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>
36 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
38 static int nfs_pagein_one(struct list_head *, struct inode *);
39 static const struct rpc_call_ops nfs_read_partial_ops;
40 static const struct rpc_call_ops nfs_read_full_ops;
42 static struct kmem_cache *nfs_rdata_cachep;
43 static mempool_t *nfs_rdata_mempool;
45 #define MIN_POOL_READ (32)
47 struct nfs_read_data *nfs_readdata_alloc(size_t len)
49 unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
50 struct nfs_read_data *p = mempool_alloc(nfs_rdata_mempool, GFP_NOFS);
53 memset(p, 0, sizeof(*p));
54 INIT_LIST_HEAD(&p->pages);
55 p->npages = pagecount;
56 if (pagecount <= ARRAY_SIZE(p->page_array))
57 p->pagevec = p->page_array;
59 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
61 mempool_free(p, nfs_rdata_mempool);
69 static void nfs_readdata_rcu_free(struct rcu_head *head)
71 struct nfs_read_data *p = container_of(head, struct nfs_read_data, task.u.tk_rcu);
72 if (p && (p->pagevec != &p->page_array[0]))
74 mempool_free(p, nfs_rdata_mempool);
77 static void nfs_readdata_free(struct nfs_read_data *rdata)
79 call_rcu_bh(&rdata->task.u.tk_rcu, nfs_readdata_rcu_free);
82 void nfs_readdata_release(void *data)
84 nfs_readdata_free(data);
88 int nfs_return_empty_page(struct page *page)
90 memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
91 SetPageUptodate(page);
96 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
98 unsigned int remainder = data->args.count - data->res.count;
99 unsigned int base = data->args.pgbase + data->res.count;
103 if (data->res.eof == 0 || remainder == 0)
106 * Note: "remainder" can never be negative, since we check for
107 * this in the XDR code.
109 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
110 base &= ~PAGE_CACHE_MASK;
111 pglen = PAGE_CACHE_SIZE - base;
113 if (remainder <= pglen) {
114 memclear_highpage_flush(*pages, base, remainder);
117 memclear_highpage_flush(*pages, base, pglen);
120 pglen = PAGE_CACHE_SIZE;
126 * Read a page synchronously.
128 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
131 unsigned int rsize = NFS_SERVER(inode)->rsize;
132 unsigned int count = PAGE_CACHE_SIZE;
133 int result = -ENOMEM;
134 struct nfs_read_data *rdata;
136 rdata = nfs_readdata_alloc(count);
140 memset(rdata, 0, sizeof(*rdata));
141 rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
142 rdata->cred = ctx->cred;
143 rdata->inode = inode;
144 INIT_LIST_HEAD(&rdata->pages);
145 rdata->args.fh = NFS_FH(inode);
146 rdata->args.context = ctx;
147 rdata->args.pages = &page;
148 rdata->args.pgbase = 0UL;
149 rdata->args.count = rsize;
150 rdata->res.fattr = &rdata->fattr;
152 dprintk("NFS: nfs_readpage_sync(%p)\n", page);
155 * This works now because the socket layer never tries to DMA
156 * into this buffer directly.
160 rdata->args.count = count;
161 rdata->res.count = rdata->args.count;
162 rdata->args.offset = page_offset(page) + rdata->args.pgbase;
164 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
165 NFS_SERVER(inode)->nfs_client->cl_hostname,
167 (long long)NFS_FILEID(inode),
168 (unsigned long long)rdata->args.pgbase,
172 result = NFS_PROTO(inode)->read(rdata);
176 * Even if we had a partial success we can't mark the page
180 if (result == -EISDIR)
185 rdata->args.pgbase += result;
186 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, result);
188 /* Note: result == 0 should only happen if we're caching
189 * a write that extends the file and punches a hole.
