xref: /openbmc/linux/fs/nfs/read.c (revision edfa0b16)
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 static 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 inode *inode = d_inode(nfs_req_openctx(req)->dentry);
114 	struct page *page = req->wb_page;
115 
116 	dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
117 		(unsigned long long)NFS_FILEID(inode), req->wb_bytes,
118 		(long long)req_offset(req));
119 
120 	if (nfs_error_is_fatal_on_server(error) && error != -ETIMEDOUT)
121 		SetPageError(page);
122 	if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
123 		struct address_space *mapping = page_file_mapping(page);
124 
125 		if (PageUptodate(page))
126 			nfs_readpage_to_fscache(inode, page, 0);
127 		else if (!PageError(page) && !PagePrivate(page))
128 			generic_error_remove_page(mapping, page);
129 		unlock_page(page);
130 	}
131 	nfs_release_request(req);
132 }
133 
134 struct nfs_readdesc {
135 	struct nfs_pageio_descriptor pgio;
136 	struct nfs_open_context *ctx;
137 };
138 
139 static void nfs_page_group_set_uptodate(struct nfs_page *req)
140 {
141 	if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
142 		SetPageUptodate(req->wb_page);
143 }
144 
145 static void nfs_read_completion(struct nfs_pgio_header *hdr)
146 {
147 	unsigned long bytes = 0;
148 	int error;
149 
150 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
151 		goto out;
152 	while (!list_empty(&hdr->pages)) {
153 		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
154 		struct page *page = req->wb_page;
155 		unsigned long start = req->wb_pgbase;
156 		unsigned long end = req->wb_pgbase + req->wb_bytes;
157 
158 		if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
159 			/* note: regions of the page not covered by a
160 			 * request are zeroed in readpage_async_filler */
161 			if (bytes > hdr->good_bytes) {
162 				/* nothing in this request was good, so zero
163 				 * the full extent of the request */
164 				zero_user_segment(page, start, end);
165 
166 			} else if (hdr->good_bytes - bytes < req->wb_bytes) {
167 				/* part of this request has good bytes, but
168 				 * not all. zero the bad bytes */
169 				start += hdr->good_bytes - bytes;
170 				WARN_ON(start < req->wb_pgbase);
171 				zero_user_segment(page, start, end);
172 			}
173 		}
174 		error = 0;
175 		bytes += req->wb_bytes;
176 		if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
177 			if (bytes <= hdr->good_bytes)
178 				nfs_page_group_set_uptodate(req);
179 			else {
180 				error = hdr->error;
181 				xchg(&nfs_req_openctx(req)->error, error);
182 			}
183 		} else
184 			nfs_page_group_set_uptodate(req);
185 		nfs_list_remove_request(req);
186 		nfs_readpage_release(req, error);
187 	}
188 out:
189 	hdr->release(hdr);
190 }
191 
192 static void nfs_initiate_read(struct nfs_pgio_header *hdr,
193 			      struct rpc_message *msg,
194 			      const struct nfs_rpc_ops *rpc_ops,
195 			      struct rpc_task_setup *task_setup_data, int how)
196 {
197 	struct inode *inode = hdr->inode;
198 	int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
199 
200 	task_setup_data->flags |= swap_flags;
201 	rpc_ops->read_setup(hdr, msg);
202 	trace_nfs_initiate_read(hdr);
203 }
204 
205 static void
206 nfs_async_read_error(struct list_head *head, int error)
207 {
208 	struct nfs_page	*req;
209 
210 	while (!list_empty(head)) {
211 		req = nfs_list_entry(head->next);
212 		nfs_list_remove_request(req);
213 		nfs_readpage_release(req, error);
214 	}
215 }
216 
217 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
218 	.error_cleanup = nfs_async_read_error,
219 	.completion = nfs_read_completion,
220 };
221 
222 /*
223  * This is the callback from RPC telling us whether a reply was
224  * received or some error occurred (timeout or socket shutdown).
