1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/fs/nfs/read.c 4 * 5 * Block I/O for NFS 6 * 7 * Partial copy of Linus' read cache modifications to fs/nfs/file.c 8 * modified for async RPC by okir@monad.swb.de 9 */ 10 11 #include <linux/time.h> 12 #include <linux/kernel.h> 13 #include <linux/errno.h> 14 #include <linux/fcntl.h> 15 #include <linux/stat.h> 16 #include <linux/mm.h> 17 #include <linux/slab.h> 18 #include <linux/task_io_accounting_ops.h> 19 #include <linux/pagemap.h> 20 #include <linux/sunrpc/clnt.h> 21 #include <linux/nfs_fs.h> 22 #include <linux/nfs_page.h> 23 #include <linux/module.h> 24 25 #include "nfs4_fs.h" 26 #include "internal.h" 27 #include "iostat.h" 28 #include "fscache.h" 29 #include "pnfs.h" 30 #include "nfstrace.h" 31 #include "delegation.h" 32 33 #define NFSDBG_FACILITY NFSDBG_PAGECACHE 34 35 const struct nfs_pgio_completion_ops nfs_async_read_completion_ops; 36 static const struct nfs_rw_ops nfs_rw_read_ops; 37 38 static struct kmem_cache *nfs_rdata_cachep; 39 40 static struct nfs_pgio_header *nfs_readhdr_alloc(void) 41 { 42 struct nfs_pgio_header *p = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL); 43 44 if (p) 45 p->rw_mode = FMODE_READ; 46 return p; 47 } 48 49 static void nfs_readhdr_free(struct nfs_pgio_header *rhdr) 50 { 51 kfree(rhdr->res.scratch); 52 kmem_cache_free(nfs_rdata_cachep, rhdr); 53 } 54 55 static int nfs_return_empty_folio(struct folio *folio) 56 { 57 folio_zero_segment(folio, 0, folio_size(folio)); 58 folio_mark_uptodate(folio); 59 if (nfs_netfs_folio_unlock(folio)) 60 folio_unlock(folio); 61 return 0; 62 } 63 64 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, 65 struct inode *inode, bool force_mds, 66 const struct nfs_pgio_completion_ops *compl_ops) 67 { 68 struct nfs_server *server = NFS_SERVER(inode); 69 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops; 70 71 #if IS_ENABLED(CONFIG_NFS_V4) 72 if (server->pnfs_curr_ld && !force_mds) 73 pg_ops = server->pnfs_curr_ld->pg_read_ops; 74 #endif /* CONFIG_NFS_V4 */ 75 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops, 76 server->rsize, 0); 77 } 78 EXPORT_SYMBOL_GPL(nfs_pageio_init_read); 79 80 void nfs_pageio_complete_read(struct nfs_pageio_descriptor *pgio) 81 { 82 struct nfs_pgio_mirror *pgm; 83 unsigned long npages; 84 85 nfs_pageio_complete(pgio); 86 87 /* It doesn't make sense to do mirrored reads! */ 88 WARN_ON_ONCE(pgio->pg_mirror_count != 1); 89 90 pgm = &pgio->pg_mirrors[0]; 91 NFS_I(pgio->pg_inode)->read_io += pgm->pg_bytes_written; 92 npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >> PAGE_SHIFT; 93 nfs_add_stats(pgio->pg_inode, NFSIOS_READPAGES, npages); 94 } 95 96 97 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio) 98 { 99 struct nfs_pgio_mirror *mirror; 100 101 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup) 102 pgio->pg_ops->pg_cleanup(pgio); 103 104 pgio->pg_ops = &nfs_pgio_rw_ops; 105 106 /* read path should never have more than one mirror */ 107 WARN_ON_ONCE(pgio->pg_mirror_count != 1); 108 109 mirror = &pgio->pg_mirrors[0]; 110 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize; 111 } 112 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds); 113 114 bool nfs_read_alloc_scratch(struct nfs_pgio_header *hdr, size_t size) 115 { 116 WARN_ON(hdr->res.scratch != NULL); 117 hdr->res.scratch = kmalloc(size, GFP_KERNEL); 118 return hdr->res.scratch != NULL; 119 } 120 EXPORT_SYMBOL_GPL(nfs_read_alloc_scratch); 