191 if (rdata->res.eof != 0 || result == 0)
194 spin_lock(&inode->i_lock);
195 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
196 spin_unlock(&inode->i_lock);
198 if (rdata->res.eof || rdata->res.count == rdata->args.count) {
199 SetPageUptodate(page);
200 if (rdata->res.eof && count != 0)
201 memclear_highpage_flush(page, rdata->args.pgbase, count);
206 nfs_readdata_free(rdata);
212 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
215 LIST_HEAD(one_request);
216 struct nfs_page *new;
219 len = nfs_page_length(page);
221 return nfs_return_empty_page(page);
222 new = nfs_create_request(ctx, inode, page, 0, len);
227 if (len < PAGE_CACHE_SIZE)
228 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
230 nfs_list_add_request(new, &one_request);
231 nfs_pagein_one(&one_request, inode);
235 static void nfs_readpage_release(struct nfs_page *req)
237 unlock_page(req->wb_page);
239 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
240 req->wb_context->dentry->d_inode->i_sb->s_id,
241 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
243 (long long)req_offset(req));
244 nfs_clear_request(req);
245 nfs_release_request(req);
249 * Set up the NFS read request struct
251 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
252 const struct rpc_call_ops *call_ops,
253 unsigned int count, unsigned int offset)
259 data->inode = inode = req->wb_context->dentry->d_inode;
260 data->cred = req->wb_context->cred;
262 data->args.fh = NFS_FH(inode);
263 data->args.offset = req_offset(req) + offset;
264 data->args.pgbase = req->wb_pgbase + offset;
265 data->args.pages = data->pagevec;
266 data->args.count = count;
267 data->args.context = req->wb_context;
269 data->res.fattr = &data->fattr;
270 data->res.count = count;
272 nfs_fattr_init(&data->fattr);
274 /* Set up the initial task struct. */
275 flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
276 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
277 NFS_PROTO(inode)->read_setup(data);
279 data->task.tk_cookie = (unsigned long)inode;
281 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
284 (long long)NFS_FILEID(inode),
286 (unsigned long long)data->args.offset);
290 nfs_async_read_error(struct list_head *head)
292 struct nfs_page *req;
294 while (!list_empty(head)) {
295 req = nfs_list_entry(head->next);
296 nfs_list_remove_request(req);
297 SetPageError(req->wb_page);
298 nfs_readpage_release(req);
303 * Start an async read operation
305 static void nfs_execute_read(struct nfs_read_data *data)
307 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
310 rpc_clnt_sigmask(clnt, &oldset);
311 rpc_execute(&data->task);
312 rpc_clnt_sigunmask(clnt, &oldset);
316 * Generate multiple requests to fill a single page.
318 * We optimize to reduce the number of read operations on the wire. If we
319 * detect that we're reading a page, or an area of a page, that is past the
320 * end of file, we do not generate NFS read operations but just clear the
321 * parts of the page that would have come back zero from the server anyway.
323 * We rely on the cached value of i_size to make this determination; another
324 * client can fill pages on the server past our cached end-of-file, but we
325 * won't see the new data until our attribute cache is updated. This is more
326 * or less conventional NFS client behavior.
328 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
330 struct nfs_page *req = nfs_list_entry(head->next);
331 struct page *page = req->wb_page;
332 struct nfs_read_data *data;
333 size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
338 nfs_list_remove_request(req);
340 nbytes = req->wb_bytes;
342 size_t len = min(nbytes,rsize);
344 data = nfs_readdata_alloc(len);
347 INIT_LIST_HEAD(&data->pages);
348 list_add(&data->pages, &list);
351 } while(nbytes != 0);
352 atomic_set(&req->wb_complete, requests);
354 ClearPageError(page);
356 nbytes = req->wb_bytes;
358 data = list_entry(list.next, struct nfs_read_data, pages);
359 list_del_init(&data->pages);
361 data->pagevec[0] = page;
363 if (nbytes > rsize) {
364 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
369 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
373 nfs_execute_read(data);
374 } while (nbytes != 0);
379 while (!list_empty(&list)) {
380 data = list_entry(list.next, struct nfs_read_data, pages);
381 list_del(&data->pages);
382 nfs_readdata_free(data);
385 nfs_readpage_release(req);
389 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
391 struct nfs_page *req;
393 struct nfs_read_data *data;
396 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
397 return nfs_pagein_multi(head, inode);
399 data = nfs_readdata_alloc(NFS_SERVER(inode)->rsize);
403 INIT_LIST_HEAD(&data->pages);
404 pages = data->pagevec;
406 while (!list_empty(head)) {
407 req = nfs_list_entry(head->next);
408 nfs_list_remove_request(req);
409 nfs_list_add_request(req, &data->pages);
410 ClearPageError(req->wb_page);
411 *pages++ = req->wb_page;
412 count += req->wb_bytes;
414 req = nfs_list_entry(data->pages.next);
416 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
418 nfs_execute_read(data);
421 nfs_async_read_error(head);
426 nfs_pagein_list(struct list_head *head, int rpages)
428 LIST_HEAD(one_request);
429 struct nfs_page *req;
431 unsigned int pages = 0;
433 while (!list_empty(head)) {
434 pages += nfs_coalesce_requests(head, &one_request, rpages);
435 req = nfs_list_entry(one_request.next);
436 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
443 nfs_async_read_error(head);
448 * This is the callback from RPC telling us whether a reply was
449 * received or some error occurred (timeout or socket shutdown).
451 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
455 dprintk("%s: %4d, (status %d)\n", __FUNCTION__, task->tk_pid,
458 status = NFS_PROTO(data->inode)->read_done(task, data);
462 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
464 if (task->tk_status == -ESTALE) {
465 set_bit(NFS_INO_STALE, &NFS_FLAGS(data->inode));
466 nfs_mark_for_revalidate(data->inode);
468 spin_lock(&data->inode->i_lock);
469 NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
470 spin_unlock(&data->inode->i_lock);
474 static int nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
476 struct nfs_readargs *argp = &data->args;
477 struct nfs_readres *resp = &data->res;
479 if (resp->eof || resp->count == argp->count)
482 /* This is a short read! */
483 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
484 /* Has the server at least made some progress? */
485 if (resp->count == 0)
488 /* Yes, so retry the read at the end of the data */
489 argp->offset += resp->count;
490 argp->pgbase += resp->count;
491 argp->count -= resp->count;
492 rpc_restart_call(task);
497 * Handle a read reply that fills part of a page.
