xref: /openbmc/linux/fs/nfs/read.c (revision f9834f18)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * linux/fs/nfs/read.c
4  *
5  * Block I/O for NFS
6  *
7  * Partial copy of Linus' read cache modifications to fs/nfs/file.c
8  * modified for async RPC by okir@monad.swb.de
9  */
10 
11 #include <linux/time.h>
12 #include <linux/kernel.h>
13 #include <linux/errno.h>
14 #include <linux/fcntl.h>
15 #include <linux/stat.h>
16 #include <linux/mm.h>
17 #include <linux/slab.h>
18 #include <linux/pagemap.h>
19 #include <linux/sunrpc/clnt.h>
20 #include <linux/nfs_fs.h>
21 #include <linux/nfs_page.h>
22 #include <linux/module.h>
23 
24 #include "nfs4_fs.h"
25 #include "internal.h"
26 #include "iostat.h"
27 #include "fscache.h"
28 #include "pnfs.h"
29 #include "nfstrace.h"
30 
31 #define NFSDBG_FACILITY		NFSDBG_PAGECACHE
32 
33 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
34 static const struct nfs_rw_ops nfs_rw_read_ops;
35 
36 static struct kmem_cache *nfs_rdata_cachep;
37 
38 static struct nfs_pgio_header *nfs_readhdr_alloc(void)
39 {
40 	struct nfs_pgio_header *p = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
41 
42 	if (p)
43 		p->rw_mode = FMODE_READ;
44 	return p;
45 }
46 
47 static void nfs_readhdr_free(struct nfs_pgio_header *rhdr)
48 {
49 	kmem_cache_free(nfs_rdata_cachep, rhdr);
50 }
51 
52 static
53 int nfs_return_empty_page(struct page *page)
54 {
55 	zero_user(page, 0, PAGE_SIZE);
56 	SetPageUptodate(page);
57 	unlock_page(page);
58 	return 0;
59 }
60 
61 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
62 			      struct inode *inode, bool force_mds,
63 			      const struct nfs_pgio_completion_ops *compl_ops)
64 {
65 	struct nfs_server *server = NFS_SERVER(inode);
66 	const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
67 
68 #ifdef CONFIG_NFS_V4_1
69 	if (server->pnfs_curr_ld && !force_mds)
70 		pg_ops = server->pnfs_curr_ld->pg_read_ops;
71 #endif
72 	nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops,
73 			server->rsize, 0);
74 }
75 EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
76 
77 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
78 {
79 	struct nfs_pgio_mirror *mirror;
80 
81 	if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
82 		pgio->pg_ops->pg_cleanup(pgio);
83 
84 	pgio->pg_ops = &nfs_pgio_rw_ops;
85 
86 	/* read path should never have more than one mirror */
87 	WARN_ON_ONCE(pgio->pg_mirror_count != 1);
88 
89 	mirror = &pgio->pg_mirrors[0];
90 	mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
91 }
92 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
93 
94 static void nfs_readpage_release(struct nfs_page *req, int error)
95 {
96 	struct inode *inode = d_inode(nfs_req_openctx(req)->dentry);
97 	struct page *page = req->wb_page;
98 
99 	dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
100 		(unsigned long long)NFS_FILEID(inode), req->wb_bytes,
101 		(long long)req_offset(req));
102 
103 	if (nfs_error_is_fatal_on_server(error) && error != -ETIMEDOUT)
104 		SetPageError(page);
105 	if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
106 		struct address_space *mapping = page_file_mapping(page);
107 
108 		if (PageUptodate(page))
109 			nfs_readpage_to_fscache(inode, page, 0);
110 		else if (!PageError(page) && !PagePrivate(page))
111 			generic_error_remove_page(mapping, page);
112 		unlock_page(page);
113 	}
114 	nfs_release_request(req);
115 }
116 
117 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
118 		       struct page *page)
119 {
120 	struct nfs_page	*new;
121 	unsigned int len;
122 	struct nfs_pageio_descriptor pgio;
123 	struct nfs_pgio_mirror *pgm;
124 
