fs/nfs/read.c

   1 /*
   2  * linux/fs/nfs/read.c
   3  *
   4  * Block I/O for NFS
   5  *
   6  * Partial copy of Linus' read cache modifications to fs/nfs/file.c
   7  * modified for async RPC by okir@monad.swb.de
   8  */
   9
  10 #include <linux/time.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/fcntl.h>
  14 #include <linux/stat.h>
  15 #include <linux/mm.h>
  16 #include <linux/slab.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/sunrpc/clnt.h>
  19 #include <linux/nfs_fs.h>
  20 #include <linux/nfs_page.h>
  21 #include <linux/module.h>
  22
  23 #include "nfs4_fs.h"
  24 #include "internal.h"
  25 #include "iostat.h"
  26 #include "fscache.h"
  27 #include "pnfs.h"
  28
  29 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  30
  31 static const struct nfs_pageio_ops nfs_pageio_read_ops;
  32 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
  33 static const struct nfs_rw_ops nfs_rw_read_ops;
  34
  35 static struct kmem_cache *nfs_rdata_cachep;
  36
  37 static struct nfs_rw_header *nfs_readhdr_alloc(void)
  38 {
  39         return kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
  40 }
  41
  42 static void nfs_readhdr_free(struct nfs_rw_header *rhdr)
  43 {
  44         kmem_cache_free(nfs_rdata_cachep, rhdr);
  45 }
  46
  47 static
  48 int nfs_return_empty_page(struct page *page)
  49 {
  50         zero_user(page, 0, PAGE_CACHE_SIZE);
  51         SetPageUptodate(page);
  52         unlock_page(page);
  53         return 0;
  54 }
  55
  56 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
  57                               struct inode *inode, bool force_mds,
  58                               const struct nfs_pgio_completion_ops *compl_ops)
  59 {
  60         struct nfs_server *server = NFS_SERVER(inode);
  61         const struct nfs_pageio_ops *pg_ops = &nfs_pageio_read_ops;
  62
  63 #ifdef CONFIG_NFS_V4_1
  64         if (server->pnfs_curr_ld && !force_mds)
  65                 pg_ops = server->pnfs_curr_ld->pg_read_ops;
  66 #endif
  67         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops,
  68                         server->rsize, 0);
  69 }
  70 EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
  71
  72 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
  73 {
  74         pgio->pg_ops = &nfs_pageio_read_ops;
  75         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
  76 }
  77 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
  78
  79 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
  80                        struct page *page)
  81 {
  82         struct nfs_page *new;
  83         unsigned int len;
  84         struct nfs_pageio_descriptor pgio;
  85
  86         len = nfs_page_length(page);
  87         if (len == 0)
  88                 return nfs_return_empty_page(page);
  89         new = nfs_create_request(ctx, inode, page, 0, len);
  90         if (IS_ERR(new)) {
  91                 unlock_page(page);
  92                 return PTR_ERR(new);
  93         }
  94         if (len < PAGE_CACHE_SIZE)
  95                 zero_user_segment(page, len, PAGE_CACHE_SIZE);
  96
  97         nfs_pageio_init_read(&pgio, inode, false,
  98                              &nfs_async_read_completion_ops);
  99         nfs_pageio_add_request(&pgio, new);
 100         nfs_pageio_complete(&pgio);
 101         NFS_I(inode)->read_io += pgio.pg_bytes_written;
 102         return 0;
 103 }
 104
 105 static void nfs_readpage_release(struct nfs_page *req)
 106 {
 107         struct inode *d_inode = req->wb_context->dentry->d_inode;
 108
 109         if (PageUptodate(req->wb_page))
 110                 nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
 111
 112         unlock_page(req->wb_page);
 113
 114         dprintk("NFS: read done (%s/%Lu %d@%Ld)\n",
 115                         req->wb_context->dentry->d_inode->i_sb->s_id,
 116                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 117                         req->wb_bytes,
 118                         (long long)req_offset(req));
 119         nfs_release_request(req);
 120 }
 121
 122 /* Note io was page aligned */
 123 static void nfs_read_completion(struct nfs_pgio_header *hdr)
 124 {
 125         unsigned long bytes = 0;
 126
 127         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 128                 goto out;
 129         while (!list_empty(&hdr->pages)) {
 130                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 131                 struct page *page = req->wb_page;
 132
 133                 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
