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 #include "nfstrace.h"
  29
  30 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  31
  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_pgio_header *nfs_readhdr_alloc(void)
  38 {
  39         struct nfs_pgio_header *p = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
  40
  41         if (p)
  42                 p->rw_mode = FMODE_READ;
  43         return p;
  44 }
  45
  46 static void nfs_readhdr_free(struct nfs_pgio_header *rhdr)
  47 {
  48         kmem_cache_free(nfs_rdata_cachep, rhdr);
  49 }
  50
  51 static
  52 int nfs_return_empty_page(struct page *page)
  53 {
  54         zero_user(page, 0, PAGE_SIZE);
  55         SetPageUptodate(page);
  56         unlock_page(page);
  57         return 0;
  58 }
  59
  60 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
  61                               struct inode *inode, bool force_mds,
  62                               const struct nfs_pgio_completion_ops *compl_ops)
  63 {
  64         struct nfs_server *server = NFS_SERVER(inode);
  65         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
  66
  67 #ifdef CONFIG_NFS_V4_1
  68         if (server->pnfs_curr_ld && !force_mds)
  69                 pg_ops = server->pnfs_curr_ld->pg_read_ops;
  70 #endif
  71         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops,
  72                         server->rsize, 0);
  73 }
  74 EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
  75
  76 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
  77 {
  78         struct nfs_pgio_mirror *mirror;
  79
  80         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
  81                 pgio->pg_ops->pg_cleanup(pgio);
  82
  83         pgio->pg_ops = &nfs_pgio_rw_ops;
  84
  85         /* read path should never have more than one mirror */
  86         WARN_ON_ONCE(pgio->pg_mirror_count != 1);
  87
  88         mirror = &pgio->pg_mirrors[0];
  89         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
  90 }
  91 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
  92
  93 static void nfs_readpage_release(struct nfs_page *req)
  94 {
  95         struct inode *inode = d_inode(req->wb_context->dentry);
  96
  97         dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
  98                 (unsigned long long)NFS_FILEID(inode), req->wb_bytes,
  99                 (long long)req_offset(req));
 100
 101         if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
 102                 if (PageUptodate(req->wb_page))
 103                         nfs_readpage_to_fscache(inode, req->wb_page, 0);
 104
 105                 unlock_page(req->wb_page);
 106         }
 107         nfs_release_request(req);
 108 }
 109
 110 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
 111                        struct page *page)
 112 {
 113         struct nfs_page *new;
 114         unsigned int len;
 115         struct nfs_pageio_descriptor pgio;
 116         struct nfs_pgio_mirror *pgm;
 117
 118         len = nfs_page_length(page);
 119         if (len == 0)
 120                 return nfs_return_empty_page(page);
 121         new = nfs_create_request(ctx, page, NULL, 0, len);
 122         if (IS_ERR(new)) {
 123                 unlock_page(page);
 124                 return PTR_ERR(new);
 125         }
 126         if (len < PAGE_SIZE)
 127                 zero_user_segment(page, len, PAGE_SIZE);
 128
 129         nfs_pageio_init_read(&pgio, inode, false,
 130                              &nfs_async_read_completion_ops);
 131         if (!nfs_pageio_add_request(&pgio, new)) {
 132                 nfs_list_remove_request(new);
 133                 nfs_readpage_release(new);
 134         }
 135         nfs_pageio_complete(&pgio);
 136
 137         /* It doesn't make sense to do mirrored reads! */
 138         WARN_ON_ONCE(pgio.pg_mirror_count != 1);
 139
 140         pgm = &pgio.pg_mirrors[0];
 141         NFS_I(inode)->read_io += pgm->pg_bytes_written;
 142
 143         return pgio.pg_error < 0 ? pgio.pg_error : 0;
 144 }
 145
 146 static void nfs_page_group_set_uptodate(struct nfs_page *req)
 147 {
 148         if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
 149                 SetPageUptodate(req->wb_page);
 150 }
 151
 152 static void nfs_read_completion(struct nfs_pgio_header *hdr)
 153 {
 154         unsigned long bytes = 0;
 155
 156         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 157                 goto out;
 158         while (!list_empty(&hdr->pages)) {
 159                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 160                 struct page *page = req->wb_page;
 161                 unsigned long start = req->wb_pgbase;
 162                 unsigned long end = req->wb_pgbase + req->wb_bytes;
 163
 164                 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
 165                         /* note: regions of the page not covered by a
 166                          * request are zeroed in nfs_readpage_async /
 167                          * readpage_async_filler */
 168                         if (bytes > hdr->good_bytes) {
 169                                 /* nothing in this request was good, so zero
 170                                  * the full extent of the request */
 171                                 zero_user_segment(page, start, end);
 172
 173                         } else if (hdr->good_bytes - bytes < req->wb_bytes) {
 174                                 /* part of this request has good bytes, but
 175                                  * not all. zero the bad bytes */
