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
35const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
36static const struct nfs_rw_ops nfs_rw_read_ops;
37
38static struct kmem_cache *nfs_rdata_cachep;
39
40static 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
49static void nfs_readhdr_free(struct nfs_pgio_header *rhdr)
50{
51 kfree(objp: rhdr->res.scratch);
52 kmem_cache_free(s: nfs_rdata_cachep, objp: rhdr);
53}
54
55static int nfs_return_empty_folio(struct folio *folio)
56{
57 folio_zero_segment(folio, start: 0, xend: folio_size(folio));
58 folio_mark_uptodate(folio);
59 if (nfs_netfs_folio_unlock(folio))
60 folio_unlock(folio);
61 return 0;
62}
63
64void 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#ifdef CONFIG_NFS_V4_1
72 if (server->pnfs_curr_ld && !force_mds)
73 pg_ops = server->pnfs_curr_ld->pg_read_ops;
74#endif
75 nfs_pageio_init(desc: pgio, inode, pg_ops, compl_ops, rw_ops: &nfs_rw_read_ops,
76 bsize: server->rsize, how: 0);
77}
78EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
79
80void 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(desc: 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(inode: pgio->pg_inode)->read_io += pgm->pg_bytes_written;
92 npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >> PAGE_SHIFT;
93 nfs_add_stats(inode: pgio->pg_inode, stat: NFSIOS_READPAGES, addend: npages);
94}
95
96
97void 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(inode: pgio->pg_inode)->rsize;
111}
112EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
113
114bool 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}
120EXPORT_SYMBOL_GPL(nfs_read_alloc_scratch);
121
122static 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
133static void nfs_page_group_set_uptodate(struct nfs_page *req)
134{
135 if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
136 folio_mark_uptodate(folio: nfs_page_to_folio(req));
137}
138
139static void nfs_read_completion(struct nfs_pgio_header *hdr)
140{
141 unsigned long bytes = 0;
142 int error;
143
144 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
145 goto out;
146 while (!list_empty(head: &hdr->pages)) {
147 struct nfs_page *req = nfs_list_entry(head: hdr->pages.next);
148 struct folio *folio = nfs_page_to_folio(req);
149 unsigned long start = req->wb_pgbase;
150 unsigned long end = req->wb_pgbase + req->wb_bytes;
151
152 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
153 /* note: regions of the page not covered by a
154 * request are zeroed in nfs_read_add_folio
155 */
156 if (bytes > hdr->good_bytes) {
157 /* nothing in this request was good, so zero
158 * the full extent of the request */
159 folio_zero_segment(folio, start, xend: end);
160
161 } else if (hdr->good_bytes - bytes < req->wb_bytes) {
162 /* part of this request has good bytes, but
163 * not all. zero the bad bytes */
164 start += hdr->good_bytes - bytes;
165 WARN_ON(start < req->wb_pgbase);
166 folio_zero_segment(folio, start, xend: end);
167 }
168 }
169 error = 0;
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 error = hdr->error;
176 xchg(&nfs_req_openctx(req)->error, error);
177 }
178 } else
179 nfs_page_group_set_uptodate(req);
180 nfs_list_remove_request(req);
181 nfs_readpage_release(req, error);
182 }
183 nfs_netfs_read_completion(hdr);
184
185out:
186 hdr->release(hdr);
187}
188
189static 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 rpc_ops->read_setup(hdr, msg);
195 nfs_netfs_initiate_read(hdr);
196 trace_nfs_initiate_read(hdr);
197}
198
199static void
200nfs_async_read_error(struct list_head *head, int error)
201{
202 struct nfs_page *req;
203
204 while (!list_empty(head)) {
205 req = nfs_list_entry(head: head->next);
206 nfs_list_remove_request(req);
207 nfs_readpage_release(req, error);
208 }
209}
210
211const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
212 .error_cleanup = nfs_async_read_error,
213 .completion = nfs_read_completion,
214};
215
216/*
217 * This is the callback from RPC telling us whether a reply was
218 * received or some error occurred (timeout or socket shutdown).