121 122 static void nfs_readpage_release(struct nfs_page *req, int error) 123 { 124 struct folio *folio = nfs_page_to_folio(req); 125 126 if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) 127 if (nfs_netfs_folio_unlock(folio)) 128 folio_unlock(folio); 129 130 nfs_release_request(req); 131 } 132 133 static void nfs_page_group_set_uptodate(struct nfs_page *req) 134 { 135 bool uptodate = false; 136 137 nfs_page_group_lock(req); 138 if (!test_bit(PG_READ_FAILED, &req->wb_head->wb_flags) && 139 nfs_page_group_sync_on_bit_locked(req, PG_UPTODATE)) 140 uptodate = true; 141 nfs_page_group_unlock(req); 142 143 if (uptodate) 144 folio_mark_uptodate(nfs_page_to_folio(req)); 145 } 146 147 static void nfs_page_group_mark_read_failed(struct nfs_page *req) 148 { 149 struct nfs_page *tmp; 150 151 nfs_page_group_lock(req); 152 set_bit(PG_READ_FAILED, &req->wb_head->wb_flags); 153 tmp = req; 154 do { 155 clear_bit(PG_UPTODATE, &tmp->wb_flags); 156 tmp = tmp->wb_this_page; 157 } while (tmp != req); 158 nfs_page_group_unlock(req); 159 } 160 161 static void nfs_read_completion(struct nfs_pgio_header *hdr) 162 { 163 unsigned long bytes = 0; 164 int error; 165 166 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) 167 goto out; 168 while (!list_empty(&hdr->pages)) { 169 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 170 struct folio *folio = nfs_page_to_folio(req); 171 unsigned long start = req->wb_pgbase; 172 unsigned long end = req->wb_pgbase + req->wb_bytes; 173 174 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) { 175 /* note: regions of the page not covered by a 176 * request are zeroed in nfs_read_add_folio 177 */ 178 if (bytes > hdr->good_bytes) { 179 /* nothing in this request was good, so zero 180 * the full extent of the request */ 181 folio_zero_segment(folio, start, end); 182 183 } else if (hdr->good_bytes - bytes < req->wb_bytes) { 184 /* part of this request has good bytes, but 185 * not all. zero the bad bytes */ 186 start += hdr->good_bytes - bytes; 187 WARN_ON(start < req->wb_pgbase); 188 folio_zero_segment(folio, start, end); 189 } 190 } 191 error = 0; 192 bytes += req->wb_bytes; 193 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { 194 if (bytes <= hdr->good_bytes) 195 nfs_page_group_set_uptodate(req); 196 else { 197 nfs_page_group_mark_read_failed(req); 198 error = hdr->error; 199 xchg(&nfs_req_openctx(req)->error, error); 200 } 201 } else 202 nfs_page_group_set_uptodate(req); 203 nfs_list_remove_request(req); 204 nfs_readpage_release(req, error); 205 } 206 nfs_netfs_read_completion(hdr); 207 208 out: 209 hdr->release(hdr); 210 } 211 212 static void nfs_initiate_read(struct nfs_pgio_header *hdr, 213 struct rpc_message *msg, 214 const struct nfs_rpc_ops *rpc_ops, 215 struct rpc_task_setup *task_setup_data, int how) 216 { 217 rpc_ops->read_setup(hdr, msg); 218 nfs_netfs_initiate_read(hdr); 219 trace_nfs_initiate_read(hdr); 220 } 221 222 static void 223 nfs_async_read_error(struct list_head *head, int error) 224 { 225 struct nfs_page *req; 226 227 while (!list_empty(head)) { 228 req = nfs_list_entry(head->next); 229 nfs_list_remove_request(req); 230 nfs_readpage_release(req, error); 231 } 232 } 233 234 const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = { 235 .error_cleanup = nfs_async_read_error, 236 .completion = nfs_read_completion, 237 }; 238 239 /* 240 * This is the callback from RPC telling us whether a reply was 241 * received or some error occurred (timeout or socket shutdown). 