6 * Partial copy of Linus' read cache modifications to fs/nfs/file.c
7 * modified for async RPC by okir@monad.swb.de
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>
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>
23 #include <asm/system.h>
31 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
33 static const struct nfs_pageio_ops nfs_pageio_read_ops;
34 static const struct rpc_call_ops nfs_read_partial_ops;
35 static const struct rpc_call_ops nfs_read_full_ops;
37 static struct kmem_cache *nfs_rdata_cachep;
39 struct nfs_read_data *nfs_readdata_alloc(unsigned int pagecount)
41 struct nfs_read_data *p;
43 p = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
45 INIT_LIST_HEAD(&p->pages);
46 p->npages = pagecount;
47 if (pagecount <= ARRAY_SIZE(p->page_array))
48 p->pagevec = p->page_array;
50 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL);
52 kmem_cache_free(nfs_rdata_cachep, p);
60 void nfs_readdata_free(struct nfs_read_data *p)
62 if (p && (p->pagevec != &p->page_array[0]))
64 kmem_cache_free(nfs_rdata_cachep, p);
67 void nfs_readdata_release(struct nfs_read_data *rdata)
69 put_lseg(rdata->lseg);
70 put_nfs_open_context(rdata->args.context);
71 nfs_readdata_free(rdata);
75 int nfs_return_empty_page(struct page *page)
77 zero_user(page, 0, PAGE_CACHE_SIZE);
78 SetPageUptodate(page);
83 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
85 unsigned int remainder = data->args.count - data->res.count;
86 unsigned int base = data->args.pgbase + data->res.count;
90 if (data->res.eof == 0 || remainder == 0)
93 * Note: "remainder" can never be negative, since we check for
94 * this in the XDR code.
96 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
97 base &= ~PAGE_CACHE_MASK;
98 pglen = PAGE_CACHE_SIZE - base;
100 if (remainder <= pglen) {
101 zero_user(*pages, base, remainder);
104 zero_user(*pages, base, pglen);
107 pglen = PAGE_CACHE_SIZE;
112 void nfs_pageio_init_read_mds(struct nfs_pageio_descriptor *pgio,
115 nfs_pageio_init(pgio, inode, &nfs_pageio_read_ops,
116 NFS_SERVER(inode)->rsize, 0);
119 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
121 pgio->pg_ops = &nfs_pageio_read_ops;
122 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
124 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
126 static void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
129 if (!pnfs_pageio_init_read(pgio, inode))
130 nfs_pageio_init_read_mds(pgio, inode);
133 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
136 struct nfs_page *new;
138 struct nfs_pageio_descriptor pgio;
140 len = nfs_page_length(page);
142 return nfs_return_empty_page(page);
143 new = nfs_create_request(ctx, inode, page, 0, len);
148 if (len < PAGE_CACHE_SIZE)
149 zero_user_segment(page, len, PAGE_CACHE_SIZE);
151 nfs_pageio_init_read(&pgio, inode);
152 nfs_pageio_add_request(&pgio, new);
153 nfs_pageio_complete(&pgio);
157 static void nfs_readpage_release(struct nfs_page *req)
159 struct inode *d_inode = req->wb_context->dentry->d_inode;
161 if (PageUptodate(req->wb_page))
162 nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
164 unlock_page(req->wb_page);
166 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
167 req->wb_context->dentry->d_inode->i_sb->s_id,
168 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
170 (long long)req_offset(req));
171 nfs_release_request(req);
174 int nfs_initiate_read(struct nfs_read_data *data, struct rpc_clnt *clnt,
175 const struct rpc_call_ops *call_ops)
177 struct inode *inode = data->inode;
178 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
179 struct rpc_task *task;
180 struct rpc_message msg = {
181 .rpc_argp = &data->args,
182 .rpc_resp = &data->res,
183 .rpc_cred = data->cred,
185 struct rpc_task_setup task_setup_data = {
189 .callback_ops = call_ops,
190 .callback_data = data,
191 .workqueue = nfsiod_workqueue,
192 .flags = RPC_TASK_ASYNC | swap_flags,
195 /* Set up the initial task struct. */
196 NFS_PROTO(inode)->read_setup(data, &msg);
198 dprintk("NFS: %5u initiated read call (req %s/%lld, %u bytes @ "
202 (long long)NFS_FILEID(inode),
204 (unsigned long long)data->args.offset);
206 task = rpc_run_task(&task_setup_data);
208 return PTR_ERR(task);
