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 int nfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
 241 {
 242         struct nfs_rw_header *rhdr;
 243         struct nfs_pgio_header *hdr;
 244         int ret;
 245
 246         rhdr = nfs_rw_header_alloc(desc->pg_rw_ops);
 247         if (!rhdr) {
 248                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 249                 return -ENOMEM;
 250         }
 251         hdr = &rhdr->header;
 252         nfs_pgheader_init(desc, hdr, nfs_rw_header_free);
 253         atomic_inc(&hdr->refcnt);
 254         ret = nfs_generic_pgio(desc, hdr);
 255         if (ret == 0)
 256                 ret = nfs_do_multiple_reads(&hdr->rpc_list,
 257                                             desc->pg_rpc_callops);
 258         if (atomic_dec_and_test(&hdr->refcnt))
 259                 hdr->completion_ops->completion(hdr);
 260         return ret;
 261 }
 262
 263 static const struct nfs_pageio_ops nfs_pageio_read_ops = {
 264         .pg_test = nfs_generic_pg_test,
 265         .pg_doio = nfs_generic_pg_readpages,
 266 };
 267
 268 /*
 269  * This is the callback from RPC telling us whether a reply was
 270  * received or some error occurred (timeout or socket shutdown).
 271  */
 272 static int nfs_readpage_done(struct rpc_task *task, struct nfs_pgio_data *data,
 273                              struct inode *inode)
 274 {
 275         int status = NFS_PROTO(inode)->read_done(task, data);
 276         if (status != 0)
 277                 return status;
 278
 279         nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, data->res.count);
 280
 281         if (task->tk_status == -ESTALE) {
 282                 set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
 283                 nfs_mark_for_revalidate(inode);
 284         }
 285         return 0;
 286 }
 287
 288 static void nfs_readpage_retry(struct rpc_task *task, struct nfs_pgio_data *data)
 289 {
 290         struct nfs_pgio_args *argp = &data->args;
 291         struct nfs_pgio_res  *resp = &data->res;
 292
 293         /* This is a short read! */
 294         nfs_inc_stats(data->header->inode, NFSIOS_SHORTREAD);
 295         /* Has the server at least made some progress? */
 296         if (resp->count == 0) {
 297                 nfs_set_pgio_error(data->header, -EIO, argp->offset);
 298                 return;
 299         }
 300         /* Yes, so retry the read at the end of the data */
 301         data->mds_offset += resp->count;
 302         argp->offset += resp->count;
 303         argp->pgbase += resp->count;
 304         argp->count -= resp->count;
 305         rpc_restart_call_prepare(task);
 306 }
 307
 308 static void nfs_readpage_result(struct rpc_task *task, struct nfs_pgio_data *data)
 309 {
 310         struct nfs_pgio_header *hdr = data->header;
 311
 312         if (data->res.eof) {
 313                 loff_t bound;
 314
 315                 bound = data->args.offset + data->res.count;
 316                 spin_lock(&hdr->lock);
 317                 if (bound < hdr->io_start + hdr->good_bytes) {
 318                         set_bit(NFS_IOHDR_EOF, &hdr->flags);
 319                         clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
 320                         hdr->good_bytes = bound - hdr->io_start;
 321                 }
 322                 spin_unlock(&hdr->lock);
 323         } else if (data->res.count != data->args.count)
 324                 nfs_readpage_retry(task, data);
 325 }
 326
 327 /*
 328  * Read a page over NFS.
 329  * We read the page synchronously in the following case:
 330  *  -   The error flag is set for this page. This happens only when a
 331  *      previous async read operation failed.
 332  */
 333 int nfs_readpage(struct file *file, struct page *page)
 334 {
 335         struct nfs_open_context *ctx;
 336         struct inode *inode = page_file_mapping(page)->host;
 337         int             error;
 338
 339         dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
 340                 page, PAGE_CACHE_SIZE, page_file_index(page));
 341         nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
 342         nfs_add_stats(inode, NFSIOS_READPAGES, 1);
 343
 344         /*
 345          * Try to flush any pending writes to the file..
 346          *
 347          * NOTE! Because we own the page lock, there cannot
 348          * be any new pending writes generated at this point
 349          * for this page (other pages can be written to).
