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