212 EXPORT_SYMBOL_GPL(nfs_initiate_read);
215 * Set up the NFS read request struct
217 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
218 unsigned int count, unsigned int offset)
220 struct inode *inode = req->wb_context->dentry->d_inode;
224 data->cred = req->wb_context->cred;
226 data->args.fh = NFS_FH(inode);
227 data->args.offset = req_offset(req) + offset;
228 data->args.pgbase = req->wb_pgbase + offset;
229 data->args.pages = data->pagevec;
230 data->args.count = count;
231 data->args.context = get_nfs_open_context(req->wb_context);
232 data->args.lock_context = req->wb_lock_context;
234 data->res.fattr = &data->fattr;
235 data->res.count = count;
237 nfs_fattr_init(&data->fattr);
240 static int nfs_do_read(struct nfs_read_data *data,
241 const struct rpc_call_ops *call_ops)
243 struct inode *inode = data->args.context->dentry->d_inode;
245 return nfs_initiate_read(data, NFS_CLIENT(inode), call_ops);
249 nfs_do_multiple_reads(struct list_head *head,
250 const struct rpc_call_ops *call_ops)
252 struct nfs_read_data *data;
255 while (!list_empty(head)) {
258 data = list_entry(head->next, struct nfs_read_data, list);
259 list_del_init(&data->list);
261 ret2 = nfs_do_read(data, call_ops);
269 nfs_async_read_error(struct list_head *head)
271 struct nfs_page *req;
273 while (!list_empty(head)) {
274 req = nfs_list_entry(head->next);
275 nfs_list_remove_request(req);
276 nfs_readpage_release(req);
281 * Generate multiple requests to fill a single page.
283 * We optimize to reduce the number of read operations on the wire. If we
284 * detect that we're reading a page, or an area of a page, that is past the
285 * end of file, we do not generate NFS read operations but just clear the
286 * parts of the page that would have come back zero from the server anyway.
288 * We rely on the cached value of i_size to make this determination; another
289 * client can fill pages on the server past our cached end-of-file, but we
290 * won't see the new data until our attribute cache is updated. This is more
291 * or less conventional NFS client behavior.
293 static int nfs_pagein_multi(struct nfs_pageio_descriptor *desc, struct list_head *res)
295 struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
296 struct page *page = req->wb_page;
297 struct nfs_read_data *data;
298 size_t rsize = desc->pg_bsize, nbytes;
303 nfs_list_remove_request(req);
306 nbytes = desc->pg_count;
308 size_t len = min(nbytes,rsize);
310 data = nfs_readdata_alloc(1);
313 data->pagevec[0] = page;
314 nfs_read_rpcsetup(req, data, len, offset);
315 list_add(&data->list, res);
319 } while(nbytes != 0);
320 atomic_set(&req->wb_complete, requests);
321 desc->pg_rpc_callops = &nfs_read_partial_ops;
324 while (!list_empty(res)) {
325 data = list_entry(res->next, struct nfs_read_data, list);
326 list_del(&data->list);
327 nfs_readdata_free(data);
329 nfs_readpage_release(req);
333 static int nfs_pagein_one(struct nfs_pageio_descriptor *desc, struct list_head *res)
335 struct nfs_page *req;
337 struct nfs_read_data *data;
338 struct list_head *head = &desc->pg_list;
341 data = nfs_readdata_alloc(nfs_page_array_len(desc->pg_base,
344 nfs_async_read_error(head);
349 pages = data->pagevec;
350 while (!list_empty(head)) {
351 req = nfs_list_entry(head->next);
352 nfs_list_remove_request(req);
353 nfs_list_add_request(req, &data->pages);
354 *pages++ = req->wb_page;
356 req = nfs_list_entry(data->pages.next);
358 nfs_read_rpcsetup(req, data, desc->pg_count, 0);
359 list_add(&data->list, res);
360 desc->pg_rpc_callops = &nfs_read_full_ops;
365 int nfs_generic_pagein(struct nfs_pageio_descriptor *desc, struct list_head *head)
367 if (desc->pg_bsize < PAGE_CACHE_SIZE)
368 return nfs_pagein_multi(desc, head);
369 return nfs_pagein_one(desc, head);
372 static int nfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
377 ret = nfs_generic_pagein(desc, &head);
379 ret = nfs_do_multiple_reads(&head, desc->pg_rpc_callops);
383 static const struct nfs_pageio_ops nfs_pageio_read_ops = {
384 .pg_test = nfs_generic_pg_test,
385 .pg_doio = nfs_generic_pg_readpages,
389 * This is the callback from RPC telling us whether a reply was
390 * received or some error occurred (timeout or socket shutdown).
