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