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 struct kmem_cache *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, GFP_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_free(struct nfs_read_data *p)
70 if (p && (p->pagevec != &p->page_array[0]))
72 mempool_free(p, nfs_rdata_mempool);
75 void nfs_readdata_release(void *data)
77 nfs_readdata_free(data);
81 unsigned int nfs_page_length(struct inode *inode, struct page *page)
83 loff_t i_size = i_size_read(inode);
88 idx = (i_size - 1) >> PAGE_CACHE_SHIFT;
89 if (page->index > idx)
91 if (page->index != idx)
92 return PAGE_CACHE_SIZE;
93 return 1 + ((i_size - 1) & (PAGE_CACHE_SIZE - 1));
97 int nfs_return_empty_page(struct page *page)
99 memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
100 SetPageUptodate(page);
105 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
107 unsigned int remainder = data->args.count - data->res.count;
108 unsigned int base = data->args.pgbase + data->res.count;
112 if (data->res.eof == 0 || remainder == 0)
115 * Note: "remainder" can never be negative, since we check for
116 * this in the XDR code.
118 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
119 base &= ~PAGE_CACHE_MASK;
120 pglen = PAGE_CACHE_SIZE - base;
122 if (remainder <= pglen) {
123 memclear_highpage_flush(*pages, base, remainder);
126 memclear_highpage_flush(*pages, base, pglen);
129 pglen = PAGE_CACHE_SIZE;
135 * Read a page synchronously.
137 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
140 unsigned int rsize = NFS_SERVER(inode)->rsize;
141 unsigned int count = PAGE_CACHE_SIZE;
143 struct nfs_read_data *rdata;
145 rdata = nfs_readdata_alloc(count);
149 memset(rdata, 0, sizeof(*rdata));
150 rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
151 rdata->cred = ctx->cred;
152 rdata->inode = inode;
153 INIT_LIST_HEAD(&rdata->pages);
154 rdata->args.fh = NFS_FH(inode);
155 rdata->args.context = ctx;
156 rdata->args.pages = &page;
157 rdata->args.pgbase = 0UL;
158 rdata->args.count = rsize;
159 rdata->res.fattr = &rdata->fattr;
161 dprintk("NFS: nfs_readpage_sync(%p)\n", page);
164 * This works now because the socket layer never tries to DMA
165 * into this buffer directly.
169 rdata->args.count = count;
170 rdata->res.count = rdata->args.count;
171 rdata->args.offset = page_offset(page) + rdata->args.pgbase;
173 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
174 NFS_SERVER(inode)->nfs_client->cl_hostname,
176 (long long)NFS_FILEID(inode),
177 (unsigned long long)rdata->args.pgbase,
181 result = NFS_PROTO(inode)->read(rdata);
185 * Even if we had a partial success we can't mark the page
189 if (result == -EISDIR)
194 rdata->args.pgbase += result;
195 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, result);
197 /* Note: result == 0 should only happen if we're caching
198 * a write that extends the file and punches a hole.
200 if (rdata->res.eof != 0 || result == 0)
203 spin_lock(&inode->i_lock);
204 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
205 spin_unlock(&inode->i_lock);
207 if (rdata->res.eof || rdata->res.count == rdata->args.count) {
208 SetPageUptodate(page);
209 if (rdata->res.eof && count != 0)
210 memclear_highpage_flush(page, rdata->args.pgbase, count);
216 nfs_readdata_free(rdata);
220 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
223 LIST_HEAD(one_request);
224 struct nfs_page *new;
227 len = nfs_page_length(inode, page);
229 return nfs_return_empty_page(page);
230 new = nfs_create_request(ctx, inode, page, 0, len);
235 if (len < PAGE_CACHE_SIZE)
236 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
238 nfs_list_add_request(new, &one_request);
239 nfs_pagein_one(&one_request, inode);
243 static void nfs_readpage_release(struct nfs_page *req)
245 unlock_page(req->wb_page);
247 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
248 req->wb_context->dentry->d_inode->i_sb->s_id,
249 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
251 (long long)req_offset(req));
252 nfs_clear_request(req);
253 nfs_release_request(req);
257 * Set up the NFS read request struct
259 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
260 const struct rpc_call_ops *call_ops,
261 unsigned int count, unsigned int offset)
267 data->inode = inode = req->wb_context->dentry->d_inode;
268 data->cred = req->wb_context->cred;
270 data->args.fh = NFS_FH(inode);
271 data->args.offset = req_offset(req) + offset;
272 data->args.pgbase = req->wb_pgbase + offset;
273 data->args.pages = data->pagevec;
274 data->args.count = count;
275 data->args.context = req->wb_context;
277 data->res.fattr = &data->fattr;
278 data->res.count = count;
280 nfs_fattr_init(&data->fattr);
282 /* Set up the initial task struct. */
283 flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
284 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
285 NFS_PROTO(inode)->read_setup(data);
287 data->task.tk_cookie = (unsigned long)inode;
289 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
292 (long long)NFS_FILEID(inode),
294 (unsigned long long)data->args.offset);
298 nfs_async_read_error(struct list_head *head)
300 struct nfs_page *req;
302 while (!list_empty(head)) {
303 req = nfs_list_entry(head->next);
304 nfs_list_remove_request(req);
305 SetPageError(req->wb_page);
306 nfs_readpage_release(req);
311 * Start an async read operation
313 static void nfs_execute_read(struct nfs_read_data *data)
315 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
318 rpc_clnt_sigmask(clnt, &oldset);
320 rpc_execute(&data->task);
322 rpc_clnt_sigunmask(clnt, &oldset);
326 * Generate multiple requests to fill a single page.
