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 folio_unlock(folio); 60 return 0; 61 } 62 63 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, 64 struct inode *inode, bool force_mds, 65 const struct nfs_pgio_completion_ops *compl_ops) 66 { 67 struct nfs_server *server = NFS_SERVER(inode); 68 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops; 69 70 #ifdef CONFIG_NFS_V4_1 71 if (server->pnfs_curr_ld && !force_mds) 72 pg_ops = server->pnfs_curr_ld->pg_read_ops; 73 #endif 74 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops, 75 server->rsize, 0); 76 } 77 EXPORT_SYMBOL_GPL(nfs_pageio_init_read); 78 79 void nfs_pageio_complete_read(struct nfs_pageio_descriptor *pgio) 80 { 81 struct nfs_pgio_mirror *pgm; 82 unsigned long npages; 83 84 nfs_pageio_complete(pgio); 85 86 /* It doesn't make sense to do mirrored reads! */ 87 WARN_ON_ONCE(pgio->pg_mirror_count != 1); 88 89 pgm = &pgio->pg_mirrors[0]; 90 NFS_I(pgio->pg_inode)->read_io += pgm->pg_bytes_written; 91 npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >> PAGE_SHIFT; 92 nfs_add_stats(pgio->pg_inode, NFSIOS_READPAGES, npages); 93 } 94 95 96 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio) 97 { 98 struct nfs_pgio_mirror *mirror; 99 100 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup) 101 pgio->pg_ops->pg_cleanup(pgio); 102 103 pgio->pg_ops = &nfs_pgio_rw_ops; 104 105 /* read path should never have more than one mirror */ 106 WARN_ON_ONCE(pgio->pg_mirror_count != 1); 107 108 mirror = &pgio->pg_mirrors[0]; 109 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize; 110 } 111 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds); 112 113 bool nfs_read_alloc_scratch(struct nfs_pgio_header *hdr, size_t size) 114 { 115 WARN_ON(hdr->res.scratch != NULL); 116 hdr->res.scratch = kmalloc(size, GFP_KERNEL); 117 return hdr->res.scratch != NULL; 118 } 119 EXPORT_SYMBOL_GPL(nfs_read_alloc_scratch); 120 121 static void nfs_readpage_release(struct nfs_page *req, int error) 122 { 123 struct folio *folio = nfs_page_to_folio(req); 124 125 if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) 126 if (nfs_netfs_folio_unlock(folio)) 127 folio_unlock(folio); 128 129 nfs_release_request(req); 130 } 131 132 static void nfs_page_group_set_uptodate(struct nfs_page *req) 133 { 134 if (nfs_page_group_sync_on_bit(req, PG_UPTODATE)) 135 folio_mark_uptodate(nfs_page_to_folio(req)); 136 } 137 138 static void nfs_read_completion(struct nfs_pgio_header *hdr) 139 { 140 unsigned long bytes = 0; 141 int error; 142 143 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) 144 goto out; 145 while (!list_empty(&hdr->pages)) { 146 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 147 struct folio *folio = nfs_page_to_folio(req); 148 unsigned long start = req->wb_pgbase; 149 unsigned long end = req->wb_pgbase + req->wb_bytes; 150 151 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) { 152 /* note: regions of the page not covered by a 153 * request are zeroed in nfs_read_add_folio 154 */ 155 if (bytes > hdr->good_bytes) { 156 /* nothing in this request was good, so zero 157 * the full extent of the request */ 158 folio_zero_segment(folio, start, end); 159 160 } else if (hdr->good_bytes - bytes < req->wb_bytes) { 161 /* part of this request has good bytes, but 162 * not all. zero the bad bytes */ 163 start += hdr->good_bytes - bytes; 164 WARN_ON(start < req->wb_pgbase); 165 folio_zero_segment(folio, start, end); 166 } 167 } 168 error = 0; 169 bytes += req->wb_bytes; 170 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { 171 if (bytes <= hdr->good_bytes) 172 nfs_page_group_set_uptodate(req); 173 else { 174 error = hdr->error; 175 xchg(&nfs_req_openctx(req)->error, error); 176 } 177 } else 178 