xref: /openbmc/linux/fs/nfs/read.c (revision 0545810f)
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/task_io_accounting_ops.h>
19 #include <linux/pagemap.h>
20 #include <linux/sunrpc/clnt.h>
21 #include <linux/nfs_fs.h>
22 #include <linux/nfs_page.h>
23 #include <linux/module.h>
24 
25 #include "nfs4_fs.h"
26 #include "internal.h"
27 #include "iostat.h"
28 #include "fscache.h"
29 #include "pnfs.h"
30 #include "nfstrace.h"
31 
32 #define NFSDBG_FACILITY		NFSDBG_PAGECACHE
33 
34 const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
35 static const struct nfs_rw_ops nfs_rw_read_ops;
36 
37 static struct kmem_cache *nfs_rdata_cachep;
38 
39 static struct nfs_pgio_header *nfs_readhdr_alloc(void)
40 {
41 	struct nfs_pgio_header *p = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
42 
43 	if (p)
44 		p->rw_mode = FMODE_READ;
45 	return p;
46 }
47 
48 static void nfs_readhdr_free(struct nfs_pgio_header *rhdr)
49 {
50 	kmem_cache_free(nfs_rdata_cachep, rhdr);
51 }
52 
53 static int nfs_return_empty_folio(struct folio *folio)
54 {
55 	folio_zero_segment(folio, 0, folio_size(folio));
56 	folio_mark_uptodate(folio);
57 	folio_unlock(folio);
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_complete_read(struct nfs_pageio_descriptor *pgio)
78 {
79 	struct nfs_pgio_mirror *pgm;
80 	unsigned long npages;
81 
82 	nfs_pageio_complete(pgio);
83 
84 	/* It doesn't make sense to do mirrored reads! */
85 	WARN_ON_ONCE(pgio->pg_mirror_count != 1);
86 
87 	pgm = &pgio->pg_mirrors[0];
88 	NFS_I(pgio->pg_inode)->read_io += pgm->pg_bytes_written;
89 	npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >> PAGE_SHIFT;
90 	nfs_add_stats(pgio->pg_inode, NFSIOS_READPAGES, npages);
91 }
92 
93 
94 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
95 {
96 	struct nfs_pgio_mirror *mirror;
97 
98 	if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
99 		pgio->pg_ops->pg_cleanup(pgio);
100 
101 	pgio->pg_ops = &nfs_pgio_rw_ops;
102 
103 	/* read path should never have more than one mirror */
104 	WARN_ON_ONCE(pgio->pg_mirror_count != 1);
105 
106 	mirror = &pgio->pg_mirrors[0];
107 	mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
108 }
109 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
110 
111 static void nfs_readpage_release(struct nfs_page *req, int error)
112 {
113 	struct folio *folio = nfs_page_to_folio(req);
114 
115 	if (nfs_error_is_fatal_on_server(error) && error != -ETIMEDOUT)
116 		folio_set_error(folio);
117 	if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE))
118 		if (nfs_netfs_folio_unlock(folio))
119 			folio_unlock(folio);
120 
121 	nfs_release_request(req);
122 }
123 
124 static void nfs_page_group_set_uptodate(struct nfs_page *req)
125 {
126 	if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
127 		folio_mark_uptodate(nfs_page_to_folio(req));
128 }
129 
130 static void nfs_read_completion(struct nfs_pgio_header *hdr)
131 {
132 	unsigned long bytes = 0;
133 	int error;
134 
135 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
136 		goto out;
137 	while (!list_empty(&hdr->pages)) {
138 		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
139 		struct folio *folio = nfs_page_to_folio(req);
140 		unsigned long start = req->wb_pgbase;
141 		unsigned long end = req->wb_pgbase + req->wb_bytes;
142 
143 		if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
144 			/* note: regions of the page not covered by a
145 			 * request are zeroed in nfs_read_add_folio
146 			 */
147 			if (bytes > hdr->good_bytes) {
148 				/* nothing in this request was good, so zero
149 				 * the full extent of the request */
150 				folio_zero_segment(folio, start, end);
151 
152 			} else if (hdr->good_bytes - bytes < req->wb_bytes) {
153 				/* part of this request has good bytes, but
154 				 * not all. zero the bad bytes */
155 				start += hdr->good_bytes - bytes;
156 				WARN_ON(start < req->wb_pgbase);
157 				folio_zero_segment(folio, start, end);
158 			}
159 		}
160 		error = 0;
161 		bytes += req->wb_bytes;
162 		if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
163 			if (bytes <= hdr->good_bytes)
164 				nfs_page_group_set_uptodate(req);
165 			else {
166 				error = hdr->error;
167 				xchg(&nfs_req_openctx(req)->error, error);
168 			}
169 		} else
170 			nfs_page_group_set_uptodate(req);
171 		nfs_list_remove_request(req);
172 		nfs_readpage_release(req, error);
173 	}
174 	nfs_netfs_read_completion(hdr);
175 
176 out:
177 	hdr->release(hdr);
178 }
179 
180 static void nfs_initiate_read(struct nfs_pgio_header *hdr,
181 			      struct rpc_message *msg,
182 			      const struct nfs_rpc_ops *rpc_ops,
183 			      struct rpc_task_setup *task_setup_data, int how)
184 {
185 	rpc_ops->read_setup(hdr, msg);
186 	nfs_netfs_initiate_read(hdr);
187 	trace_nfs_initiate_read(hdr);
188 }
189 
190 static void
191 nfs_async_read_error(struct list_head *head, int error)
192 {
193 	struct nfs_page	*req;
194 
195 	while (!list_empty(head)) {
196 		req = nfs_list_entry(head->next);
197 		nfs_list_remove_request(req);
198 		nfs_readpage_release(req, error);
199 	}
200 }
201 
202 const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
203 	.error_cleanup = nfs_async_read_error,
204 	.completion = nfs_read_completion,
205 };
206 
207 /*
208  * This is the callback from RPC telling us whether a reply was
209  * received or some error occurred (timeout or socket shutdown).