225  */
226 static int nfs_readpage_done(struct rpc_task *task,
227 			     struct nfs_pgio_header *hdr,
228 			     struct inode *inode)
229 {
230 	int status = NFS_PROTO(inode)->read_done(task, hdr);
231 	if (status != 0)
232 		return status;
233 
234 	nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
235 	trace_nfs_readpage_done(task, hdr);
236 
237 	if (task->tk_status == -ESTALE) {
238 		nfs_set_inode_stale(inode);
239 		nfs_mark_for_revalidate(inode);
240 	}
241 	return 0;
242 }
243 
244 static void nfs_readpage_retry(struct rpc_task *task,
245 			       struct nfs_pgio_header *hdr)
246 {
247 	struct nfs_pgio_args *argp = &hdr->args;
248 	struct nfs_pgio_res  *resp = &hdr->res;
249 
250 	/* This is a short read! */
251 	nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
252 	trace_nfs_readpage_short(task, hdr);
253 
254 	/* Has the server at least made some progress? */
255 	if (resp->count == 0) {
256 		nfs_set_pgio_error(hdr, -EIO, argp->offset);
257 		return;
258 	}
259 
260 	/* For non rpc-based layout drivers, retry-through-MDS */
261 	if (!task->tk_ops) {
262 		hdr->pnfs_error = -EAGAIN;
263 		return;
264 	}
265 
266 	/* Yes, so retry the read at the end of the hdr */
267 	hdr->mds_offset += resp->count;
268 	argp->offset += resp->count;
269 	argp->pgbase += resp->count;
270 	argp->count -= resp->count;
271 	resp->count = 0;
272 	resp->eof = 0;
273 	rpc_restart_call_prepare(task);
274 }
275 
276 static void nfs_readpage_result(struct rpc_task *task,
277 				struct nfs_pgio_header *hdr)
278 {
279 	if (hdr->res.eof) {
280 		loff_t pos = hdr->args.offset + hdr->res.count;
281 		unsigned int new = pos - hdr->io_start;
282 
283 		if (hdr->good_bytes > new) {
284 			hdr->good_bytes = new;
285 			set_bit(NFS_IOHDR_EOF, &hdr->flags);
286 			clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
287 		}
288 	} else if (hdr->res.count < hdr->args.count)
289 		nfs_readpage_retry(task, hdr);
290 }
291 
292 static int
293 readpage_async_filler(void *data, struct page *page)
294 {
295 	struct nfs_readdesc *desc = data;
296 	struct inode *inode = page_file_mapping(page)->host;
297 	unsigned int rsize = NFS_SERVER(inode)->rsize;
298 	struct nfs_page *new;
299 	unsigned int len, aligned_len;
300 	int error;
301 
302 	len = nfs_page_length(page);
303 	if (len == 0)
304 		return nfs_return_empty_page(page);
305 
306 	aligned_len = min_t(unsigned int, ALIGN(len, rsize), PAGE_SIZE);
307 
308 	new = nfs_create_request(desc->ctx, page, 0, aligned_len);
309 	if (IS_ERR(new))
310 		goto out_error;
311 
312 	if (len < PAGE_SIZE)
313 		zero_user_segment(page, len, PAGE_SIZE);
314 	if (!nfs_pageio_add_request(&desc->pgio, new)) {
315 		nfs_list_remove_request(new);
316 		error = desc->pgio.pg_error;
317 		nfs_readpage_release(new, error);
318 		goto out;
319 	}
320 	return 0;
321 out_error:
322 	error = PTR_ERR(new);
323 	unlock_page(page);
324 out:
325 	return error;
326 }
327 
328 /*
329  * Read a page over NFS.
330  * We read the page synchronously in the following case:
331  *  -	The error flag is set for this page. This happens only when a
332  *	previous async read operation failed.
333  */
334 int nfs_readpage(struct file *file, struct page *page)
335 {
336 	struct nfs_readdesc desc;
337 	struct inode *inode = page_file_mapping(page)->host;
338 	int ret;
339 
340 	trace_nfs_aop_readpage(inode, page);
341 	nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
342 
343 	/*
344 	 * Try to flush any pending writes to the file..
345 	 *
346 	 * NOTE! Because we own the page lock, there cannot
347 	 * be any new pending writes generated at this point
348 	 * for this page (other pages can be written to).
