 230                 nfs_iocounter_dec(&l_ctx->io_count);
 231                 nfs_put_lock_context(l_ctx);
 232                 req->wb_lock_context = NULL;
 233         }
 234         if (ctx != NULL) {
 235                 put_nfs_open_context(ctx);
 236                 req->wb_context = NULL;
 237         }
 238 }
 239
 240
 241 /**
 242  * nfs_release_request - Release the count on an NFS read/write request
 243  * @req: request to release
 244  *
 245  * Note: Should never be called with the spinlock held!
 246  */
 247 static void nfs_free_request(struct kref *kref)
 248 {
 249         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 250
 251         /* Release struct file and open context */
 252         nfs_clear_request(req);
 253         nfs_page_free(req);
 254 }
 255
 256 void nfs_release_request(struct nfs_page *req)
 257 {
 258         kref_put(&req->wb_kref, nfs_free_request);
 259 }
 260
 261 static int nfs_wait_bit_uninterruptible(void *word)
 262 {
 263         io_schedule();
 264         return 0;
 265 }
 266
 267 /**
 268  * nfs_wait_on_request - Wait for a request to complete.
 269  * @req: request to wait upon.
 270  *
 271  * Interruptible by fatal signals only.
 272  * The user is responsible for holding a count on the request.
 273  */
 274 int
 275 nfs_wait_on_request(struct nfs_page *req)
 276 {
 277         return wait_on_bit(&req->wb_flags, PG_BUSY,
 278                         nfs_wait_bit_uninterruptible,
 279                         TASK_UNINTERRUPTIBLE);
 280 }
 281
 282 bool nfs_generic_pg_test(struct nfs_pageio_descriptor *desc, struct nfs_page *prev, struct nfs_page *req)
 283 {
 284         /*
 285          * FIXME: ideally we should be able to coalesce all requests
 286          * that are not block boundary aligned, but currently this
 287          * is problematic for the case of bsize < PAGE_CACHE_SIZE,
 288          * since nfs_flush_multi and nfs_pagein_multi assume you
 289          * can have only one struct nfs_page.
 290          */
 291         if (desc->pg_bsize < PAGE_SIZE)
 292                 return 0;
 293
 294         return desc->pg_count + req->wb_bytes <= desc->pg_bsize;
 295 }
 296 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
 297
 298 static inline struct nfs_rw_header *NFS_RW_HEADER(struct nfs_pgio_header *hdr)
 299 {
 300         return container_of(hdr, struct nfs_rw_header, header);
 301 }
 302
 303 /**
 304  * nfs_pgio_data_alloc - Allocate pageio data
 305  * @hdr: The header making a request
 306  * @pagecount: Number of pages to create
 307  */
 308 struct nfs_pgio_data *nfs_pgio_data_alloc(struct nfs_pgio_header *hdr,
 309                                           unsigned int pagecount)
 310 {
 311         struct nfs_pgio_data *data, *prealloc;
 312
 313         prealloc = &NFS_RW_HEADER(hdr)->rpc_data;
 314         if (prealloc->header == NULL)
 315                 data = prealloc;
 316         else
 317                 data = kzalloc(sizeof(*data), GFP_KERNEL);
 318         if (!data)
 319                 goto out;
 320
 321         if (nfs_pgarray_set(&data->pages, pagecount)) {
 322                 data->header = hdr;
 323                 atomic_inc(&hdr->refcnt);
 324         } else {
 325                 if (data != prealloc)
 326                         kfree(data);
 327                 data = NULL;
 328         }
 329 out:
 330         return data;
 331 }
 332
 333 /**
 334  * nfs_pgio_data_release - Properly free pageio data
 335  * @data: The data to release
 336  */
 337 void nfs_pgio_data_release(struct nfs_pgio_data *data)
 338 {
 339         struct nfs_pgio_header *hdr = data->header;
 340         struct nfs_rw_header *pageio_header = NFS_RW_HEADER(hdr);
 341
 342         put_nfs_open_context(data->args.context);
 343         if (data->pages.pagevec != data->pages.page_array)
 344                 kfree(data->pages.pagevec);
 345         if (data == &pageio_header->rpc_data) {
 346                 data->header = NULL;
 347                 data = NULL;
 348         }
 349         if (atomic_dec_and_test(&hdr->refcnt))
 350                 hdr->completion_ops->completion(hdr);
 351         /* Note: we only free the rpc_task after callbacks are done.
