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/config.h>
19 #include <linux/time.h>
20 #include <linux/kernel.h>
21 #include <linux/errno.h>
22 #include <linux/fcntl.h>
23 #include <linux/stat.h>
25 #include <linux/slab.h>
26 #include <linux/pagemap.h>
27 #include <linux/sunrpc/clnt.h>
28 #include <linux/nfs_fs.h>
29 #include <linux/nfs_page.h>
30 #include <linux/smp_lock.h>
32 #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 kmem_cache_t *nfs_rdata_cachep;
43 static mempool_t *nfs_rdata_mempool;
45 #define MIN_POOL_READ (32)
47 struct nfs_read_data *nfs_readdata_alloc(unsigned int pagecount)
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 if (pagecount < NFS_PAGEVEC_SIZE)
55 p->pagevec = &p->page_array[0];
57 size_t size = ++pagecount * sizeof(struct page *);
58 p->pagevec = kmalloc(size, GFP_NOFS);
60 memset(p->pagevec, 0, size);
62 mempool_free(p, nfs_rdata_mempool);
70 void nfs_readdata_free(struct nfs_read_data *p)
72 if (p && (p->pagevec != &p->page_array[0]))
74 mempool_free(p, nfs_rdata_mempool);
77 void nfs_readdata_release(void *data)
79 nfs_readdata_free(data);
83 unsigned int nfs_page_length(struct inode *inode, struct page *page)
85 loff_t i_size = i_size_read(inode);
90 idx = (i_size - 1) >> PAGE_CACHE_SHIFT;
91 if (page->index > idx)
93 if (page->index != idx)
94 return PAGE_CACHE_SIZE;
95 return 1 + ((i_size - 1) & (PAGE_CACHE_SIZE - 1));
99 int nfs_return_empty_page(struct page *page)
101 memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
102 SetPageUptodate(page);
107 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
109 unsigned int remainder = data->args.count - data->res.count;
110 unsigned int base = data->args.pgbase + data->res.count;
114 if (data->res.eof == 0 || remainder == 0)
117 * Note: "remainder" can never be negative, since we check for
118 * this in the XDR code.
120 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
121 base &= ~PAGE_CACHE_MASK;
122 pglen = PAGE_CACHE_SIZE - base;
123 if (pglen < remainder)
124 memclear_highpage_flush(*pages, base, pglen);
126 memclear_highpage_flush(*pages, base, remainder);
130 * Read a page synchronously.
132 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
135 unsigned int rsize = NFS_SERVER(inode)->rsize;
136 unsigned int count = PAGE_CACHE_SIZE;
138 struct nfs_read_data *rdata;
140 rdata = nfs_readdata_alloc(1);
144 memset(rdata, 0, sizeof(*rdata));
145 rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
146 rdata->cred = ctx->cred;
147 rdata->inode = inode;
148 INIT_LIST_HEAD(&rdata->pages);
149 rdata->args.fh = NFS_FH(inode);
150 rdata->args.context = ctx;
151 rdata->args.pages = &page;
152 rdata->args.pgbase = 0UL;
153 rdata->args.count = rsize;
154 rdata->res.fattr = &rdata->fattr;
156 dprintk("NFS: nfs_readpage_sync(%p)\n", page);
159 * This works now because the socket layer never tries to DMA
160 * into this buffer directly.
164 rdata->args.count = count;
165 rdata->res.count = rdata->args.count;
166 rdata->args.offset = page_offset(page) + rdata->args.pgbase;
168 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
169 NFS_SERVER(inode)->hostname,
171 (long long)NFS_FILEID(inode),
172 (unsigned long long)rdata->args.pgbase,
176 result = NFS_PROTO(inode)->read(rdata);
180 * Even if we had a partial success we can't mark the page
184 if (result == -EISDIR)
189 rdata->args.pgbase += result;
190 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, result);
192 /* Note: result == 0 should only happen if we're caching
193 * a write that extends the file and punches a hole.