 134                         if (bytes > hdr->good_bytes)
 135                                 zero_user(page, 0, PAGE_SIZE);
 136                         else if (hdr->good_bytes - bytes < PAGE_SIZE)
 137                                 zero_user_segment(page,
 138                                         hdr->good_bytes & ~PAGE_MASK,
 139                                         PAGE_SIZE);
 140                 }
 141                 bytes += req->wb_bytes;
 142                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 143                         if (bytes <= hdr->good_bytes)
 144                                 SetPageUptodate(page);
 145                 } else
 146                         SetPageUptodate(page);
 147                 nfs_list_remove_request(req);
 148                 nfs_readpage_release(req);
 149         }
 150 out:
 151         hdr->release(hdr);
 152 }
 153
 154 int nfs_initiate_read(struct rpc_clnt *clnt,
 155                       struct nfs_pgio_data *data,
 156                       const struct rpc_call_ops *call_ops, int flags)
 157 {
 158         struct inode *inode = data->header->inode;
 159         int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
 160         struct rpc_task *task;
 161         struct rpc_message msg = {
 162                 .rpc_argp = &data->args,
 163                 .rpc_resp = &data->res,
 164                 .rpc_cred = data->header->cred,
 165         };
 166         struct rpc_task_setup task_setup_data = {
 167                 .task = &data->task,
 168                 .rpc_client = clnt,
 169                 .rpc_message = &msg,
 170                 .callback_ops = call_ops,
 171                 .callback_data = data,
 172                 .workqueue = nfsiod_workqueue,
 173                 .flags = RPC_TASK_ASYNC | swap_flags | flags,
 174         };
 175
 176         /* Set up the initial task struct. */
 177         NFS_PROTO(inode)->read_setup(data, &msg);
 178
 179         dprintk("NFS: %5u initiated read call (req %s/%llu, %u bytes @ "
 180                         "offset %llu)\n",
 181                         data->task.tk_pid,
 182                         inode->i_sb->s_id,
 183                         (unsigned long long)NFS_FILEID(inode),
 184                         data->args.count,
 185                         (unsigned long long)data->args.offset);
 186
 187         task = rpc_run_task(&task_setup_data);
 188         if (IS_ERR(task))
 189                 return PTR_ERR(task);
 190         rpc_put_task(task);
 191         return 0;
 192 }
 193 EXPORT_SYMBOL_GPL(nfs_initiate_read);
 194
 195 static int nfs_do_read(struct nfs_pgio_data *data,
 196                 const struct rpc_call_ops *call_ops)
 197 {
 198         struct inode *inode = data->header->inode;
 199
 200         return nfs_initiate_read(NFS_CLIENT(inode), data, call_ops, 0);
 201 }
 202
 203 static int
 204 nfs_do_multiple_reads(struct list_head *head,
 205                 const struct rpc_call_ops *call_ops)
 206 {
 207         struct nfs_pgio_data *data;
 208         int ret = 0;
 209
 210         while (!list_empty(head)) {
 211                 int ret2;
 212
 213                 data = list_first_entry(head, struct nfs_pgio_data, list);
 214                 list_del_init(&data->list);
 215
 216                 ret2 = nfs_do_read(data, call_ops);
 217                 if (ret == 0)
 218                         ret = ret2;
 219         }
 220         return ret;
 221 }
 222
 223 static void
 224 nfs_async_read_error(struct list_head *head)
 225 {
 226         struct nfs_page *req;
 227
 228         while (!list_empty(head)) {
 229                 req = nfs_list_entry(head->next);
 230                 nfs_list_remove_request(req);
 231                 nfs_readpage_release(req);
 232         }
 233 }
 234
 235 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
 236         .error_cleanup = nfs_async_read_error,
 237         .completion = nfs_read_completion,
 238 };
 239
 240 static void nfs_pagein_error(struct nfs_pageio_descriptor *desc,
 241                 struct nfs_pgio_header *hdr)
 242 {
 243         set_bit(NFS_IOHDR_REDO, &hdr->flags);
 244         while (!list_empty(&hdr->rpc_list)) {
 245                 struct nfs_pgio_data *data = list_first_entry(&hdr->rpc_list,
 246                                 struct nfs_pgio_data, list);
 247                 list_del(&data->list);
 248                 nfs_pgio_data_release(data);
 249         }
 250         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 251 }
 252
 253 /*
 254  * Generate multiple requests to fill a single page.