210  */
211 static int nfs_readpage_done(struct rpc_task *task,
212 			     struct nfs_pgio_header *hdr,
213 			     struct inode *inode)
214 {
215 	int status = NFS_PROTO(inode)->read_done(task, hdr);
216 	if (status != 0)
217 		return status;
218 
219 	nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
220 	trace_nfs_readpage_done(task, hdr);
221 
222 	if (task->tk_status == -ESTALE) {
223 		nfs_set_inode_stale(inode);
224 		nfs_mark_for_revalidate(inode);
225 	}
226 	return 0;
227 }
228 
229 static void nfs_readpage_retry(struct rpc_task *task,
230 			       struct nfs_pgio_header *hdr)
231 {
232 	struct nfs_pgio_args *argp = &hdr->args;
233 	struct nfs_pgio_res  *resp = &hdr->res;
234 
235 	/* This is a short read! */
236 	nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
237 	trace_nfs_readpage_short(task, hdr);
238 
239 	/* Has the server at least made some progress? */
240 	if (resp->count == 0) {
241 		nfs_set_pgio_error(hdr, -EIO, argp->offset);
242 		return;
243 	}
244 
245 	/* For non rpc-based layout drivers, retry-through-MDS */
246 	if (!task->tk_ops) {
247 		hdr->pnfs_error = -EAGAIN;
248 		return;
249 	}
250 
251 	/* Yes, so retry the read at the end of the hdr */
252 	hdr->mds_offset += resp->count;
253 	argp->offset += resp->count;
254 	argp->pgbase += resp->count;
255 	argp->count -= resp->count;
256 	resp->count = 0;
257 	resp->eof = 0;
258 	rpc_restart_call_prepare(task);
259 }
260 
261 static void nfs_readpage_result(struct rpc_task *task,
262 				struct nfs_pgio_header *hdr)
263 {
264 	if (hdr->res.eof) {
265 		loff_t pos = hdr->args.offset + hdr->res.count;
266 		unsigned int new = pos - hdr->io_start;
267 
268 		if (hdr->good_bytes > new) {
269 			hdr->good_bytes = new;
270 			set_bit(NFS_IOHDR_EOF, &hdr->flags);
271 			clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
272 		}
273 	} else if (hdr->res.count < hdr->args.count)
274 		nfs_readpage_retry(task, hdr);
275 }
276 
277 int nfs_read_add_folio(struct nfs_pageio_descriptor *pgio,
278 		       struct nfs_open_context *ctx,
279 		       struct folio *folio)
280 {
281 	struct inode *inode = folio_file_mapping(folio)->host;
282 	struct nfs_server *server = NFS_SERVER(inode);
283 	size_t fsize = folio_size(folio);
284 	unsigned int rsize = server->rsize;
285 	struct nfs_page *new;
286 	unsigned int len, aligned_len;
287 	int error;
288 
289 	len = nfs_folio_length(folio);
290 	if (len == 0)
291 		return nfs_return_empty_folio(folio);
292 
293 	aligned_len = min_t(unsigned int, ALIGN(len, rsize), fsize);
294 
295 	new = nfs_page_create_from_folio(ctx, folio, 0, aligned_len);
296 	if (IS_ERR(new)) {
297 		error = PTR_ERR(new);
298 		goto out;
299 	}
300 
301 	if (len < fsize)
302 		folio_zero_segment(folio, len, fsize);
303 	if (!nfs_pageio_add_request(pgio, new)) {
304 		nfs_list_remove_request(new);
305 		error = pgio->pg_error;
306 		nfs_readpage_release(new, error);
307 		goto out;
308 	}
309 	return 0;
310 out:
311 	return error;
312 }
313 
314 /*
315  * Read a page over NFS.
