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