 350          */
 351         error = nfs_wb_page(inode, page);
 352         if (error)
 353                 goto out_unlock;
 354         if (PageUptodate(page))
 355                 goto out_unlock;
 356
 357         error = -ESTALE;
 358         if (NFS_STALE(inode))
 359                 goto out_unlock;
 360
 361         if (file == NULL) {
 362                 error = -EBADF;
 363                 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 364                 if (ctx == NULL)
 365                         goto out_unlock;
 366         } else
 367                 ctx = get_nfs_open_context(nfs_file_open_context(file));
 368
 369         if (!IS_SYNC(inode)) {
 370                 error = nfs_readpage_from_fscache(ctx, inode, page);
 371                 if (error == 0)
 372                         goto out;
 373         }
 374
 375         error = nfs_readpage_async(ctx, inode, page);
 376
 377 out:
 378         put_nfs_open_context(ctx);
 379         return error;
 380 out_unlock:
 381         unlock_page(page);
 382         return error;
 383 }
 384
 385 struct nfs_readdesc {
 386         struct nfs_pageio_descriptor *pgio;
 387         struct nfs_open_context *ctx;
 388 };
 389
 390 static int
 391 readpage_async_filler(void *data, struct page *page)
 392 {
 393         struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
 394         struct inode *inode = page_file_mapping(page)->host;
 395         struct nfs_page *new;
 396         unsigned int len;
 397         int error;
 398
 399         len = nfs_page_length(page);
 400         if (len == 0)
 401                 return nfs_return_empty_page(page);
 402
 403         new = nfs_create_request(desc->ctx, inode, page, 0, len);
 404         if (IS_ERR(new))
 405                 goto out_error;
 406
 407         if (len < PAGE_CACHE_SIZE)
 408                 zero_user_segment(page, len, PAGE_CACHE_SIZE);
 409         if (!nfs_pageio_add_request(desc->pgio, new)) {
 410                 error = desc->pgio->pg_error;
 411                 goto out_unlock;
 412         }
 413         return 0;
 414 out_error:
 415         error = PTR_ERR(new);
 416 out_unlock:
 417         unlock_page(page);
 418         return error;
 419 }
 420
 421 int nfs_readpages(struct file *filp, struct address_space *mapping,
 422                 struct list_head *pages, unsigned nr_pages)
 423 {
 424         struct nfs_pageio_descriptor pgio;
 425         struct nfs_readdesc desc = {
 426                 .pgio = &pgio,
 427         };
 428         struct inode *inode = mapping->host;
 429         unsigned long npages;
 430         int ret = -ESTALE;
 431
 432         dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
 433                         inode->i_sb->s_id,
 434                         (unsigned long long)NFS_FILEID(inode),
 435                         nr_pages);
 436         nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
 437
 438         if (NFS_STALE(inode))
 439                 goto out;
 440
 441         if (filp == NULL) {
 442                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 443                 if (desc.ctx == NULL)
 444                         return -EBADF;
 445         } else
 446                 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
 447
 448         /* attempt to read as many of the pages as possible from the cache
 449          * - this returns -ENOBUFS immediately if the cookie is negative
 450          */
 451         ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
 452                                          pages, &nr_pages);
 453         if (ret == 0)
 454                 goto read_complete; /* all pages were read */
 455
 456         nfs_pageio_init_read(&pgio, inode, false,
 457                              &nfs_async_read_completion_ops);
 458
 459         ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
 460
 461         nfs_pageio_complete(&pgio);
 462         NFS_I(inode)->read_io += pgio.pg_bytes_written;
 463         npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 464         nfs_add_stats(inode, NFSIOS_READPAGES, npages);
 465 read_complete:
 466         put_nfs_open_context(desc.ctx);
 467 out:
 468         return ret;
 469 }
 470
 471 int __init nfs_init_readpagecache(void)
 472 {
 473         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
 474                                              sizeof(struct nfs_rw_header),
 475                                              0, SLAB_HWCACHE_ALIGN,
 476                                              NULL);
 477         if (nfs_rdata_cachep == NULL)
 478                 return -ENOMEM;
 479
 480         return 0;
 481 }
 482
 483 void nfs_destroy_readpagecache(void)
 484 {
 485         kmem_cache_destroy(nfs_rdata_cachep);
 486 }
 487
 488 static const struct nfs_rw_ops nfs_rw_read_ops = {
 489         .rw_mode                = FMODE_READ,
 490         .rw_alloc_header        = nfs_readhdr_alloc,
 491         .rw_free_header         = nfs_readhdr_free,
 492         .rw_done                = nfs_readpage_done,
 493         .rw_result              = nfs_readpage_result,
 494 };