195 if (rdata->res.eof != 0 || result == 0)
198 spin_lock(&inode->i_lock);
199 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
200 spin_unlock(&inode->i_lock);
202 nfs_readpage_truncate_uninitialised_page(rdata);
203 if (rdata->res.eof || rdata->res.count == rdata->args.count)
204 SetPageUptodate(page);
209 nfs_readdata_free(rdata);
213 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
216 LIST_HEAD(one_request);
217 struct nfs_page *new;
220 len = nfs_page_length(inode, page);
222 return nfs_return_empty_page(page);
223 new = nfs_create_request(ctx, inode, page, 0, len);
228 if (len < PAGE_CACHE_SIZE)
229 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
231 nfs_list_add_request(new, &one_request);
232 nfs_pagein_one(&one_request, inode);
236 static void nfs_readpage_release(struct nfs_page *req)
238 unlock_page(req->wb_page);
240 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
241 req->wb_context->dentry->d_inode->i_sb->s_id,
242 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
244 (long long)req_offset(req));
245 nfs_clear_request(req);
246 nfs_release_request(req);
250 * Set up the NFS read request struct
252 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
253 const struct rpc_call_ops *call_ops,
254 unsigned int count, unsigned int offset)
260 data->inode = inode = req->wb_context->dentry->d_inode;
261 data->cred = req->wb_context->cred;
263 data->args.fh = NFS_FH(inode);
264 data->args.offset = req_offset(req) + offset;
265 data->args.pgbase = req->wb_pgbase + offset;
266 data->args.pages = data->pagevec;
267 data->args.count = count;
268 data->args.context = req->wb_context;
270 data->res.fattr = &data->fattr;
271 data->res.count = count;
273 nfs_fattr_init(&data->fattr);
275 /* Set up the initial task struct. */
276 flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
277 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
278 NFS_PROTO(inode)->read_setup(data);
280 data->task.tk_cookie = (unsigned long)inode;
282 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
285 (long long)NFS_FILEID(inode),
287 (unsigned long long)data->args.offset);
291 nfs_async_read_error(struct list_head *head)
293 struct nfs_page *req;
295 while (!list_empty(head)) {
296 req = nfs_list_entry(head->next);
297 nfs_list_remove_request(req);
298 SetPageError(req->wb_page);
299 nfs_readpage_release(req);
304 * Start an async read operation
306 static void nfs_execute_read(struct nfs_read_data *data)
308 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
311 rpc_clnt_sigmask(clnt, &oldset);
313 rpc_execute(&data->task);
315 rpc_clnt_sigunmask(clnt, &oldset);
319 * Generate multiple requests to fill a single page.
321 * We optimize to reduce the number of read operations on the wire. If we
322 * detect that we're reading a page, or an area of a page, that is past the
323 * end of file, we do not generate NFS read operations but just clear the
324 * parts of the page that would have come back zero from the server anyway.
326 * We rely on the cached value of i_size to make this determination; another
327 * client can fill pages on the server past our cached end-of-file, but we
328 * won't see the new data until our attribute cache is updated. This is more
329 * or less conventional NFS client behavior.
331 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
333 struct nfs_page *req = nfs_list_entry(head->next);
334 struct page *page = req->wb_page;
335 struct nfs_read_data *data;
336 unsigned int rsize = NFS_SERVER(inode)->rsize;
337 unsigned int nbytes, offset;
341 nfs_list_remove_request(req);
343 nbytes = req->wb_bytes;
345 data = nfs_readdata_alloc(1);
348 INIT_LIST_HEAD(&data->pages);
349 list_add(&data->pages, &list);
355 atomic_set(&req->wb_complete, requests);
357 ClearPageError(page);
359 nbytes = req->wb_bytes;
361 data = list_entry(list.next, struct nfs_read_data, pages);
362 list_del_init(&data->pages);
364 data->pagevec[0] = page;
366 if (nbytes > rsize) {
367 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
372 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
376 nfs_execute_read(data);
377 } while (nbytes != 0);
382 while (!list_empty(&list)) {
383 data = list_entry(list.next, struct nfs_read_data, pages);
384 list_del(&data->pages);
385 nfs_readdata_free(data);
388 nfs_readpage_release(req);
392 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
394 struct nfs_page *req;
396 struct nfs_read_data *data;
399 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
400 return nfs_pagein_multi(head, inode);
402 data = nfs_readdata_alloc(NFS_SERVER(inode)->rpages);
406 INIT_LIST_HEAD(&data->pages);
407 pages = data->pagevec;
409 while (!list_empty(head)) {
410 req = nfs_list_entry(head->next);
411 nfs_list_remove_request(req);
412 nfs_list_add_request(req, &data->pages);
413 ClearPageError(req->wb_page);
414 *pages++ = req->wb_page;
415 count += req->wb_bytes;
417 req = nfs_list_entry(data->pages.next);
419 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
421 nfs_execute_read(data);
424 nfs_async_read_error(head);
429 nfs_pagein_list(struct list_head *head, int rpages)
431 LIST_HEAD(one_request);
432 struct nfs_page *req;
434 unsigned int pages = 0;
436 while (!list_empty(head)) {
437 pages += nfs_coalesce_requests(head, &one_request, rpages);
438 req = nfs_list_entry(one_request.next);
439 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
446 nfs_async_read_error(head);
451 * Handle a read reply that fills part of a page.
