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  * We do an ugly hack here in order to return proper error codes to the
  10  * user program when a read request failed: since generic_file_read
  11  * only checks the return value of inode->i_op->readpage() which is always 0
  12  * for async RPC, we set the error bit of the page to 1 when an error occurs,
  13  * and make nfs_readpage transmit requests synchronously when encountering this.
  14  * This is only a small problem, though, since we now retry all operations
  15  * within the RPC code when root squashing is suspected.
  16  */
  17
  18 #include <linux/config.h>
  19 #include <linux/time.h>
  20 #include <linux/kernel.h>
  21 #include <linux/errno.h>
  22 #include <linux/fcntl.h>
  23 #include <linux/stat.h>
  24 #include <linux/mm.h>
  25 #include <linux/slab.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/mempool.h>
  28 #include <linux/sunrpc/clnt.h>
  29 #include <linux/nfs_fs.h>
  30 #include <linux/nfs_page.h>
  31 #include <linux/smp_lock.h>
  32
  33 #include <asm/system.h>
  34
  35 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  36
  37 static int nfs_pagein_one(struct list_head *, struct inode *);
  38 static void nfs_readpage_result_partial(struct nfs_read_data *, int);
  39 static void nfs_readpage_result_full(struct nfs_read_data *, int);
  40
  41 static kmem_cache_t *nfs_rdata_cachep;
  42 static mempool_t *nfs_rdata_mempool;
  43
  44 #define MIN_POOL_READ   (32)
  45
  46 static __inline__ struct nfs_read_data *nfs_readdata_alloc(void)
  47 {
  48         struct nfs_read_data   *p;
  49         p = (struct nfs_read_data *)mempool_alloc(nfs_rdata_mempool, SLAB_NOFS);
  50         if (p) {
  51                 memset(p, 0, sizeof(*p));
  52                 INIT_LIST_HEAD(&p->pages);
  53         }
  54         return p;
  55 }
  56
  57 static __inline__ void nfs_readdata_free(struct nfs_read_data *p)
  58 {
  59         mempool_free(p, nfs_rdata_mempool);
  60 }
  61
  62 static void nfs_readdata_release(struct rpc_task *task)
  63 {
  64         struct nfs_read_data   *data = (struct nfs_read_data *)task->tk_calldata;
  65         nfs_readdata_free(data);
  66 }
  67
  68 static
  69 unsigned int nfs_page_length(struct inode *inode, struct page *page)
  70 {
  71         loff_t i_size = i_size_read(inode);
  72         unsigned long idx;
  73
  74         if (i_size <= 0)
  75                 return 0;
  76         idx = (i_size - 1) >> PAGE_CACHE_SHIFT;
  77         if (page->index > idx)
  78                 return 0;
  79         if (page->index != idx)
  80                 return PAGE_CACHE_SIZE;
  81         return 1 + ((i_size - 1) & (PAGE_CACHE_SIZE - 1));
  82 }
  83
  84 static
  85 int nfs_return_empty_page(struct page *page)
  86 {
  87         memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
  88         SetPageUptodate(page);
  89         unlock_page(page);
  90         return 0;
  91 }
  92
  93 /*
  94  * Read a page synchronously.
  95  */
  96 static int
  97 nfs_readpage_sync(struct file *file, struct inode *inode, struct page *page)
  98 {
  99         unsigned int    rsize = NFS_SERVER(inode)->rsize;
 100         unsigned int    count = PAGE_CACHE_SIZE;
 101         int             result;
 102         struct nfs_read_data    rdata = {
 103                 .flags          = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0),
 104                 .cred           = NULL,
 105                 .inode          = inode,
 106                 .args           = {
 107                         .fh             = NFS_FH(inode),
 108                         .lockowner      = current->files,
 109                         .pages          = &page,
 110                         .pgbase         = 0UL,
 111                         .count          = rsize,
 112                 },
 113                 .res            = {
 114                         .fattr          = &rdata.fattr,
 115                 }
 116         };
 117
 118         dprintk("NFS: nfs_readpage_sync(%p)\n", page);
 119
 120         /*
 121          * This works now because the socket layer never tries to DMA
 122          * into this buffer directly.
