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 static 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;
120 if (remainder <= pglen) {
121 memclear_highpage_flush(*pages, base, remainder);
124 memclear_highpage_flush(*pages, base, pglen);
127 pglen = PAGE_CACHE_SIZE;
133 * Read a page synchronously.
135 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
138 unsigned int rsize = NFS_SERVER(inode)->rsize;
139 unsigned int count = PAGE_CACHE_SIZE;
141 struct nfs_read_data *rdata;
143 rdata = nfs_readdata_alloc(1);
147 memset(rdata, 0, sizeof(*rdata));
148 rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
149 rdata->cred = ctx->cred;
150 rdata->inode = inode;
151 INIT_LIST_HEAD(&rdata->pages);
152 rdata->args.fh = NFS_FH(inode);
153 rdata->args.context = ctx;
154 rdata->args.pages = &page;
155 rdata->args.pgbase = 0UL;
156 rdata->args.count = rsize;
157 rdata->res.fattr = &rdata->fattr;
159 dprintk("NFS: nfs_readpage_sync(%p)\n", page);
162 * This works now because the socket layer never tries to DMA
163 * into this buffer directly.
167 rdata->args.count = count;
168 rdata->res.count = rdata->args.count;
169 rdata->args.offset = page_offset(page) + rdata->args.pgbase;
171 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
172 NFS_SERVER(inode)->hostname,
174 (long long)NFS_FILEID(inode),
175 (unsigned long long)rdata->args.pgbase,
179 result = NFS_PROTO(inode)->read(rdata);
183 * Even if we had a partial success we can't mark the page
187 if (result == -EISDIR)
192 rdata->args.pgbase += result;
193 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, result);
195 /* Note: result == 0 should only happen if we're caching
196 * a write that extends the file and punches a hole.
198 if (rdata->res.eof != 0 || result == 0)
201 spin_lock(&inode->i_lock);
202 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
203 spin_unlock(&inode->i_lock);
205 nfs_readpage_truncate_uninitialised_page(rdata);
206 if (rdata->res.eof || rdata->res.count == rdata->args.count)
207 SetPageUptodate(page);
212 nfs_readdata_free(rdata);
216 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
219 LIST_HEAD(one_request);
220 struct nfs_page *new;
223 len = nfs_page_length(inode, page);
225 return nfs_return_empty_page(page);
226 new = nfs_create_request(ctx, inode, page, 0, len);
231 if (len < PAGE_CACHE_SIZE)
232 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
234 nfs_list_add_request(new, &one_request);
235 nfs_pagein_one(&one_request, inode);
239 static void nfs_readpage_release(struct nfs_page *req)
241 unlock_page(req->wb_page);
243 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
244 req->wb_context->dentry->d_inode->i_sb->s_id,
245 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
247 (long long)req_offset(req));
248 nfs_clear_request(req);
249 nfs_release_request(req);
253 * Set up the NFS read request struct
255 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
256 const struct rpc_call_ops *call_ops,
257 unsigned int count, unsigned int offset)
263 data->inode = inode = req->wb_context->dentry->d_inode;
264 data->cred = req->wb_context->cred;
266 data->args.fh = NFS_FH(inode);
267 data->args.offset = req_offset(req) + offset;
268 data->args.pgbase = req->wb_pgbase + offset;
269 data->args.pages = data->pagevec;
270 data->args.count = count;
271 data->args.context = req->wb_context;
273 data->res.fattr = &data->fattr;
274 data->res.count = count;
276 nfs_fattr_init(&data->fattr);
278 /* Set up the initial task struct. */
279 flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
280 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
281 NFS_PROTO(inode)->read_setup(data);
283 data->task.tk_cookie = (unsigned long)inode;
285 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
288 (long long)NFS_FILEID(inode),
290 (unsigned long long)data->args.offset);
294 nfs_async_read_error(struct list_head *head)
296 struct nfs_page *req;
298 while (!list_empty(head)) {
299 req = nfs_list_entry(head->next);
300 nfs_list_remove_request(req);
301 SetPageError(req->wb_page);
302 nfs_readpage_release(req);
307 * Start an async read operation
309 static void nfs_execute_read(struct nfs_read_data *data)
311 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
314 rpc_clnt_sigmask(clnt, &oldset);
316 rpc_execute(&data->task);
318 rpc_clnt_sigunmask(clnt, &oldset);
322 * Generate multiple requests to fill a single page.