453 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
455 struct nfs_read_data *data = calldata;
456 struct nfs_page *req = data->req;
457 struct page *page = req->wb_page;
459 if (likely(task->tk_status >= 0))
460 nfs_readpage_truncate_uninitialised_page(data);
463 if (nfs_readpage_result(task, data) != 0)
465 if (atomic_dec_and_test(&req->wb_complete)) {
466 if (!PageError(page))
467 SetPageUptodate(page);
468 nfs_readpage_release(req);
472 static const struct rpc_call_ops nfs_read_partial_ops = {
473 .rpc_call_done = nfs_readpage_result_partial,
474 .rpc_release = nfs_readdata_release,
477 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
479 unsigned int count = data->res.count;
480 unsigned int base = data->args.pgbase;
483 if (unlikely(count == 0))
485 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
486 base &= ~PAGE_CACHE_MASK;
488 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
489 SetPageUptodate(*pages);
491 * Was this an eof or a short read? If the latter, don't mark the page
494 if (count > 0 && (data->res.eof || data->args.count == data->res.count))
495 SetPageUptodate(*pages);
498 static void nfs_readpage_set_pages_error(struct nfs_read_data *data)
500 unsigned int count = data->args.count;
501 unsigned int base = data->args.pgbase;
504 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
505 base &= ~PAGE_CACHE_MASK;
507 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
508 SetPageError(*pages);
512 * This is the callback from RPC telling us whether a reply was
513 * received or some error occurred (timeout or socket shutdown).
515 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
517 struct nfs_read_data *data = calldata;
520 * Note: nfs_readpage_result may change the values of
521 * data->args. In the multi-page case, we therefore need
522 * to ensure that we call the next nfs_readpage_set_page_uptodate()
523 * first in the multi-page case.
525 if (likely(task->tk_status >= 0)) {
526 nfs_readpage_truncate_uninitialised_page(data);
527 nfs_readpage_set_pages_uptodate(data);
529 nfs_readpage_set_pages_error(data);
530 if (nfs_readpage_result(task, data) != 0)
532 while (!list_empty(&data->pages)) {
533 struct nfs_page *req = nfs_list_entry(data->pages.next);
535 nfs_list_remove_request(req);
536 nfs_readpage_release(req);
540 static const struct rpc_call_ops nfs_read_full_ops = {
541 .rpc_call_done = nfs_readpage_result_full,
542 .rpc_release = nfs_readdata_release,
546 * This is the callback from RPC telling us whether a reply was
547 * received or some error occurred (timeout or socket shutdown).
549 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
551 struct nfs_readargs *argp = &data->args;
552 struct nfs_readres *resp = &data->res;
555 dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
556 task->tk_pid, task->tk_status);
558 status = NFS_PROTO(data->inode)->read_done(task, data);
562 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, resp->count);
564 /* Is this a short read? */
565 if (task->tk_status >= 0 && resp->count < argp->count && !resp->eof) {
566 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
567 /* Has the server at least made some progress? */
568 if (resp->count != 0) {
569 /* Yes, so retry the read at the end of the data */
570 argp->offset += resp->count;
571 argp->pgbase += resp->count;
572 argp->count -= resp->count;
573 rpc_restart_call(task);
576 task->tk_status = -EIO;
578 spin_lock(&data->inode->i_lock);
579 NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
580 spin_unlock(&data->inode->i_lock);
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 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
617 ctx = get_nfs_open_context((struct nfs_open_context *)
619 if (!IS_SYNC(inode)) {
620 error = nfs_readpage_async(ctx, inode, page);
624 error = nfs_readpage_sync(ctx, inode, page);
625 if (error < 0 && IS_SWAPFILE(inode))
626 printk("Aiee.. nfs swap-in of page failed!\n");
628 put_nfs_open_context(ctx);
636 struct nfs_readdesc {
637 struct list_head *head;
638 struct nfs_open_context *ctx;
642 readpage_async_filler(void *data, struct page *page)
644 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
645 struct inode *inode = page->mapping->host;
646 struct nfs_page *new;
649 nfs_wb_page(inode, page);
650 len = nfs_page_length(inode, page);
652 return nfs_return_empty_page(page);
653 new = nfs_create_request(desc->ctx, inode, page, 0, len);
659 if (len < PAGE_CACHE_SIZE)
660 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
661 nfs_list_add_request(new, desc->head);
665 int nfs_readpages(struct file *filp, struct address_space *mapping,
666 struct list_head *pages, unsigned nr_pages)
669 struct nfs_readdesc desc = {
672 struct inode *inode = mapping->host;
673 struct nfs_server *server = NFS_SERVER(inode);
676 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
678 (long long)NFS_FILEID(inode),
680 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
683 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
684 if (desc.ctx == NULL)
687 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
689 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
690 if (!list_empty(&head)) {
691 int err = nfs_pagein_list(&head, server->rpages);
693 nfs_add_stats(inode, NFSIOS_READPAGES, err);
696 put_nfs_open_context(desc.ctx);
700 int nfs_init_readpagecache(void)
702 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
703 sizeof(struct nfs_read_data),
704 0, SLAB_HWCACHE_ALIGN,
706 if (nfs_rdata_cachep == NULL)
709 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
711 if (nfs_rdata_mempool == NULL)
717 void nfs_destroy_readpagecache(void)
719 mempool_destroy(nfs_rdata_mempool);
720 if (kmem_cache_destroy(nfs_rdata_cachep))
721 printk(KERN_INFO "nfs_read_data: not all structures were freed\n");