499 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
501 struct nfs_read_data *data = calldata;
502 struct nfs_page *req = data->req;
503 struct page *page = req->wb_page;
505 if (nfs_readpage_result(task, data) != 0)
508 if (likely(task->tk_status >= 0)) {
509 nfs_readpage_truncate_uninitialised_page(data);
510 if (nfs_readpage_retry(task, data) != 0)
513 if (unlikely(task->tk_status < 0))
515 if (atomic_dec_and_test(&req->wb_complete)) {
516 if (!PageError(page))
517 SetPageUptodate(page);
518 nfs_readpage_release(req);
522 static const struct rpc_call_ops nfs_read_partial_ops = {
523 .rpc_call_done = nfs_readpage_result_partial,
524 .rpc_release = nfs_readdata_release,
527 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
529 unsigned int count = data->res.count;
530 unsigned int base = data->args.pgbase;
534 count = data->args.count;
535 if (unlikely(count == 0))
537 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
538 base &= ~PAGE_CACHE_MASK;
540 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
541 SetPageUptodate(*pages);
544 /* Was this a short read? */
545 if (data->res.eof || data->res.count == data->args.count)
546 SetPageUptodate(*pages);
550 * This is the callback from RPC telling us whether a reply was
551 * received or some error occurred (timeout or socket shutdown).
553 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
555 struct nfs_read_data *data = calldata;
557 if (nfs_readpage_result(task, data) != 0)
560 * Note: nfs_readpage_retry may change the values of
561 * data->args. In the multi-page case, we therefore need
562 * to ensure that we call nfs_readpage_set_pages_uptodate()
565 if (likely(task->tk_status >= 0)) {
566 nfs_readpage_truncate_uninitialised_page(data);
567 nfs_readpage_set_pages_uptodate(data);
568 if (nfs_readpage_retry(task, data) != 0)
571 while (!list_empty(&data->pages)) {
572 struct nfs_page *req = nfs_list_entry(data->pages.next);
574 nfs_list_remove_request(req);
575 nfs_readpage_release(req);
579 static const struct rpc_call_ops nfs_read_full_ops = {
580 .rpc_call_done = nfs_readpage_result_full,
581 .rpc_release = nfs_readdata_release,
585 * Read a page over NFS.
586 * We read the page synchronously in the following case:
587 * - The error flag is set for this page. This happens only when a
588 * previous async read operation failed.
590 int nfs_readpage(struct file *file, struct page *page)
592 struct nfs_open_context *ctx;
593 struct inode *inode = page->mapping->host;
596 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
597 page, PAGE_CACHE_SIZE, page->index);
598 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
599 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
602 * Try to flush any pending writes to the file..
604 * NOTE! Because we own the page lock, there cannot
605 * be any new pending writes generated at this point
606 * for this page (other pages can be written to).
608 error = nfs_wb_page(inode, page);
613 if (NFS_STALE(inode))
618 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
622 ctx = get_nfs_open_context((struct nfs_open_context *)
624 if (!IS_SYNC(inode)) {
625 error = nfs_readpage_async(ctx, inode, page);
629 error = nfs_readpage_sync(ctx, inode, page);
630 if (error < 0 && IS_SWAPFILE(inode))
631 printk("Aiee.. nfs swap-in of page failed!\n");
633 put_nfs_open_context(ctx);
641 struct nfs_readdesc {
642 struct list_head *head;
643 struct nfs_open_context *ctx;
647 readpage_async_filler(void *data, struct page *page)
649 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
650 struct inode *inode = page->mapping->host;
651 struct nfs_page *new;
654 nfs_wb_page(inode, page);
655 len = nfs_page_length(page);
657 return nfs_return_empty_page(page);
658 new = nfs_create_request(desc->ctx, inode, page, 0, len);
664 if (len < PAGE_CACHE_SIZE)
665 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
666 nfs_list_add_request(new, desc->head);
670 int nfs_readpages(struct file *filp, struct address_space *mapping,
671 struct list_head *pages, unsigned nr_pages)
674 struct nfs_readdesc desc = {
677 struct inode *inode = mapping->host;
678 struct nfs_server *server = NFS_SERVER(inode);
681 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
683 (long long)NFS_FILEID(inode),
685 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
687 if (NFS_STALE(inode))
691 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
692 if (desc.ctx == NULL)
695 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
697 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
698 if (!list_empty(&head)) {
699 int err = nfs_pagein_list(&head, server->rpages);
701 nfs_add_stats(inode, NFSIOS_READPAGES, err);
704 put_nfs_open_context(desc.ctx);
709 int __init nfs_init_readpagecache(void)
711 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
712 sizeof(struct nfs_read_data),
713 0, SLAB_HWCACHE_ALIGN,
715 if (nfs_rdata_cachep == NULL)
718 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
720 if (nfs_rdata_mempool == NULL)
726 void nfs_destroy_readpagecache(void)
728 mempool_destroy(nfs_rdata_mempool);
729 kmem_cache_destroy(nfs_rdata_cachep);