xref: /openbmc/linux/fs/nfs/read.c (revision d5cb9783536a41df9f9cba5b0a1d78047ed787f7)
1 /*
2  * linux/fs/nfs/read.c
3  *
4  * Block I/O for NFS
5  *
6  * Partial copy of Linus' read cache modifications to fs/nfs/file.c
7  * modified for async RPC by okir@monad.swb.de
8  *
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.
16  */
17 
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>
24 #include <linux/mm.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>
31 
32 #include <asm/system.h>
33 
34 #define NFSDBG_FACILITY		NFSDBG_PAGECACHE
35 
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);
39 
40 static kmem_cache_t *nfs_rdata_cachep;
41 mempool_t *nfs_rdata_mempool;
42 
43 #define MIN_POOL_READ	(32)
44 
45 void nfs_readdata_release(struct rpc_task *task)
46 {
47         struct nfs_read_data   *data = (struct nfs_read_data *)task->tk_calldata;
48         nfs_readdata_free(data);
49 }
50 
51 static
52 unsigned int nfs_page_length(struct inode *inode, struct page *page)
53 {
54 	loff_t i_size = i_size_read(inode);
55 	unsigned long idx;
56 
57 	if (i_size <= 0)
58 		return 0;
59 	idx = (i_size - 1) >> PAGE_CACHE_SHIFT;
60 	if (page->index > idx)
61 		return 0;
62 	if (page->index != idx)
63 		return PAGE_CACHE_SIZE;
64 	return 1 + ((i_size - 1) & (PAGE_CACHE_SIZE - 1));
65 }
66 
67 static
68 int nfs_return_empty_page(struct page *page)
69 {
70 	memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
71 	SetPageUptodate(page);
72 	unlock_page(page);
73 	return 0;
74 }
75 
76 /*
77  * Read a page synchronously.
78  */
79 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
80 		struct page *page)
81 {
82 	unsigned int	rsize = NFS_SERVER(inode)->rsize;
83 	unsigned int	count = PAGE_CACHE_SIZE;
84 	int		result;
85 	struct nfs_read_data *rdata;
86 
87 	rdata = nfs_readdata_alloc();
88 	if (!rdata)
89 		return -ENOMEM;
90 
91 	memset(rdata, 0, sizeof(*rdata));
92 	rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
93 	rdata->cred = ctx->cred;
94 	rdata->inode = inode;
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;
102 
103 	dprintk("NFS: nfs_readpage_sync(%p)\n", page);
104 
105 	/*
106 	 * This works now because the socket layer never tries to DMA
107 	 * into this buffer directly.
108 	 */
109 	do {
110 		if (count < rsize)
111 			rdata->args.count = count;
112 		rdata->res.count = rdata->args.count;
113 		rdata->args.offset = page_offset(page) + rdata->args.pgbase;
114 
115 		dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
116 			NFS_SERVER(inode)->hostname,
117 			inode->i_sb->s_id,
118 			(long long)NFS_FILEID(inode),
119 			(unsigned long long)rdata->args.pgbase,
120 			rdata->args.count);
121 
122 		lock_kernel();
123 		result = NFS_PROTO(inode)->read(rdata);
124 		unlock_kernel();
125 
126 		/*
127 		 * Even if we had a partial success we can't mark the page
128 		 * cache valid.
129 		 */
130 		if (result < 0) {
131 			if (result == -EISDIR)
132 				result = -EINVAL;
133 			goto io_error;
134 		}
135 		count -= result;
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.