 352          * See the comment in rpc_free_task() for why
 353          */
 354         kfree(data);
 355 }
 356 EXPORT_SYMBOL_GPL(nfs_pgio_data_release);
 357
 358 /**
 359  * nfs_pageio_init - initialise a page io descriptor
 360  * @desc: pointer to descriptor
 361  * @inode: pointer to inode
 362  * @doio: pointer to io function
 363  * @bsize: io block size
 364  * @io_flags: extra parameters for the io function
 365  */
 366 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 367                      struct inode *inode,
 368                      const struct nfs_pageio_ops *pg_ops,
 369                      const struct nfs_pgio_completion_ops *compl_ops,
 370                      size_t bsize,
 371                      int io_flags)
 372 {
 373         INIT_LIST_HEAD(&desc->pg_list);
 374         desc->pg_bytes_written = 0;
 375         desc->pg_count = 0;
 376         desc->pg_bsize = bsize;
 377         desc->pg_base = 0;
 378         desc->pg_moreio = 0;
 379         desc->pg_recoalesce = 0;
 380         desc->pg_inode = inode;
 381         desc->pg_ops = pg_ops;
 382         desc->pg_completion_ops = compl_ops;
 383         desc->pg_ioflags = io_flags;
 384         desc->pg_error = 0;
 385         desc->pg_lseg = NULL;
 386         desc->pg_dreq = NULL;
 387         desc->pg_layout_private = NULL;
 388 }
 389 EXPORT_SYMBOL_GPL(nfs_pageio_init);
 390
 391 static bool nfs_match_open_context(const struct nfs_open_context *ctx1,
 392                 const struct nfs_open_context *ctx2)
 393 {
 394         return ctx1->cred == ctx2->cred && ctx1->state == ctx2->state;
 395 }
 396
 397 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
 398                 const struct nfs_lock_context *l2)
 399 {
 400         return l1->lockowner.l_owner == l2->lockowner.l_owner
 401                 && l1->lockowner.l_pid == l2->lockowner.l_pid;
 402 }
 403
 404 /**
 405  * nfs_can_coalesce_requests - test two requests for compatibility
 406  * @prev: pointer to nfs_page
 407  * @req: pointer to nfs_page
 408  *
 409  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
 410  * page data area they describe is contiguous, and that their RPC
 411  * credentials, NFSv4 open state, and lockowners are the same.
 412  *
 413  * Return 'true' if this is the case, else return 'false'.
 414  */
 415 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
 416                                       struct nfs_page *req,
 417                                       struct nfs_pageio_descriptor *pgio)
 418 {
 419         if (!nfs_match_open_context(req->wb_context, prev->wb_context))
 420                 return false;
 421         if (req->wb_context->dentry->d_inode->i_flock != NULL &&
 422             !nfs_match_lock_context(req->wb_lock_context, prev->wb_lock_context))
 423                 return false;
 424         if (req->wb_pgbase != 0)
 425                 return false;
 426         if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
 427                 return false;
 428         if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
 429                 return false;
 430         return pgio->pg_ops->pg_test(pgio, prev, req);
 431 }
 432
 433 /**
 434  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
 435  * @desc: destination io descriptor
 436  * @req: request
 437  *
 438  * Returns true if the request 'req' was successfully coalesced into the
 439  * existing list of pages 'desc'.
 440  */
 441 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
 442                                      struct nfs_page *req)
 443 {
 444         if (desc->pg_count != 0) {
 445                 struct nfs_page *prev;
 446
 447                 prev = nfs_list_entry(desc->pg_list.prev);
 448                 if (!nfs_can_coalesce_requests(prev, req, desc))
 449                         return 0;
 450         } else {
 451                 if (desc->pg_ops->pg_init)
 452                         desc->pg_ops->pg_init(desc, req);
 453                 desc->pg_base = req->wb_pgbase;
 454         }
 455         nfs_list_remove_request(req);
 456         nfs_list_add_request(req, &desc->pg_list);
 457         desc->pg_count += req->wb_bytes;
 458         return 1;
 459 }
 460
 461 /*
 462  * Helper for nfs_pageio_add_request and nfs_pageio_complete
 463  */
 464 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
 465 {
 466         if (!list_empty(&desc->pg_list)) {
 467                 int error = desc->pg_ops->pg_doio(desc);
 468                 if (error < 0)
 469                         desc->pg_error = error;
 470                 else
 471                         desc->pg_bytes_written += desc->pg_count;
 472         }
 473         if (list_empty(&desc->pg_list)) {
 474                 desc->pg_count = 0;
 475                 desc->pg_base = 0;
 476         }
 477 }
 478
 479 /**
 480  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
 481  * @desc: destination io descriptor
 482  * @req: request
 483  *
 484  * Returns true if the request 'req' was successfully coalesced into the
 485  * existing list of pages 'desc'.