328 * We optimize to reduce the number of read operations on the wire. If we
329 * detect that we're reading a page, or an area of a page, that is past the
330 * end of file, we do not generate NFS read operations but just clear the
331 * parts of the page that would have come back zero from the server anyway.
333 * We rely on the cached value of i_size to make this determination; another
334 * client can fill pages on the server past our cached end-of-file, but we
335 * won't see the new data until our attribute cache is updated. This is more
336 * or less conventional NFS client behavior.
338 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
340 struct nfs_page *req = nfs_list_entry(head->next);
341 struct page *page = req->wb_page;
342 struct nfs_read_data *data;
343 size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
348 nfs_list_remove_request(req);
350 nbytes = req->wb_bytes;
352 size_t len = min(nbytes,rsize);
354 data = nfs_readdata_alloc(len);
357 INIT_LIST_HEAD(&data->pages);
358 list_add(&data->pages, &list);
361 } while(nbytes != 0);
362 atomic_set(&req->wb_complete, requests);
364 ClearPageError(page);
366 nbytes = req->wb_bytes;
368 data = list_entry(list.next, struct nfs_read_data, pages);
369 list_del_init(&data->pages);
371 data->pagevec[0] = page;
373 if (nbytes > rsize) {
374 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
379 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
383 nfs_execute_read(data);
384 } while (nbytes != 0);
389 while (!list_empty(&list)) {
390 data = list_entry(list.next, struct nfs_read_data, pages);
391 list_del(&data->pages);
392 nfs_readdata_free(data);
395 nfs_readpage_release(req);
399 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
401 struct nfs_page *req;
403 struct nfs_read_data *data;
406 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
407 return nfs_pagein_multi(head, inode);
409 data = nfs_readdata_alloc(NFS_SERVER(inode)->rsize);
413 INIT_LIST_HEAD(&data->pages);
414 pages = data->pagevec;
416 while (!list_empty(head)) {
417 req = nfs_list_entry(head->next);
418 nfs_list_remove_request(req);
419 nfs_list_add_request(req, &data->pages);
420 ClearPageError(req->wb_page);
421 *pages++ = req->wb_page;
422 count += req->wb_bytes;
424 req = nfs_list_entry(data->pages.next);
426 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
428 nfs_execute_read(data);
431 nfs_async_read_error(head);
436 nfs_pagein_list(struct list_head *head, int rpages)
438 LIST_HEAD(one_request);
439 struct nfs_page *req;
441 unsigned int pages = 0;
443 while (!list_empty(head)) {
444 pages += nfs_coalesce_requests(head, &one_request, rpages);
445 req = nfs_list_entry(one_request.next);
446 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
453 nfs_async_read_error(head);
458 * Handle a read reply that fills part of a page.
460 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
462 struct nfs_read_data *data = calldata;
463 struct nfs_page *req = data->req;
464 struct page *page = req->wb_page;
466 if (likely(task->tk_status >= 0))
467 nfs_readpage_truncate_uninitialised_page(data);
470 if (nfs_readpage_result(task, data) != 0)
472 if (atomic_dec_and_test(&req->wb_complete)) {
473 if (!PageError(page))
474 SetPageUptodate(page);
475 nfs_readpage_release(req);
479 static const struct rpc_call_ops nfs_read_partial_ops = {
480 .rpc_call_done = nfs_readpage_result_partial,
481 .rpc_release = nfs_readdata_release,
484 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
486 unsigned int count = data->res.count;
487 unsigned int base = data->args.pgbase;
491 count = data->args.count;
492 if (unlikely(count == 0))
494 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
495 base &= ~PAGE_CACHE_MASK;
497 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
498 SetPageUptodate(*pages);
500 SetPageUptodate(*pages);
503 static void nfs_readpage_set_pages_error(struct nfs_read_data *data)
505 unsigned int count = data->args.count;
506 unsigned int base = data->args.pgbase;
509 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
510 base &= ~PAGE_CACHE_MASK;
512 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
513 SetPageError(*pages);
515 SetPageError(*pages);
519 * This is the callback from RPC telling us whether a reply was
520 * received or some error occurred (timeout or socket shutdown).
