Loading...
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
28#define NFSDBG_FACILITY NFSDBG_PAGECACHE
29
30static const struct nfs_pageio_ops nfs_pageio_read_ops;
31static const struct rpc_call_ops nfs_read_common_ops;
32static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
33
34static struct kmem_cache *nfs_rdata_cachep;
35
36struct nfs_read_header *nfs_readhdr_alloc(void)
37{
38 struct nfs_read_header *rhdr;
39
40 rhdr = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
41 if (rhdr) {
42 struct nfs_pgio_header *hdr = &rhdr->header;
43
44 INIT_LIST_HEAD(&hdr->pages);
45 INIT_LIST_HEAD(&hdr->rpc_list);
46 spin_lock_init(&hdr->lock);
47 atomic_set(&hdr->refcnt, 0);
48 }
49 return rhdr;
50}
51EXPORT_SYMBOL_GPL(nfs_readhdr_alloc);
52
53static struct nfs_read_data *nfs_readdata_alloc(struct nfs_pgio_header *hdr,
54 unsigned int pagecount)
55{
56 struct nfs_read_data *data, *prealloc;
57
58 prealloc = &container_of(hdr, struct nfs_read_header, header)->rpc_data;
59 if (prealloc->header == NULL)
60 data = prealloc;
61 else
62 data = kzalloc(sizeof(*data), GFP_KERNEL);
63 if (!data)
64 goto out;
65
66 if (nfs_pgarray_set(&data->pages, pagecount)) {
67 data->header = hdr;
68 atomic_inc(&hdr->refcnt);
69 } else {
70 if (data != prealloc)
71 kfree(data);
72 data = NULL;
73 }
74out:
75 return data;
76}
77
78void nfs_readhdr_free(struct nfs_pgio_header *hdr)
79{
80 struct nfs_read_header *rhdr = container_of(hdr, struct nfs_read_header, header);
81
82 kmem_cache_free(nfs_rdata_cachep, rhdr);
83}
84EXPORT_SYMBOL_GPL(nfs_readhdr_free);
85
86void nfs_readdata_release(struct nfs_read_data *rdata)
87{
88 struct nfs_pgio_header *hdr = rdata->header;
89 struct nfs_read_header *read_header = container_of(hdr, struct nfs_read_header, header);
90
91 put_nfs_open_context(rdata->args.context);
92 if (rdata->pages.pagevec != rdata->pages.page_array)
93 kfree(rdata->pages.pagevec);
94 if (rdata == &read_header->rpc_data) {
95 rdata->header = NULL;
96 rdata = NULL;
97 }
98 if (atomic_dec_and_test(&hdr->refcnt))
99 hdr->completion_ops->completion(hdr);
100 /* Note: we only free the rpc_task after callbacks are done.
101 * See the comment in rpc_free_task() for why
102 */
103 kfree(rdata);
104}
105EXPORT_SYMBOL_GPL(nfs_readdata_release);
106
107static
108int nfs_return_empty_page(struct page *page)
109{
110 zero_user(page, 0, PAGE_CACHE_SIZE);
111 SetPageUptodate(page);
112 unlock_page(page);
113 return 0;
114}
115
116void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
117 struct inode *inode,
118 const struct nfs_pgio_completion_ops *compl_ops)
119{
120 nfs_pageio_init(pgio, inode, &nfs_pageio_read_ops, compl_ops,
121 NFS_SERVER(inode)->rsize, 0);
122}
123EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
124
125void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
126{
127 pgio->pg_ops = &nfs_pageio_read_ops;
128 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
129}
130EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
131
132int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
133 struct page *page)
134{
135 struct nfs_page *new;
136 unsigned int len;
137 struct nfs_pageio_descriptor pgio;
138
139 len = nfs_page_length(page);
140 if (len == 0)
141 return nfs_return_empty_page(page);
142 new = nfs_create_request(ctx, inode, page, 0, len);
143 if (IS_ERR(new)) {
144 unlock_page(page);
145 return PTR_ERR(new);
146 }
147 if (len < PAGE_CACHE_SIZE)
148 zero_user_segment(page, len, PAGE_CACHE_SIZE);
149
150 NFS_PROTO(inode)->read_pageio_init(&pgio, inode, &nfs_async_read_completion_ops);
151 nfs_pageio_add_request(&pgio, new);
152 nfs_pageio_complete(&pgio);
153 NFS_I(inode)->read_io += pgio.pg_bytes_written;
154 return 0;
155}
156
157static void nfs_readpage_release(struct nfs_page *req)
158{
159 struct inode *d_inode = req->wb_context->dentry->d_inode;
160
161 if (PageUptodate(req->wb_page))
162 nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
163
164 unlock_page(req->wb_page);
165
166 dprintk("NFS: read done (%s/%Lu %d@%Ld)\n",
167 req->wb_context->dentry->d_inode->i_sb->s_id,
168 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
169 req->wb_bytes,
170 (long long)req_offset(req));
171 nfs_release_request(req);
172}
173
174/* Note io was page aligned */