 123          */
 124         do {
 125                 if (count < rsize)
 126                         rdata.args.count = count;
 127                 rdata.res.count = rdata.args.count;
 128                 rdata.args.offset = page_offset(page) + rdata.args.pgbase;
 129
 130                 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
 131                         NFS_SERVER(inode)->hostname,
 132                         inode->i_sb->s_id,
 133                         (long long)NFS_FILEID(inode),
 134                         (unsigned long long)rdata.args.pgbase,
 135                         rdata.args.count);
 136
 137                 lock_kernel();
 138                 result = NFS_PROTO(inode)->read(&rdata, file);
 139                 unlock_kernel();
 140
 141                 /*
 142                  * Even if we had a partial success we can't mark the page
 143                  * cache valid.
 144                  */
 145                 if (result < 0) {
 146                         if (result == -EISDIR)
 147                                 result = -EINVAL;
 148                         goto io_error;
 149                 }
 150                 count -= result;
 151                 rdata.args.pgbase += result;
 152                 /* Note: result == 0 should only happen if we're caching
 153                  * a write that extends the file and punches a hole.
 154                  */
 155                 if (rdata.res.eof != 0 || result == 0)
 156                         break;
 157         } while (count);
 158         NFS_FLAGS(inode) |= NFS_INO_INVALID_ATIME;
 159
 160         if (count)
 161                 memclear_highpage_flush(page, rdata.args.pgbase, count);
 162         SetPageUptodate(page);
 163         if (PageError(page))
 164                 ClearPageError(page);
 165         result = 0;
 166
 167 io_error:
 168         unlock_page(page);
 169         return result;
 170 }
 171
 172 static int
 173 nfs_readpage_async(struct file *file, struct inode *inode, struct page *page)
 174 {
 175         LIST_HEAD(one_request);
 176         struct nfs_page *new;
 177         unsigned int len;
 178
 179         len = nfs_page_length(inode, page);
 180         if (len == 0)
 181                 return nfs_return_empty_page(page);
 182         new = nfs_create_request(file, inode, page, 0, len);
 183         if (IS_ERR(new)) {
 184                 unlock_page(page);
 185                 return PTR_ERR(new);
 186         }
 187         if (len < PAGE_CACHE_SIZE)
 188                 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
 189
 190         nfs_lock_request(new);
 191         nfs_list_add_request(new, &one_request);
 192         nfs_pagein_one(&one_request, inode);
 193         return 0;
 194 }
 195
 196 static void nfs_readpage_release(struct nfs_page *req)
 197 {
 198         unlock_page(req->wb_page);
 199
 200         nfs_clear_request(req);
 201         nfs_release_request(req);
 202         nfs_unlock_request(req);
 203
 204         dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
 205                         req->wb_inode->i_sb->s_id,
 206                         (long long)NFS_FILEID(req->wb_inode),
 207                         req->wb_bytes,
 208                         (long long)req_offset(req));
 209 }
 210
 211 /*
 212  * Set up the NFS read request struct
 213  */
 214 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
 215                 unsigned int count, unsigned int offset)
 216 {
 217         struct inode            *inode;
 218
 219         data->req         = req;
 220         data->inode       = inode = req->wb_inode;
 221         data->cred        = req->wb_cred;
 222
 223         data->args.fh     = NFS_FH(inode);
 224         data->args.offset = req_offset(req) + offset;
 225         data->args.pgbase = req->wb_pgbase + offset;
 226         data->args.pages  = data->pagevec;
 227         data->args.count  = count;
 228         data->args.lockowner = req->wb_lockowner;
 229         data->args.state  = req->wb_state;
 230
 231         data->res.fattr   = &data->fattr;
 232         data->res.count   = count;
 233         data->res.eof     = 0;
 234
 235         NFS_PROTO(inode)->read_setup(data);
 236
 237         data->task.tk_cookie = (unsigned long)inode;
 238         data->task.tk_calldata = data;
 239         /* Release requests */
 240         data->task.tk_release = nfs_readdata_release;
 241
 242         dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 243                         data->task.tk_pid,
 244                         inode->i_sb->s_id,
 245                         (long long)NFS_FILEID(inode),
 246                         count,
 247                         (unsigned long long)data->args.offset);
 248 }
 249
 250 static void
 251 nfs_async_read_error(struct list_head *head)
 252 {
 253         struct nfs_page *req;
 254
 255         while (!list_empty(head)) {
 256                 req = nfs_list_entry(head->next);
 257                 nfs_list_remove_request(req);
 258                 SetPageError(req->wb_page);
 259                 nfs_readpage_release(req);
 260         }
 261 }
 262
 263 /*
 264  * Start an async read operation
 265  */
 266 static void nfs_execute_read(struct nfs_read_data *data)
 267 {
 268         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 269         sigset_t oldset;
 270
 271         rpc_clnt_sigmask(clnt, &oldset);
 272         lock_kernel();
 273         rpc_execute(&data->task);
 274         unlock_kernel();
 275         rpc_clnt_sigunmask(clnt, &oldset);
 276 }
 277
 278 /*
 279  * Generate multiple requests to fill a single page.
