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>
34 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
36 static int nfs_pagein_one(struct list_head *, struct inode *);
37 static void nfs_readpage_result_partial(struct nfs_read_data *, int);
38 static void nfs_readpage_result_full(struct nfs_read_data *, int);
40 static kmem_cache_t *nfs_rdata_cachep;
41 mempool_t *nfs_rdata_mempool;
43 #define MIN_POOL_READ (32)
45 void nfs_readdata_release(struct rpc_task *task)
47 struct nfs_read_data *data = (struct nfs_read_data *)task->tk_calldata;
48 nfs_readdata_free(data);
52 unsigned int nfs_page_length(struct inode *inode, struct page *page)
54 loff_t i_size = i_size_read(inode);
59 idx = (i_size - 1) >> PAGE_CACHE_SHIFT;
60 if (page->index > idx)
62 if (page->index != idx)
63 return PAGE_CACHE_SIZE;
64 return 1 + ((i_size - 1) & (PAGE_CACHE_SIZE - 1));
68 int nfs_return_empty_page(struct page *page)
70 memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
71 SetPageUptodate(page);
77 * Read a page synchronously.
79 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
82 unsigned int rsize = NFS_SERVER(inode)->rsize;
83 unsigned int count = PAGE_CACHE_SIZE;
85 struct nfs_read_data *rdata;
87 rdata = nfs_readdata_alloc();
91 memset(rdata, 0, sizeof(*rdata));
92 rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
93 rdata->cred = ctx->cred;
95 INIT_LIST_HEAD(&rdata->pages);
96 rdata->args.fh = NFS_FH(inode);
97 rdata->args.context = ctx;
98 rdata->args.pages = &page;
99 rdata->args.pgbase = 0UL;
100 rdata->args.count = rsize;
101 rdata->res.fattr = &rdata->fattr;
103 dprintk("NFS: nfs_readpage_sync(%p)\n", page);
106 * This works now because the socket layer never tries to DMA
107 * into this buffer directly.
111 rdata->args.count = count;
112 rdata->res.count = rdata->args.count;
113 rdata->args.offset = page_offset(page) + rdata->args.pgbase;
115 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
116 NFS_SERVER(inode)->hostname,
118 (long long)NFS_FILEID(inode),
119 (unsigned long long)rdata->args.pgbase,
123 result = NFS_PROTO(inode)->read(rdata);
127 * Even if we had a partial success we can't mark the page
131 if (result == -EISDIR)
136 rdata->args.pgbase += result;
137 /* Note: result == 0 should only happen if we're caching
138 * a write that extends the file and punches a hole.
140 if (rdata->res.eof != 0 || result == 0)
143 spin_lock(&inode->i_lock);
144 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
145 spin_unlock(&inode->i_lock);
148 memclear_highpage_flush(page, rdata->args.pgbase, count);
149 SetPageUptodate(page);
151 ClearPageError(page);
156 nfs_readdata_free(rdata);
160 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
163 LIST_HEAD(one_request);
164 struct nfs_page *new;
167 len = nfs_page_length(inode, page);
169 return nfs_return_empty_page(page);
170 new = nfs_create_request(ctx, inode, page, 0, len);
175 if (len < PAGE_CACHE_SIZE)
176 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
178 nfs_list_add_request(new, &one_request);
179 nfs_pagein_one(&one_request, inode);
183 static void nfs_readpage_release(struct nfs_page *req)
185 unlock_page(req->wb_page);
187 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
188 req->wb_context->dentry->d_inode->i_sb->s_id,
189 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
191 (long long)req_offset(req));
192 nfs_clear_request(req);
193 nfs_release_request(req);
197 * Set up the NFS read request struct
199 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
200 unsigned int count, unsigned int offset)
205 data->inode = inode = req->wb_context->dentry->d_inode;
206 data->cred = req->wb_context->cred;
208 data->args.fh = NFS_FH(inode);
209 data->args.offset = req_offset(req) + offset;
210 data->args.pgbase = req->wb_pgbase + offset;
211 data->args.pages = data->pagevec;
212 data->args.count = count;
213 data->args.context = req->wb_context;
215 data->res.fattr = &data->fattr;
216 data->res.count = count;
218 nfs_fattr_init(&data->fattr);
220 NFS_PROTO(inode)->read_setup(data);
222 data->task.tk_cookie = (unsigned long)inode;
223 data->task.tk_calldata = data;
224 /* Release requests */
225 data->task.tk_release = nfs_readdata_release;
227 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
230 (long long)NFS_FILEID(inode),
232 (unsigned long long)data->args.offset);
236 nfs_async_read_error(struct list_head *head)
238 struct nfs_page *req;
240 while (!list_empty(head)) {
241 req = nfs_list_entry(head->next);
242 nfs_list_remove_request(req);
243 SetPageError(req->wb_page);
244 nfs_readpage_release(req);
249 * Start an async read operation
251 static void nfs_execute_read(struct nfs_read_data *data)
253 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
256 rpc_clnt_sigmask(clnt, &oldset);
258 rpc_execute(&data->task);
260 rpc_clnt_sigunmask(clnt, &oldset);
264 * Generate multiple requests to fill a single page.