nfs_page_group_set_uptodate(req); 179 nfs_list_remove_request(req); 180 nfs_readpage_release(req, error); 181 } 182 nfs_netfs_read_completion(hdr); 183 184 out: 185 hdr->release(hdr); 186 } 187 188 static void nfs_initiate_read(struct nfs_pgio_header *hdr, 189 struct rpc_message *msg, 190 const struct nfs_rpc_ops *rpc_ops, 191 struct rpc_task_setup *task_setup_data, int how) 192 { 193 rpc_ops->read_setup(hdr, msg); 194 nfs_netfs_initiate_read(hdr); 195 trace_nfs_initiate_read(hdr); 196 } 197 198 static void 199 nfs_async_read_error(struct list_head *head, int error) 200 { 201 struct nfs_page *req; 202 203 while (!list_empty(head)) { 204 req = nfs_list_entry(head->next); 205 nfs_list_remove_request(req); 206 nfs_readpage_release(req, error); 207 } 208 } 209 210 const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = { 211 .error_cleanup = nfs_async_read_error, 212 .completion = nfs_read_completion, 213 }; 214 215 /* 216 * This is the callback from RPC telling us whether a reply was 217 * received or some error occurred (timeout or socket shutdown). 218 */ 219 static int nfs_readpage_done(struct rpc_task *task, 220 struct nfs_pgio_header *hdr, 221 struct inode *inode) 222 { 223 int status = NFS_PROTO(inode)->read_done(task, hdr); 224 if (status != 0) 225 return status; 226 227 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count); 228 trace_nfs_readpage_done(task, hdr); 229 230 if (task->tk_status == -ESTALE) { 231 nfs_set_inode_stale(inode); 232 nfs_mark_for_revalidate(inode); 233 } 234 return 0; 235 } 236 237 static void nfs_readpage_retry(struct rpc_task *task, 238 struct nfs_pgio_header *hdr) 239 { 240 struct nfs_pgio_args *argp = &hdr->args; 241 struct nfs_pgio_res *resp = &hdr->res; 242 243 /* This is a short read! */ 244 nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD); 245 trace_nfs_readpage_short(task, hdr); 246 247 /* 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 resp->count = 0; 265 resp->eof = 0; 266 rpc_restart_call_prepare(task); 267 } 268 269 static void nfs_readpage_result(struct rpc_task *task, 270 struct nfs_pgio_header *hdr) 271 { 272 if (hdr->res.eof) { 273 loff_t pos = hdr->args.offset + hdr->res.count; 274 unsigned int new = pos - hdr->io_start; 275 276 if (hdr->good_bytes > new) { 277 hdr->good_bytes = new; 278 set_bit(NFS_IOHDR_EOF, &hdr->flags); 279 clear_bit(NFS_IOHDR_ERROR, &hdr->flags); 280 } 281 } else if (hdr->res.count < hdr->args.count) 282 nfs_readpage_retry(task, hdr); 283 } 284 285 int nfs_read_add_folio(struct nfs_pageio_descriptor *pgio, 286 struct nfs_open_context *ctx, 287 struct folio *folio) 288 { 289 struct inode *inode = folio->mapping->host; 290 struct nfs_server *server = NFS_SERVER(inode); 291 size_t fsize = folio_size(folio); 292 unsigned int rsize = server->rsize; 293 struct nfs_page *new; 294 unsigned int len, aligned_len; 295 int error; 296 297 len = nfs_folio_length(folio); 298 if (len == 0) 299 return nfs_return_empty_folio(folio); 300 301 aligned_len = min_t(unsigned int, ALIGN(len, rsize), fsize); 302 303 new = nfs_page_create_from_folio(ctx, folio, 0, aligned_len); 304 if (IS_ERR(new)) { 305 error = PTR_ERR(new); 306 if (nfs_netfs_folio_unlock(folio)) 307 folio_unlock(folio); 308 goto out; 309 } 310 311 if (len < fsize) 312 folio_zero_segment(folio, len, fsize); 313 if (!nfs_pageio_add_request(pgio, new)) { 314 nfs_list_remove_request(new); 315 error = pgio->pg_error; 316 nfs_readpage_release(new, error); 317 goto out; 318 } 319 return 0; 320 out: 321 return error; 322 } 323 324 /* 325 * Actually read a folio over the wire. 326 */ 327 static int nfs_do_read_folio(struct file *file, struct folio *folio) 328 { 329 struct inode *inode = file_inode(file); 330 struct nfs_pageio_descriptor pgio; 331 