 280  *
 281  * We optimize to reduce the number of read operations on the wire.  If we
 282  * detect that we're reading a page, or an area of a page, that is past the
 283  * end of file, we do not generate NFS read operations but just clear the
 284  * parts of the page that would have come back zero from the server anyway.
 285  *
 286  * We rely on the cached value of i_size to make this determination; another
 287  * client can fill pages on the server past our cached end-of-file, but we
 288  * won't see the new data until our attribute cache is updated.  This is more
 289  * or less conventional NFS client behavior.
 290  */
 291 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
 292 {
 293         struct nfs_page *req = nfs_list_entry(head->next);
 294         struct page *page = req->wb_page;
 295         struct nfs_read_data *data;
 296         unsigned int rsize = NFS_SERVER(inode)->rsize;
 297         unsigned int nbytes, offset;
 298         int requests = 0;
 299         LIST_HEAD(list);
 300
 301         nfs_list_remove_request(req);
 302
 303         nbytes = req->wb_bytes;
 304         for(;;) {
 305                 data = nfs_readdata_alloc();
 306                 if (!data)
 307                         goto out_bad;
 308                 list_add(&data->pages, &list);
 309                 requests++;
 310                 if (nbytes <= rsize)
 311                         break;
 312                 nbytes -= rsize;
 313         }
 314         atomic_set(&req->wb_complete, requests);
 315
 316         ClearPageError(page);
 317         offset = 0;
 318         nbytes = req->wb_bytes;
 319         do {
 320                 data = list_entry(list.next, struct nfs_read_data, pages);
 321                 list_del_init(&data->pages);
 322
 323                 data->pagevec[0] = page;
 324                 data->complete = nfs_readpage_result_partial;
 325
 326                 if (nbytes > rsize) {
 327                         nfs_read_rpcsetup(req, data, rsize, offset);
 328                         offset += rsize;
 329                         nbytes -= rsize;
 330                 } else {
 331                         nfs_read_rpcsetup(req, data, nbytes, offset);
 332                         nbytes = 0;
 333                 }
 334                 nfs_execute_read(data);
 335         } while (nbytes != 0);
 336
 337         return 0;
 338
 339 out_bad:
 340         while (!list_empty(&list)) {
 341                 data = list_entry(list.next, struct nfs_read_data, pages);
 342                 list_del(&data->pages);
 343                 nfs_readdata_free(data);
 344         }
 345         SetPageError(page);
 346         nfs_readpage_release(req);
 347         return -ENOMEM;
 348 }
 349
 350 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
 351 {
 352         struct nfs_page         *req;
 353         struct page             **pages;
 354         struct nfs_read_data    *data;
 355         unsigned int            count;
 356
 357         if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
 358                 return nfs_pagein_multi(head, inode);
 359
 360         data = nfs_readdata_alloc();
 361         if (!data)
 362                 goto out_bad;
 363
 364         pages = data->pagevec;
 365         count = 0;
 366         while (!list_empty(head)) {
 367                 req = nfs_list_entry(head->next);
 368                 nfs_list_remove_request(req);
 369                 nfs_list_add_request(req, &data->pages);
 370                 ClearPageError(req->wb_page);
 371                 *pages++ = req->wb_page;
 372                 count += req->wb_bytes;
 373         }
 374         req = nfs_list_entry(data->pages.next);
 375
 376         data->complete = nfs_readpage_result_full;
 377         nfs_read_rpcsetup(req, data, count, 0);
 378
 379         nfs_execute_read(data);
 380         return 0;
 381 out_bad:
 382         nfs_async_read_error(head);
 383         return -ENOMEM;
 384 }
 385
 386 int
 387 nfs_pagein_list(struct list_head *head, int rpages)
 388 {
 389         LIST_HEAD(one_request);
 390         struct nfs_page         *req;
 391         int                     error = 0;
 392         unsigned int            pages = 0;
 393
 394         while (!list_empty(head)) {
 395                 pages += nfs_coalesce_requests(head, &one_request, rpages);
 396                 req = nfs_list_entry(one_request.next);
 397                 error = nfs_pagein_one(&one_request, req->wb_inode);
 398                 if (error < 0)
 399                         break;
 400         }
 401         if (error >= 0)
 402                 return pages;
 403
 404         nfs_async_read_error(head);
 405         return error;
 406 }
 407
 408 /*
 409  * Handle a read reply that fills part of a page.
