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_pgio_completion_ops nfs_async_read_completion_ops;
  32 static const struct nfs_rw_ops nfs_rw_read_ops;
  33
  34 static struct kmem_cache *nfs_rdata_cachep;
  35
  36 static struct nfs_pgio_header *nfs_readhdr_alloc(void)
  37 {
  38         struct nfs_pgio_header *p = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
  39
  40         if (p)
  41                 p->rw_mode = FMODE_READ;
  42         return p;
  43 }
  44
  45 static void nfs_readhdr_free(struct nfs_pgio_header *rhdr)
  46 {
  47         kmem_cache_free(nfs_rdata_cachep, rhdr);
  48 }
  49
  50 static
  51 int nfs_return_empty_page(struct page *page)
  52 {
  53         zero_user(page, 0, PAGE_SIZE);
  54         SetPageUptodate(page);
  55         unlock_page(page);
  56         return 0;
  57 }
  58
  59 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
  60                               struct inode *inode, bool force_mds,
  61                               const struct nfs_pgio_completion_ops *compl_ops)
  62 {
  63         struct nfs_server *server = NFS_SERVER(inode);
  64         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
  65
  66 #ifdef CONFIG_NFS_V4_1
  67         if (server->pnfs_curr_ld && !force_mds)
  68                 pg_ops = server->pnfs_curr_ld->pg_read_ops;
  69 #endif
  70         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops,
  71                         server->rsize, 0, GFP_KERNEL);
  72 }
  73 EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
  74
  75 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
  76 {
  77         struct nfs_pgio_mirror *mirror;
  78
  79         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
  80                 pgio->pg_ops->pg_cleanup(pgio);
  81
  82         pgio->pg_ops = &nfs_pgio_rw_ops;
  83
  84         /* read path should never have more than one mirror */
  85         WARN_ON_ONCE(pgio->pg_mirror_count != 1);
  86
  87         mirror = &pgio->pg_mirrors[0];
  88         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
  89 }
  90 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
  91
  92 static void nfs_readpage_release(struct nfs_page *req)
  93 {
  94         struct inode *inode = d_inode(req->wb_context->dentry);
  95
  96         dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
  97                 (unsigned long long)NFS_FILEID(inode), req->wb_bytes,
  98                 (long long)req_offset(req));
  99
 100         if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
 101                 if (PageUptodate(req->wb_page))
 102                         nfs_readpage_to_fscache(inode, req->wb_page, 0);
 103
 104                 unlock_page(req->wb_page);
 105         }
 106         nfs_release_request(req);
 107 }
 108
 109 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
 110                        struct page *page)
 111 {
 112         struct nfs_page *new;
 113         unsigned int len;
 114         struct nfs_pageio_descriptor pgio;
 115         struct nfs_pgio_mirror *pgm;
 116
 117         len = nfs_page_length(page);
 118         if (len == 0)
 119                 return nfs_return_empty_page(page);
 120         new = nfs_create_request(ctx, page, NULL, 0, len);
 121         if (IS_ERR(new)) {
 122                 unlock_page(page);
 123                 return PTR_ERR(new);
 124         }
 125         if (len < PAGE_SIZE)
 126                 zero_user_segment(page, len, PAGE_SIZE);
 127
 128         nfs_pageio_init_read(&pgio, inode, false,
 129                              &nfs_async_read_completion_ops);
 130         if (!nfs_pageio_add_request(&pgio, new)) {
 131                 nfs_list_remove_request(new);
 132                 nfs_readpage_release(new);
 133         }
 134         nfs_pageio_complete(&pgio);
 135
 136         /* It doesn't make sense to do mirrored reads! */
 137         WARN_ON_ONCE(pgio.pg_mirror_count != 1);
 138
 139         pgm = &pgio.pg_mirrors[0];
 140         NFS_I(inode)->read_io += pgm->pg_bytes_written;
 141
 142         return pgio.pg_error < 0 ? pgio.pg_error : 0;
 143 }
 144
 145 static void nfs_page_group_set_uptodate(struct nfs_page *req)
 146 {
 147         if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
 148                 SetPageUptodate(req->wb_page);
 149 }
 150
 151 static void nfs_read_completion(struct nfs_pgio_header *hdr)
 152 {
 153         unsigned long bytes = 0;
 154
 155         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 156                 goto out;
 157         while (!list_empty(&hdr->pages)) {
 158                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 159                 struct page *page = req->wb_page;
 160                 unsigned long start = req->wb_pgbase;
 161                 unsigned long end = req->wb_pgbase + req->wb_bytes;
 162
 163                 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
 164                         /* note: regions of the page not covered by a
 165                          * request are zeroed in nfs_readpage_async /
 166                          * readpage_async_filler */
 167                         if (bytes > hdr->good_bytes) {
 168                                 /* nothing in this request was good, so zero
 169                                  * the full extent of the request */
 170                                 zero_user_segment(page, start, end);
 171
 172                         } else if (hdr->good_bytes - bytes < req->wb_bytes) {
