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_nfs_open_context(rdata->args.context);
70 nfs_readdata_free(rdata);
74 int nfs_return_empty_page(struct page *page)
76 zero_user(page, 0, PAGE_CACHE_SIZE);
77 SetPageUptodate(page);
82 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
84 unsigned int remainder = data->args.count - data->res.count;
85 unsigned int base = data->args.pgbase + data->res.count;
89 if (data->res.eof == 0 || remainder == 0)
92 * Note: "remainder" can never be negative, since we check for
93 * this in the XDR code.
95 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
96 base &= ~PAGE_CACHE_MASK;
97 pglen = PAGE_CACHE_SIZE - base;
99 if (remainder <= pglen) {
100 zero_user(*pages, base, remainder);
103 zero_user(*pages, base, pglen);
106 pglen = PAGE_CACHE_SIZE;
111 void nfs_pageio_init_read_mds(struct nfs_pageio_descriptor *pgio,
114 nfs_pageio_init(pgio, inode, &nfs_pageio_read_ops,
115 NFS_SERVER(inode)->rsize, 0);
118 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
120 pgio->pg_ops = &nfs_pageio_read_ops;
121 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
123 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
125 static void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
128 if (!pnfs_pageio_init_read(pgio, inode))
129 nfs_pageio_init_read_mds(pgio, inode);
132 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
135 struct nfs_page *new;
137 struct nfs_pageio_descriptor pgio;
139 len = nfs_page_length(page);
141 return nfs_return_empty_page(page);
142 new = nfs_create_request(ctx, inode, page, 0, len);
147 if (len < PAGE_CACHE_SIZE)
148 zero_user_segment(page, len, PAGE_CACHE_SIZE);
150 nfs_pageio_init_read(&pgio, inode);
151 nfs_pageio_add_request(&pgio, new);
152 nfs_pageio_complete(&pgio);
156 static void nfs_readpage_release(struct nfs_page *req)
158 struct inode *d_inode = req->wb_context->dentry->d_inode;
160 if (PageUptodate(req->wb_page))
161 nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
163 unlock_page(req->wb_page);
165 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
166 req->wb_context->dentry->d_inode->i_sb->s_id,
167 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
169 (long long)req_offset(req));
170 nfs_release_request(req);
173 int nfs_initiate_read(struct nfs_read_data *data, struct rpc_clnt *clnt,
174 const struct rpc_call_ops *call_ops)
176 struct inode *inode = data->inode;
177 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
178 struct rpc_task *task;
179 struct rpc_message msg = {
180 .rpc_argp = &data->args,
181 .rpc_resp = &data->res,
182 .rpc_cred = data->cred,
184 struct rpc_task_setup task_setup_data = {
188 .callback_ops = call_ops,
189 .callback_data = data,
190 .workqueue = nfsiod_workqueue,
191 .flags = RPC_TASK_ASYNC | swap_flags,
194 /* Set up the initial task struct. */
195 NFS_PROTO(inode)->read_setup(data, &msg);
197 dprintk("NFS: %5u initiated read call (req %s/%lld, %u bytes @ "
201 (long long)NFS_FILEID(inode),
203 (unsigned long long)data->args.offset);
205 task = rpc_run_task(&task_setup_data);
207 return PTR_ERR(task);
211 EXPORT_SYMBOL_GPL(nfs_initiate_read);
214 * Set up the NFS read request struct
216 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
217 unsigned int count, unsigned int offset)
219 struct inode *inode = req->wb_context->dentry->d_inode;
223 data->cred = req->wb_context->cred;
225 data->args.fh = NFS_FH(inode);
226 data->args.offset = req_offset(req) + offset;
227 data->args.pgbase = req->wb_pgbase + offset;
228 data->args.pages = data->pagevec;
229 data->args.count = count;
230 data->args.context = get_nfs_open_context(req->wb_context);
231 data->args.lock_context = req->wb_lock_context;
233 data->res.fattr = &data->fattr;
234 data->res.count = count;
236 nfs_fattr_init(&data->fattr);
239 static int nfs_do_read(struct nfs_read_data *data,
240 const struct rpc_call_ops *call_ops)
242 struct inode *inode = data->args.context->dentry->d_inode;
244 return nfs_initiate_read(data, NFS_CLIENT(inode), call_ops);
248 nfs_do_multiple_reads(struct list_head *head,
249 const struct rpc_call_ops *call_ops)
251 struct nfs_read_data *data;
254 while (!list_empty(head)) {
257 data = list_entry(head->next, struct nfs_read_data, list);
258 list_del_init(&data->list);
260 ret2 = nfs_do_read(data, call_ops);
268 nfs_async_read_error(struct list_head *head)
270 struct nfs_page *req;
272 while (!list_empty(head)) {
273 req = nfs_list_entry(head->next);
274 nfs_list_remove_request(req);
275 nfs_readpage_release(req);
280 * Generate multiple requests to fill a single page.
