6 * Partial copy of Linus' read cache modifications to fs/nfs/file.c
7 * modified for async RPC by okir@monad.swb.de
9 * We do an ugly hack here in order to return proper error codes to the
10 * user program when a read request failed: since generic_file_read
11 * only checks the return value of inode->i_op->readpage() which is always 0
12 * for async RPC, we set the error bit of the page to 1 when an error occurs,
13 * and make nfs_readpage transmit requests synchronously when encountering this.
14 * This is only a small problem, though, since we now retry all operations
15 * within the RPC code when root squashing is suspected.
18 #include <linux/time.h>
19 #include <linux/kernel.h>
20 #include <linux/errno.h>
21 #include <linux/fcntl.h>
22 #include <linux/stat.h>
24 #include <linux/slab.h>
25 #include <linux/pagemap.h>
26 #include <linux/sunrpc/clnt.h>
27 #include <linux/nfs_fs.h>
28 #include <linux/nfs_page.h>
29 #include <linux/smp_lock.h>
31 #include <asm/system.h>
35 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
37 static int nfs_pagein_one(struct list_head *, struct inode *);
38 static const struct rpc_call_ops nfs_read_partial_ops;
39 static const struct rpc_call_ops nfs_read_full_ops;
41 static kmem_cache_t *nfs_rdata_cachep;
42 static mempool_t *nfs_rdata_mempool;
44 #define MIN_POOL_READ (32)
46 struct nfs_read_data *nfs_readdata_alloc(size_t len)
48 unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
49 struct nfs_read_data *p = mempool_alloc(nfs_rdata_mempool, SLAB_NOFS);
52 memset(p, 0, sizeof(*p));
53 INIT_LIST_HEAD(&p->pages);
54 p->npages = pagecount;
55 if (pagecount <= ARRAY_SIZE(p->page_array))
56 p->pagevec = p->page_array;
58 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
60 mempool_free(p, nfs_rdata_mempool);
68 static void nfs_readdata_rcu_free(struct rcu_head *head)
70 struct nfs_read_data *p = container_of(head, struct nfs_read_data, task.u.tk_rcu);
71 if (p && (p->pagevec != &p->page_array[0]))
73 mempool_free(p, nfs_rdata_mempool);
76 static void nfs_readdata_free(struct nfs_read_data *rdata)
78 call_rcu_bh(&rdata->task.u.tk_rcu, nfs_readdata_rcu_free);
81 void nfs_readdata_release(void *data)
83 nfs_readdata_free(data);
87 unsigned int nfs_page_length(struct inode *inode, struct page *page)
89 loff_t i_size = i_size_read(inode);
94 idx = (i_size - 1) >> PAGE_CACHE_SHIFT;
95 if (page->index > idx)
97 if (page->index != idx)
98 return PAGE_CACHE_SIZE;
99 return 1 + ((i_size - 1) & (PAGE_CACHE_SIZE - 1));
103 int nfs_return_empty_page(struct page *page)
105 memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
106 SetPageUptodate(page);
111 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
113 unsigned int remainder = data->args.count - data->res.count;
114 unsigned int base = data->args.pgbase + data->res.count;
118 if (data->res.eof == 0 || remainder == 0)
121 * Note: "remainder" can never be negative, since we check for
122 * this in the XDR code.
124 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
125 base &= ~PAGE_CACHE_MASK;
126 pglen = PAGE_CACHE_SIZE - base;
128 if (remainder <= pglen) {
129 memclear_highpage_flush(*pages, base, remainder);
132 memclear_highpage_flush(*pages, base, pglen);
135 pglen = PAGE_CACHE_SIZE;
141 * Read a page synchronously.
143 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
146 unsigned int rsize = NFS_SERVER(inode)->rsize;
147 unsigned int count = PAGE_CACHE_SIZE;
148 int result = -ENOMEM;
149 struct nfs_read_data *rdata;
151 rdata = nfs_readdata_alloc(count);
155 memset(rdata, 0, sizeof(*rdata));
156 rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
157 rdata->cred = ctx->cred;
158 rdata->inode = inode;
159 INIT_LIST_HEAD(&rdata->pages);
160 rdata->args.fh = NFS_FH(inode);
161 rdata->args.context = ctx;
162 rdata->args.pages = &page;
163 rdata->args.pgbase = 0UL;
164 rdata->args.count = rsize;
165 rdata->res.fattr = &rdata->fattr;
167 dprintk("NFS: nfs_readpage_sync(%p)\n", page);
170 * This works now because the socket layer never tries to DMA
171 * into this buffer directly.