 255  *
 256  * We optimize to reduce the number of read operations on the wire.  If we
 257  * detect that we're reading a page, or an area of a page, that is past the
 258  * end of file, we do not generate NFS read operations but just clear the
 259  * parts of the page that would have come back zero from the server anyway.
 260  *
 261  * We rely on the cached value of i_size to make this determination; another
 262  * client can fill pages on the server past our cached end-of-file, but we
 263  * won't see the new data until our attribute cache is updated.  This is more
 264  * or less conventional NFS client behavior.
 265  */
 266 static int nfs_pagein_multi(struct nfs_pageio_descriptor *desc,
 267                             struct nfs_pgio_header *hdr)
 268 {
 269         struct nfs_page *req = hdr->req;
 270         struct page *page = req->wb_page;
 271         struct nfs_pgio_data *data;
 272         size_t rsize = desc->pg_bsize, nbytes;
 273         unsigned int offset;
 274
 275         offset = 0;
 276         nbytes = desc->pg_count;
 277         do {
 278                 size_t len = min(nbytes,rsize);
 279
 280                 data = nfs_pgio_data_alloc(hdr, 1);
 281                 if (!data) {
 282                         nfs_pagein_error(desc, hdr);
 283                         return -ENOMEM;
 284                 }
 285                 data->pages.pagevec[0] = page;
 286                 nfs_pgio_rpcsetup(data, len, offset, 0, NULL);
 287                 list_add(&data->list, &hdr->rpc_list);
 288                 nbytes -= len;
 289                 offset += len;
 290         } while (nbytes != 0);
 291
 292         nfs_list_remove_request(req);
 293         nfs_list_add_request(req, &hdr->pages);
 294         desc->pg_rpc_callops = &nfs_pgio_common_ops;
 295         return 0;
 296 }
 297
 298 static int nfs_pagein_one(struct nfs_pageio_descriptor *desc,
 299                           struct nfs_pgio_header *hdr)
 300 {
 301         struct nfs_page         *req;
 302         struct page             **pages;
 303         struct nfs_pgio_data    *data;
 304         struct list_head *head = &desc->pg_list;
 305
 306         data = nfs_pgio_data_alloc(hdr, nfs_page_array_len(desc->pg_base,
 307                                                           desc->pg_count));
 308         if (!data) {
 309                 nfs_pagein_error(desc, hdr);
 310                 return -ENOMEM;
 311         }
 312
 313         pages = data->pages.pagevec;
 314         while (!list_empty(head)) {
 315                 req = nfs_list_entry(head->next);
 316                 nfs_list_remove_request(req);
 317                 nfs_list_add_request(req, &hdr->pages);
 318                 *pages++ = req->wb_page;
 319         }
 320
 321         nfs_pgio_rpcsetup(data, desc->pg_count, 0, 0, NULL);
 322         list_add(&data->list, &hdr->rpc_list);
 323         desc->pg_rpc_callops = &nfs_pgio_common_ops;
 324         return 0;
 325 }
 326
 327 int nfs_generic_pagein(struct nfs_pageio_descriptor *desc,
 328                        struct nfs_pgio_header *hdr)
 329 {
 330         if (desc->pg_bsize < PAGE_CACHE_SIZE)
 331                 return nfs_pagein_multi(desc, hdr);
 332         return nfs_pagein_one(desc, hdr);
 333 }
 334 EXPORT_SYMBOL_GPL(nfs_generic_pagein);
 335
 336 static int nfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
 337 {
 338         struct nfs_rw_header *rhdr;
 339         struct nfs_pgio_header *hdr;
 340         int ret;
 341
 342         rhdr = nfs_rw_header_alloc(desc->pg_rw_ops);
 343         if (!rhdr) {
 344                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 345                 return -ENOMEM;
 346         }
 347         hdr = &rhdr->header;
 348         nfs_pgheader_init(desc, hdr, nfs_rw_header_free);
 349         atomic_inc(&hdr->refcnt);
 350         ret = nfs_generic_pagein(desc, hdr);
 351         if (ret == 0)
 352                 ret = nfs_do_multiple_reads(&hdr->rpc_list,
 353                                             desc->pg_rpc_callops);
 354         if (atomic_dec_and_test(&hdr->refcnt))
 355                 hdr->completion_ops->completion(hdr);
 356         return ret;
 357 }
 358
 359 static const struct nfs_pageio_ops nfs_pageio_read_ops = {
 360         .pg_test = nfs_generic_pg_test,
 361         .pg_doio = nfs_generic_pg_readpages,
 362 };
 363
 364 /*
 365  * This is the callback from RPC telling us whether a reply was
 366  * received or some error occurred (timeout or socket shutdown).