139 		 */
140 		if (rdata->res.eof != 0 || result == 0)
141 			break;
142 	} while (count);
143 	spin_lock(&inode->i_lock);
144 	NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
145 	spin_unlock(&inode->i_lock);
146 
147 	if (count)
148 		memclear_highpage_flush(page, rdata->args.pgbase, count);
149 	SetPageUptodate(page);
150 	if (PageError(page))
151 		ClearPageError(page);
152 	result = 0;
153 
154 io_error:
155 	unlock_page(page);
156 	nfs_readdata_free(rdata);
157 	return result;
158 }
159 
160 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
161 		struct page *page)
162 {
163 	LIST_HEAD(one_request);
164 	struct nfs_page	*new;
165 	unsigned int len;
166 
167 	len = nfs_page_length(inode, page);
168 	if (len == 0)
169 		return nfs_return_empty_page(page);
170 	new = nfs_create_request(ctx, inode, page, 0, len);
171 	if (IS_ERR(new)) {
172 		unlock_page(page);
173 		return PTR_ERR(new);
174 	}
175 	if (len < PAGE_CACHE_SIZE)
176 		memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
177 
178 	nfs_list_add_request(new, &one_request);
179 	nfs_pagein_one(&one_request, inode);
180 	return 0;
181 }
182 
183 static void nfs_readpage_release(struct nfs_page *req)
184 {
185 	unlock_page(req->wb_page);
186 
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),
190 			req->wb_bytes,
191 			(long long)req_offset(req));
192 	nfs_clear_request(req);
193 	nfs_release_request(req);
194 }
195 
196 /*
197  * Set up the NFS read request struct
198  */
199 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
200 		unsigned int count, unsigned int offset)
201 {
202 	struct inode		*inode;
203 
204 	data->req	  = req;
205 	data->inode	  = inode = req->wb_context->dentry->d_inode;
206 	data->cred	  = req->wb_context->cred;
207 
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;
214 
215 	data->res.fattr   = &data->fattr;
216 	data->res.count   = count;
217 	data->res.eof     = 0;
218 	nfs_fattr_init(&data->fattr);
219 
220 	NFS_PROTO(inode)->read_setup(data);
221 
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;
226 
227 	dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
228 			data->task.tk_pid,
229 			inode->i_sb->s_id,
230 			(long long)NFS_FILEID(inode),
231 			count,
232 			(unsigned long long)data->args.offset);
233 }
234 
235 static void
236 nfs_async_read_error(struct list_head *head)
237 {
238 	struct nfs_page	*req;
239 
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);
245 	}
246 }
247 
248 /*
249  * Start an async read operation
250  */
251 static void nfs_execute_read(struct nfs_read_data *data)
252 {
253 	struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
254 	sigset_t oldset;
255 
256 	rpc_clnt_sigmask(clnt, &oldset);
257 	lock_kernel();
258 	rpc_execute(&data->task);
259 	unlock_kernel();
260 	rpc_clnt_sigunmask(clnt, &oldset);
261 }
262 
263 /*
264  * Generate multiple requests to fill a single page.
265  *
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.
270  *
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.
275  */
276 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
277 {
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;
283 	int requests = 0;
284 	LIST_HEAD(list);
285 
286 	nfs_list_remove_request(req);
287 
288 	nbytes = req->wb_bytes;
289 	for(;;) {
290 		data = nfs_readdata_alloc();
291 		if (!data)
292 			goto out_bad;
293 		INIT_LIST_HEAD(&data->pages);
294 		list_add(&data->pages, &list);
295 		requests++;
296 		if (nbytes <= rsize)
297 			break;
298 		nbytes -= rsize;
299 	}
300 	atomic_set(&req->wb_complete, requests);
301 
302 	ClearPageError(page);
303 	offset = 0;
304 	nbytes = req->wb_bytes;
305 	do {
306 		data = list_entry(list.next, struct nfs_read_data, pages);
307 		list_del_init(&data->pages);
308 
309 		data->pagevec[0] = page;
310 		data->complete = nfs_readpage_result_partial;
311 
312 		if (nbytes > rsize) {
313 			nfs_read_rpcsetup(req, data, rsize, offset);
314 			offset += rsize;
315 			nbytes -= rsize;
316 		} else {
317 			nfs_read_rpcsetup(req, data, nbytes, offset);
318 			nbytes = 0;
319 		}
320 		nfs_execute_read(data);
321 	} while (nbytes != 0);
322 
323 	return 0;
324 
325 out_bad:
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);
330 	}
331 	SetPageError(page);
332 	nfs_readpage_release(req);
333 	return -ENOMEM;
334 }
335 
336 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
337 {
338 	struct nfs_page		*req;
339 	struct page		**pages;
340 	struct nfs_read_data	*data;
341 	unsigned int		count;
342 
343 	if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
344 		return nfs_pagein_multi(head, inode);
345 
346 	data = nfs_readdata_alloc();
347 	if (!data)
348 		goto out_bad;
349 
350 	INIT_LIST_HEAD(&data->pages);
351 	pages = data->pagevec;
352 	count = 0;
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;
360 	}
361 	req = nfs_list_entry(data->pages.next);
362 
363 	data->complete = nfs_readpage_result_full;
364 	nfs_read_rpcsetup(req, data, count, 0);
365 
366 	nfs_execute_read(data);
367 	return 0;
368 out_bad:
369 	nfs_async_read_error(head);
370 	return -ENOMEM;
371 }
372 
373 static int
374 nfs_pagein_list(struct list_head *head, int rpages)
375 {
376 	LIST_HEAD(one_request);
377 	struct nfs_page		*req;
378 	int			error = 0;
379 	unsigned int		pages = 0;
380 
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);
385 		if (error < 0)
386 			break;
387 	}
388 	if (error >= 0)
389 		return pages;
390 
391 	nfs_async_read_error(head);
392 	return error;
393 }
394 
395 /*
396  * Handle a read reply that fills part of a page.
