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(unsigned int pagecount)
48 struct nfs_read_data *p = mempool_alloc(nfs_rdata_mempool, SLAB_NOFS);
51 memset(p, 0, sizeof(*p));
52 INIT_LIST_HEAD(&p->pages);
53 if (pagecount <= ARRAY_SIZE(p->page_array))
54 p->pagevec = p->page_array;
56 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
58 mempool_free(p, nfs_rdata_mempool);
66 void nfs_readdata_free(struct nfs_read_data *p)
68 if (p && (p->pagevec != &p->page_array[0]))
70 mempool_free(p, nfs_rdata_mempool);
73 void nfs_readdata_release(void *data)
75 nfs_readdata_free(data);
79 unsigned int nfs_page_length(struct inode *inode, struct page *page)
81 loff_t i_size = i_size_read(inode);
86 idx = (i_size - 1) >> PAGE_CACHE_SHIFT;
87 if (page->index > idx)
89 if (page->index != idx)
90 return PAGE_CACHE_SIZE;
91 return 1 + ((i_size - 1) & (PAGE_CACHE_SIZE - 1));
95 int nfs_return_empty_page(struct page *page)
97 memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
98 SetPageUptodate(page);
103 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
105 unsigned int remainder = data->args.count - data->res.count;
106 unsigned int base = data->args.pgbase + data->res.count;
110 if (data->res.eof == 0 || remainder == 0)
113 * Note: "remainder" can never be negative, since we check for
114 * this in the XDR code.
116 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
117 base &= ~PAGE_CACHE_MASK;
118 pglen = PAGE_CACHE_SIZE - base;
119 if (pglen < remainder)
120 memclear_highpage_flush(*pages, base, pglen);
122 memclear_highpage_flush(*pages, base, remainder);
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;
134 struct nfs_read_data *rdata;
136 rdata = nfs_readdata_alloc(1);
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)->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 nfs_readpage_truncate_uninitialised_page(rdata);
199 if (rdata->res.eof || rdata->res.count == rdata->args.count)
200 SetPageUptodate(page);
205 nfs_readdata_free(rdata);
209 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
212 LIST_HEAD(one_request);
213 struct nfs_page *new;
216 len = nfs_page_length(inode, page);
218 return nfs_return_empty_page(page);
219 new = nfs_create_request(ctx, inode, page, 0, len);
224 if (len < PAGE_CACHE_SIZE)
225 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
227 nfs_list_add_request(new, &one_request);
228 nfs_pagein_one(&one_request, inode);
232 static void nfs_readpage_release(struct nfs_page *req)
234 unlock_page(req->wb_page);
236 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
237 req->wb_context->dentry->d_inode->i_sb->s_id,
238 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
240 (long long)req_offset(req));
241 nfs_clear_request(req);
242 nfs_release_request(req);
246 * Set up the NFS read request struct
248 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
249 const struct rpc_call_ops *call_ops,
250 unsigned int count, unsigned int offset)
256 data->inode = inode = req->wb_context->dentry->d_inode;
257 data->cred = req->wb_context->cred;
259 data->args.fh = NFS_FH(inode);
260 data->args.offset = req_offset(req) + offset;
261 data->args.pgbase = req->wb_pgbase + offset;
262 data->args.pages = data->pagevec;
263 data->args.count = count;
264 data->args.context = req->wb_context;
266 data->res.fattr = &data->fattr;
267 data->res.count = count;
269 nfs_fattr_init(&data->fattr);
271 /* Set up the initial task struct. */
272 flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
273 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
274 NFS_PROTO(inode)->read_setup(data);
276 data->task.tk_cookie = (unsigned long)inode;
278 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
281 (long long)NFS_FILEID(inode),
283 (unsigned long long)data->args.offset);
287 nfs_async_read_error(struct list_head *head)
289 struct nfs_page *req;
291 while (!list_empty(head)) {
292 req = nfs_list_entry(head->next);
293 nfs_list_remove_request(req);
294 SetPageError(req->wb_page);
295 nfs_readpage_release(req);
300 * Start an async read operation
302 static void nfs_execute_read(struct nfs_read_data *data)
304 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
307 rpc_clnt_sigmask(clnt, &oldset);
309 rpc_execute(&data->task);
311 rpc_clnt_sigunmask(clnt, &oldset);
315 * Generate multiple requests to fill a single page.