 367  */
 368 static int nfs_readpage_done(struct rpc_task *task, struct nfs_pgio_data *data,
 369                              struct inode *inode)
 370 {
 371         int status = NFS_PROTO(inode)->read_done(task, data);
 372         if (status != 0)
 373                 return status;
 374
 375         nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, data->res.count);
 376
 377         if (task->tk_status == -ESTALE) {
 378                 set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
 379                 nfs_mark_for_revalidate(inode);
 380         }
 381         return 0;
 382 }
 383
 384 static void nfs_readpage_retry(struct rpc_task *task, struct nfs_pgio_data *data)
 385 {
 386         struct nfs_pgio_args *argp = &data->args;
 387         struct nfs_pgio_res  *resp = &data->res;
 388
 389         /* This is a short read! */
 390         nfs_inc_stats(data->header->inode, NFSIOS_SHORTREAD);
 391         /* Has the server at least made some progress? */
 392         if (resp->count == 0) {
 393                 nfs_set_pgio_error(data->header, -EIO, argp->offset);
 394                 return;
 395         }
 396         /* Yes, so retry the read at the end of the data */
 397         data->mds_offset += resp->count;
 398         argp->offset += resp->count;
 399         argp->pgbase += resp->count;
 400         argp->count -= resp->count;
 401         rpc_restart_call_prepare(task);
 402 }
 403
 404 static void nfs_readpage_result(struct rpc_task *task, struct nfs_pgio_data *data)
 405 {
 406         struct nfs_pgio_header *hdr = data->header;
 407
 408         if (data->res.eof) {
 409                 loff_t bound;
 410
 411                 bound = data->args.offset + data->res.count;
 412                 spin_lock(&hdr->lock);
 413                 if (bound < hdr->io_start + hdr->good_bytes) {
 414                         set_bit(NFS_IOHDR_EOF, &hdr->flags);
 415                         clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
 416                         hdr->good_bytes = bound - hdr->io_start;
 417                 }
 418                 spin_unlock(&hdr->lock);
 419         } else if (data->res.count != data->args.count)
 420                 nfs_readpage_retry(task, data);
 421 }
 422
 423 /*
 424  * Read a page over NFS.
 425  * We read the page synchronously in the following case:
 426  *  -   The error flag is set for this page. This happens only when a
 427  *      previous async read operation failed.
 428  */
 429 int nfs_readpage(struct file *file, struct page *page)
 430 {
 431         struct nfs_open_context *ctx;
 432         struct inode *inode = page_file_mapping(page)->host;
 433         int             error;
 434
 435         dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
 436                 page, PAGE_CACHE_SIZE, page_file_index(page));
 437         nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
 438         nfs_add_stats(inode, NFSIOS_READPAGES, 1);
 439
 440         /*
 441          * Try to flush any pending writes to the file..
 442          *
 443          * NOTE! Because we own the page lock, there cannot
 444          * be any new pending writes generated at this point
 445          * for this page (other pages can be written to).