242 */ 243 static int nfs_readpage_done(struct rpc_task *task, 244 struct nfs_pgio_header *hdr, 245 struct inode *inode) 246 { 247 int status = NFS_PROTO(inode)->read_done(task, hdr); 248 if (status != 0) 249 return status; 250 251 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count); 252 trace_nfs_readpage_done(task, hdr); 253 254 if (task->tk_status == -ESTALE) { 255 nfs_set_inode_stale(inode); 256 nfs_mark_for_revalidate(inode); 257 } 258 return 0; 259 } 260 261 static void nfs_readpage_retry(struct rpc_task *task, 262 struct nfs_pgio_header *hdr) 263 { 264 struct nfs_pgio_args *argp = &hdr->args; 265 struct nfs_pgio_res *resp = &hdr->res; 266 267 /* This is a short read! */ 268 nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD); 269 trace_nfs_readpage_short(task, hdr); 270 271 /* Has the server at least made some progress? */ 272 if (resp->count == 0) { 273 nfs_set_pgio_error(hdr, -EIO, argp->offset); 274 return; 275 } 276 277 /* For non rpc-based layout drivers, retry-through-MDS */ 278 if (!task->tk_ops) { 279 hdr->pnfs_error = -EAGAIN; 280 return; 281 } 282 283 /* Yes, so retry the read at the end of the hdr */ 284 hdr->mds_offset += resp->count; 285 argp->offset += resp->count; 286 argp->pgbase += resp->count; 287 argp->count -= resp->count; 288 resp->count = 0; 289 resp->eof = 0; 290 rpc_restart_call_prepare(task); 291 } 292 293 static void nfs_readpage_result(struct rpc_task *task, 294 struct nfs_pgio_header *hdr) 295 { 296 if (hdr->res.eof) { 297 loff_t pos = hdr->args.offset + hdr->res.count; 298 unsigned int new = pos - hdr->io_start; 299 300 if (hdr->good_bytes > new) { 301 hdr->good_bytes = new; 302 set_bit(NFS_IOHDR_EOF, &hdr->flags); 303 clear_bit(NFS_IOHDR_ERROR, &hdr->flags); 304 } 305 } else if (hdr->res.count < hdr->args.count) 306 nfs_readpage_retry(task, hdr); 307 } 308 309 int nfs_read_add_folio(struct nfs_pageio_descriptor *pgio, 310 struct nfs_open_context *ctx, 311 struct folio *folio) 312 { 313 struct inode *inode = folio->mapping->host; 314 struct nfs_server *server = NFS_SERVER(inode); 315 size_t fsize = folio_size(folio); 316 unsigned int rsize = server->rsize; 317 struct nfs_page *new; 318 unsigned int len, aligned_len; 319 int error; 320 321 len = nfs_folio_length(folio); 322 if (len == 0) 323 return nfs_return_empty_folio(folio); 324 325 aligned_len = min_t(unsigned int, ALIGN(len, rsize), fsize); 326 327 new = nfs_page_create_from_folio(ctx, folio, 0, aligned_len); 328 if (IS_ERR(new)) { 329 error = PTR_ERR(new); 330 if (nfs_netfs_folio_unlock(folio)) 331 folio_unlock(folio); 332 goto out; 333 } 334 335 if (len < fsize) 336 folio_zero_segment(folio, len, fsize); 337 if (!nfs_pageio_add_request(pgio, new)) { 338 nfs_list_remove_request(new); 339 error = pgio->pg_error; 340 nfs_readpage_release(new, error); 341 goto out; 342 } 343 return 0; 344 out: 345 return error; 346 } 347 348 /* 349 * Actually read a folio over the wire. 350 */ 351 static int nfs_do_read_folio(struct file *file, struct folio *folio) 352 { 353 struct inode *inode = file_inode(file); 354 struct nfs_pageio_descriptor pgio; 355 struct nfs_open_context *ctx; 356 int ret; 357 358 ctx = get_nfs_open_context(nfs_file_open_context(file)); 359 360 xchg(&ctx->error, 0); 361 nfs_pageio_init_read(&pgio, inode, false, 362 &nfs_async_read_completion_ops); 363 364 ret = nfs_read_add_folio(&pgio, ctx, folio); 365 if (ret) 366 goto out_put; 367 368 nfs_pageio_complete_read(&pgio); 369 nfs_update_delegated_atime(inode); 370 if (pgio.pg_error < 0) { 371 ret = pgio.pg_error; 372 goto out_put; 373 } 374 375 ret = folio_wait_locked_killable(folio); 376 if (!folio_test_uptodate(folio) && !ret) 377 ret = xchg(&ctx->error, 0); 378 379 out_put: 380 put_nfs_open_context(ctx); 381 return ret; 382 } 383 384 /* 385 * Synchronously read a folio. 