 486  */
 487 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 488                            struct nfs_page *req)
 489 {
 490         while (!nfs_pageio_do_add_request(desc, req)) {
 491                 desc->pg_moreio = 1;
 492                 nfs_pageio_doio(desc);
 493                 if (desc->pg_error < 0)
 494                         return 0;
 495                 desc->pg_moreio = 0;
 496                 if (desc->pg_recoalesce)
 497                         return 0;
 498         }
 499         return 1;
 500 }
 501
 502 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
 503 {
 504         LIST_HEAD(head);
 505
 506         do {
 507                 list_splice_init(&desc->pg_list, &head);
 508                 desc->pg_bytes_written -= desc->pg_count;
 509                 desc->pg_count = 0;
 510                 desc->pg_base = 0;
 511                 desc->pg_recoalesce = 0;
 512
 513                 while (!list_empty(&head)) {
 514                         struct nfs_page *req;
 515
 516                         req = list_first_entry(&head, struct nfs_page, wb_list);
 517                         nfs_list_remove_request(req);
 518                         if (__nfs_pageio_add_request(desc, req))
 519                                 continue;
 520                         if (desc->pg_error < 0)
 521                                 return 0;
 522                         break;
 523                 }
 524         } while (desc->pg_recoalesce);
 525         return 1;
 526 }
 527
 528 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 529                 struct nfs_page *req)
 530 {
 531         int ret;
 532
 533         do {
 534                 ret = __nfs_pageio_add_request(desc, req);
 535                 if (ret)
 536                         break;
 537                 if (desc->pg_error < 0)
 538                         break;
 539                 ret = nfs_do_recoalesce(desc);
 540         } while (ret);
 541         return ret;
 542 }
 543 EXPORT_SYMBOL_GPL(nfs_pageio_add_request);
 544
 545 /**
 546  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
 547  * @desc: pointer to io descriptor
 548  */
 549 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
 550 {
 551         for (;;) {
 552                 nfs_pageio_doio(desc);
 553                 if (!desc->pg_recoalesce)
 554                         break;
 555                 if (!nfs_do_recoalesce(desc))
 556                         break;
 557         }
 558 }
 559 EXPORT_SYMBOL_GPL(nfs_pageio_complete);
 560
 561 /**
 562  * nfs_pageio_cond_complete - Conditional I/O completion
 563  * @desc: pointer to io descriptor
 564  * @index: page index
 565  *
 566  * It is important to ensure that processes don't try to take locks
 567  * on non-contiguous ranges of pages as that might deadlock. This
 568  * function should be called before attempting to wait on a locked
 569  * nfs_page. It will complete the I/O if the page index 'index'
 570  * is not contiguous with the existing list of pages in 'desc'.
 571  */
 572 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
 573 {
 574         if (!list_empty(&desc->pg_list)) {
 575                 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
 576                 if (index != prev->wb_index + 1)
 577                         nfs_pageio_complete(desc);
 578         }
 579 }
 580
 581 int __init nfs_init_nfspagecache(void)
 582 {
 583         nfs_page_cachep = kmem_cache_create("nfs_page",
 584                                             sizeof(struct nfs_page),
 585                                             0, SLAB_HWCACHE_ALIGN,
 586                                             NULL);
 587         if (nfs_page_cachep == NULL)
 588                 return -ENOMEM;
 589
 590         return 0;
 591 }
 592
 593 void nfs_destroy_nfspagecache(void)
 594 {
 595         kmem_cache_destroy(nfs_page_cachep);
 596 }
 597