397  */
398 static void nfs_readpage_result_partial(struct nfs_read_data *data, int status)
399 {
400 	struct nfs_page *req = data->req;
401 	struct page *page = req->wb_page;
402 
403 	if (status >= 0) {
404 		unsigned int request = data->args.count;
405 		unsigned int result = data->res.count;
406 
407 		if (result < request) {
408 			memclear_highpage_flush(page,
409 						data->args.pgbase + result,
410 						request - result);
411 		}
412 	} else
413 		SetPageError(page);
414 
415 	if (atomic_dec_and_test(&req->wb_complete)) {
416 		if (!PageError(page))
417 			SetPageUptodate(page);
418 		nfs_readpage_release(req);
419 	}
420 }
421 
422 /*
423  * This is the callback from RPC telling us whether a reply was
424  * received or some error occurred (timeout or socket shutdown).
425  */
426 static void nfs_readpage_result_full(struct nfs_read_data *data, int status)
427 {
428 	unsigned int count = data->res.count;
429 
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);
434 
435 		if (status >= 0) {
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);
441 				count = 0;
442 			} else
443 				count -= PAGE_CACHE_SIZE;
444 			SetPageUptodate(page);
445 		} else
446 			SetPageError(page);
447 		nfs_readpage_release(req);
448 	}
449 }
450 
451 /*
452  * This is the callback from RPC telling us whether a reply was
453  * received or some error occurred (timeout or socket shutdown).
454  */
455 void nfs_readpage_result(struct rpc_task *task)
456 {
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;
461 
462 	dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
463 		task->tk_pid, status);
464 
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);
474 			return;
475 		}
476 		task->tk_status = -EIO;
477 	}
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);
482 }
483 
484 /*
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.
489  */
490 int nfs_readpage(struct file *file, struct page *page)
491 {
492 	struct nfs_open_context *ctx;
493 	struct inode *inode = page->mapping->host;
494 	int		error;
495 
496 	dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
497 		page, PAGE_CACHE_SIZE, page->index);
498 	/*
499 	 * Try to flush any pending writes to the file..
500 	 *
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).
504 	 */
505 	error = nfs_wb_page(inode, page);
506 	if (error)
507 		goto out_error;
508 
509 	if (file == NULL) {
510 		ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
511 		if (ctx == NULL)
512 			return -EBADF;
513 	} else
514 		ctx = get_nfs_open_context((struct nfs_open_context *)
515 				file->private_data);
516 	if (!IS_SYNC(inode)) {
517 		error = nfs_readpage_async(ctx, inode, page);
518 		goto out;
519 	}
520 
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");
524 out:
525 	put_nfs_open_context(ctx);
526 	return error;
527 
528 out_error:
529 	unlock_page(page);
530 	return error;
531 }
532 
533 struct nfs_readdesc {
534 	struct list_head *head;
535 	struct nfs_open_context *ctx;
536 };
537 
538 static int
539 readpage_async_filler(void *data, struct page *page)
540 {
541 	struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
542 	struct inode *inode = page->mapping->host;
543 	struct nfs_page *new;
544 	unsigned int len;
545 
546 	nfs_wb_page(inode, page);
547 	len = nfs_page_length(inode, page);
548 	if (len == 0)
549 		return nfs_return_empty_page(page);
550 	new = nfs_create_request(desc->ctx, inode, page, 0, len);
551 	if (IS_ERR(new)) {
552 			SetPageError(page);
553 			unlock_page(page);
554 			return PTR_ERR(new);
555 	}
556 	if (len < PAGE_CACHE_SIZE)
557 		memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
558 	nfs_list_add_request(new, desc->head);
559 	return 0;
560 }
561 
562 int nfs_readpages(struct file *filp, struct address_space *mapping,
563 		struct list_head *pages, unsigned nr_pages)
564 {
565 	LIST_HEAD(head);
566 	struct nfs_readdesc desc = {
567 		.head		= &head,
568 	};
569 	struct inode *inode = mapping->host;
570 	struct nfs_server *server = NFS_SERVER(inode);
571 	int ret;
572 
573 	dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
574 			inode->i_sb->s_id,
575 			(long long)NFS_FILEID(inode),
576 			nr_pages);
577 
578 	if (filp == NULL) {
579 		desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
580 		if (desc.ctx == NULL)
581 			return -EBADF;
582 	} else
583 		desc.ctx = get_nfs_open_context((struct nfs_open_context *)
584 				filp->private_data);
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);
588 		if (!ret)
589 			ret = err;
590 	}
591 	put_nfs_open_context(desc.ctx);
592 	return ret;
593 }
594 
595 int nfs_init_readpagecache(void)
596 {
597 	nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
598 					     sizeof(struct nfs_read_data),
599 					     0, SLAB_HWCACHE_ALIGN,
600 					     NULL, NULL);
601 	if (nfs_rdata_cachep == NULL)
602 		return -ENOMEM;
603 
604 	nfs_rdata_mempool = mempool_create(MIN_POOL_READ,
605 					   mempool_alloc_slab,
606 					   mempool_free_slab,
607 					   nfs_rdata_cachep);
608 	if (nfs_rdata_mempool == NULL)
609 		return -ENOMEM;
610 
611 	return 0;
612 }
613 
614 void nfs_destroy_readpagecache(void)
615 {
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");
619 }
620