266 * We optimize to reduce the number of read operations on the wire. If we
267 * detect that we're reading a page, or an area of a page, that is past the
268 * end of file, we do not generate NFS read operations but just clear the
269 * parts of the page that would have come back zero from the server anyway.
271 * We rely on the cached value of i_size to make this determination; another
272 * client can fill pages on the server past our cached end-of-file, but we
273 * won't see the new data until our attribute cache is updated. This is more
274 * or less conventional NFS client behavior.
276 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
278 struct nfs_page *req = nfs_list_entry(head->next);
279 struct page *page = req->wb_page;
280 struct nfs_read_data *data;
281 unsigned int rsize = NFS_SERVER(inode)->rsize;
282 unsigned int nbytes, offset;
286 nfs_list_remove_request(req);
288 nbytes = req->wb_bytes;
290 data = nfs_readdata_alloc();
293 INIT_LIST_HEAD(&data->pages);
294 list_add(&data->pages, &list);
300 atomic_set(&req->wb_complete, requests);
302 ClearPageError(page);
304 nbytes = req->wb_bytes;
306 data = list_entry(list.next, struct nfs_read_data, pages);
307 list_del_init(&data->pages);
309 data->pagevec[0] = page;
310 data->complete = nfs_readpage_result_partial;
312 if (nbytes > rsize) {
313 nfs_read_rpcsetup(req, data, rsize, offset);
317 nfs_read_rpcsetup(req, data, nbytes, offset);
320 nfs_execute_read(data);
321 } while (nbytes != 0);
326 while (!list_empty(&list)) {
327 data = list_entry(list.next, struct nfs_read_data, pages);
328 list_del(&data->pages);
329 nfs_readdata_free(data);
332 nfs_readpage_release(req);
336 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
338 struct nfs_page *req;
340 struct nfs_read_data *data;
343 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
344 return nfs_pagein_multi(head, inode);
346 data = nfs_readdata_alloc();
350 INIT_LIST_HEAD(&data->pages);
351 pages = data->pagevec;
353 while (!list_empty(head)) {
354 req = nfs_list_entry(head->next);
355 nfs_list_remove_request(req);
356 nfs_list_add_request(req, &data->pages);
357 ClearPageError(req->wb_page);
358 *pages++ = req->wb_page;
359 count += req->wb_bytes;
361 req = nfs_list_entry(data->pages.next);
363 data->complete = nfs_readpage_result_full;
364 nfs_read_rpcsetup(req, data, count, 0);
366 nfs_execute_read(data);
369 nfs_async_read_error(head);
374 nfs_pagein_list(struct list_head *head, int rpages)
376 LIST_HEAD(one_request);
377 struct nfs_page *req;
379 unsigned int pages = 0;
381 while (!list_empty(head)) {
382 pages += nfs_coalesce_requests(head, &one_request, rpages);
383 req = nfs_list_entry(one_request.next);
384 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
391 nfs_async_read_error(head);
396 * Handle a read reply that fills part of a page.
398 static void nfs_readpage_result_partial(struct nfs_read_data *data, int status)
400 struct nfs_page *req = data->req;
401 struct page *page = req->wb_page;
404 unsigned int request = data->args.count;
405 unsigned int result = data->res.count;
407 if (result < request) {
408 memclear_highpage_flush(page,
409 data->args.pgbase + result,
415 if (atomic_dec_and_test(&req->wb_complete)) {
416 if (!PageError(page))
417 SetPageUptodate(page);
418 nfs_readpage_release(req);
423 * This is the callback from RPC telling us whether a reply was
424 * received or some error occurred (timeout or socket shutdown).