219 */
220static int nfs_readpage_done(struct rpc_task *task,
221 struct nfs_pgio_header *hdr,
222 struct inode *inode)
223{
224 int status = NFS_PROTO(inode)->read_done(task, hdr);
225 if (status != 0)
226 return status;
227
228 nfs_add_stats(inode, stat: NFSIOS_SERVERREADBYTES, addend: hdr->res.count);
229 trace_nfs_readpage_done(task, hdr);
230
231 if (task->tk_status == -ESTALE) {
232 nfs_set_inode_stale(inode);
233 nfs_mark_for_revalidate(inode);
234 }
235 return 0;
236}
237
238static void nfs_readpage_retry(struct rpc_task *task,
239 struct nfs_pgio_header *hdr)
240{
241 struct nfs_pgio_args *argp = &hdr->args;
242 struct nfs_pgio_res *resp = &hdr->res;
243
244 /* This is a short read! */
245 nfs_inc_stats(inode: hdr->inode, stat: NFSIOS_SHORTREAD);
246 trace_nfs_readpage_short(task, hdr);
247
248 /* Has the server at least made some progress? */
249 if (resp->count == 0) {
250 nfs_set_pgio_error(hdr, error: -EIO, pos: argp->offset);
251 return;
252 }
253
254 /* For non rpc-based layout drivers, retry-through-MDS */
255 if (!task->tk_ops) {
256 hdr->pnfs_error = -EAGAIN;
257 return;
258 }
259
260 /* Yes, so retry the read at the end of the hdr */
261 hdr->mds_offset += resp->count;
262 argp->offset += resp->count;
263 argp->pgbase += resp->count;
264 argp->count -= resp->count;
265 resp->count = 0;
266 resp->eof = 0;
267 rpc_restart_call_prepare(task);
268}
269
270static void nfs_readpage_result(struct rpc_task *task,
271 struct nfs_pgio_header *hdr)
272{
273 if (hdr->res.eof) {
274 loff_t pos = hdr->args.offset + hdr->res.count;
275 unsigned int new = pos - hdr->io_start;
276
277 if (hdr->good_bytes > new) {
278 hdr->good_bytes = new;
279 set_bit(nr: NFS_IOHDR_EOF, addr: &hdr->flags);
280 clear_bit(nr: NFS_IOHDR_ERROR, addr: &hdr->flags);
281 }
282 } else if (hdr->res.count < hdr->args.count)
283 nfs_readpage_retry(task, hdr);
284}
285
286int nfs_read_add_folio(struct nfs_pageio_descriptor *pgio,
287 struct nfs_open_context *ctx,
288 struct folio *folio)
289{
290 struct inode *inode = folio->mapping->host;
291 struct nfs_server *server = NFS_SERVER(inode);
292 size_t fsize = folio_size(folio);
293 unsigned int rsize = server->rsize;
294 struct nfs_page *new;
295 unsigned int len, aligned_len;
296 int error;
297
298 len = nfs_folio_length(folio);
299 if (len == 0)
300 return nfs_return_empty_folio(folio);
301
302 aligned_len = min_t(unsigned int, ALIGN(len, rsize), fsize);
303
304 new = nfs_page_create_from_folio(ctx, folio, offset: 0, count: aligned_len);
305 if (IS_ERR(ptr: new)) {
306 error = PTR_ERR(ptr: new);
307 if (nfs_netfs_folio_unlock(folio))
308 folio_unlock(folio);
309 goto out;
310 }
311
312 if (len < fsize)
313 folio_zero_segment(folio, start: len, xend: fsize);
314 if (!nfs_pageio_add_request(pgio, new)) {
315 nfs_list_remove_request(req: new);
316 error = pgio->pg_error;
317 nfs_readpage_release(req: new, error);
318 goto out;
319 }
320 return 0;
321out:
322 return error;
323}
324
325/*
326 * Actually read a folio over the wire.
327 */
328static int nfs_do_read_folio(struct file *file, struct folio *folio)
329{
330 struct inode *inode = file_inode(f: file);
331 struct nfs_pageio_descriptor pgio;
332 struct nfs_open_context *ctx;
333 int ret;
334
335 ctx = get_nfs_open_context(ctx: nfs_file_open_context(filp: file));
336
337 xchg(&ctx->error, 0);
338 nfs_pageio_init_read(&pgio, inode, false,
339 &nfs_async_read_completion_ops);
340
341 ret = nfs_read_add_folio(pgio: &pgio, ctx, folio);
342 if (ret)
343 goto out_put;
344
345 nfs_pageio_complete_read(pgio: &pgio);
346 nfs_update_delegated_atime(inode);
347 if (pgio.pg_error < 0) {
348 ret = pgio.pg_error;
349 goto out_put;
350 }
351
352 ret = folio_wait_locked_killable(folio);
353 if (!folio_test_uptodate(folio) && !ret)
354 ret = xchg(&ctx->error, 0);
355
356out_put:
357 put_nfs_open_context(ctx);
358 return ret;
359}
360
361/*
362 * Synchronously read a folio.