324 * We optimize to reduce the number of read operations on the wire. If we
325 * detect that we're reading a page, or an area of a page, that is past the
326 * end of file, we do not generate NFS read operations but just clear the
327 * parts of the page that would have come back zero from the server anyway.
329 * We rely on the cached value of i_size to make this determination; another
330 * client can fill pages on the server past our cached end-of-file, but we
331 * won't see the new data until our attribute cache is updated. This is more
332 * or less conventional NFS client behavior.
334 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
336 struct nfs_page *req = nfs_list_entry(head->next);
337 struct page *page = req->wb_page;
338 struct nfs_read_data *data;
339 unsigned int rsize = NFS_SERVER(inode)->rsize;
340 unsigned int nbytes, offset;
344 nfs_list_remove_request(req);
346 nbytes = req->wb_bytes;
348 data = nfs_readdata_alloc(1);
351 INIT_LIST_HEAD(&data->pages);
352 list_add(&data->pages, &list);
358 atomic_set(&req->wb_complete, requests);
360 ClearPageError(page);
362 nbytes = req->wb_bytes;
364 data = list_entry(list.next, struct nfs_read_data, pages);
365 list_del_init(&data->pages);
367 data->pagevec[0] = page;
369 if (nbytes > rsize) {
370 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
375 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
379 nfs_execute_read(data);
380 } while (nbytes != 0);
385 while (!list_empty(&list)) {
386 data = list_entry(list.next, struct nfs_read_data, pages);
387 list_del(&data->pages);
388 nfs_readdata_free(data);
391 nfs_readpage_release(req);
395 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
397 struct nfs_page *req;
399 struct nfs_read_data *data;
402 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
403 return nfs_pagein_multi(head, inode);
405 data = nfs_readdata_alloc(NFS_SERVER(inode)->rpages);
409 INIT_LIST_HEAD(&data->pages);
410 pages = data->pagevec;
412 while (!list_empty(head)) {
413 req = nfs_list_entry(head->next);
414 nfs_list_remove_request(req);
415 nfs_list_add_request(req, &data->pages);
416 ClearPageError(req->wb_page);
417 *pages++ = req->wb_page;
418 count += req->wb_bytes;
420 req = nfs_list_entry(data->pages.next);
422 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
424 nfs_execute_read(data);
427 nfs_async_read_error(head);
432 nfs_pagein_list(struct list_head *head, int rpages)
434 LIST_HEAD(one_request);
435 struct nfs_page *req;
437 unsigned int pages = 0;
439 while (!list_empty(head)) {
440 pages += nfs_coalesce_requests(head, &one_request, rpages);
441 req = nfs_list_entry(one_request.next);
442 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
449 nfs_async_read_error(head);
454 * Handle a read reply that fills part of a page.
456 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
458 struct nfs_read_data *data = calldata;
459 struct nfs_page *req = data->req;
460 struct page *page = req->wb_page;
462 if (likely(task->tk_status >= 0))
463 nfs_readpage_truncate_uninitialised_page(data);
466 if (nfs_readpage_result(task, data) != 0)
468 if (atomic_dec_and_test(&req->wb_complete)) {
469 if (!PageError(page))
470 SetPageUptodate(page);
471 nfs_readpage_release(req);
475 static const struct rpc_call_ops nfs_read_partial_ops = {
476 .rpc_call_done = nfs_readpage_result_partial,
477 .rpc_release = nfs_readdata_release,
480 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
482 unsigned int count = data->res.count;
483 unsigned int base = data->args.pgbase;
487 count = data->args.count;
488 if (unlikely(count == 0))
490 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
491 base &= ~PAGE_CACHE_MASK;
493 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
494 SetPageUptodate(*pages);
496 SetPageUptodate(*pages);
499 static void nfs_readpage_set_pages_error(struct nfs_read_data *data)
501 unsigned int count = data->args.count;
502 unsigned int base = data->args.pgbase;
505 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
506 base &= ~PAGE_CACHE_MASK;
508 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
509 SetPageError(*pages);
511 SetPageError(*pages);
515 * This is the callback from RPC telling us whether a reply was
516 * received or some error occurred (timeout or socket shutdown).
518 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
520 struct nfs_read_data *data = calldata;
523 * Note: nfs_readpage_result may change the values of
524 * data->args. In the multi-page case, we therefore need
525 * to ensure that we call the next nfs_readpage_set_page_uptodate()
526 * first in the multi-page case.