522 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
524 struct nfs_read_data *data = calldata;
527 * Note: nfs_readpage_result may change the values of
528 * data->args. In the multi-page case, we therefore need
529 * to ensure that we call the next nfs_readpage_set_page_uptodate()
530 * first in the multi-page case.
532 if (likely(task->tk_status >= 0)) {
533 nfs_readpage_truncate_uninitialised_page(data);
534 nfs_readpage_set_pages_uptodate(data);
536 nfs_readpage_set_pages_error(data);
537 if (nfs_readpage_result(task, data) != 0)
539 while (!list_empty(&data->pages)) {
540 struct nfs_page *req = nfs_list_entry(data->pages.next);
542 nfs_list_remove_request(req);
543 nfs_readpage_release(req);
547 static const struct rpc_call_ops nfs_read_full_ops = {
548 .rpc_call_done = nfs_readpage_result_full,
549 .rpc_release = nfs_readdata_release,
553 * This is the callback from RPC telling us whether a reply was
554 * received or some error occurred (timeout or socket shutdown).
556 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
558 struct nfs_readargs *argp = &data->args;
559 struct nfs_readres *resp = &data->res;
562 dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
563 task->tk_pid, task->tk_status);
565 status = NFS_PROTO(data->inode)->read_done(task, data);
569 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, resp->count);
571 if (task->tk_status < 0) {
572 if (task->tk_status == -ESTALE) {
573 set_bit(NFS_INO_STALE, &NFS_FLAGS(data->inode));
574 nfs_mark_for_revalidate(data->inode);
576 } else if (resp->count < argp->count && !resp->eof) {
577 /* This is a short read! */
578 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
579 /* Has the server at least made some progress? */
580 if (resp->count != 0) {
581 /* Yes, so retry the read at the end of the data */
582 argp->offset += resp->count;
583 argp->pgbase += resp->count;
584 argp->count -= resp->count;
585 rpc_restart_call(task);
588 task->tk_status = -EIO;
590 spin_lock(&data->inode->i_lock);
591 NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
592 spin_unlock(&data->inode->i_lock);
597 * Read a page over NFS.
598 * We read the page synchronously in the following case:
599 * - The error flag is set for this page. This happens only when a
600 * previous async read operation failed.
602 int nfs_readpage(struct file *file, struct page *page)
604 struct nfs_open_context *ctx;
605 struct inode *inode = page->mapping->host;
608 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
609 page, PAGE_CACHE_SIZE, page->index);
610 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
611 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
614 * Try to flush any pending writes to the file..
616 * NOTE! Because we own the page lock, there cannot
617 * be any new pending writes generated at this point
618 * for this page (other pages can be written to).
620 error = nfs_wb_page(inode, page);
625 if (NFS_STALE(inode))
629 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
633 ctx = get_nfs_open_context((struct nfs_open_context *)
635 if (!IS_SYNC(inode)) {
636 error = nfs_readpage_async(ctx, inode, page);
640 error = nfs_readpage_sync(ctx, inode, page);
641 if (error < 0 && IS_SWAPFILE(inode))
642 printk("Aiee.. nfs swap-in of page failed!\n");
644 put_nfs_open_context(ctx);
652 struct nfs_readdesc {
653 struct list_head *head;
654 struct nfs_open_context *ctx;
658 readpage_async_filler(void *data, struct page *page)
660 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
661 struct inode *inode = page->mapping->host;
662 struct nfs_page *new;
665 nfs_wb_page(inode, page);
666 len = nfs_page_length(inode, page);
668 return nfs_return_empty_page(page);
669 new = nfs_create_request(desc->ctx, inode, page, 0, len);
675 if (len < PAGE_CACHE_SIZE)
676 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
677 nfs_list_add_request(new, desc->head);
681 int nfs_readpages(struct file *filp, struct address_space *mapping,
682 struct list_head *pages, unsigned nr_pages)
685 struct nfs_readdesc desc = {
688 struct inode *inode = mapping->host;
689 struct nfs_server *server = NFS_SERVER(inode);
692 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
694 (long long)NFS_FILEID(inode),
696 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
698 if (NFS_STALE(inode))
702 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
703 if (desc.ctx == NULL)
706 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
708 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
709 if (!list_empty(&head)) {
710 int err = nfs_pagein_list(&head, server->rpages);
712 nfs_add_stats(inode, NFSIOS_READPAGES, err);
715 put_nfs_open_context(desc.ctx);
720 int __init nfs_init_readpagecache(void)
722 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
723 sizeof(struct nfs_read_data),
724 0, SLAB_HWCACHE_ALIGN,
726 if (nfs_rdata_cachep == NULL)
729 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
731 if (nfs_rdata_mempool == NULL)
737 void nfs_destroy_readpagecache(void)
739 mempool_destroy(nfs_rdata_mempool);
740 kmem_cache_destroy(nfs_rdata_cachep);