 410  */
 411 static void nfs_readpage_result_partial(struct nfs_read_data *data, int status)
 412 {
 413         struct nfs_page *req = data->req;
 414         struct page *page = req->wb_page;
 415
 416         if (status >= 0) {
 417                 unsigned int request = data->args.count;
 418                 unsigned int result = data->res.count;
 419
 420                 if (result < request) {
 421                         memclear_highpage_flush(page,
 422                                                 data->args.pgbase + result,
 423                                                 request - result);
 424                 }
 425         } else
 426                 SetPageError(page);
 427
 428         if (atomic_dec_and_test(&req->wb_complete)) {
 429                 if (!PageError(page))
 430                         SetPageUptodate(page);
 431                 nfs_readpage_release(req);
 432         }
 433 }
 434
 435 /*
 436  * This is the callback from RPC telling us whether a reply was
 437  * received or some error occurred (timeout or socket shutdown).
 438  */
 439 static void nfs_readpage_result_full(struct nfs_read_data *data, int status)
 440 {
 441         unsigned int count = data->res.count;
 442
 443         while (!list_empty(&data->pages)) {
 444                 struct nfs_page *req = nfs_list_entry(data->pages.next);
 445                 struct page *page = req->wb_page;
 446                 nfs_list_remove_request(req);
 447
 448                 if (status >= 0) {
 449                         if (count < PAGE_CACHE_SIZE) {
 450                                 if (count < req->wb_bytes)
 451                                         memclear_highpage_flush(page,
 452                                                         req->wb_pgbase + count,
 453                                                         req->wb_bytes - count);
 454                                 count = 0;
 455                         } else
 456                                 count -= PAGE_CACHE_SIZE;
 457                         SetPageUptodate(page);
 458                 } else
 459                         SetPageError(page);
 460                 nfs_readpage_release(req);
 461         }
 462 }
 463
 464 /*
 465  * This is the callback from RPC telling us whether a reply was
 466  * received or some error occurred (timeout or socket shutdown).
 467  */
 468 void nfs_readpage_result(struct rpc_task *task)
 469 {
 470         struct nfs_read_data *data = (struct nfs_read_data *)task->tk_calldata;
 471         struct nfs_readargs *argp = &data->args;
 472         struct nfs_readres *resp = &data->res;
 473         int status = task->tk_status;
 474
 475         dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
 476                 task->tk_pid, status);
 477
 478         /* Is this a short read? */
 479         if (task->tk_status >= 0 && resp->count < argp->count && !resp->eof) {
 480                 /* Has the server at least made some progress? */
 481                 if (resp->count != 0) {
 482                         /* Yes, so retry the read at the end of the data */
 483                         argp->offset += resp->count;
 484                         argp->pgbase += resp->count;
 485                         argp->count -= resp->count;
 486                         rpc_restart_call(task);
 487                         return;
 488                 }
 489                 task->tk_status = -EIO;
 490         }
 491         NFS_FLAGS(data->inode) |= NFS_INO_INVALID_ATIME;
 492         data->complete(data, status);
 493 }
 494
 495 /*
 496  * Read a page over NFS.