175static void nfs_read_completion(struct nfs_pgio_header *hdr)
176{
177 unsigned long bytes = 0;
178
179 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
180 goto out;
181 while (!list_empty(&hdr->pages)) {
182 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
183 struct page *page = req->wb_page;
184
185 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
186 if (bytes > hdr->good_bytes)
187 zero_user(page, 0, PAGE_SIZE);
188 else if (hdr->good_bytes - bytes < PAGE_SIZE)
189 zero_user_segment(page,
190 hdr->good_bytes & ~PAGE_MASK,
191 PAGE_SIZE);
192 }
193 bytes += req->wb_bytes;
194 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
195 if (bytes <= hdr->good_bytes)
196 SetPageUptodate(page);
197 } else
198 SetPageUptodate(page);
199 nfs_list_remove_request(req);
200 nfs_readpage_release(req);
201 }
202out:
203 hdr->release(hdr);
204}
205
206int nfs_initiate_read(struct rpc_clnt *clnt,
207 struct nfs_read_data *data,
208 const struct rpc_call_ops *call_ops, int flags)
209{
210 struct inode *inode = data->header->inode;
211 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
212 struct rpc_task *task;
213 struct rpc_message msg = {
214 .rpc_argp = &data->args,
215 .rpc_resp = &data->res,
216 .rpc_cred = data->header->cred,
217 };
218 struct rpc_task_setup task_setup_data = {
219 .task = &data->task,
220 .rpc_client = clnt,
221 .rpc_message = &msg,
222 .callback_ops = call_ops,
223 .callback_data = data,
224 .workqueue = nfsiod_workqueue,
225 .flags = RPC_TASK_ASYNC | swap_flags | flags,
226 };
227
228 /* Set up the initial task struct. */
229 NFS_PROTO(inode)->read_setup(data, &msg);
230
231 dprintk("NFS: %5u initiated read call (req %s/%llu, %u bytes @ "
232 "offset %llu)\n",
233 data->task.tk_pid,
234 inode->i_sb->s_id,
235 (unsigned long long)NFS_FILEID(inode),
236 data->args.count,
237 (unsigned long long)data->args.offset);
238
239 task = rpc_run_task(&task_setup_data);
240 if (IS_ERR(task))
241 return PTR_ERR(task);
242 rpc_put_task(task);
243 return 0;
244}
245EXPORT_SYMBOL_GPL(nfs_initiate_read);
246
247/*
248 * Set up the NFS read request struct
249 */
250static void nfs_read_rpcsetup(struct nfs_read_data *data,
251 unsigned int count, unsigned int offset)
252{
253 struct nfs_page *req = data->header->req;
254
255 data->args.fh = NFS_FH(data->header->inode);
256 data->args.offset = req_offset(req) + offset;
257 data->args.pgbase = req->wb_pgbase + offset;
258 data->args.pages = data->pages.pagevec;
259 data->args.count = count;
260 data->args.context = get_nfs_open_context(req->wb_context);
261 data->args.lock_context = req->wb_lock_context;
262
263 data->res.fattr = &data->fattr;
264 data->res.count = count;
265 data->res.eof = 0;
266 nfs_fattr_init(&data->fattr);
267}
268
269static int nfs_do_read(struct nfs_read_data *data,
270 const struct rpc_call_ops *call_ops)
271{
272 struct inode *inode = data->header->inode;
273
274 return nfs_initiate_read(NFS_CLIENT(inode), data, call_ops, 0);
275}
276
277static int
278nfs_do_multiple_reads(struct list_head *head,
279 const struct rpc_call_ops *call_ops)
280{
281 struct nfs_read_data *data;
282 int ret = 0;
283
284 while (!list_empty(head)) {
285 int ret2;
286
287 data = list_first_entry(head, struct nfs_read_data, list);
288 list_del_init(&data->list);
289
290 ret2 = nfs_do_read(data, call_ops);
291 if (ret == 0)
292 ret = ret2;
293 }
294 return ret;
295}
296
297static void
298nfs_async_read_error(struct list_head *head)
299{
300 struct nfs_page *req;
301
302 while (!list_empty(head)) {
303 req = nfs_list_entry(head->next);
304 nfs_list_remove_request(req);
305 nfs_readpage_release(req);
306 }
307}
308
309static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
310 .error_cleanup = nfs_async_read_error,
311 .completion = nfs_read_completion,
312};
313
314static void nfs_pagein_error(struct nfs_pageio_descriptor *desc,
315 struct nfs_pgio_header *hdr)
316{
317 set_bit(NFS_IOHDR_REDO, &hdr->flags);
318 while (!list_empty(&hdr->rpc_list)) {
319 struct nfs_read_data *data = list_first_entry(&hdr->rpc_list,
320 struct nfs_read_data, list);
321 list_del(&data->list);
322 nfs_readdata_release(data);
323 }
324 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
325}
326
327/*
328 * Generate multiple requests to fill a single page.