 173                                 /* part of this request has good bytes, but
 174                                  * not all. zero the bad bytes */
 175                                 start += hdr->good_bytes - bytes;
 176                                 WARN_ON(start < req->wb_pgbase);
 177                                 zero_user_segment(page, start, end);
 178                         }
 179                 }
 180                 bytes += req->wb_bytes;
 181                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 182                         if (bytes <= hdr->good_bytes)
 183                                 nfs_page_group_set_uptodate(req);
 184                 } else
 185                         nfs_page_group_set_uptodate(req);
 186                 nfs_list_remove_request(req);
 187                 nfs_readpage_release(req);
 188         }
 189 out:
 190         hdr->release(hdr);
 191 }
 192
 193 static void nfs_initiate_read(struct nfs_pgio_header *hdr,
 194                               struct rpc_message *msg,
 195                               const struct nfs_rpc_ops *rpc_ops,
 196                               struct rpc_task_setup *task_setup_data, int how)
 197 {
 198         struct inode *inode = hdr->inode;
 199         int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
 200
 201         task_setup_data->flags |= swap_flags;
 202         rpc_ops->read_setup(hdr, msg);
 203 }
 204
 205 static void
 206 nfs_async_read_error(struct list_head *head)
 207 {
 208         struct nfs_page *req;
 209
 210         while (!list_empty(head)) {
 211                 req = nfs_list_entry(head->next);
 212                 nfs_list_remove_request(req);
 213                 nfs_readpage_release(req);
 214         }
 215 }
 216
 217 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
 218         .error_cleanup = nfs_async_read_error,
 219         .completion = nfs_read_completion,
 220 };
 221
 222 /*
 223  * This is the callback from RPC telling us whether a reply was
 224  * received or some error occurred (timeout or socket shutdown).
 225  */
 226 static int nfs_readpage_done(struct rpc_task *task,
 227                              struct nfs_pgio_header *hdr,
 228                              struct inode *inode)
 229 {
 230         int status = NFS_PROTO(inode)->read_done(task, hdr);
 231         if (status != 0)
 232                 return status;
 233
 234         nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
 235
 236         if (task->tk_status == -ESTALE) {
 237                 set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
 238                 nfs_mark_for_revalidate(inode);
 239         }
 240         return 0;
 241 }
 242
 243 static void nfs_readpage_retry(struct rpc_task *task,
 244                                struct nfs_pgio_header *hdr)
 245 {
 246         struct nfs_pgio_args *argp = &hdr->args;
 247         struct nfs_pgio_res  *resp = &hdr->res;
 248
 249         /* This is a short read! */
 250         nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
 251         /* Has the server at least made some progress? */
 252         if (resp->count == 0) {
 253                 nfs_set_pgio_error(hdr, -EIO, argp->offset);
 254                 return;
 255         }
 256
 257         /* For non rpc-based layout drivers, retry-through-MDS */
 258         if (!task->tk_ops) {
 259                 hdr->pnfs_error = -EAGAIN;
 260                 return;
 261         }
 262
 263         /* Yes, so retry the read at the end of the hdr */
 264         hdr->mds_offset += resp->count;
 265         argp->offset += resp->count;
 266         argp->pgbase += resp->count;
 267         argp->count -= resp->count;
 268         rpc_restart_call_prepare(task);
 269 }
 270
 271 static void nfs_readpage_result(struct rpc_task *task,
 272                                 struct nfs_pgio_header *hdr)
 273 {
 274         if (hdr->res.eof) {
 275                 loff_t bound;
 276
 277                 bound = hdr->args.offset + hdr->res.count;
 278                 spin_lock(&hdr->lock);
 279                 if (bound < hdr->io_start + hdr->good_bytes) {
 280                         set_bit(NFS_IOHDR_EOF, &hdr->flags);
 281                         clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
 282                         hdr->good_bytes = bound - hdr->io_start;
 283                 }
 284                 spin_unlock(&hdr->lock);
 285         } else if (hdr->res.count < hdr->args.count)
 286                 nfs_readpage_retry(task, hdr);
 287 }
 288
 289 /*
 290  * Read a page over NFS.
 291  * We read the page synchronously in the following case:
 292  *  -   The error flag is set for this page. This happens only when a
 293  *      previous async read operation failed.
 294  */
 295 int nfs_readpage(struct file *file, struct page *page)
 296 {
 297         struct nfs_open_context *ctx;
 298         struct inode *inode = page_file_mapping(page)->host;
 299         int             error;
 300
 301         dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
 302                 page, PAGE_SIZE, page_index(page));
 303         nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
 304         nfs_add_stats(inode, NFSIOS_READPAGES, 1);
 305
 306         /*
 307          * Try to flush any pending writes to the file..
 308          *
 309          * NOTE! Because we own the page lock, there cannot
 310          * be any new pending writes generated at this point
 311          * for this page (other pages can be written to).
