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