317 * We optimize to reduce the number of read operations on the wire. If we
318 * detect that we're reading a page, or an area of a page, that is past the
319 * end of file, we do not generate NFS read operations but just clear the
320 * parts of the page that would have come back zero from the server anyway.
322 * We rely on the cached value of i_size to make this determination; another
323 * client can fill pages on the server past our cached end-of-file, but we
324 * won't see the new data until our attribute cache is updated. This is more
325 * or less conventional NFS client behavior.
327 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
329 struct nfs_page *req = nfs_list_entry(head->next);
330 struct page *page = req->wb_page;
331 struct nfs_read_data *data;
332 unsigned int rsize = NFS_SERVER(inode)->rsize;
333 unsigned int nbytes, offset;
337 nfs_list_remove_request(req);
339 nbytes = req->wb_bytes;
341 data = nfs_readdata_alloc(1);
344 INIT_LIST_HEAD(&data->pages);
345 list_add(&data->pages, &list);
351 atomic_set(&req->wb_complete, requests);
353 ClearPageError(page);
355 nbytes = req->wb_bytes;
357 data = list_entry(list.next, struct nfs_read_data, pages);
358 list_del_init(&data->pages);
360 data->pagevec[0] = page;
362 if (nbytes > rsize) {
363 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
368 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
372 nfs_execute_read(data);
373 } while (nbytes != 0);
378 while (!list_empty(&list)) {
379 data = list_entry(list.next, struct nfs_read_data, pages);
380 list_del(&data->pages);
381 nfs_readdata_free(data);
384 nfs_readpage_release(req);
388 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
390 struct nfs_page *req;
392 struct nfs_read_data *data;
395 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
396 return nfs_pagein_multi(head, inode);
398 data = nfs_readdata_alloc(NFS_SERVER(inode)->rpages);
402 INIT_LIST_HEAD(&data->pages);
403 pages = data->pagevec;
405 while (!list_empty(head)) {
406 req = nfs_list_entry(head->next);
407 nfs_list_remove_request(req);
408 nfs_list_add_request(req, &data->pages);
409 ClearPageError(req->wb_page);
410 *pages++ = req->wb_page;
411 count += req->wb_bytes;
413 req = nfs_list_entry(data->pages.next);
415 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
417 nfs_execute_read(data);
420 nfs_async_read_error(head);
425 nfs_pagein_list(struct list_head *head, int rpages)
427 LIST_HEAD(one_request);
428 struct nfs_page *req;
430 unsigned int pages = 0;
432 while (!list_empty(head)) {
433 pages += nfs_coalesce_requests(head, &one_request, rpages);
434 req = nfs_list_entry(one_request.next);
435 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
442 nfs_async_read_error(head);
447 * Handle a read reply that fills part of a page.
449 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
451 struct nfs_read_data *data = calldata;
452 struct nfs_page *req = data->req;
453 struct page *page = req->wb_page;
455 if (likely(task->tk_status >= 0))
456 nfs_readpage_truncate_uninitialised_page(data);
459 if (nfs_readpage_result(task, data) != 0)
461 if (atomic_dec_and_test(&req->wb_complete)) {
462 if (!PageError(page))
463 SetPageUptodate(page);
464 nfs_readpage_release(req);
468 static const struct rpc_call_ops nfs_read_partial_ops = {
469 .rpc_call_done = nfs_readpage_result_partial,
470 .rpc_release = nfs_readdata_release,
473 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
475 unsigned int count = data->res.count;
476 unsigned int base = data->args.pgbase;
479 if (unlikely(count == 0))
481 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
482 base &= ~PAGE_CACHE_MASK;
484 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
485 SetPageUptodate(*pages);
487 * Was this an eof or a short read? If the latter, don't mark the page
490 if (count > 0 && (data->res.eof || data->args.count == data->res.count))
491 SetPageUptodate(*pages);
494 static void nfs_readpage_set_pages_error(struct nfs_read_data *data)
496 unsigned int count = data->args.count;
497 unsigned int base = data->args.pgbase;
500 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
501 base &= ~PAGE_CACHE_MASK;
503 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
504 SetPageError(*pages);
508 * This is the callback from RPC telling us whether a reply was
509 * received or some error occurred (timeout or socket shutdown).