349 	 */
350 	ret = nfs_wb_page(inode, page);
351 	if (ret)
352 		goto out_unlock;
353 	if (PageUptodate(page))
354 		goto out_unlock;
355 
356 	ret = -ESTALE;
357 	if (NFS_STALE(inode))
358 		goto out_unlock;
359 
360 	if (file == NULL) {
361 		ret = -EBADF;
362 		desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
363 		if (desc.ctx == NULL)
364 			goto out_unlock;
365 	} else
366 		desc.ctx = get_nfs_open_context(nfs_file_open_context(file));
367 
368 	xchg(&desc.ctx->error, 0);
369 	if (!IS_SYNC(inode)) {
370 		ret = nfs_readpage_from_fscache(desc.ctx, inode, page);
371 		if (ret == 0)
372 			goto out_wait;
373 	}
374 
375 	nfs_pageio_init_read(&desc.pgio, inode, false,
376 			     &nfs_async_read_completion_ops);
377 
378 	ret = readpage_async_filler(&desc, page);
379 	if (ret)
380 		goto out;
381 
382 	nfs_pageio_complete_read(&desc.pgio);
383 	ret = desc.pgio.pg_error < 0 ? desc.pgio.pg_error : 0;
384 out_wait:
385 	if (!ret) {
386 		ret = wait_on_page_locked_killable(page);
387 		if (!PageUptodate(page) && !ret)
388 			ret = xchg(&desc.ctx->error, 0);
389 	}
390 out:
391 	put_nfs_open_context(desc.ctx);
392 	trace_nfs_aop_readpage_done(inode, page, ret);
393 	return ret;
394 out_unlock:
395 	unlock_page(page);
396 	trace_nfs_aop_readpage_done(inode, page, ret);
397 	return ret;
398 }
399 
400 int nfs_readpages(struct file *file, struct address_space *mapping,
401 		struct list_head *pages, unsigned nr_pages)
402 {
403 	struct nfs_readdesc desc;
404 	struct inode *inode = mapping->host;
405 	int ret;
406 
407 	trace_nfs_aop_readahead(inode, lru_to_page(pages), nr_pages);
408 	nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
409 
410 	ret = -ESTALE;
411 	if (NFS_STALE(inode))
412 		goto out;
413 
414 	if (file == NULL) {
415 		ret = -EBADF;
416 		desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
417 		if (desc.ctx == NULL)
418 			goto out;
419 	} else
420 		desc.ctx = get_nfs_open_context(nfs_file_open_context(file));
421 
422 	/* attempt to read as many of the pages as possible from the cache
423 	 * - this returns -ENOBUFS immediately if the cookie is negative
424 	 */
425 	ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
426 					 pages, &nr_pages);
427 	if (ret == 0)
428 		goto read_complete; /* all pages were read */
429 
430 	nfs_pageio_init_read(&desc.pgio, inode, false,
431 			     &nfs_async_read_completion_ops);
432 
433 	ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
434 
435 	nfs_pageio_complete_read(&desc.pgio);
436 
437 read_complete:
438 	put_nfs_open_context(desc.ctx);
439 out:
440 	trace_nfs_aop_readahead_done(inode, nr_pages, ret);
441 	return ret;
442 }
443 
444 int __init nfs_init_readpagecache(void)
445 {
446 	nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
447 					     sizeof(struct nfs_pgio_header),
448 					     0, SLAB_HWCACHE_ALIGN,
449 					     NULL);
450 	if (nfs_rdata_cachep == NULL)
451 		return -ENOMEM;
452 
453 	return 0;
454 }
455 
456 void nfs_destroy_readpagecache(void)
457 {
458 	kmem_cache_destroy(nfs_rdata_cachep);
459 }
460 
461 static const struct nfs_rw_ops nfs_rw_read_ops = {
462 	.rw_alloc_header	= nfs_readhdr_alloc,
463 	.rw_free_header		= nfs_readhdr_free,
464 	.rw_done		= nfs_readpage_done,
465 	.rw_result		= nfs_readpage_result,
466 	.rw_initiate		= nfs_initiate_read,
467 };
468