329 *
330 * We optimize to reduce the number of read operations on the wire. If we
331 * detect that we're reading a page, or an area of a page, that is past the
332 * end of file, we do not generate NFS read operations but just clear the
333 * parts of the page that would have come back zero from the server anyway.
334 *
335 * We rely on the cached value of i_size to make this determination; another
336 * client can fill pages on the server past our cached end-of-file, but we
337 * won't see the new data until our attribute cache is updated. This is more
338 * or less conventional NFS client behavior.
339 */
340static int nfs_pagein_multi(struct nfs_pageio_descriptor *desc,
341 struct nfs_pgio_header *hdr)
342{
343 struct nfs_page *req = hdr->req;
344 struct page *page = req->wb_page;
345 struct nfs_read_data *data;
346 size_t rsize = desc->pg_bsize, nbytes;
347 unsigned int offset;
348
349 offset = 0;
350 nbytes = desc->pg_count;
351 do {
352 size_t len = min(nbytes,rsize);
353
354 data = nfs_readdata_alloc(hdr, 1);
355 if (!data) {
356 nfs_pagein_error(desc, hdr);
357 return -ENOMEM;
358 }
359 data->pages.pagevec[0] = page;
360 nfs_read_rpcsetup(data, len, offset);
361 list_add(&data->list, &hdr->rpc_list);
362 nbytes -= len;
363 offset += len;
364 } while (nbytes != 0);
365
366 nfs_list_remove_request(req);
367 nfs_list_add_request(req, &hdr->pages);
368 desc->pg_rpc_callops = &nfs_read_common_ops;
369 return 0;
370}
371
372static int nfs_pagein_one(struct nfs_pageio_descriptor *desc,
373 struct nfs_pgio_header *hdr)
374{
375 struct nfs_page *req;
376 struct page **pages;
377 struct nfs_read_data *data;
378 struct list_head *head = &desc->pg_list;
379
380 data = nfs_readdata_alloc(hdr, nfs_page_array_len(desc->pg_base,
381 desc->pg_count));
382 if (!data) {
383 nfs_pagein_error(desc, hdr);
384 return -ENOMEM;
385 }
386
387 pages = data->pages.pagevec;
388 while (!list_empty(head)) {
389 req = nfs_list_entry(head->next);
390 nfs_list_remove_request(req);
391 nfs_list_add_request(req, &hdr->pages);
392 *pages++ = req->wb_page;
393 }
394
395 nfs_read_rpcsetup(data, desc->pg_count, 0);
396 list_add(&data->list, &hdr->rpc_list);
397 desc->pg_rpc_callops = &nfs_read_common_ops;
398 return 0;
399}
400
401int nfs_generic_pagein(struct nfs_pageio_descriptor *desc,
402 struct nfs_pgio_header *hdr)
403{
404 if (desc->pg_bsize < PAGE_CACHE_SIZE)
405 return nfs_pagein_multi(desc, hdr);
406 return nfs_pagein_one(desc, hdr);
407}
408EXPORT_SYMBOL_GPL(nfs_generic_pagein);
409
410static int nfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
411{
412 struct nfs_read_header *rhdr;
413 struct nfs_pgio_header *hdr;
414 int ret;
415
416 rhdr = nfs_readhdr_alloc();
417 if (!rhdr) {
418 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
419 return -ENOMEM;
420 }
421 hdr = &rhdr->header;
422 nfs_pgheader_init(desc, hdr, nfs_readhdr_free);
423 atomic_inc(&hdr->refcnt);
424 ret = nfs_generic_pagein(desc, hdr);
425 if (ret == 0)
426 ret = nfs_do_multiple_reads(&hdr->rpc_list,
427 desc->pg_rpc_callops);
428 if (atomic_dec_and_test(&hdr->refcnt))
429 hdr->completion_ops->completion(hdr);
430 return ret;
431}
432
433static const struct nfs_pageio_ops nfs_pageio_read_ops = {
434 .pg_test = nfs_generic_pg_test,
435 .pg_doio = nfs_generic_pg_readpages,
436};
437
438/*
439 * This is the callback from RPC telling us whether a reply was
440 * received or some error occurred (timeout or socket shutdown).