175 rdata->args.count = count;
176 rdata->res.count = rdata->args.count;
177 rdata->args.offset = page_offset(page) + rdata->args.pgbase;
179 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
180 NFS_SERVER(inode)->nfs_client->cl_hostname,
182 (long long)NFS_FILEID(inode),
183 (unsigned long long)rdata->args.pgbase,
187 result = NFS_PROTO(inode)->read(rdata);
191 * Even if we had a partial success we can't mark the page
195 if (result == -EISDIR)
200 rdata->args.pgbase += result;
201 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, result);
203 /* Note: result == 0 should only happen if we're caching
204 * a write that extends the file and punches a hole.
206 if (rdata->res.eof != 0 || result == 0)
209 spin_lock(&inode->i_lock);
210 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
211 spin_unlock(&inode->i_lock);
213 if (rdata->res.eof || rdata->res.count == rdata->args.count) {
214 SetPageUptodate(page);
215 if (rdata->res.eof && count != 0)
216 memclear_highpage_flush(page, rdata->args.pgbase, count);
221 nfs_readdata_free(rdata);
227 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
230 LIST_HEAD(one_request);
231 struct nfs_page *new;
234 len = nfs_page_length(inode, page);
236 return nfs_return_empty_page(page);
237 new = nfs_create_request(ctx, inode, page, 0, len);
242 if (len < PAGE_CACHE_SIZE)
243 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
245 nfs_list_add_request(new, &one_request);
246 nfs_pagein_one(&one_request, inode);
250 static void nfs_readpage_release(struct nfs_page *req)
252 unlock_page(req->wb_page);
254 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
255 req->wb_context->dentry->d_inode->i_sb->s_id,
256 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
258 (long long)req_offset(req));
259 nfs_clear_request(req);
260 nfs_release_request(req);
264 * Set up the NFS read request struct
266 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
267 const struct rpc_call_ops *call_ops,
268 unsigned int count, unsigned int offset)
274 data->inode = inode = req->wb_context->dentry->d_inode;
275 data->cred = req->wb_context->cred;
277 data->args.fh = NFS_FH(inode);
278 data->args.offset = req_offset(req) + offset;
279 data->args.pgbase = req->wb_pgbase + offset;
280 data->args.pages = data->pagevec;
281 data->args.count = count;
282 data->args.context = req->wb_context;
284 data->res.fattr = &data->fattr;
285 data->res.count = count;
287 nfs_fattr_init(&data->fattr);
289 /* Set up the initial task struct. */
290 flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
291 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
292 NFS_PROTO(inode)->read_setup(data);
294 data->task.tk_cookie = (unsigned long)inode;
296 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
299 (long long)NFS_FILEID(inode),
301 (unsigned long long)data->args.offset);
305 nfs_async_read_error(struct list_head *head)
307 struct nfs_page *req;
309 while (!list_empty(head)) {
310 req = nfs_list_entry(head->next);
311 nfs_list_remove_request(req);
312 SetPageError(req->wb_page);
313 nfs_readpage_release(req);
318 * Start an async read operation
320 static void nfs_execute_read(struct nfs_read_data *data)
322 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
325 rpc_clnt_sigmask(clnt, &oldset);
326 rpc_execute(&data->task);
327 rpc_clnt_sigunmask(clnt, &oldset);
331 * Generate multiple requests to fill a single page.
333 * We optimize to reduce the number of read operations on the wire. If we
334 * detect that we're reading a page, or an area of a page, that is past the
335 * end of file, we do not generate NFS read operations but just clear the
336 * parts of the page that would have come back zero from the server anyway.
338 * We rely on the cached value of i_size to make this determination; another
339 * client can fill pages on the server past our cached end-of-file, but we
340 * won't see the new data until our attribute cache is updated. This is more
341 * or less conventional NFS client behavior.