316  * We read the page synchronously in the following case:
317  *  -	The error flag is set for this page. This happens only when a
318  *	previous async read operation failed.
319  */
320 int nfs_read_folio(struct file *file, struct folio *folio)
321 {
322 	struct inode *inode = file_inode(file);
323 	struct nfs_pageio_descriptor pgio;
324 	struct nfs_open_context *ctx;
325 	int ret;
326 
327 	trace_nfs_aop_readpage(inode, folio);
328 	nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
329 	task_io_account_read(folio_size(folio));
330 
331 	/*
332 	 * Try to flush any pending writes to the file..
333 	 *
334 	 * NOTE! Because we own the folio lock, there cannot
335 	 * be any new pending writes generated at this point
336 	 * for this folio (other folios can be written to).
337 	 */
338 	ret = nfs_wb_folio(inode, folio);
339 	if (ret)
340 		goto out_unlock;
341 	if (folio_test_uptodate(folio))
342 		goto out_unlock;
343 
344 	ret = -ESTALE;
345 	if (NFS_STALE(inode))
346 		goto out_unlock;
347 
348 	ret = nfs_netfs_read_folio(file, folio);
349 	if (!ret)
350 		goto out;
351 
352 	ctx = get_nfs_open_context(nfs_file_open_context(file));
353 
354 	xchg(&ctx->error, 0);
355 	nfs_pageio_init_read(&pgio, inode, false,
356 			     &nfs_async_read_completion_ops);
357 
358 	ret = nfs_read_add_folio(&pgio, ctx, folio);
359 	if (ret)
360 		goto out_put;
361 
362 	nfs_pageio_complete_read(&pgio);
363 	ret = pgio.pg_error < 0 ? pgio.pg_error : 0;
364 	if (!ret) {
365 		ret = folio_wait_locked_killable(folio);
366 		if (!folio_test_uptodate(folio) && !ret)
367 			ret = xchg(&ctx->error, 0);
368 	}
369 out_put:
370 	put_nfs_open_context(ctx);
371 out:
372 	trace_nfs_aop_readpage_done(inode, folio, ret);
373 	return ret;
374 out_unlock:
375 	folio_unlock(folio);
376 	goto out;
377 }
378 
379 void nfs_readahead(struct readahead_control *ractl)
380 {
381 	struct nfs_pageio_descriptor pgio;
382 	struct nfs_open_context *ctx;
383 	unsigned int nr_pages = readahead_count(ractl);
384 	struct file *file = ractl->file;
385 	struct inode *inode = ractl->mapping->host;
386 	struct folio *folio;
387 	int ret;
388 
389 	trace_nfs_aop_readahead(inode, readahead_pos(ractl), nr_pages);
390 	nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
391 	task_io_account_read(readahead_length(ractl));
392 
393 	ret = -ESTALE;
394 	if (NFS_STALE(inode))
395 		goto out;
396 
397 	ret = nfs_netfs_readahead(ractl);
398 	if (!ret)
399 		goto out;
400 
401 	if (file == NULL) {
402 		ret = -EBADF;
403 		ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
404 		if (ctx == NULL)
405 			goto out;
406 	} else
407 		ctx = get_nfs_open_context(nfs_file_open_context(file));
408 
409 	nfs_pageio_init_read(&pgio, inode, false,
410 			     &nfs_async_read_completion_ops);
411 
412 	while ((folio = readahead_folio(ractl)) != NULL) {
413 		ret = nfs_read_add_folio(&pgio, ctx, folio);
414 		if (ret)
415 			break;
416 	}
417 
418 	nfs_pageio_complete_read(&pgio);
419 
420 	put_nfs_open_context(ctx);
421 out:
422 	trace_nfs_aop_readahead_done(inode, nr_pages, ret);
423 }
424 
425 int __init nfs_init_readpagecache(void)
426 {
427 	nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
428 					     sizeof(struct nfs_pgio_header),
429 					     0, SLAB_HWCACHE_ALIGN,
430 					     NULL);
431 	if (nfs_rdata_cachep == NULL)
432 		return -ENOMEM;
433 
434 	return 0;
435 }
436 
437 void nfs_destroy_readpagecache(void)
438 {
439 	kmem_cache_destroy(nfs_rdata_cachep);
440 }
441 
442 static const struct nfs_rw_ops nfs_rw_read_ops = {
443 	.rw_alloc_header	= nfs_readhdr_alloc,
444 	.rw_free_header		= nfs_readhdr_free,
445 	.rw_done		= nfs_readpage_done,
446 	.rw_result		= nfs_readpage_result,
447 	.rw_initiate		= nfs_initiate_read,
448 };
449