441 */
442int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
443{
444 struct inode *inode = data->header->inode;
445 int status;
446
447 dprintk("NFS: %s: %5u, (status %d)\n", __func__, task->tk_pid,
448 task->tk_status);
449
450 status = NFS_PROTO(inode)->read_done(task, data);
451 if (status != 0)
452 return status;
453
454 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, data->res.count);
455
456 if (task->tk_status == -ESTALE) {
457 set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
458 nfs_mark_for_revalidate(inode);
459 }
460 return 0;
461}
462
463static void nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
464{
465 struct nfs_readargs *argp = &data->args;
466 struct nfs_readres *resp = &data->res;
467
468 /* This is a short read! */
469 nfs_inc_stats(data->header->inode, NFSIOS_SHORTREAD);
470 /* Has the server at least made some progress? */
471 if (resp->count == 0) {
472 nfs_set_pgio_error(data->header, -EIO, argp->offset);
473 return;
474 }
475 /* Yes, so retry the read at the end of the data */
476 data->mds_offset += resp->count;
477 argp->offset += resp->count;
478 argp->pgbase += resp->count;
479 argp->count -= resp->count;
480 rpc_restart_call_prepare(task);
481}
482
483static void nfs_readpage_result_common(struct rpc_task *task, void *calldata)
484{
485 struct nfs_read_data *data = calldata;
486 struct nfs_pgio_header *hdr = data->header;
487
488 /* Note the only returns of nfs_readpage_result are 0 and -EAGAIN */
489 if (nfs_readpage_result(task, data) != 0)
490 return;
491 if (task->tk_status < 0)
492 nfs_set_pgio_error(hdr, task->tk_status, data->args.offset);
493 else if (data->res.eof) {
494 loff_t bound;
495
496 bound = data->args.offset + data->res.count;
497 spin_lock(&hdr->lock);
498 if (bound < hdr->io_start + hdr->good_bytes) {
499 set_bit(NFS_IOHDR_EOF, &hdr->flags);
500 clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
501 hdr->good_bytes = bound - hdr->io_start;
502 }
503 spin_unlock(&hdr->lock);
504 } else if (data->res.count != data->args.count)
505 nfs_readpage_retry(task, data);
506}
507
508static void nfs_readpage_release_common(void *calldata)
509{
510 nfs_readdata_release(calldata);
511}
512
513void nfs_read_prepare(struct rpc_task *task, void *calldata)
514{
515 struct nfs_read_data *data = calldata;
516 int err;
517 err = NFS_PROTO(data->header->inode)->read_rpc_prepare(task, data);
518 if (err)
519 rpc_exit(task, err);
520}
521
522static const struct rpc_call_ops nfs_read_common_ops = {
523 .rpc_call_prepare = nfs_read_prepare,
524 .rpc_call_done = nfs_readpage_result_common,
525 .rpc_release = nfs_readpage_release_common,
526};
527
528/*
529 * Read a page over NFS.
530 * We read the page synchronously in the following case:
531 * - The error flag is set for this page. This happens only when a
532 * previous async read operation failed.
533 */
534int nfs_readpage(struct file *file, struct page *page)
535{
536 struct nfs_open_context *ctx;
537 struct inode *inode = page_file_mapping(page)->host;
538 int error;
539
540 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
541 page, PAGE_CACHE_SIZE, page_file_index(page));
542 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
543 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
544
545 /*
546 * Try to flush any pending writes to the file..
547 *
548 * NOTE! Because we own the page lock, there cannot
549 * be any new pending writes generated at this point
550 * for this page (other pages can be written to).