528 if (likely(task->tk_status >= 0)) {
529 nfs_readpage_truncate_uninitialised_page(data);
530 nfs_readpage_set_pages_uptodate(data);
532 nfs_readpage_set_pages_error(data);
533 if (nfs_readpage_result(task, data) != 0)
535 while (!list_empty(&data->pages)) {
536 struct nfs_page *req = nfs_list_entry(data->pages.next);
538 nfs_list_remove_request(req);
539 nfs_readpage_release(req);
543 static const struct rpc_call_ops nfs_read_full_ops = {
544 .rpc_call_done = nfs_readpage_result_full,
545 .rpc_release = nfs_readdata_release,
549 * This is the callback from RPC telling us whether a reply was
550 * received or some error occurred (timeout or socket shutdown).
552 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
554 struct nfs_readargs *argp = &data->args;
555 struct nfs_readres *resp = &data->res;
558 dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
559 task->tk_pid, task->tk_status);
561 status = NFS_PROTO(data->inode)->read_done(task, data);
565 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, resp->count);
567 /* Is this a short read? */
568 if (task->tk_status >= 0 && resp->count < argp->count && !resp->eof) {
569 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
570 /* Has the server at least made some progress? */
571 if (resp->count != 0) {
572 /* Yes, so retry the read at the end of the data */
573 argp->offset += resp->count;
574 argp->pgbase += resp->count;
575 argp->count -= resp->count;
576 rpc_restart_call(task);
579 task->tk_status = -EIO;
581 spin_lock(&data->inode->i_lock);
582 NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
583 spin_unlock(&data->inode->i_lock);
588 * Read a page over NFS.
589 * We read the page synchronously in the following case:
590 * - The error flag is set for this page. This happens only when a
591 * previous async read operation failed.
593 int nfs_readpage(struct file *file, struct page *page)
595 struct nfs_open_context *ctx;
596 struct inode *inode = page->mapping->host;
599 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
600 page, PAGE_CACHE_SIZE, page->index);
601 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
602 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
605 * Try to flush any pending writes to the file..
607 * NOTE! Because we own the page lock, there cannot
608 * be any new pending writes generated at this point
609 * for this page (other pages can be written to).
611 error = nfs_wb_page(inode, page);
616 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
620 ctx = get_nfs_open_context((struct nfs_open_context *)
622 if (!IS_SYNC(inode)) {
623 error = nfs_readpage_async(ctx, inode, page);
627 error = nfs_readpage_sync(ctx, inode, page);
628 if (error < 0 && IS_SWAPFILE(inode))
629 printk("Aiee.. nfs swap-in of page failed!\n");
631 put_nfs_open_context(ctx);
639 struct nfs_readdesc {
640 struct list_head *head;
641 struct nfs_open_context *ctx;
645 readpage_async_filler(void *data, struct page *page)
647 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
648 struct inode *inode = page->mapping->host;
649 struct nfs_page *new;
652 nfs_wb_page(inode, page);
653 len = nfs_page_length(inode, page);
655 return nfs_return_empty_page(page);
656 new = nfs_create_request(desc->ctx, inode, page, 0, len);
662 if (len < PAGE_CACHE_SIZE)
663 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
664 nfs_list_add_request(new, desc->head);
668 int nfs_readpages(struct file *filp, struct address_space *mapping,
669 struct list_head *pages, unsigned nr_pages)
672 struct nfs_readdesc desc = {
675 struct inode *inode = mapping->host;
676 struct nfs_server *server = NFS_SERVER(inode);
679 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
681 (long long)NFS_FILEID(inode),
683 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
686 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
687 if (desc.ctx == NULL)
690 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
692 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
693 if (!list_empty(&head)) {
694 int err = nfs_pagein_list(&head, server->rpages);
696 nfs_add_stats(inode, NFSIOS_READPAGES, err);
699 put_nfs_open_context(desc.ctx);
703 int __init nfs_init_readpagecache(void)
705 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
706 sizeof(struct nfs_read_data),
707 0, SLAB_HWCACHE_ALIGN,
709 if (nfs_rdata_cachep == NULL)
712 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
714 if (nfs_rdata_mempool == NULL)
720 void nfs_destroy_readpagecache(void)
722 mempool_destroy(nfs_rdata_mempool);
723 if (kmem_cache_destroy(nfs_rdata_cachep))
724 printk(KERN_INFO "nfs_read_data: not all structures were freed\n");