 176                                 start += hdr->good_bytes - bytes;
 177                                 WARN_ON(start < req->wb_pgbase);
 178                                 zero_user_segment(page, start, end);
 179                         }
 180                 }
 181                 bytes += req->wb_bytes;
 182                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 183                         if (bytes <= hdr->good_bytes)
 184                                 nfs_page_group_set_uptodate(req);
 185                 } else
 186                         nfs_page_group_set_uptodate(req);
 187                 nfs_list_remove_request(req);
 188                 nfs_readpage_release(req);
 189         }
 190 out:
 191         hdr->release(hdr);
 192 }
 193
 194 static void nfs_initiate_read(struct nfs_pgio_header *hdr,
 195                               struct rpc_message *msg,
 196                               const struct nfs_rpc_ops *rpc_ops,
 197                               struct rpc_task_setup *task_setup_data, int how)
 198 {
 199         struct inode *inode = hdr->inode;
 200         int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
 201
 202         task_setup_data->flags |= swap_flags;
 203         rpc_ops->read_setup(hdr, msg);
 204         trace_nfs_initiate_read(inode, hdr->io_start, hdr->good_bytes);
 205 }
 206
 207 static void
 208 nfs_async_read_error(struct list_head *head)
 209 {
 210         struct nfs_page *req;
 211
 212         while (!list_empty(head)) {
 213                 req = nfs_list_entry(head->next);
 214                 nfs_list_remove_request(req);
 215                 nfs_readpage_release(req);
 216         }
 217 }
 218
 219 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
 220         .error_cleanup = nfs_async_read_error,
 221         .completion = nfs_read_completion,
 222 };
 223
 224 /*
 225  * This is the callback from RPC telling us whether a reply was
 226  * received or some error occurred (timeout or socket shutdown).
 227  */
 228 static int nfs_readpage_done(struct rpc_task *task,
 229                              struct nfs_pgio_header *hdr,
 230                              struct inode *inode)
 231 {
 232         int status = NFS_PROTO(inode)->read_done(task, hdr);
 233         if (status != 0)
 234                 return status;
 235
 236         nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
 237         trace_nfs_readpage_done(inode, task->tk_status,
 238                                 hdr->args.offset, hdr->res.eof);
 239
 240         if (task->tk_status == -ESTALE) {
 241                 set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
 242                 nfs_mark_for_revalidate(inode);
 243         }
 244         return 0;
 245 }
 246
 247 static void nfs_readpage_retry(struct rpc_task *task,
 248                                struct nfs_pgio_header *hdr)
 249 {
 250         struct nfs_pgio_args *argp = &hdr->args;
 251         struct nfs_pgio_res  *resp = &hdr->res;
 252
 253         /* This is a short read! */
 254         nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
 255         /* Has the server at least made some progress? */
 256         if (resp->count == 0) {
 257                 nfs_set_pgio_error(hdr, -EIO, argp->offset);
 258                 return;
 259         }
 260
 261         /* For non rpc-based layout drivers, retry-through-MDS */
 262         if (!task->tk_ops) {
 263                 hdr->pnfs_error = -EAGAIN;
 264                 return;
 265         }
 266
 267         /* Yes, so retry the read at the end of the hdr */
 268         hdr->mds_offset += resp->count;
 269         argp->offset += resp->count;
 270         argp->pgbase += resp->count;
 271         argp->count -= resp->count;
 272         rpc_restart_call_prepare(task);
 273 }
 274
 275 static void nfs_readpage_result(struct rpc_task *task,
 276                                 struct nfs_pgio_header *hdr)
 277 {
 278         if (hdr->res.eof) {
 279                 loff_t bound;
 280
 281                 bound = hdr->args.offset + hdr->res.count;
 282                 spin_lock(&hdr->lock);
 283                 if (bound < hdr->io_start + hdr->good_bytes) {
 284                         set_bit(NFS_IOHDR_EOF, &hdr->flags);
 285                         clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
 286                         hdr->good_bytes = bound - hdr->io_start;
 287                 }
 288                 spin_unlock(&hdr->lock);
 289         } else if (hdr->res.count < hdr->args.count)
 290                 nfs_readpage_retry(task, hdr);
 291 }
 292
 293 /*
 294  * Read a page over NFS.
 295  * We read the page synchronously in the following case:
 296  *  -   The error flag is set for this page. This happens only when a
 297  *      previous async read operation failed.
 298  */
 299 int nfs_readpage(struct file *file, struct page *page)
 300 {
 301         struct nfs_open_context *ctx;
 302         struct inode *inode = page_file_mapping(page)->host;
 303         int             error;
 304
 305         dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
 306                 page, PAGE_SIZE, page_index(page));
 307         nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
 308         nfs_add_stats(inode, NFSIOS_READPAGES, 1);
 309
 310         /*
 311          * Try to flush any pending writes to the file..
 312          *
 313          * NOTE! Because we own the page lock, there cannot
 314          * be any new pending writes generated at this point
 315          * for this page (other pages can be written to).