 446          */
 447         error = nfs_wb_page(inode, page);
 448         if (error)
 449                 goto out_unlock;
 450         if (PageUptodate(page))
 451                 goto out_unlock;
 452
 453         error = -ESTALE;
 454         if (NFS_STALE(inode))
 455                 goto out_unlock;
 456
 457         if (file == NULL) {
 458                 error = -EBADF;
 459                 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 460                 if (ctx == NULL)
 461                         goto out_unlock;
 462         } else
 463                 ctx = get_nfs_open_context(nfs_file_open_context(file));
 464
 465         if (!IS_SYNC(inode)) {
 466                 error = nfs_readpage_from_fscache(ctx, inode, page);
 467                 if (error == 0)
 468                         goto out;
 469         }
 470
 471         error = nfs_readpage_async(ctx, inode, page);
 472
 473 out:
 474         put_nfs_open_context(ctx);
 475         return error;
 476 out_unlock:
 477         unlock_page(page);
 478         return error;
 479 }
 480
 481 struct nfs_readdesc {
 482         struct nfs_pageio_descriptor *pgio;
 483         struct nfs_open_context *ctx;
 484 };
 485
 486 static int
 487 readpage_async_filler(void *data, struct page *page)
 488 {
 489         struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
 490         struct inode *inode = page_file_mapping(page)->host;
 491         struct nfs_page *new;
 492         unsigned int len;
 493         int error;
 494
 495         len = nfs_page_length(page);
 496         if (len == 0)
 497                 return nfs_return_empty_page(page);
 498
 499         new = nfs_create_request(desc->ctx, inode, page, 0, len);
 500         if (IS_ERR(new))
 501                 goto out_error;
 502
 503         if (len < PAGE_CACHE_SIZE)
 504                 zero_user_segment(page, len, PAGE_CACHE_SIZE);
 505         if (!nfs_pageio_add_request(desc->pgio, new)) {
 506                 error = desc->pgio->pg_error;
 507                 goto out_unlock;
 508         }
 509         return 0;
 510 out_error:
 511         error = PTR_ERR(new);
 512 out_unlock:
 513         unlock_page(page);
 514         return error;
 515 }
 516
 517 int nfs_readpages(struct file *filp, struct address_space *mapping,
 518                 struct list_head *pages, unsigned nr_pages)
 519 {
 520         struct nfs_pageio_descriptor pgio;
 521         struct nfs_readdesc desc = {
 522                 .pgio = &pgio,
 523         };
 524         struct inode *inode = mapping->host;
 525         unsigned long npages;
 526         int ret = -ESTALE;
 527
 528         dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
 529                         inode->i_sb->s_id,
 530                         (unsigned long long)NFS_FILEID(inode),
 531                         nr_pages);
 532         nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
 533
 534         if (NFS_STALE(inode))
 535                 goto out;
 536
 537         if (filp == NULL) {
 538                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 539                 if (desc.ctx == NULL)
 540                         return -EBADF;
 541         } else
 542                 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
 543
 544         /* attempt to read as many of the pages as possible from the cache
 545          * - this returns -ENOBUFS immediately if the cookie is negative
 546          */
 547         ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
 548                                          pages, &nr_pages);
 549         if (ret == 0)
 550                 goto read_complete; /* all pages were read */
 551
 552         nfs_pageio_init_read(&pgio, inode, false,
 553                              &nfs_async_read_completion_ops);
 554
 555         ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
 556
 557         nfs_pageio_complete(&pgio);
 558         NFS_I(inode)->read_io += pgio.pg_bytes_written;
 559         npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 560         nfs_add_stats(inode, NFSIOS_READPAGES, npages);
 561 read_complete:
 562         put_nfs_open_context(desc.ctx);
 563 out:
 564         return ret;
 565 }
 566
 567 int __init nfs_init_readpagecache(void)
 568 {
 569         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
 570                                              sizeof(struct nfs_rw_header),
 571                                              0, SLAB_HWCACHE_ALIGN,
 572                                              NULL);
 573         if (nfs_rdata_cachep == NULL)
 574                 return -ENOMEM;
 575
 576         return 0;
 577 }
 578
 579 void nfs_destroy_readpagecache(void)
 580 {
 581         kmem_cache_destroy(nfs_rdata_cachep);
 582 }
 583
 584 static const struct nfs_rw_ops nfs_rw_read_ops = {
 585         .rw_mode                = FMODE_READ,
 586         .rw_alloc_header        = nfs_readhdr_alloc,
 587         .rw_free_header         = nfs_readhdr_free,
 588         .rw_done                = nfs_readpage_done,
 589         .rw_result              = nfs_readpage_result,
 590 };