363 *
364 * This is not heavily used as most users to try an asynchronous
365 * large read through ->readahead first.
366 */
367int nfs_read_folio(struct file *file, struct folio *folio)
368{
369 struct inode *inode = file_inode(f: file);
370 loff_t pos = folio_pos(folio);
371 size_t len = folio_size(folio);
372 int ret;
373
374 trace_nfs_aop_readpage(inode, offset: pos, count: len);
375 nfs_inc_stats(inode, stat: NFSIOS_VFSREADPAGE);
376 task_io_account_read(bytes: len);
377
378 /*
379 * Try to flush any pending writes to the file..
380 *
381 * NOTE! Because we own the folio lock, there cannot
382 * be any new pending writes generated at this point
383 * for this folio (other folios can be written to).
384 */
385 ret = nfs_wb_folio(inode, folio);
386 if (ret)
387 goto out_unlock;
388 if (folio_test_uptodate(folio))
389 goto out_unlock;
390
391 ret = -ESTALE;
392 if (NFS_STALE(inode))
393 goto out_unlock;
394
395 ret = nfs_netfs_read_folio(file, folio);
396 if (ret)
397 ret = nfs_do_read_folio(file, folio);
398out:
399 trace_nfs_aop_readpage_done(inode, offset: pos, count: len, ret);
400 return ret;
401out_unlock:
402 folio_unlock(folio);
403 goto out;
404}
405
406void nfs_readahead(struct readahead_control *ractl)
407{
408 struct nfs_pageio_descriptor pgio;
409 struct nfs_open_context *ctx;
410 unsigned int nr_pages = readahead_count(rac: ractl);
411 struct file *file = ractl->file;
412 struct inode *inode = ractl->mapping->host;
413 struct folio *folio;
414 int ret;
415
416 trace_nfs_aop_readahead(inode, pos: readahead_pos(rac: ractl), nr_pages);
417 nfs_inc_stats(inode, stat: NFSIOS_VFSREADPAGES);
418 task_io_account_read(bytes: readahead_length(rac: ractl));
419
420 ret = -ESTALE;
421 if (NFS_STALE(inode))
422 goto out;
423
424 ret = nfs_netfs_readahead(ractl);
425 if (!ret)
426 goto out;
427
428 if (file == NULL) {
429 ret = -EBADF;
430 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
431 if (ctx == NULL)
432 goto out;
433 } else
434 ctx = get_nfs_open_context(ctx: nfs_file_open_context(filp: file));
435
436 nfs_pageio_init_read(&pgio, inode, false,
437 &nfs_async_read_completion_ops);
438
439 while ((folio = readahead_folio(ractl)) != NULL) {
440 ret = nfs_read_add_folio(pgio: &pgio, ctx, folio);
441 if (ret)
442 break;
443 }
444
445 nfs_pageio_complete_read(pgio: &pgio);
446 nfs_update_delegated_atime(inode);
447
448 put_nfs_open_context(ctx);
449out:
450 trace_nfs_aop_readahead_done(inode, nr_pages, ret);
451}
452
453int __init nfs_init_readpagecache(void)
454{
455 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
456 sizeof(struct nfs_pgio_header),
457 0, SLAB_HWCACHE_ALIGN,
458 NULL);
459 if (nfs_rdata_cachep == NULL)
460 return -ENOMEM;
461
462 return 0;
463}
464
465void nfs_destroy_readpagecache(void)
466{
467 kmem_cache_destroy(s: nfs_rdata_cachep);
468}
469
470static const struct nfs_rw_ops nfs_rw_read_ops = {
471 .rw_alloc_header = nfs_readhdr_alloc,
472 .rw_free_header = nfs_readhdr_free,
473 .rw_done = nfs_readpage_done,
474 .rw_result = nfs_readpage_result,
475 .rw_initiate = nfs_initiate_read,
476};
477

source code of linux/fs/nfs/read.c