392 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
396 dprintk("NFS: %s: %5u, (status %d)\n", __func__, task->tk_pid,
399 status = NFS_PROTO(data->inode)->read_done(task, data);
403 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
405 if (task->tk_status == -ESTALE) {
406 set_bit(NFS_INO_STALE, &NFS_I(data->inode)->flags);
407 nfs_mark_for_revalidate(data->inode);
412 static void nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
414 struct nfs_readargs *argp = &data->args;
415 struct nfs_readres *resp = &data->res;
417 if (resp->eof || resp->count == argp->count)
420 /* This is a short read! */
421 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
422 /* Has the server at least made some progress? */
423 if (resp->count == 0)
426 /* Yes, so retry the read at the end of the data */
427 data->mds_offset += resp->count;
428 argp->offset += resp->count;
429 argp->pgbase += resp->count;
430 argp->count -= resp->count;
431 rpc_restart_call_prepare(task);
435 * Handle a read reply that fills part of a page.
437 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
439 struct nfs_read_data *data = calldata;
441 if (nfs_readpage_result(task, data) != 0)
443 if (task->tk_status < 0)
446 nfs_readpage_truncate_uninitialised_page(data);
447 nfs_readpage_retry(task, data);
450 static void nfs_readpage_release_partial(void *calldata)
452 struct nfs_read_data *data = calldata;
453 struct nfs_page *req = data->req;
454 struct page *page = req->wb_page;
455 int status = data->task.tk_status;
458 set_bit(PG_PARTIAL_READ_FAILED, &req->wb_flags);
460 if (atomic_dec_and_test(&req->wb_complete)) {
461 if (!test_bit(PG_PARTIAL_READ_FAILED, &req->wb_flags))
462 SetPageUptodate(page);
463 nfs_readpage_release(req);
465 nfs_readdata_release(calldata);
468 #if defined(CONFIG_NFS_V4_1)
469 void nfs_read_prepare(struct rpc_task *task, void *calldata)
471 struct nfs_read_data *data = calldata;
473 if (nfs4_setup_sequence(NFS_SERVER(data->inode),
474 &data->args.seq_args, &data->res.seq_res,
477 rpc_call_start(task);
479 #endif /* CONFIG_NFS_V4_1 */
481 static const struct rpc_call_ops nfs_read_partial_ops = {
482 #if defined(CONFIG_NFS_V4_1)
483 .rpc_call_prepare = nfs_read_prepare,
484 #endif /* CONFIG_NFS_V4_1 */
485 .rpc_call_done = nfs_readpage_result_partial,
486 .rpc_release = nfs_readpage_release_partial,
489 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
491 unsigned int count = data->res.count;
492 unsigned int base = data->args.pgbase;
496 count = data->args.count;
497 if (unlikely(count == 0))
499 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
500 base &= ~PAGE_CACHE_MASK;
502 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
503 SetPageUptodate(*pages);
506 /* Was this a short read? */
507 if (data->res.eof || data->res.count == data->args.count)
508 SetPageUptodate(*pages);
512 * This is the callback from RPC telling us whether a reply was
513 * received or some error occurred (timeout or socket shutdown).