282 * We optimize to reduce the number of read operations on the wire. If we
283 * detect that we're reading a page, or an area of a page, that is past the
284 * end of file, we do not generate NFS read operations but just clear the
285 * parts of the page that would have come back zero from the server anyway.
287 * We rely on the cached value of i_size to make this determination; another
288 * client can fill pages on the server past our cached end-of-file, but we
289 * won't see the new data until our attribute cache is updated. This is more
290 * or less conventional NFS client behavior.
292 static int nfs_pagein_multi(struct nfs_pageio_descriptor *desc, struct list_head *res)
294 struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
295 struct page *page = req->wb_page;
296 struct nfs_read_data *data;
297 size_t rsize = desc->pg_bsize, nbytes;
302 nfs_list_remove_request(req);
305 nbytes = desc->pg_count;
307 size_t len = min(nbytes,rsize);
309 data = nfs_readdata_alloc(1);
312 data->pagevec[0] = page;
313 nfs_read_rpcsetup(req, data, len, offset);
314 list_add(&data->list, res);
318 } while(nbytes != 0);
319 atomic_set(&req->wb_complete, requests);
320 desc->pg_rpc_callops = &nfs_read_partial_ops;
323 while (!list_empty(res)) {
324 data = list_entry(res->next, struct nfs_read_data, list);
325 list_del(&data->list);
326 nfs_readdata_free(data);
328 nfs_readpage_release(req);
332 static int nfs_pagein_one(struct nfs_pageio_descriptor *desc, struct list_head *res)
334 struct nfs_page *req;
336 struct nfs_read_data *data;
337 struct list_head *head = &desc->pg_list;
340 data = nfs_readdata_alloc(nfs_page_array_len(desc->pg_base,
343 nfs_async_read_error(head);
348 pages = data->pagevec;
349 while (!list_empty(head)) {
350 req = nfs_list_entry(head->next);
351 nfs_list_remove_request(req);
352 nfs_list_add_request(req, &data->pages);
353 *pages++ = req->wb_page;
355 req = nfs_list_entry(data->pages.next);
357 nfs_read_rpcsetup(req, data, desc->pg_count, 0);
358 list_add(&data->list, res);
359 desc->pg_rpc_callops = &nfs_read_full_ops;
364 int nfs_generic_pagein(struct nfs_pageio_descriptor *desc, struct list_head *head)
366 if (desc->pg_bsize < PAGE_CACHE_SIZE)
367 return nfs_pagein_multi(desc, head);
368 return nfs_pagein_one(desc, head);
371 static int nfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
376 ret = nfs_generic_pagein(desc, &head);
378 ret = nfs_do_multiple_reads(&head, desc->pg_rpc_callops);
382 static const struct nfs_pageio_ops nfs_pageio_read_ops = {
383 .pg_test = nfs_generic_pg_test,
384 .pg_doio = nfs_generic_pg_readpages,
388 * This is the callback from RPC telling us whether a reply was
389 * received or some error occurred (timeout or socket shutdown).
391 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
395 dprintk("NFS: %s: %5u, (status %d)\n", __func__, task->tk_pid,
398 status = NFS_PROTO(data->inode)->read_done(task, data);
402 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
404 if (task->tk_status == -ESTALE) {
405 set_bit(NFS_INO_STALE, &NFS_I(data->inode)->flags);
406 nfs_mark_for_revalidate(data->inode);
411 static void nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
413 struct nfs_readargs *argp = &data->args;
414 struct nfs_readres *resp = &data->res;
416 if (resp->eof || resp->count == argp->count)
419 /* This is a short read! */
420 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
421 /* Has the server at least made some progress? */
422 if (resp->count == 0)
425 /* Yes, so retry the read at the end of the data */
426 data->mds_offset += resp->count;
427 argp->offset += resp->count;
428 argp->pgbase += resp->count;
429 argp->count -= resp->count;
430 rpc_restart_call_prepare(task);
434 * Handle a read reply that fills part of a page.
436 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
438 struct nfs_read_data *data = calldata;
440 if (nfs_readpage_result(task, data) != 0)
442 if (task->tk_status < 0)
445 nfs_readpage_truncate_uninitialised_page(data);
446 nfs_readpage_retry(task, data);
449 static void nfs_readpage_release_partial(void *calldata)
451 struct nfs_read_data *data = calldata;
452 struct nfs_page *req = data->req;
453 struct page *page = req->wb_page;
454 int status = data->task.tk_status;
457 set_bit(PG_PARTIAL_READ_FAILED, &req->wb_flags);
459 if (atomic_dec_and_test(&req->wb_complete)) {
460 if (!test_bit(PG_PARTIAL_READ_FAILED, &req->wb_flags))
461 SetPageUptodate(page);
462 nfs_readpage_release(req);
464 nfs_readdata_release(calldata);
467 void nfs_read_prepare(struct rpc_task *task, void *calldata)
469 struct nfs_read_data *data = calldata;
470 NFS_PROTO(data->inode)->read_rpc_prepare(task, data);
473 static const struct rpc_call_ops nfs_read_partial_ops = {
474 .rpc_call_prepare = nfs_read_prepare,
475 .rpc_call_done = nfs_readpage_result_partial,
476 .rpc_release = nfs_readpage_release_partial,
479 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
481 unsigned int count = data->res.count;
482 unsigned int base = data->args.pgbase;
486 count = data->args.count;
487 if (unlikely(count == 0))
489 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
490 base &= ~PAGE_CACHE_MASK;
492 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
493 SetPageUptodate(*pages);
496 /* Was this a short read? */
497 if (data->res.eof || data->res.count == data->args.count)
498 SetPageUptodate(*pages);
502 * This is the callback from RPC telling us whether a reply was
503 * received or some error occurred (timeout or socket shutdown).