 497  * We read the page synchronously in the following case:
 498  *  -   The error flag is set for this page. This happens only when a
 499  *      previous async read operation failed.
 500  */
 501 int
 502 nfs_readpage(struct file *file, struct page *page)
 503 {
 504         struct inode *inode = page->mapping->host;
 505         int             error;
 506
 507         dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
 508                 page, PAGE_CACHE_SIZE, page->index);
 509         /*
 510          * Try to flush any pending writes to the file..
 511          *
 512          * NOTE! Because we own the page lock, there cannot
 513          * be any new pending writes generated at this point
 514          * for this page (other pages can be written to).
 515          */
 516         error = nfs_wb_page(inode, page);
 517         if (error)
 518                 goto out_error;
 519
 520         if (!IS_SYNC(inode)) {
 521                 error = nfs_readpage_async(file, inode, page);
 522                 goto out;
 523         }
 524
 525         error = nfs_readpage_sync(file, inode, page);
 526         if (error < 0 && IS_SWAPFILE(inode))
 527                 printk("Aiee.. nfs swap-in of page failed!\n");
 528 out:
 529         return error;
 530
 531 out_error:
 532         unlock_page(page);
 533         goto out;
 534 }
 535
 536 struct nfs_readdesc {
 537         struct list_head *head;
 538         struct file *filp;
 539 };
 540
 541 static int
 542 readpage_async_filler(void *data, struct page *page)
 543 {
 544         struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
 545         struct inode *inode = page->mapping->host;
 546         struct nfs_page *new;
 547         unsigned int len;
 548
 549         nfs_wb_page(inode, page);
 550         len = nfs_page_length(inode, page);
 551         if (len == 0)
 552                 return nfs_return_empty_page(page);
 553         new = nfs_create_request(desc->filp, inode, page, 0, len);
 554         if (IS_ERR(new)) {
 555                         SetPageError(page);
 556                         unlock_page(page);
 557                         return PTR_ERR(new);
 558         }
 559         if (len < PAGE_CACHE_SIZE)
 560                 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
 561         nfs_lock_request(new);
 562         nfs_list_add_request(new, desc->head);
 563         return 0;
 564 }
 565
 566 int
 567 nfs_readpages(struct file *filp, struct address_space *mapping,
 568                 struct list_head *pages, unsigned nr_pages)
 569 {
 570         LIST_HEAD(head);
 571         struct nfs_readdesc desc = {
 572                 .filp           = filp,
 573                 .head           = &head,
 574         };
 575         struct inode *inode = mapping->host;
 576         struct nfs_server *server = NFS_SERVER(inode);
 577         int ret;
 578
 579         dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
 580                         inode->i_sb->s_id,
 581                         (long long)NFS_FILEID(inode),
 582                         nr_pages);
 583
 584         ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
 585         if (!list_empty(&head)) {
 586                 int err = nfs_pagein_list(&head, server->rpages);
 587                 if (!ret)
 588                         ret = err;
 589         }
 590         return ret;
 591 }
 592
 593 int nfs_init_readpagecache(void)
 594 {
 595         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
 596                                              sizeof(struct nfs_read_data),
 597                                              0, SLAB_HWCACHE_ALIGN,
 598                                              NULL, NULL);
 599         if (nfs_rdata_cachep == NULL)
 600                 return -ENOMEM;
 601
 602         nfs_rdata_mempool = mempool_create(MIN_POOL_READ,
 603                                            mempool_alloc_slab,
 604                                            mempool_free_slab,
 605                                            nfs_rdata_cachep);
 606         if (nfs_rdata_mempool == NULL)
 607                 return -ENOMEM;
 608
 609         return 0;
 610 }
 611
 612 void nfs_destroy_readpagecache(void)
 613 {
 614         mempool_destroy(nfs_rdata_mempool);
 615         if (kmem_cache_destroy(nfs_rdata_cachep))
 616                 printk(KERN_INFO "nfs_read_data: not all structures were freed\n");
 617 }