125 	len = nfs_page_length(page);
126 	if (len == 0)
127 		return nfs_return_empty_page(page);
128 	new = nfs_create_request(ctx, page, 0, len);
129 	if (IS_ERR(new)) {
130 		unlock_page(page);
131 		return PTR_ERR(new);
132 	}
133 	if (len < PAGE_SIZE)
134 		zero_user_segment(page, len, PAGE_SIZE);
135 
136 	nfs_pageio_init_read(&pgio, inode, false,
137 			     &nfs_async_read_completion_ops);
138 	if (!nfs_pageio_add_request(&pgio, new)) {
139 		nfs_list_remove_request(new);
140 		nfs_readpage_release(new, pgio.pg_error);
141 	}
142 	nfs_pageio_complete(&pgio);
143 
144 	/* It doesn't make sense to do mirrored reads! */
145 	WARN_ON_ONCE(pgio.pg_mirror_count != 1);
146 
147 	pgm = &pgio.pg_mirrors[0];
148 	NFS_I(inode)->read_io += pgm->pg_bytes_written;
149 
150 	return pgio.pg_error < 0 ? pgio.pg_error : 0;
151 }
152 
153 static void nfs_page_group_set_uptodate(struct nfs_page *req)
154 {
155 	if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
156 		SetPageUptodate(req->wb_page);
157 }
158 
159 static void nfs_read_completion(struct nfs_pgio_header *hdr)
160 {
161 	unsigned long bytes = 0;
162 	int error;
163 
164 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
165 		goto out;
166 	while (!list_empty(&hdr->pages)) {
167 		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
168 		struct page *page = req->wb_page;
169 		unsigned long start = req->wb_pgbase;
170 		unsigned long end = req->wb_pgbase + req->wb_bytes;
171 
172 		if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
173 			/* note: regions of the page not covered by a
174 			 * request are zeroed in nfs_readpage_async /
175 			 * readpage_async_filler */
176 			if (bytes > hdr->good_bytes) {
177 				/* nothing in this request was good, so zero
178 				 * the full extent of the request */
179 				zero_user_segment(page, start, end);
180 
181 			} else if (hdr->good_bytes - bytes < req->wb_bytes) {
182 				/* part of this request has good bytes, but
183 				 * not all. zero the bad bytes */
184 				start += hdr->good_bytes - bytes;
185 				WARN_ON(start < req->wb_pgbase);
186 				zero_user_segment(page, start, end);
187 			}
188 		}
189 		error = 0;
190 		bytes += req->wb_bytes;
191 		if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
192 			if (bytes <= hdr->good_bytes)
193 				nfs_page_group_set_uptodate(req);
194 			else {
195 				error = hdr->error;
196 				xchg(&nfs_req_openctx(req)->error, error);
197 			}
198 		} else
199 			nfs_page_group_set_uptodate(req);
200 		nfs_list_remove_request(req);
201 		nfs_readpage_release(req, error);
202 	}
203 out:
204 	hdr->release(hdr);
205 }
206 
207 static void nfs_initiate_read(struct nfs_pgio_header *hdr,
208 			      struct rpc_message *msg,
209 			      const struct nfs_rpc_ops *rpc_ops,
210 			      struct rpc_task_setup *task_setup_data, int how)
211 {
212 	struct inode *inode = hdr->inode;
213 	int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
214 
215 	task_setup_data->flags |= swap_flags;
216 	rpc_ops->read_setup(hdr, msg);
217 	trace_nfs_initiate_read(hdr);
218 }
219 
220 static void
221 nfs_async_read_error(struct list_head *head, int error)
222 {
223 	struct nfs_page	*req;
224 
225 	while (!list_empty(head)) {
226 		req = nfs_list_entry(head->next);
227 		nfs_list_remove_request(req);
228 		nfs_readpage_release(req, error);
229 	}
230 }
231 
232 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
233 	.error_cleanup = nfs_async_read_error,
234 	.completion = nfs_read_completion,
235 };
236 
237 /*
238  * This is the callback from RPC telling us whether a reply was
239  * received or some error occurred (timeout or socket shutdown).
