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);
108 * Read a page synchronously.
110 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
113 unsigned int rsize = NFS_SERVER(inode)->rsize;
114 unsigned int count = PAGE_CACHE_SIZE;
116 struct nfs_read_data *rdata;
118 rdata = nfs_readdata_alloc(1);
122 memset(rdata, 0, sizeof(*rdata));
123 rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
124 rdata->cred = ctx->cred;
125 rdata->inode = inode;
126 INIT_LIST_HEAD(&rdata->pages);
127 rdata->args.fh = NFS_FH(inode);
128 rdata->args.context = ctx;
129 rdata->args.pages = &page;
130 rdata->args.pgbase = 0UL;
131 rdata->args.count = rsize;
132 rdata->res.fattr = &rdata->fattr;
134 dprintk("NFS: nfs_readpage_sync(%p)\n", page);
137 * This works now because the socket layer never tries to DMA
138 * into this buffer directly.
142 rdata->args.count = count;
143 rdata->res.count = rdata->args.count;
144 rdata->args.offset = page_offset(page) + rdata->args.pgbase;
146 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
147 NFS_SERVER(inode)->hostname,
149 (long long)NFS_FILEID(inode),
150 (unsigned long long)rdata->args.pgbase,
154 result = NFS_PROTO(inode)->read(rdata);
158 * Even if we had a partial success we can't mark the page
162 if (result == -EISDIR)
167 rdata->args.pgbase += result;
168 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, result);
170 /* Note: result == 0 should only happen if we're caching
171 * a write that extends the file and punches a hole.
173 if (rdata->res.eof != 0 || result == 0)
176 spin_lock(&inode->i_lock);
177 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
178 spin_unlock(&inode->i_lock);
181 memclear_highpage_flush(page, rdata->args.pgbase, count);
182 SetPageUptodate(page);
184 ClearPageError(page);
189 nfs_readdata_free(rdata);
193 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
196 LIST_HEAD(one_request);
197 struct nfs_page *new;
200 len = nfs_page_length(inode, page);
202 return nfs_return_empty_page(page);
203 new = nfs_create_request(ctx, inode, page, 0, len);
208 if (len < PAGE_CACHE_SIZE)
209 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
211 nfs_list_add_request(new, &one_request);
212 nfs_pagein_one(&one_request, inode);
216 static void nfs_readpage_release(struct nfs_page *req)
218 unlock_page(req->wb_page);
220 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
221 req->wb_context->dentry->d_inode->i_sb->s_id,
222 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
224 (long long)req_offset(req));
225 nfs_clear_request(req);
226 nfs_release_request(req);
230 * Set up the NFS read request struct
232 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
233 const struct rpc_call_ops *call_ops,
234 unsigned int count, unsigned int offset)
240 data->inode = inode = req->wb_context->dentry->d_inode;
241 data->cred = req->wb_context->cred;
243 data->args.fh = NFS_FH(inode);
244 data->args.offset = req_offset(req) + offset;
245 data->args.pgbase = req->wb_pgbase + offset;
246 data->args.pages = data->pagevec;
247 data->args.count = count;
248 data->args.context = req->wb_context;
250 data->res.fattr = &data->fattr;
251 data->res.count = count;
253 nfs_fattr_init(&data->fattr);
255 /* Set up the initial task struct. */
256 flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
257 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
258 NFS_PROTO(inode)->read_setup(data);
260 data->task.tk_cookie = (unsigned long)inode;
262 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
265 (long long)NFS_FILEID(inode),
267 (unsigned long long)data->args.offset);
271 nfs_async_read_error(struct list_head *head)
273 struct nfs_page *req;
275 while (!list_empty(head)) {
276 req = nfs_list_entry(head->next);
277 nfs_list_remove_request(req);
278 SetPageError(req->wb_page);
279 nfs_readpage_release(req);
284 * Start an async read operation
286 static void nfs_execute_read(struct nfs_read_data *data)
288 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
291 rpc_clnt_sigmask(clnt, &oldset);
293 rpc_execute(&data->task);
295 rpc_clnt_sigunmask(clnt, &oldset);
299 * Generate multiple requests to fill a single page.