551 */
552 error = nfs_wb_page(inode, page);
553 if (error)
554 goto out_unlock;
555 if (PageUptodate(page))
556 goto out_unlock;
557
558 error = -ESTALE;
559 if (NFS_STALE(inode))
560 goto out_unlock;
561
562 if (file == NULL) {
563 error = -EBADF;
564 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
565 if (ctx == NULL)
566 goto out_unlock;
567 } else
568 ctx = get_nfs_open_context(nfs_file_open_context(file));
569
570 if (!IS_SYNC(inode)) {
571 error = nfs_readpage_from_fscache(ctx, inode, page);
572 if (error == 0)
573 goto out;
574 }
575
576 error = nfs_readpage_async(ctx, inode, page);
577
578out:
579 put_nfs_open_context(ctx);
580 return error;
581out_unlock:
582 unlock_page(page);
583 return error;
584}
585
586struct nfs_readdesc {
587 struct nfs_pageio_descriptor *pgio;
588 struct nfs_open_context *ctx;
589};
590
591static int
592readpage_async_filler(void *data, struct page *page)
593{
594 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
595 struct inode *inode = page_file_mapping(page)->host;
596 struct nfs_page *new;
597 unsigned int len;
598 int error;
599
600 len = nfs_page_length(page);
601 if (len == 0)
602 return nfs_return_empty_page(page);
603
604 new = nfs_create_request(desc->ctx, inode, page, 0, len);
605 if (IS_ERR(new))
606 goto out_error;
607
608 if (len < PAGE_CACHE_SIZE)
609 zero_user_segment(page, len, PAGE_CACHE_SIZE);
610 if (!nfs_pageio_add_request(desc->pgio, new)) {
611 error = desc->pgio->pg_error;
612 goto out_unlock;
613 }
614 return 0;
615out_error:
616 error = PTR_ERR(new);
617out_unlock:
618 unlock_page(page);
619 return error;
620}
621
622int nfs_readpages(struct file *filp, struct address_space *mapping,
623 struct list_head *pages, unsigned nr_pages)
624{
625 struct nfs_pageio_descriptor pgio;
626 struct nfs_readdesc desc = {
627 .pgio = &pgio,
628 };
629 struct inode *inode = mapping->host;
630 unsigned long npages;
631 int ret = -ESTALE;
632
633 dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
634 inode->i_sb->s_id,
635 (unsigned long long)NFS_FILEID(inode),
636 nr_pages);
637 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
638
639 if (NFS_STALE(inode))
640 goto out;
641
642 if (filp == NULL) {
643 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
644 if (desc.ctx == NULL)
645 return -EBADF;
646 } else
647 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
648
649 /* attempt to read as many of the pages as possible from the cache
650 * - this returns -ENOBUFS immediately if the cookie is negative
651 */
652 ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
653 pages, &nr_pages);
654 if (ret == 0)
655 goto read_complete; /* all pages were read */
656
657 NFS_PROTO(inode)->read_pageio_init(&pgio, inode, &nfs_async_read_completion_ops);
658
659 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
660
661 nfs_pageio_complete(&pgio);
662 NFS_I(inode)->read_io += pgio.pg_bytes_written;
663 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
664 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
665read_complete:
666 put_nfs_open_context(desc.ctx);
667out:
668 return ret;
669}
670
671int __init nfs_init_readpagecache(void)
672{
673 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
674 sizeof(struct nfs_read_header),
675 0, SLAB_HWCACHE_ALIGN,
676 NULL);
677 if (nfs_rdata_cachep == NULL)
678 return -ENOMEM;
679
680 return 0;
681}
682
683void nfs_destroy_readpagecache(void)
684{
685 kmem_cache_destroy(nfs_rdata_cachep);
686}
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/pagemap.h>
19#include <linux/sunrpc/clnt.h>
20#include <linux/nfs_fs.h>
21#include <linux/nfs_page.h>
22#include <linux/module.h>
23
24#include "nfs4_fs.h"
25#include "internal.h"
26#include "iostat.h"
27#include "fscache.h"
28#include "pnfs.h"
29#include "nfstrace.h"
30
31#define NFSDBG_FACILITY NFSDBG_PAGECACHE
32
33static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
34static const struct nfs_rw_ops nfs_rw_read_ops;
35
36static struct kmem_cache *nfs_rdata_cachep;
37
38static struct nfs_pgio_header *nfs_readhdr_alloc(void)
39{
40 struct nfs_pgio_header *p = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
41
42 if (p)
43 p->rw_mode = FMODE_READ;
44 return p;