240  */
241 static int nfs_readpage_done(struct rpc_task *task,
242 			     struct nfs_pgio_header *hdr,
243 			     struct inode *inode)
244 {
245 	int status = NFS_PROTO(inode)->read_done(task, hdr);
246 	if (status != 0)
247 		return status;
248 
249 	nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
250 	trace_nfs_readpage_done(task, hdr);
251 
252 	if (task->tk_status == -ESTALE) {
253 		set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
254 		nfs_mark_for_revalidate(inode);
255 	}
256 	return 0;
257 }
258 
259 static void nfs_readpage_retry(struct rpc_task *task,
260 			       struct nfs_pgio_header *hdr)
261 {
262 	struct nfs_pgio_args *argp = &hdr->args;
263 	struct nfs_pgio_res  *resp = &hdr->res;
264 
265 	/* This is a short read! */
266 	nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
267 	/* Has the server at least made some progress? */
268 	if (resp->count == 0) {
269 		nfs_set_pgio_error(hdr, -EIO, argp->offset);
270 		return;
271 	}
272 
273 	/* For non rpc-based layout drivers, retry-through-MDS */
274 	if (!task->tk_ops) {
275 		hdr->pnfs_error = -EAGAIN;
276 		return;
277 	}
278 
279 	/* Yes, so retry the read at the end of the hdr */
280 	hdr->mds_offset += resp->count;
281 	argp->offset += resp->count;
282 	argp->pgbase += resp->count;
283 	argp->count -= resp->count;
284 	resp->count = 0;
285 	resp->eof = 0;
286 	rpc_restart_call_prepare(task);
287 }
288 
289 static void nfs_readpage_result(struct rpc_task *task,
290 				struct nfs_pgio_header *hdr)
291 {
292 	if (hdr->res.eof) {
293 		loff_t pos = hdr->args.offset + hdr->res.count;
294 		unsigned int new = pos - hdr->io_start;
295 
296 		if (hdr->good_bytes > new) {
297 			hdr->good_bytes = new;
298 			set_bit(NFS_IOHDR_EOF, &hdr->flags);
299 			clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
300 		}
301 	} else if (hdr->res.count < hdr->args.count)
302 		nfs_readpage_retry(task, hdr);
303 }
304 
305 /*
306  * Read a page over NFS.
307  * We read the page synchronously in the following case:
308  *  -	The error flag is set for this page. This happens only when a
309  *	previous async read operation failed.
310  */
311 int nfs_readpage(struct file *file, struct page *page)
312 {
313 	struct nfs_open_context *ctx;
314 	struct inode *inode = page_file_mapping(page)->host;
315 	int		error;
316 
317 	dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
318 		page, PAGE_SIZE, page_index(page));
319 	nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
320 	nfs_add_stats(inode, NFSIOS_READPAGES, 1);
321 
322 	/*
323 	 * Try to flush any pending writes to the file..
324 	 *
325 	 * NOTE! Because we own the page lock, there cannot
326 	 * be any new pending writes generated at this point
327 	 * for this page (other pages can be written to).