301 * We optimize to reduce the number of read operations on the wire. If we
302 * detect that we're reading a page, or an area of a page, that is past the
303 * end of file, we do not generate NFS read operations but just clear the
304 * parts of the page that would have come back zero from the server anyway.
306 * We rely on the cached value of i_size to make this determination; another
307 * client can fill pages on the server past our cached end-of-file, but we
308 * won't see the new data until our attribute cache is updated. This is more
309 * or less conventional NFS client behavior.
311 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
313 struct nfs_page *req = nfs_list_entry(head->next);
314 struct page *page = req->wb_page;
315 struct nfs_read_data *data;
316 unsigned int rsize = NFS_SERVER(inode)->rsize;
317 unsigned int nbytes, offset;
321 nfs_list_remove_request(req);
323 nbytes = req->wb_bytes;
325 data = nfs_readdata_alloc(1);
328 INIT_LIST_HEAD(&data->pages);
329 list_add(&data->pages, &list);
335 atomic_set(&req->wb_complete, requests);
337 ClearPageError(page);
339 nbytes = req->wb_bytes;
341 data = list_entry(list.next, struct nfs_read_data, pages);
342 list_del_init(&data->pages);
344 data->pagevec[0] = page;
346 if (nbytes > rsize) {
347 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
352 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
356 nfs_execute_read(data);
357 } while (nbytes != 0);
362 while (!list_empty(&list)) {
363 data = list_entry(list.next, struct nfs_read_data, pages);
364 list_del(&data->pages);
365 nfs_readdata_free(data);
368 nfs_readpage_release(req);
372 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
374 struct nfs_page *req;
376 struct nfs_read_data *data;
379 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
380 return nfs_pagein_multi(head, inode);
382 data = nfs_readdata_alloc(NFS_SERVER(inode)->rpages);
386 INIT_LIST_HEAD(&data->pages);
387 pages = data->pagevec;
389 while (!list_empty(head)) {
390 req = nfs_list_entry(head->next);
391 nfs_list_remove_request(req);
392 nfs_list_add_request(req, &data->pages);
393 ClearPageError(req->wb_page);
394 *pages++ = req->wb_page;
395 count += req->wb_bytes;
397 req = nfs_list_entry(data->pages.next);
399 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
401 nfs_execute_read(data);
404 nfs_async_read_error(head);
409 nfs_pagein_list(struct list_head *head, int rpages)
411 LIST_HEAD(one_request);
412 struct nfs_page *req;
414 unsigned int pages = 0;
416 while (!list_empty(head)) {
417 pages += nfs_coalesce_requests(head, &one_request, rpages);
418 req = nfs_list_entry(one_request.next);
419 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
426 nfs_async_read_error(head);
431 * Handle a read reply that fills part of a page.
433 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
435 struct nfs_read_data *data = calldata;
436 struct nfs_page *req = data->req;
437 struct page *page = req->wb_page;
439 if (nfs_readpage_result(task, data) != 0)
441 if (task->tk_status >= 0) {
442 unsigned int request = data->args.count;
443 unsigned int result = data->res.count;
445 if (result < request) {
446 memclear_highpage_flush(page,
447 data->args.pgbase + result,
453 if (atomic_dec_and_test(&req->wb_complete)) {
454 if (!PageError(page))
455 SetPageUptodate(page);
456 nfs_readpage_release(req);
460 static const struct rpc_call_ops nfs_read_partial_ops = {
461 .rpc_call_done = nfs_readpage_result_partial,
462 .rpc_release = nfs_readdata_release,
466 * This is the callback from RPC telling us whether a reply was
467 * received or some error occurred (timeout or socket shutdown).
469 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
471 struct nfs_read_data *data = calldata;
472 unsigned int count = data->res.count;
474 if (nfs_readpage_result(task, data) != 0)
476 while (!list_empty(&data->pages)) {
477 struct nfs_page *req = nfs_list_entry(data->pages.next);
478 struct page *page = req->wb_page;
479 nfs_list_remove_request(req);
481 if (task->tk_status >= 0) {
482 if (count < PAGE_CACHE_SIZE) {
483 if (count < req->wb_bytes)
484 memclear_highpage_flush(page,
485 req->wb_pgbase + count,
486 req->wb_bytes - count);
489 count -= PAGE_CACHE_SIZE;
490 SetPageUptodate(page);
493 nfs_readpage_release(req);
497 static const struct rpc_call_ops nfs_read_full_ops = {
498 .rpc_call_done = nfs_readpage_result_full,
499 .rpc_release = nfs_readdata_release,
503 * This is the callback from RPC telling us whether a reply was
504 * received or some error occurred (timeout or socket shutdown).