45}
46
47static void nfs_readhdr_free(struct nfs_pgio_header *rhdr)
48{
49 kmem_cache_free(nfs_rdata_cachep, rhdr);
50}
51
52static
53int nfs_return_empty_page(struct page *page)
54{
55 zero_user(page, 0, PAGE_SIZE);
56 SetPageUptodate(page);
57 unlock_page(page);
58 return 0;
59}
60
61void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
62 struct inode *inode, bool force_mds,
63 const struct nfs_pgio_completion_ops *compl_ops)
64{
65 struct nfs_server *server = NFS_SERVER(inode);
66 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
67
68#ifdef CONFIG_NFS_V4_1
69 if (server->pnfs_curr_ld && !force_mds)
70 pg_ops = server->pnfs_curr_ld->pg_read_ops;
71#endif
72 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops,
73 server->rsize, 0);
74}
75EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
76
77void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
78{
79 struct nfs_pgio_mirror *mirror;
80
81 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
82 pgio->pg_ops->pg_cleanup(pgio);
83
84 pgio->pg_ops = &nfs_pgio_rw_ops;
85
86 /* read path should never have more than one mirror */
87 WARN_ON_ONCE(pgio->pg_mirror_count != 1);
88
89 mirror = &pgio->pg_mirrors[0];
90 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
91}
92EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
93
94static void nfs_readpage_release(struct nfs_page *req, int error)
95{
96 struct inode *inode = d_inode(nfs_req_openctx(req)->dentry);
97 struct page *page = req->wb_page;
98
99 dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
100 (unsigned long long)NFS_FILEID(inode), req->wb_bytes,
101 (long long)req_offset(req));
102
103 if (nfs_error_is_fatal_on_server(error) && error != -ETIMEDOUT)
104 SetPageError(page);
105 if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
106 struct address_space *mapping = page_file_mapping(page);
107
108 if (PageUptodate(page))
109 nfs_readpage_to_fscache(inode, page, 0);
110 else if (!PageError(page) && !PagePrivate(page))
111 generic_error_remove_page(mapping, page);
112 unlock_page(page);
113 }
114 nfs_release_request(req);
115}
116
117int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
118 struct page *page)
119{
120 struct nfs_page *new;
121 unsigned int len;
122 struct nfs_pageio_descriptor pgio;
123 struct nfs_pgio_mirror *pgm;
124
125 len = nfs_page_length(page);
126 if (len == 0)
127 return nfs_return_empty_page(page);
128 new = nfs_create_request(ctx, page, 0, len);
129 if (IS_ERR(new)) {
130 unlock_page(page);
131 return PTR_ERR(new);
132 }
133 if (len < PAGE_SIZE)
134 zero_user_segment(page, len, PAGE_SIZE);
135
136 nfs_pageio_init_read(&pgio, inode, false,
137 &nfs_async_read_completion_ops);
138 if (!nfs_pageio_add_request(&pgio, new)) {
139 nfs_list_remove_request(new);
140 nfs_readpage_release(new, pgio.pg_error);
141 }
142 nfs_pageio_complete(&pgio);
143
144 /* It doesn't make sense to do mirrored reads! */
145 WARN_ON_ONCE(pgio.pg_mirror_count != 1);
146
147 pgm = &pgio.pg_mirrors[0];
148 NFS_I(inode)->read_io += pgm->pg_bytes_written;
149
150 return pgio.pg_error < 0 ? pgio.pg_error : 0;
151}
152
153static void nfs_page_group_set_uptodate(struct nfs_page *req)
154{
155 if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
156 SetPageUptodate(req->wb_page);
157}
158
159static void nfs_read_completion(struct nfs_pgio_header *hdr)
160{
161 unsigned long bytes = 0;
162 int error;
163
164 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
165 goto out;
166 while (!list_empty(&hdr->pages)) {
167 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
168 struct page *page = req->wb_page;
169 unsigned long start = req->wb_pgbase;
170 unsigned long end = req->wb_pgbase + req->wb_bytes;
171
172 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
173 /* note: regions of the page not covered by a
174 * request are zeroed in nfs_readpage_async /
175 * readpage_async_filler */
176 if (bytes > hdr->good_bytes) {
177 /* nothing in this request was good, so zero
178 * the full extent of the request */
179 zero_user_segment(page, start, end);
180
181 } else if (hdr->good_bytes - bytes < req->wb_bytes) {
182 /* part of this request has good bytes, but
183 * not all. zero the bad bytes */