515 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
517 struct nfs_read_data *data = calldata;
519 if (nfs_readpage_result(task, data) != 0)
521 if (task->tk_status < 0)
524 * Note: nfs_readpage_retry may change the values of
525 * data->args. In the multi-page case, we therefore need
526 * to ensure that we call nfs_readpage_set_pages_uptodate()
529 nfs_readpage_truncate_uninitialised_page(data);
530 nfs_readpage_set_pages_uptodate(data);
531 nfs_readpage_retry(task, data);
534 static void nfs_readpage_release_full(void *calldata)
536 struct nfs_read_data *data = calldata;
538 while (!list_empty(&data->pages)) {
539 struct nfs_page *req = nfs_list_entry(data->pages.next);
541 nfs_list_remove_request(req);
542 nfs_readpage_release(req);
544 nfs_readdata_release(calldata);
547 static const struct rpc_call_ops nfs_read_full_ops = {
548 #if defined(CONFIG_NFS_V4_1)
549 .rpc_call_prepare = nfs_read_prepare,
550 #endif /* CONFIG_NFS_V4_1 */
551 .rpc_call_done = nfs_readpage_result_full,
552 .rpc_release = nfs_readpage_release_full,
556 * Read a page over NFS.
557 * We read the page synchronously in the following case:
558 * - The error flag is set for this page. This happens only when a
559 * previous async read operation failed.
561 int nfs_readpage(struct file *file, struct page *page)
563 struct nfs_open_context *ctx;
564 struct inode *inode = page->mapping->host;
567 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
568 page, PAGE_CACHE_SIZE, page->index);
569 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
570 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
573 * Try to flush any pending writes to the file..
575 * NOTE! Because we own the page lock, there cannot
576 * be any new pending writes generated at this point
577 * for this page (other pages can be written to).
579 error = nfs_wb_page(inode, page);
582 if (PageUptodate(page))
586 if (NFS_STALE(inode))
591 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
595 ctx = get_nfs_open_context(nfs_file_open_context(file));
597 if (!IS_SYNC(inode)) {
598 error = nfs_readpage_from_fscache(ctx, inode, page);
603 error = nfs_readpage_async(ctx, inode, page);
606 put_nfs_open_context(ctx);
613 struct nfs_readdesc {
614 struct nfs_pageio_descriptor *pgio;
615 struct nfs_open_context *ctx;
619 readpage_async_filler(void *data, struct page *page)
621 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
622 struct inode *inode = page->mapping->host;
623 struct nfs_page *new;
627 len = nfs_page_length(page);
629 return nfs_return_empty_page(page);
631 new = nfs_create_request(desc->ctx, inode, page, 0, len);
635 if (len < PAGE_CACHE_SIZE)
636 zero_user_segment(page, len, PAGE_CACHE_SIZE);
637 if (!nfs_pageio_add_request(desc->pgio, new)) {
638 error = desc->pgio->pg_error;
643 error = PTR_ERR(new);
649 int nfs_readpages(struct file *filp, struct address_space *mapping,
650 struct list_head *pages, unsigned nr_pages)
652 struct nfs_pageio_descriptor pgio;
653 struct nfs_readdesc desc = {
656 struct inode *inode = mapping->host;
657 unsigned long npages;
660 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
662 (long long)NFS_FILEID(inode),
664 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
666 if (NFS_STALE(inode))
670 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
671 if (desc.ctx == NULL)
674 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
676 /* attempt to read as many of the pages as possible from the cache
677 * - this returns -ENOBUFS immediately if the cookie is negative
679 ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
682 goto read_complete; /* all pages were read */
684 nfs_pageio_init_read(&pgio, inode);
686 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
688 nfs_pageio_complete(&pgio);
689 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
690 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
692 put_nfs_open_context(desc.ctx);
697 int __init nfs_init_readpagecache(void)
699 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
700 sizeof(struct nfs_read_data),
701 0, SLAB_HWCACHE_ALIGN,
703 if (nfs_rdata_cachep == NULL)
709 void nfs_destroy_readpagecache(void)
711 kmem_cache_destroy(nfs_rdata_cachep);