506 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
508 struct nfs_readargs *argp = &data->args;
509 struct nfs_readres *resp = &data->res;
512 dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
513 task->tk_pid, task->tk_status);
515 status = NFS_PROTO(data->inode)->read_done(task, data);
519 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, resp->count);
521 /* Is this a short read? */
522 if (task->tk_status >= 0 && resp->count < argp->count && !resp->eof) {
523 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
524 /* Has the server at least made some progress? */
525 if (resp->count != 0) {
526 /* Yes, so retry the read at the end of the data */
527 argp->offset += resp->count;
528 argp->pgbase += resp->count;
529 argp->count -= resp->count;
530 rpc_restart_call(task);
533 task->tk_status = -EIO;
535 spin_lock(&data->inode->i_lock);
536 NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
537 spin_unlock(&data->inode->i_lock);
542 * Read a page over NFS.
543 * We read the page synchronously in the following case:
544 * - The error flag is set for this page. This happens only when a
545 * previous async read operation failed.
547 int nfs_readpage(struct file *file, struct page *page)
549 struct nfs_open_context *ctx;
550 struct inode *inode = page->mapping->host;
553 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
554 page, PAGE_CACHE_SIZE, page->index);
555 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
556 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
559 * Try to flush any pending writes to the file..
561 * NOTE! Because we own the page lock, there cannot
562 * be any new pending writes generated at this point
563 * for this page (other pages can be written to).
565 error = nfs_wb_page(inode, page);
570 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
574 ctx = get_nfs_open_context((struct nfs_open_context *)
576 if (!IS_SYNC(inode)) {
577 error = nfs_readpage_async(ctx, inode, page);
581 error = nfs_readpage_sync(ctx, inode, page);
582 if (error < 0 && IS_SWAPFILE(inode))
583 printk("Aiee.. nfs swap-in of page failed!\n");
585 put_nfs_open_context(ctx);
593 struct nfs_readdesc {
594 struct list_head *head;
595 struct nfs_open_context *ctx;
599 readpage_async_filler(void *data, struct page *page)
601 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
602 struct inode *inode = page->mapping->host;
603 struct nfs_page *new;
606 nfs_wb_page(inode, page);
607 len = nfs_page_length(inode, page);
609 return nfs_return_empty_page(page);
610 new = nfs_create_request(desc->ctx, inode, page, 0, len);
616 if (len < PAGE_CACHE_SIZE)
617 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
618 nfs_list_add_request(new, desc->head);
622 int nfs_readpages(struct file *filp, struct address_space *mapping,
623 struct list_head *pages, unsigned nr_pages)
626 struct nfs_readdesc desc = {
629 struct inode *inode = mapping->host;
630 struct nfs_server *server = NFS_SERVER(inode);
633 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
635 (long long)NFS_FILEID(inode),
637 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
640 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
641 if (desc.ctx == NULL)
644 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
646 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
647 if (!list_empty(&head)) {
648 int err = nfs_pagein_list(&head, server->rpages);
650 nfs_add_stats(inode, NFSIOS_READPAGES, err);
653 put_nfs_open_context(desc.ctx);
657 int nfs_init_readpagecache(void)
659 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
660 sizeof(struct nfs_read_data),
661 0, SLAB_HWCACHE_ALIGN,
663 if (nfs_rdata_cachep == NULL)
666 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
668 if (nfs_rdata_mempool == NULL)
674 void nfs_destroy_readpagecache(void)
676 mempool_destroy(nfs_rdata_mempool);
677 if (kmem_cache_destroy(nfs_rdata_cachep))
678 printk(KERN_INFO "nfs_read_data: not all structures were freed\n");