511 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
513 struct nfs_read_data *data = calldata;
516 * Note: nfs_readpage_result may change the values of
517 * data->args. In the multi-page case, we therefore need
518 * to ensure that we call the next nfs_readpage_set_page_uptodate()
519 * first in the multi-page case.
521 if (likely(task->tk_status >= 0)) {
522 nfs_readpage_truncate_uninitialised_page(data);
523 nfs_readpage_set_pages_uptodate(data);
525 nfs_readpage_set_pages_error(data);
526 if (nfs_readpage_result(task, data) != 0)
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);
536 static const struct rpc_call_ops nfs_read_full_ops = {
537 .rpc_call_done = nfs_readpage_result_full,
538 .rpc_release = nfs_readdata_release,
542 * This is the callback from RPC telling us whether a reply was
543 * received or some error occurred (timeout or socket shutdown).
545 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
547 struct nfs_readargs *argp = &data->args;
548 struct nfs_readres *resp = &data->res;
551 dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
552 task->tk_pid, task->tk_status);
554 status = NFS_PROTO(data->inode)->read_done(task, data);
558 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, resp->count);
560 /* Is this a short read? */
561 if (task->tk_status >= 0 && resp->count < argp->count && !resp->eof) {
562 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
563 /* Has the server at least made some progress? */
564 if (resp->count != 0) {
565 /* Yes, so retry the read at the end of the data */
566 argp->offset += resp->count;
567 argp->pgbase += resp->count;
568 argp->count -= resp->count;
569 rpc_restart_call(task);
572 task->tk_status = -EIO;
574 spin_lock(&data->inode->i_lock);
575 NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
576 spin_unlock(&data->inode->i_lock);
581 * Read a page over NFS.
582 * We read the page synchronously in the following case:
583 * - The error flag is set for this page. This happens only when a
584 * previous async read operation failed.
586 int nfs_readpage(struct file *file, struct page *page)
588 struct nfs_open_context *ctx;
589 struct inode *inode = page->mapping->host;
592 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
593 page, PAGE_CACHE_SIZE, page->index);
594 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
595 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
598 * Try to flush any pending writes to the file..
600 * NOTE! Because we own the page lock, there cannot
601 * be any new pending writes generated at this point
602 * for this page (other pages can be written to).
604 error = nfs_wb_page(inode, page);
609 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
613 ctx = get_nfs_open_context((struct nfs_open_context *)
615 if (!IS_SYNC(inode)) {
616 error = nfs_readpage_async(ctx, inode, page);
620 error = nfs_readpage_sync(ctx, inode, page);
621 if (error < 0 && IS_SWAPFILE(inode))
622 printk("Aiee.. nfs swap-in of page failed!\n");
624 put_nfs_open_context(ctx);
632 struct nfs_readdesc {
633 struct list_head *head;
634 struct nfs_open_context *ctx;
638 readpage_async_filler(void *data, struct page *page)
640 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
641 struct inode *inode = page->mapping->host;
642 struct nfs_page *new;
645 nfs_wb_page(inode, page);
646 len = nfs_page_length(inode, page);
648 return nfs_return_empty_page(page);
649 new = nfs_create_request(desc->ctx, inode, page, 0, len);
655 if (len < PAGE_CACHE_SIZE)
656 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
657 nfs_list_add_request(new, desc->head);
661 int nfs_readpages(struct file *filp, struct address_space *mapping,
662 struct list_head *pages, unsigned nr_pages)
665 struct nfs_readdesc desc = {
668 struct inode *inode = mapping->host;
669 struct nfs_server *server = NFS_SERVER(inode);
672 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
674 (long long)NFS_FILEID(inode),
676 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
679 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
680 if (desc.ctx == NULL)
683 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
685 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
686 if (!list_empty(&head)) {
687 int err = nfs_pagein_list(&head, server->rpages);
689 nfs_add_stats(inode, NFSIOS_READPAGES, err);
692 put_nfs_open_context(desc.ctx);
696 int __init nfs_init_readpagecache(void)
698 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
699 sizeof(struct nfs_read_data),
700 0, SLAB_HWCACHE_ALIGN,
702 if (nfs_rdata_cachep == NULL)
705 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
707 if (nfs_rdata_mempool == NULL)
713 void nfs_destroy_readpagecache(void)
715 mempool_destroy(nfs_rdata_mempool);
716 if (kmem_cache_destroy(nfs_rdata_cachep))
717 printk(KERN_INFO "nfs_read_data: not all structures were freed\n");