184 start += hdr->good_bytes - bytes;
185 WARN_ON(start < req->wb_pgbase);
186 zero_user_segment(page, start, end);
187 }
188 }
189 error = 0;
190 bytes += req->wb_bytes;
191 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
192 if (bytes <= hdr->good_bytes)
193 nfs_page_group_set_uptodate(req);
194 else {
195 error = hdr->error;
196 xchg(&nfs_req_openctx(req)->error, error);
197 }
198 } else
199 nfs_page_group_set_uptodate(req);
200 nfs_list_remove_request(req);
201 nfs_readpage_release(req, error);
202 }
203out:
204 hdr->release(hdr);
205}
206
207static void nfs_initiate_read(struct nfs_pgio_header *hdr,
208 struct rpc_message *msg,
209 const struct nfs_rpc_ops *rpc_ops,
210 struct rpc_task_setup *task_setup_data, int how)
211{
212 struct inode *inode = hdr->inode;
213 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
214
215 task_setup_data->flags |= swap_flags;
216 rpc_ops->read_setup(hdr, msg);
217 trace_nfs_initiate_read(hdr);
218}
219
220static void
221nfs_async_read_error(struct list_head *head, int error)
222{
223 struct nfs_page *req;
224
225 while (!list_empty(head)) {
226 req = nfs_list_entry(head->next);
227 nfs_list_remove_request(req);
228 nfs_readpage_release(req, error);
229 }
230}
231
232static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
233 .error_cleanup = nfs_async_read_error,
234 .completion = nfs_read_completion,
235};
236
237/*
238 * This is the callback from RPC telling us whether a reply was
239 * received or some error occurred (timeout or socket shutdown).
240 */
241static int nfs_readpage_done(struct rpc_task *task,
242 struct nfs_pgio_header *hdr,
243 struct inode *inode)
244{
245 int status = NFS_PROTO(inode)->read_done(task, hdr);
246 if (status != 0)
247 return status;
248
249 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
250 trace_nfs_readpage_done(task, hdr);
251
252 if (task->tk_status == -ESTALE) {
253 nfs_set_inode_stale(inode);
254 nfs_mark_for_revalidate(inode);
255 }
256 return 0;
257}
258
259static void nfs_readpage_retry(struct rpc_task *task,
260 struct nfs_pgio_header *hdr)
261{
262 struct nfs_pgio_args *argp = &hdr->args;
263 struct nfs_pgio_res *resp = &hdr->res;
264
265 /* This is a short read! */
266 nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
267 trace_nfs_readpage_short(task, hdr);
268
269 /* Has the server at least made some progress? */
270 if (resp->count == 0) {
271 nfs_set_pgio_error(hdr, -EIO, argp->offset);
272 return;
273 }
274
275 /* For non rpc-based layout drivers, retry-through-MDS */
276 if (!task->tk_ops) {
277 hdr->pnfs_error = -EAGAIN;
278 return;
279 }
280
281 /* Yes, so retry the read at the end of the hdr */
282 hdr->mds_offset += resp->count;
283 argp->offset += resp->count;
284 argp->pgbase += resp->count;
285 argp->count -= resp->count;
286 resp->count = 0;
287 resp->eof = 0;
288 rpc_restart_call_prepare(task);
289}
290
291static void nfs_readpage_result(struct rpc_task *task,
292 struct nfs_pgio_header *hdr)
293{
294 if (hdr->res.eof) {
295 loff_t pos = hdr->args.offset + hdr->res.count;
296 unsigned int new = pos - hdr->io_start;
297
298 if (hdr->good_bytes > new) {
299 hdr->good_bytes = new;
300 set_bit(NFS_IOHDR_EOF, &hdr->flags);
301 clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
302 }
303 } else if (hdr->res.count < hdr->args.count)
304 nfs_readpage_retry(task, hdr);
305}
306
307/*
308 * Read a page over NFS.
309 * We read the page synchronously in the following case:
310 * - The error flag is set for this page. This happens only when a
311 * previous async read operation failed.
312 */
313int nfs_readpage(struct file *file, struct page *page)
314{
315 struct nfs_open_context *ctx;
316 struct inode *inode = page_file_mapping(page)->host;
317 int error;
318
319 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
320 page, PAGE_SIZE, page_index(page));
321 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
322 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
323
324 /*
325 * Try to flush any pending writes to the file..
326 *
327 * NOTE! Because we own the page lock, there cannot
328 * be any new pending writes generated at this point
329 * for this page (other pages can be written to).