 316          */
 317         error = nfs_wb_page(inode, page);
 318         if (error)
 319                 goto out_unlock;
 320         if (PageUptodate(page))
 321                 goto out_unlock;
 322
 323         error = -ESTALE;
 324         if (NFS_STALE(inode))
 325                 goto out_unlock;
 326
 327         if (file == NULL) {
 328                 error = -EBADF;
 329                 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 330                 if (ctx == NULL)
 331                         goto out_unlock;
 332         } else
 333                 ctx = get_nfs_open_context(nfs_file_open_context(file));
 334
 335         if (!IS_SYNC(inode)) {
 336                 error = nfs_readpage_from_fscache(ctx, inode, page);
 337                 if (error == 0)
 338                         goto out;
 339         }
 340
 341         error = nfs_readpage_async(ctx, inode, page);
 342
 343 out:
 344         put_nfs_open_context(ctx);
 345         return error;
 346 out_unlock:
 347         unlock_page(page);
 348         return error;
 349 }
 350
 351 struct nfs_readdesc {
 352         struct nfs_pageio_descriptor *pgio;
 353         struct nfs_open_context *ctx;
 354 };
 355
 356 static int
 357 readpage_async_filler(void *data, struct page *page)
 358 {
 359         struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
 360         struct nfs_page *new;
 361         unsigned int len;
 362         int error;
 363
 364         len = nfs_page_length(page);
 365         if (len == 0)
 366                 return nfs_return_empty_page(page);
 367
 368         new = nfs_create_request(desc->ctx, page, NULL, 0, len);
 369         if (IS_ERR(new))
 370                 goto out_error;
 371
 372         if (len < PAGE_SIZE)
 373                 zero_user_segment(page, len, PAGE_SIZE);
 374         if (!nfs_pageio_add_request(desc->pgio, new)) {
 375                 nfs_list_remove_request(new);
 376                 nfs_readpage_release(new);
 377                 error = desc->pgio->pg_error;
 378                 goto out;
 379         }
 380         return 0;
 381 out_error:
 382         error = PTR_ERR(new);
 383         unlock_page(page);
 384 out:
 385         return error;
 386 }
 387
 388 int nfs_readpages(struct file *filp, struct address_space *mapping,
 389                 struct list_head *pages, unsigned nr_pages)
 390 {
 391         struct nfs_pageio_descriptor pgio;
 392         struct nfs_pgio_mirror *pgm;
 393         struct nfs_readdesc desc = {
 394                 .pgio = &pgio,
 395         };
 396         struct inode *inode = mapping->host;
 397         unsigned long npages;
 398         int ret = -ESTALE;
 399
 400         dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
 401                         inode->i_sb->s_id,
 402                         (unsigned long long)NFS_FILEID(inode),
 403                         nr_pages);
 404         nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
 405
 406         if (NFS_STALE(inode))
 407                 goto out;
 408
 409         if (filp == NULL) {
 410                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 411                 if (desc.ctx == NULL)
 412                         return -EBADF;
 413         } else
 414                 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
 415
 416         /* attempt to read as many of the pages as possible from the cache
 417          * - this returns -ENOBUFS immediately if the cookie is negative
 418          */
 419         ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
 420                                          pages, &nr_pages);
 421         if (ret == 0)
 422                 goto read_complete; /* all pages were read */
 423
 424         nfs_pageio_init_read(&pgio, inode, false,
 425                              &nfs_async_read_completion_ops);
 426
 427         ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
 428         nfs_pageio_complete(&pgio);
 429
 430         /* It doesn't make sense to do mirrored reads! */
 431         WARN_ON_ONCE(pgio.pg_mirror_count != 1);
 432
 433         pgm = &pgio.pg_mirrors[0];
 434         NFS_I(inode)->read_io += pgm->pg_bytes_written;
 435         npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >>
 436                  PAGE_SHIFT;
 437         nfs_add_stats(inode, NFSIOS_READPAGES, npages);
 438 read_complete:
 439         put_nfs_open_context(desc.ctx);
 440 out:
 441         return ret;
 442 }
 443
 444 int __init nfs_init_readpagecache(void)
 445 {
 446         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
 447                                              sizeof(struct nfs_pgio_header),
 448                                              0, SLAB_HWCACHE_ALIGN,
 449                                              NULL);
 450         if (nfs_rdata_cachep == NULL)
 451                 return -ENOMEM;
 452
 453         return 0;
 454 }
 455
 456 void nfs_destroy_readpagecache(void)
 457 {
 458         kmem_cache_destroy(nfs_rdata_cachep);
 459 }
 460
 461 static const struct nfs_rw_ops nfs_rw_read_ops = {
 462         .rw_alloc_header        = nfs_readhdr_alloc,
 463         .rw_free_header         = nfs_readhdr_free,
 464         .rw_done                = nfs_readpage_done,
 465         .rw_result              = nfs_readpage_result,
 466         .rw_initiate            = nfs_initiate_read,
 467 };