505 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
507 struct nfs_read_data *data = calldata;
509 if (nfs_readpage_result(task, data) != 0)
511 if (task->tk_status < 0)
514 * Note: nfs_readpage_retry may change the values of
515 * data->args. In the multi-page case, we therefore need
516 * to ensure that we call nfs_readpage_set_pages_uptodate()
519 nfs_readpage_truncate_uninitialised_page(data);
520 nfs_readpage_set_pages_uptodate(data);
521 nfs_readpage_retry(task, data);
524 static void nfs_readpage_release_full(void *calldata)
526 struct nfs_read_data *data = calldata;
528 while (!list_empty(&data->pages)) {
529 struct nfs_page *req = nfs_list_entry(data->pages.next);
531 nfs_list_remove_request(req);
532 nfs_readpage_release(req);
534 nfs_readdata_release(calldata);
537 static const struct rpc_call_ops nfs_read_full_ops = {
538 .rpc_call_prepare = nfs_read_prepare,
539 .rpc_call_done = nfs_readpage_result_full,
540 .rpc_release = nfs_readpage_release_full,
544 * Read a page over NFS.
545 * We read the page synchronously in the following case:
546 * - The error flag is set for this page. This happens only when a
547 * previous async read operation failed.
549 int nfs_readpage(struct file *file, struct page *page)
551 struct nfs_open_context *ctx;
552 struct inode *inode = page->mapping->host;
555 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
556 page, PAGE_CACHE_SIZE, page->index);
557 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
558 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
561 * Try to flush any pending writes to the file..
563 * NOTE! Because we own the page lock, there cannot
564 * be any new pending writes generated at this point
565 * for this page (other pages can be written to).
567 error = nfs_wb_page(inode, page);
570 if (PageUptodate(page))
574 if (NFS_STALE(inode))
579 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
583 ctx = get_nfs_open_context(nfs_file_open_context(file));
585 if (!IS_SYNC(inode)) {
586 error = nfs_readpage_from_fscache(ctx, inode, page);
591 error = nfs_readpage_async(ctx, inode, page);
594 put_nfs_open_context(ctx);
601 struct nfs_readdesc {
602 struct nfs_pageio_descriptor *pgio;
603 struct nfs_open_context *ctx;
607 readpage_async_filler(void *data, struct page *page)
609 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
610 struct inode *inode = page->mapping->host;
611 struct nfs_page *new;
615 len = nfs_page_length(page);
617 return nfs_return_empty_page(page);
619 new = nfs_create_request(desc->ctx, inode, page, 0, len);
623 if (len < PAGE_CACHE_SIZE)
624 zero_user_segment(page, len, PAGE_CACHE_SIZE);
625 if (!nfs_pageio_add_request(desc->pgio, new)) {
626 error = desc->pgio->pg_error;
631 error = PTR_ERR(new);
637 int nfs_readpages(struct file *filp, struct address_space *mapping,
638 struct list_head *pages, unsigned nr_pages)
640 struct nfs_pageio_descriptor pgio;
641 struct nfs_readdesc desc = {
644 struct inode *inode = mapping->host;
645 unsigned long npages;
648 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
650 (long long)NFS_FILEID(inode),
652 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
654 if (NFS_STALE(inode))
658 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
659 if (desc.ctx == NULL)
662 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
664 /* attempt to read as many of the pages as possible from the cache
665 * - this returns -ENOBUFS immediately if the cookie is negative
667 ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
670 goto read_complete; /* all pages were read */
672 nfs_pageio_init_read(&pgio, inode);
674 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
676 nfs_pageio_complete(&pgio);
677 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
678 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
680 put_nfs_open_context(desc.ctx);
685 int __init nfs_init_readpagecache(void)
687 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
688 sizeof(struct nfs_read_data),
689 0, SLAB_HWCACHE_ALIGN,
691 if (nfs_rdata_cachep == NULL)
697 void nfs_destroy_readpagecache(void)
699 kmem_cache_destroy(nfs_rdata_cachep);