330 */
331 error = nfs_wb_page(inode, page);
332 if (error)
333 goto out_unlock;
334 if (PageUptodate(page))
335 goto out_unlock;
336
337 error = -ESTALE;
338 if (NFS_STALE(inode))
339 goto out_unlock;
340
341 if (file == NULL) {
342 error = -EBADF;
343 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
344 if (ctx == NULL)
345 goto out_unlock;
346 } else
347 ctx = get_nfs_open_context(nfs_file_open_context(file));
348
349 if (!IS_SYNC(inode)) {
350 error = nfs_readpage_from_fscache(ctx, inode, page);
351 if (error == 0)
352 goto out;
353 }
354
355 xchg(&ctx->error, 0);
356 error = nfs_readpage_async(ctx, inode, page);
357 if (!error) {
358 error = wait_on_page_locked_killable(page);
359 if (!PageUptodate(page) && !error)
360 error = xchg(&ctx->error, 0);
361 }
362out:
363 put_nfs_open_context(ctx);
364 return error;
365out_unlock:
366 unlock_page(page);
367 return error;
368}
369
370struct nfs_readdesc {
371 struct nfs_pageio_descriptor *pgio;
372 struct nfs_open_context *ctx;
373};
374
375static int
376readpage_async_filler(void *data, struct page *page)
377{
378 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
379 struct nfs_page *new;
380 unsigned int len;
381 int error;
382
383 len = nfs_page_length(page);
384 if (len == 0)
385 return nfs_return_empty_page(page);
386
387 new = nfs_create_request(desc->ctx, page, 0, len);
388 if (IS_ERR(new))
389 goto out_error;
390
391 if (len < PAGE_SIZE)
392 zero_user_segment(page, len, PAGE_SIZE);
393 if (!nfs_pageio_add_request(desc->pgio, new)) {
394 nfs_list_remove_request(new);
395 error = desc->pgio->pg_error;
396 nfs_readpage_release(new, error);
397 goto out;
398 }
399 return 0;
400out_error:
401 error = PTR_ERR(new);
402 unlock_page(page);
403out:
404 return error;
405}
406
407int nfs_readpages(struct file *filp, struct address_space *mapping,
408 struct list_head *pages, unsigned nr_pages)
409{
410 struct nfs_pageio_descriptor pgio;
411 struct nfs_pgio_mirror *pgm;
412 struct nfs_readdesc desc = {
413 .pgio = &pgio,
414 };
415 struct inode *inode = mapping->host;
416 unsigned long npages;
417 int ret = -ESTALE;
418
419 dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
420 inode->i_sb->s_id,
421 (unsigned long long)NFS_FILEID(inode),
422 nr_pages);
423 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
424
425 if (NFS_STALE(inode))
426 goto out;
427
428 if (filp == NULL) {
429 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
430 if (desc.ctx == NULL)
431 return -EBADF;
432 } else
433 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
434
435 /* attempt to read as many of the pages as possible from the cache
436 * - this returns -ENOBUFS immediately if the cookie is negative
437 */
438 ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
439 pages, &nr_pages);
440 if (ret == 0)
441 goto read_complete; /* all pages were read */
442
443 nfs_pageio_init_read(&pgio, inode, false,
444 &nfs_async_read_completion_ops);
445
446 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
447 nfs_pageio_complete(&pgio);
448
449 /* It doesn't make sense to do mirrored reads! */
450 WARN_ON_ONCE(pgio.pg_mirror_count != 1);
451
452 pgm = &pgio.pg_mirrors[0];
453 NFS_I(inode)->read_io += pgm->pg_bytes_written;
454 npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >>
455 PAGE_SHIFT;
456 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
457read_complete:
458 put_nfs_open_context(desc.ctx);
459out:
460 return ret;
461}
462
463int __init nfs_init_readpagecache(void)
464{
465 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
466 sizeof(struct nfs_pgio_header),
467 0, SLAB_HWCACHE_ALIGN,
468 NULL);
469 if (nfs_rdata_cachep == NULL)
470 return -ENOMEM;
471
472 return 0;
473}
474
475void nfs_destroy_readpagecache(void)
476{
477 kmem_cache_destroy(nfs_rdata_cachep);
478}
479
480static const struct nfs_rw_ops nfs_rw_read_ops = {
481 .rw_alloc_header = nfs_readhdr_alloc,
482 .rw_free_header = nfs_readhdr_free,
483 .rw_done = nfs_readpage_done,
484 .rw_result = nfs_readpage_result,
485 .rw_initiate = nfs_initiate_read,
486};