343 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
345 struct nfs_page *req = nfs_list_entry(head->next);
346 struct page *page = req->wb_page;
347 struct nfs_read_data *data;
348 size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
353 nfs_list_remove_request(req);
355 nbytes = req->wb_bytes;
357 size_t len = min(nbytes,rsize);
359 data = nfs_readdata_alloc(len);
362 INIT_LIST_HEAD(&data->pages);
363 list_add(&data->pages, &list);
366 } while(nbytes != 0);
367 atomic_set(&req->wb_complete, requests);
369 ClearPageError(page);
371 nbytes = req->wb_bytes;
373 data = list_entry(list.next, struct nfs_read_data, pages);
374 list_del_init(&data->pages);
376 data->pagevec[0] = page;
378 if (nbytes > rsize) {
379 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
384 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
388 nfs_execute_read(data);
389 } while (nbytes != 0);
394 while (!list_empty(&list)) {
395 data = list_entry(list.next, struct nfs_read_data, pages);
396 list_del(&data->pages);
397 nfs_readdata_free(data);
400 nfs_readpage_release(req);
404 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
406 struct nfs_page *req;
408 struct nfs_read_data *data;
411 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
412 return nfs_pagein_multi(head, inode);
414 data = nfs_readdata_alloc(NFS_SERVER(inode)->rsize);
418 INIT_LIST_HEAD(&data->pages);
419 pages = data->pagevec;
421 while (!list_empty(head)) {
422 req = nfs_list_entry(head->next);
423 nfs_list_remove_request(req);
424 nfs_list_add_request(req, &data->pages);
425 ClearPageError(req->wb_page);
426 *pages++ = req->wb_page;
427 count += req->wb_bytes;
429 req = nfs_list_entry(data->pages.next);
431 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
433 nfs_execute_read(data);
436 nfs_async_read_error(head);
441 nfs_pagein_list(struct list_head *head, int rpages)
443 LIST_HEAD(one_request);
444 struct nfs_page *req;
446 unsigned int pages = 0;
448 while (!list_empty(head)) {
449 pages += nfs_coalesce_requests(head, &one_request, rpages);
450 req = nfs_list_entry(one_request.next);
451 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
458 nfs_async_read_error(head);
463 * This is the callback from RPC telling us whether a reply was
464 * received or some error occurred (timeout or socket shutdown).
466 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
470 dprintk("%s: %4d, (status %d)\n", __FUNCTION__, task->tk_pid,
473 status = NFS_PROTO(data->inode)->read_done(task, data);
477 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
479 if (task->tk_status == -ESTALE) {
480 set_bit(NFS_INO_STALE, &NFS_FLAGS(data->inode));
481 nfs_mark_for_revalidate(data->inode);
483 spin_lock(&data->inode->i_lock);
484 NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
485 spin_unlock(&data->inode->i_lock);
489 static int nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
491 struct nfs_readargs *argp = &data->args;
492 struct nfs_readres *resp = &data->res;
494 if (resp->eof || resp->count == argp->count)
497 /* This is a short read! */
498 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
499 /* Has the server at least made some progress? */
500 if (resp->count == 0)
503 /* Yes, so retry the read at the end of the data */
504 argp->offset += resp->count;
505 argp->pgbase += resp->count;
506 argp->count -= resp->count;
507 rpc_restart_call(task);
512 * Handle a read reply that fills part of a page.
514 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
516 struct nfs_read_data *data = calldata;
517 struct nfs_page *req = data->req;
518 struct page *page = req->wb_page;
520 if (nfs_readpage_result(task, data) != 0)
523 if (likely(task->tk_status >= 0)) {
524 nfs_readpage_truncate_uninitialised_page(data);
525 if (nfs_readpage_retry(task, data) != 0)
528 if (unlikely(task->tk_status < 0))
530 if (atomic_dec_and_test(&req->wb_complete)) {
531 if (!PageError(page))
532 SetPageUptodate(page);
533 nfs_readpage_release(req);
537 static const struct rpc_call_ops nfs_read_partial_ops = {
538 .rpc_call_done = nfs_readpage_result_partial,
539 .rpc_release = nfs_readdata_release,
542 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
544 unsigned int count = data->res.count;
545 unsigned int base = data->args.pgbase;
549 count = data->args.count;
550 if (unlikely(count == 0))
552 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
553 base &= ~PAGE_CACHE_MASK;
555 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
556 SetPageUptodate(*pages);
559 /* Was this a short read? */
560 if (data->res.eof || data->res.count == data->args.count)
561 SetPageUptodate(*pages);
565 * This is the callback from RPC telling us whether a reply was
566 * received or some error occurred (timeout or socket shutdown).