328 	 */
329 	error = nfs_wb_page(inode, page);
330 	if (error)
331 		goto out_unlock;
332 	if (PageUptodate(page))
333 		goto out_unlock;
334 
335 	error = -ESTALE;
336 	if (NFS_STALE(inode))
337 		goto out_unlock;
338 
339 	if (file == NULL) {
340 		error = -EBADF;
341 		ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
342 		if (ctx == NULL)
343 			goto out_unlock;
344 	} else
345 		ctx = get_nfs_open_context(nfs_file_open_context(file));
346 
347 	if (!IS_SYNC(inode)) {
348 		error = nfs_readpage_from_fscache(ctx, inode, page);
349 		if (error == 0)
350 			goto out;
351 	}
352 
353 	xchg(&ctx->error, 0);
354 	error = nfs_readpage_async(ctx, inode, page);
355 	if (!error) {
356 		error = wait_on_page_locked_killable(page);
357 		if (!PageUptodate(page) && !error)
358 			error = xchg(&ctx->error, 0);
359 	}
360 out:
361 	put_nfs_open_context(ctx);
362 	return error;
363 out_unlock:
364 	unlock_page(page);
365 	return error;
366 }
367 
368 struct nfs_readdesc {
369 	struct nfs_pageio_descriptor *pgio;
370 	struct nfs_open_context *ctx;
371 };
372 
373 static int
374 readpage_async_filler(void *data, struct page *page)
375 {
376 	struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
377 	struct nfs_page *new;
378 	unsigned int len;
379 	int error;
380 
381 	len = nfs_page_length(page);
382 	if (len == 0)
383 		return nfs_return_empty_page(page);
384 
385 	new = nfs_create_request(desc->ctx, page, 0, len);
386 	if (IS_ERR(new))
387 		goto out_error;
388 
389 	if (len < PAGE_SIZE)
390 		zero_user_segment(page, len, PAGE_SIZE);
391 	if (!nfs_pageio_add_request(desc->pgio, new)) {
392 		nfs_list_remove_request(new);
393 		error = desc->pgio->pg_error;
394 		nfs_readpage_release(new, error);
395 		goto out;
396 	}
397 	return 0;
398 out_error:
399 	error = PTR_ERR(new);
400 	unlock_page(page);
401 out:
402 	return error;
403 }
404 
405 int nfs_readpages(struct file *filp, struct address_space *mapping,
406 		struct list_head *pages, unsigned nr_pages)
407 {
408 	struct nfs_pageio_descriptor pgio;
409 	struct nfs_pgio_mirror *pgm;
410 	struct nfs_readdesc desc = {
411 		.pgio = &pgio,
412 	};
413 	struct inode *inode = mapping->host;
414 	unsigned long npages;
415 	int ret = -ESTALE;
416 
417 	dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
418 			inode->i_sb->s_id,
419 			(unsigned long long)NFS_FILEID(inode),
420 			nr_pages);
421 	nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
422 
423 	if (NFS_STALE(inode))
424 		goto out;
425 
426 	if (filp == NULL) {
427 		desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
428 		if (desc.ctx == NULL)
429 			return -EBADF;
430 	} else
431 		desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
432 
433 	/* attempt to read as many of the pages as possible from the cache
434 	 * - this returns -ENOBUFS immediately if the cookie is negative
435 	 */
436 	ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
437 					 pages, &nr_pages);
438 	if (ret == 0)
439 		goto read_complete; /* all pages were read */
440 
441 	nfs_pageio_init_read(&pgio, inode, false,
442 			     &nfs_async_read_completion_ops);
443 
444 	ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
445 	nfs_pageio_complete(&pgio);
446 
447 	/* It doesn't make sense to do mirrored reads! */
448 	WARN_ON_ONCE(pgio.pg_mirror_count != 1);
449 
450 	pgm = &pgio.pg_mirrors[0];
451 	NFS_I(inode)->read_io += pgm->pg_bytes_written;
452 	npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >>
453 		 PAGE_SHIFT;
454 	nfs_add_stats(inode, NFSIOS_READPAGES, npages);
455 read_complete:
456 	put_nfs_open_context(desc.ctx);
457 out:
458 	return ret;
459 }
460 
461 int __init nfs_init_readpagecache(void)
462 {
463 	nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
464 					     sizeof(struct nfs_pgio_header),
465 					     0, SLAB_HWCACHE_ALIGN,
466 					     NULL);
467 	if (nfs_rdata_cachep == NULL)
468 		return -ENOMEM;
469 
470 	return 0;
471 }
472 
473 void nfs_destroy_readpagecache(void)
474 {
475 	kmem_cache_destroy(nfs_rdata_cachep);
476 }
477 
478 static const struct nfs_rw_ops nfs_rw_read_ops = {
479 	.rw_alloc_header	= nfs_readhdr_alloc,
480 	.rw_free_header		= nfs_readhdr_free,
481 	.rw_done		= nfs_readpage_done,
482 	.rw_result		= nfs_readpage_result,
483 	.rw_initiate		= nfs_initiate_read,
484 };
485