 312          */
 313         error = nfs_wb_page(inode, page);
 314         if (error)
 315                 goto out_unlock;
 316         if (PageUptodate(page))
 317                 goto out_unlock;
 318
 319         error = -ESTALE;
 320         if (NFS_STALE(inode))
 321                 goto out_unlock;
 322
 323         if (file == NULL) {
 324                 error = -EBADF;
 325                 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 326                 if (ctx == NULL)
 327                         goto out_unlock;
 328         } else
 329                 ctx = get_nfs_open_context(nfs_file_open_context(file));
 330
 331         if (!IS_SYNC(inode)) {
 332                 error = nfs_readpage_from_fscache(ctx, inode, page);
 333                 if (error == 0)
 334                         goto out;
 335         }
 336
 337         error = nfs_readpage_async(ctx, inode, page);
 338
 339 out:
 340         put_nfs_open_context(ctx);
 341         return error;
 342 out_unlock:
 343         unlock_page(page);
 344         return error;
 345 }
 346
 347 struct nfs_readdesc {
 348         struct nfs_pageio_descriptor *pgio;
 349         struct nfs_open_context *ctx;
 350 };
 351
 352 static int
 353 readpage_async_filler(void *data, struct page *page)
 354 {
 355         struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
 356         struct nfs_page *new;
 357         unsigned int len;
 358         int error;
 359
 360         len = nfs_page_length(page);
 361         if (len == 0)
 362                 return nfs_return_empty_page(page);
 363
 364         new = nfs_create_request(desc->ctx, page, NULL, 0, len);
 365         if (IS_ERR(new))
 366                 goto out_error;
 367
 368         if (len < PAGE_SIZE)
 369                 zero_user_segment(page, len, PAGE_SIZE);
 370         if (!nfs_pageio_add_request(desc->pgio, new)) {
 371                 nfs_list_remove_request(new);
 372                 nfs_readpage_release(new);
 373                 error = desc->pgio->pg_error;
 374                 goto out;
 375         }
 376         return 0;
 377 out_error:
 378         error = PTR_ERR(new);
 379         unlock_page(page);
 380 out:
 381         return error;
 382 }
 383
 384 int nfs_readpages(struct file *filp, struct address_space *mapping,
 385                 struct list_head *pages, unsigned nr_pages)
 386 {
 387         struct nfs_pageio_descriptor pgio;
 388         struct nfs_pgio_mirror *pgm;
 389         struct nfs_readdesc desc = {
 390                 .pgio = &pgio,
 391         };
 392         struct inode *inode = mapping->host;
 393         unsigned long npages;
 394         int ret = -ESTALE;
 395
 396         dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
 397                         inode->i_sb->s_id,
 398                         (unsigned long long)NFS_FILEID(inode),
 399                         nr_pages);
 400         nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
 401
 402         if (NFS_STALE(inode))
 403                 goto out;
 404
 405         if (filp == NULL) {
 406                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 407                 if (desc.ctx == NULL)
 408                         return -EBADF;
 409         } else
 410                 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
 411
 412         /* attempt to read as many of the pages as possible from the cache
 413          * - this returns -ENOBUFS immediately if the cookie is negative
 414          */
 415         ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
 416                                          pages, &nr_pages);
 417         if (ret == 0)
 418                 goto read_complete; /* all pages were read */
 419
 420         nfs_pageio_init_read(&pgio, inode, false,
 421                              &nfs_async_read_completion_ops);
 422
 423         ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
 424         nfs_pageio_complete(&pgio);
 425
 426         /* It doesn't make sense to do mirrored reads! */
 427         WARN_ON_ONCE(pgio.pg_mirror_count != 1);
 428
 429         pgm = &pgio.pg_mirrors[0];
 430         NFS_I(inode)->read_io += pgm->pg_bytes_written;
 431         npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >>
 432                  PAGE_SHIFT;
 433         nfs_add_stats(inode, NFSIOS_READPAGES, npages);
 434 read_complete:
 435         put_nfs_open_context(desc.ctx);
 436 out:
 437         return ret;
 438 }
 439
 440 int __init nfs_init_readpagecache(void)
 441 {
 442         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
 443                                              sizeof(struct nfs_pgio_header),
 444                                              0, SLAB_HWCACHE_ALIGN,
 445                                              NULL);
 446         if (nfs_rdata_cachep == NULL)
 447                 return -ENOMEM;
 448
 449         return 0;
 450 }
 451
 452 void nfs_destroy_readpagecache(void)
 453 {
 454         kmem_cache_destroy(nfs_rdata_cachep);
 455 }
 456
 457 static const struct nfs_rw_ops nfs_rw_read_ops = {
 458         .rw_alloc_header        = nfs_readhdr_alloc,
 459         .rw_free_header         = nfs_readhdr_free,
 460         .rw_done                = nfs_readpage_done,
 461         .rw_result              = nfs_readpage_result,
 462         .rw_initiate            = nfs_initiate_read,
 463 };