426 static void nfs_readpage_result_full(struct nfs_read_data *data, int status)
428 unsigned int count = data->res.count;
430 while (!list_empty(&data->pages)) {
431 struct nfs_page *req = nfs_list_entry(data->pages.next);
432 struct page *page = req->wb_page;
433 nfs_list_remove_request(req);
436 if (count < PAGE_CACHE_SIZE) {
437 if (count < req->wb_bytes)
438 memclear_highpage_flush(page,
439 req->wb_pgbase + count,
440 req->wb_bytes - count);
443 count -= PAGE_CACHE_SIZE;
444 SetPageUptodate(page);
447 nfs_readpage_release(req);
452 * This is the callback from RPC telling us whether a reply was
453 * received or some error occurred (timeout or socket shutdown).
455 void nfs_readpage_result(struct rpc_task *task)
457 struct nfs_read_data *data = (struct nfs_read_data *)task->tk_calldata;
458 struct nfs_readargs *argp = &data->args;
459 struct nfs_readres *resp = &data->res;
460 int status = task->tk_status;
462 dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
463 task->tk_pid, status);
465 /* Is this a short read? */
466 if (task->tk_status >= 0 && resp->count < argp->count && !resp->eof) {
467 /* Has the server at least made some progress? */
468 if (resp->count != 0) {
469 /* Yes, so retry the read at the end of the data */
470 argp->offset += resp->count;
471 argp->pgbase += resp->count;
472 argp->count -= resp->count;
473 rpc_restart_call(task);
476 task->tk_status = -EIO;
478 spin_lock(&data->inode->i_lock);
479 NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
480 spin_unlock(&data->inode->i_lock);
481 data->complete(data, status);
485 * Read a page over NFS.
486 * We read the page synchronously in the following case:
487 * - The error flag is set for this page. This happens only when a
488 * previous async read operation failed.
490 int nfs_readpage(struct file *file, struct page *page)
492 struct nfs_open_context *ctx;
493 struct inode *inode = page->mapping->host;
496 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
497 page, PAGE_CACHE_SIZE, page->index);
499 * Try to flush any pending writes to the file..
501 * NOTE! Because we own the page lock, there cannot
502 * be any new pending writes generated at this point
503 * for this page (other pages can be written to).
505 error = nfs_wb_page(inode, page);
510 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
514 ctx = get_nfs_open_context((struct nfs_open_context *)
516 if (!IS_SYNC(inode)) {
517 error = nfs_readpage_async(ctx, inode, page);
521 error = nfs_readpage_sync(ctx, inode, page);
522 if (error < 0 && IS_SWAPFILE(inode))
523 printk("Aiee.. nfs swap-in of page failed!\n");
525 put_nfs_open_context(ctx);
533 struct nfs_readdesc {
534 struct list_head *head;
535 struct nfs_open_context *ctx;
539 readpage_async_filler(void *data, struct page *page)
541 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
542 struct inode *inode = page->mapping->host;
543 struct nfs_page *new;
546 nfs_wb_page(inode, page);
547 len = nfs_page_length(inode, page);
549 return nfs_return_empty_page(page);
550 new = nfs_create_request(desc->ctx, inode, page, 0, len);
556 if (len < PAGE_CACHE_SIZE)
557 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
558 nfs_list_add_request(new, desc->head);
562 int nfs_readpages(struct file *filp, struct address_space *mapping,
563 struct list_head *pages, unsigned nr_pages)
566 struct nfs_readdesc desc = {
569 struct inode *inode = mapping->host;
570 struct nfs_server *server = NFS_SERVER(inode);
573 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
575 (long long)NFS_FILEID(inode),
579 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
580 if (desc.ctx == NULL)
583 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
585 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
586 if (!list_empty(&head)) {
587 int err = nfs_pagein_list(&head, server->rpages);
591 put_nfs_open_context(desc.ctx);
595 int nfs_init_readpagecache(void)
597 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
598 sizeof(struct nfs_read_data),
599 0, SLAB_HWCACHE_ALIGN,
601 if (nfs_rdata_cachep == NULL)
604 nfs_rdata_mempool = mempool_create(MIN_POOL_READ,
608 if (nfs_rdata_mempool == NULL)
614 void nfs_destroy_readpagecache(void)
616 mempool_destroy(nfs_rdata_mempool);
617 if (kmem_cache_destroy(nfs_rdata_cachep))
618 printk(KERN_INFO "nfs_read_data: not all structures were freed\n");