struct nfs_open_context *ctx; 332 int ret; 333 334 ctx = get_nfs_open_context(nfs_file_open_context(file)); 335 336 xchg(&ctx->error, 0); 337 nfs_pageio_init_read(&pgio, inode, false, 338 &nfs_async_read_completion_ops); 339 340 ret = nfs_read_add_folio(&pgio, ctx, folio); 341 if (ret) 342 goto out_put; 343 344 nfs_pageio_complete_read(&pgio); 345 nfs_update_delegated_atime(inode); 346 if (pgio.pg_error < 0) { 347 ret = pgio.pg_error; 348 goto out_put; 349 } 350 351 ret = folio_wait_locked_killable(folio); 352 if (!folio_test_uptodate(folio) && !ret) 353 ret = xchg(&ctx->error, 0); 354 355 out_put: 356 put_nfs_open_context(ctx); 357 return ret; 358 } 359 360 /* 361 * Synchronously read a folio. 362 * 363 * This is not heavily used as most users to try an asynchronous 364 * large read through ->readahead first. 365 */ 366 int nfs_read_folio(struct file *file, struct folio *folio) 367 { 368 struct inode *inode = file_inode(file); 369 loff_t pos = folio_pos(folio); 370 size_t len = folio_size(folio); 371 int ret; 372 373 trace_nfs_aop_readpage(inode, pos, len); 374 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE); 375 task_io_account_read(len); 376 377 /* 378 * Try to flush any pending writes to the file.. 379 * 380 * NOTE! Because we own the folio lock, there cannot 381 * be any new pending writes generated at this point 382 * for this folio (other folios can be written to). 383 */ 384 ret = nfs_wb_folio(inode, folio); 385 if (ret) 386 goto out_unlock; 387 if (folio_test_uptodate(folio)) 388 goto out_unlock; 389 390 ret = -ESTALE; 391 if (NFS_STALE(inode)) 392 goto out_unlock; 393 394 ret = nfs_netfs_read_folio(file, folio); 395 if (ret) 396 ret = nfs_do_read_folio(file, folio); 397 out: 398 trace_nfs_aop_readpage_done(inode, pos, len, ret); 399 return ret; 400 out_unlock: 401 folio_unlock(folio); 402 goto out; 403 } 404 405 void nfs_readahead(struct readahead_control *ractl) 406 { 407 struct nfs_pageio_descriptor pgio; 408 struct nfs_open_context *ctx; 409 unsigned int nr_pages = readahead_count(ractl); 410 struct file *file = ractl->file; 411 struct inode *inode = ractl->mapping->host; 412 struct folio *folio; 413 int ret; 414 415 trace_nfs_aop_readahead(inode, readahead_pos(ractl), nr_pages); 416 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES); 417 task_io_account_read(readahead_length(ractl)); 418 419 ret = -ESTALE; 420 if (NFS_STALE(inode)) 421 goto out; 422 423 ret = nfs_netfs_readahead(ractl); 424 if (!ret) 425 goto out; 426 427 if (file == NULL) { 428 ret = -EBADF; 429 ctx = nfs_find_open_context(inode, NULL, FMODE_READ); 430 if (ctx == NULL) 431 goto out; 432 } else 433 ctx = get_nfs_open_context(nfs_file_open_context(file)); 434 435 nfs_pageio_init_read(&pgio, inode, false, 436 &nfs_async_read_completion_ops); 437 438 while ((folio = readahead_folio(ractl)) != NULL) { 439 ret = nfs_read_add_folio(&pgio, ctx, folio); 440 if (ret) 441 break; 442 } 443 444 nfs_pageio_complete_read(&pgio); 445 nfs_update_delegated_atime(inode); 446 447 put_nfs_open_context(ctx); 448 out: 449 trace_nfs_aop_readahead_done(inode, nr_pages, ret); 450 } 451 452 int __init nfs_init_readpagecache(void) 453 { 454 nfs_rdata_cachep = kmem_cache_create("nfs_read_data", 455 sizeof(struct nfs_pgio_header), 456 0, SLAB_HWCACHE_ALIGN, 457 NULL); 458 if (nfs_rdata_cachep == NULL) 459 return -ENOMEM; 460 461 return 0; 462 } 463 464 void nfs_destroy_readpagecache(void) 465 { 466 kmem_cache_destroy(nfs_rdata_cachep); 467 } 468 469 static const struct nfs_rw_ops nfs_rw_read_ops = { 470 .rw_alloc_header = nfs_readhdr_alloc, 471 .rw_free_header = nfs_readhdr_free, 472 .rw_done = nfs_readpage_done, 473 .rw_result = nfs_readpage_result, 474 .rw_initiate = nfs_initiate_read, 475 }; 476