386 * 387 * This is not heavily used as most users to try an asynchronous 388 * large read through ->readahead first. 389 */ 390 int nfs_read_folio(struct file *file, struct folio *folio) 391 { 392 struct inode *inode = file_inode(file); 393 loff_t pos = folio_pos(folio); 394 size_t len = folio_size(folio); 395 int ret; 396 397 trace_nfs_aop_readpage(inode, pos, len); 398 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE); 399 task_io_account_read(len); 400 401 /* 402 * Try to flush any pending writes to the file.. 403 * 404 * NOTE! Because we own the folio lock, there cannot 405 * be any new pending writes generated at this point 406 * for this folio (other folios can be written to). 407 */ 408 ret = nfs_wb_folio(inode, folio); 409 if (ret) 410 goto out_unlock; 411 if (folio_test_uptodate(folio)) 412 goto out_unlock; 413 414 ret = -ESTALE; 415 if (NFS_STALE(inode)) 416 goto out_unlock; 417 418 ret = nfs_netfs_read_folio(file, folio); 419 if (ret) 420 ret = nfs_do_read_folio(file, folio); 421 out: 422 trace_nfs_aop_readpage_done(inode, pos, len, ret); 423 return ret; 424 out_unlock: 425 folio_unlock(folio); 426 goto out; 427 } 428 429 void nfs_readahead(struct readahead_control *ractl) 430 { 431 struct nfs_pageio_descriptor pgio; 432 struct nfs_open_context *ctx; 433 unsigned int nr_pages = readahead_count(ractl); 434 struct file *file = ractl->file; 435 struct inode *inode = ractl->mapping->host; 436 struct folio *folio; 437 int ret; 438 439 trace_nfs_aop_readahead(inode, readahead_pos(ractl), nr_pages); 440 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES); 441 task_io_account_read(readahead_length(ractl)); 442 443 ret = -ESTALE; 444 if (NFS_STALE(inode)) 445 goto out; 446 447 ret = nfs_netfs_readahead(ractl); 448 if (!ret) 449 goto out; 450 451 if (file == NULL) { 452 ret = -EBADF; 453 ctx = nfs_find_open_context(inode, NULL, FMODE_READ); 454 if (ctx == NULL) 455 goto out; 456 } else 457 ctx = get_nfs_open_context(nfs_file_open_context(file)); 458 459 nfs_pageio_init_read(&pgio, inode, false, 460 &nfs_async_read_completion_ops); 461 462 while ((folio = readahead_folio(ractl)) != NULL) { 463 ret = nfs_read_add_folio(&pgio, ctx, folio); 464 if (ret) 465 break; 466 } 467 468 nfs_pageio_complete_read(&pgio); 469 nfs_update_delegated_atime(inode); 470 471 put_nfs_open_context(ctx); 472 out: 473 trace_nfs_aop_readahead_done(inode, nr_pages, ret); 474 } 475 476 int __init nfs_init_readpagecache(void) 477 { 478 nfs_rdata_cachep = kmem_cache_create("nfs_read_data", 479 sizeof(struct nfs_pgio_header), 480 0, SLAB_HWCACHE_ALIGN, 481 NULL); 482 if (nfs_rdata_cachep == NULL) 483 return -ENOMEM; 484 485 return 0; 486 } 487 488 void nfs_destroy_readpagecache(void) 489 { 490 kmem_cache_destroy(nfs_rdata_cachep); 491 } 492 493 static const struct nfs_rw_ops nfs_rw_read_ops = { 494 .rw_alloc_header = nfs_readhdr_alloc, 495 .rw_free_header = nfs_readhdr_free, 496 .rw_done = nfs_readpage_done, 497 .rw_result = nfs_readpage_result, 498 .rw_initiate = nfs_initiate_read, 499 }; 500