568 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
570 struct nfs_read_data *data = calldata;
572 if (nfs_readpage_result(task, data) != 0)
575 * Note: nfs_readpage_retry may change the values of
576 * data->args. In the multi-page case, we therefore need
577 * to ensure that we call nfs_readpage_set_pages_uptodate()
580 if (likely(task->tk_status >= 0)) {
581 nfs_readpage_truncate_uninitialised_page(data);
582 nfs_readpage_set_pages_uptodate(data);
583 if (nfs_readpage_retry(task, data) != 0)
586 while (!list_empty(&data->pages)) {
587 struct nfs_page *req = nfs_list_entry(data->pages.next);
589 nfs_list_remove_request(req);
590 nfs_readpage_release(req);
594 static const struct rpc_call_ops nfs_read_full_ops = {
595 .rpc_call_done = nfs_readpage_result_full,
596 .rpc_release = nfs_readdata_release,
600 * Read a page over NFS.
601 * We read the page synchronously in the following case:
602 * - The error flag is set for this page. This happens only when a
603 * previous async read operation failed.
605 int nfs_readpage(struct file *file, struct page *page)
607 struct nfs_open_context *ctx;
608 struct inode *inode = page->mapping->host;
611 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
612 page, PAGE_CACHE_SIZE, page->index);
613 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
614 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
617 * Try to flush any pending writes to the file..
619 * NOTE! Because we own the page lock, there cannot
620 * be any new pending writes generated at this point
621 * for this page (other pages can be written to).
623 error = nfs_wb_page(inode, page);
628 if (NFS_STALE(inode))
633 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
637 ctx = get_nfs_open_context((struct nfs_open_context *)
639 if (!IS_SYNC(inode)) {
640 error = nfs_readpage_async(ctx, inode, page);
644 error = nfs_readpage_sync(ctx, inode, page);
645 if (error < 0 && IS_SWAPFILE(inode))
646 printk("Aiee.. nfs swap-in of page failed!\n");
648 put_nfs_open_context(ctx);
656 struct nfs_readdesc {
657 struct list_head *head;
658 struct nfs_open_context *ctx;
662 readpage_async_filler(void *data, struct page *page)
664 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
665 struct inode *inode = page->mapping->host;
666 struct nfs_page *new;
669 nfs_wb_page(inode, page);
670 len = nfs_page_length(inode, page);
672 return nfs_return_empty_page(page);
673 new = nfs_create_request(desc->ctx, inode, page, 0, len);
679 if (len < PAGE_CACHE_SIZE)
680 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
681 nfs_list_add_request(new, desc->head);
685 int nfs_readpages(struct file *filp, struct address_space *mapping,
686 struct list_head *pages, unsigned nr_pages)
689 struct nfs_readdesc desc = {
692 struct inode *inode = mapping->host;
693 struct nfs_server *server = NFS_SERVER(inode);
696 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
698 (long long)NFS_FILEID(inode),
700 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
702 if (NFS_STALE(inode))
706 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
707 if (desc.ctx == NULL)
710 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
712 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
713 if (!list_empty(&head)) {
714 int err = nfs_pagein_list(&head, server->rpages);
716 nfs_add_stats(inode, NFSIOS_READPAGES, err);
719 put_nfs_open_context(desc.ctx);
724 int __init nfs_init_readpagecache(void)
726 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
727 sizeof(struct nfs_read_data),
728 0, SLAB_HWCACHE_ALIGN,
730 if (nfs_rdata_cachep == NULL)
733 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
735 if (nfs_rdata_mempool == NULL)
741 void nfs_destroy_readpagecache(void)
743 mempool_destroy(nfs_rdata_mempool);
744 kmem_cache_destroy(nfs_rdata_cachep);