xref: /openbmc/linux/fs/nfs/read.c (revision 59bd9ded)
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)
95 {
96 	struct inode *inode = d_inode(nfs_req_openctx(req)->dentry);
97 
98 	dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
99 		(unsigned long long)NFS_FILEID(inode), req->wb_bytes,
100 		(long long)req_offset(req));
101 
102 	if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
103 		if (PageUptodate(req->wb_page))
104 			nfs_readpage_to_fscache(inode, req->wb_page, 0);
105 
106 		unlock_page(req->wb_page);
107 	}
108 	nfs_release_request(req);
109 }
110 
111 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
112 		       struct page *page)
113 {
114 	struct nfs_page	*new;
115 	unsigned int len;
116 	struct nfs_pageio_descriptor pgio;
117 	struct nfs_pgio_mirror *pgm;
118 
119 	len = nfs_page_length(page);
120 	if (len == 0)
121 		return nfs_return_empty_page(page);
122 	new = nfs_create_request(ctx, page, 0, len);
123 	if (IS_ERR(new)) {
124 		unlock_page(page);
125 		return PTR_ERR(new);
126 	}
127 	if (len < PAGE_SIZE)
128 		zero_user_segment(page, len, PAGE_SIZE);
129 
130 	nfs_pageio_init_read(&pgio, inode, false,
131 			     &nfs_async_read_completion_ops);
132 	if (!nfs_pageio_add_request(&pgio, new)) {
133 		nfs_list_remove_request(new);
134 		nfs_readpage_release(new);
135 	}
136 	nfs_pageio_complete(&pgio);
137 
138 	/* It doesn't make sense to do mirrored reads! */
139 	WARN_ON_ONCE(pgio.pg_mirror_count != 1);
140 
141 	pgm = &pgio.pg_mirrors[0];
142 	NFS_I(inode)->read_io += pgm->pg_bytes_written;
143 
144 	return pgio.pg_error < 0 ? pgio.pg_error : 0;
145 }
146 
147 static void nfs_page_group_set_uptodate(struct nfs_page *req)
148 {
149 	if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
150 		SetPageUptodate(req->wb_page);
151 }
152 
153 static void nfs_read_completion(struct nfs_pgio_header *hdr)
154 {
155 	unsigned long bytes = 0;
156 
157 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
158 		goto out;
159 	while (!list_empty(&hdr->pages)) {
160 		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
161 		struct page *page = req->wb_page;
162 		unsigned long start = req->wb_pgbase;
163 		unsigned long end = req->wb_pgbase + req->wb_bytes;
164 
165 		if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
166 			/* note: regions of the page not covered by a
167 			 * request are zeroed in nfs_readpage_async /
168 			 * readpage_async_filler */
169 			if (bytes > hdr->good_bytes) {
170 				/* nothing in this request was good, so zero
171 				 * the full extent of the request */
172 				zero_user_segment(page, start, end);
173 
174 			} else if (hdr->good_bytes - bytes < req->wb_bytes) {
175 				/* part of this request has good bytes, but
176 				 * not all. zero the bad bytes */
177 				start += hdr->good_bytes - bytes;
178 				WARN_ON(start < req->wb_pgbase);
179 				zero_user_segment(page, start, end);
180 			}
181 		}
182 		bytes += req->wb_bytes;
183 		if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
184 			if (bytes <= hdr->good_bytes)
185 				nfs_page_group_set_uptodate(req);
186 		} else
187 			nfs_page_group_set_uptodate(req);
188 		nfs_list_remove_request(req);
189 		nfs_readpage_release(req);
190 	}
191 out:
192 	hdr->release(hdr);
193 }
194 
195 static void nfs_initiate_read(struct nfs_pgio_header *hdr,
196 			      struct rpc_message *msg,
197 			      const struct nfs_rpc_ops *rpc_ops,
198 			      struct rpc_task_setup *task_setup_data, int how)
199 {
200 	struct inode *inode = hdr->inode;
201 	int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
202 
203 	task_setup_data->flags |= swap_flags;
204 	rpc_ops->read_setup(hdr, msg);
205 	trace_nfs_initiate_read(inode, hdr->io_start, hdr->good_bytes);
206 }
207 
208 static void
209 nfs_async_read_error(struct list_head *head, int error)
210 {
211 	struct nfs_page	*req;
212 
213 	while (!list_empty(head)) {
214 		req = nfs_list_entry(head->next);
215 		nfs_list_remove_request(req);
216 		nfs_readpage_release(req);
217 	}
218 }
219 
220 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
221 	.error_cleanup = nfs_async_read_error,
222 	.completion = nfs_read_completion,
223 };
224 
225 /*
226  * This is the callback from RPC telling us whether a reply was
227  * received or some error occurred (timeout or socket shutdown).
228  */
229 static int nfs_readpage_done(struct rpc_task *task,
230 			     struct nfs_pgio_header *hdr,
231 			     struct inode *inode)
232 {
233 	int status = NFS_PROTO(inode)->read_done(task, hdr);
234 	if (status != 0)
235 		return status;
236 
237 	nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
238 	trace_nfs_readpage_done(inode, task->tk_status,
239 				hdr->args.offset, hdr->res.eof);
240 
241 	if (task->tk_status == -ESTALE) {
242 		set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
243 		nfs_mark_for_revalidate(inode);
244 	}
245 	return 0;
246 }
247 
248 static void nfs_readpage_retry(struct rpc_task *task,
249 			       struct nfs_pgio_header *hdr)
250 {
251 	struct nfs_pgio_args *argp = &hdr->args;
252 	struct nfs_pgio_res  *resp = &hdr->res;
253 
254 	/* This is a short read! */
255 	nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
256 	/* Has the server at least made some progress? */
257 	if (resp->count == 0) {
258 		nfs_set_pgio_error(hdr, -EIO, argp->offset);
259 		return;
260 	}
261 
262 	/* For non rpc-based layout drivers, retry-through-MDS */
263 	if (!task->tk_ops) {
264 		hdr->pnfs_error = -EAGAIN;
265 		return;
266 	}
267 
268 	/* Yes, so retry the read at the end of the hdr */
269 	hdr->mds_offset += resp->count;
270 	argp->offset += resp->count;
271 	argp->pgbase += resp->count;
272 	argp->count -= resp->count;
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 /*
293  * Read a page over NFS.
294  * We read the page synchronously in the following case:
295  *  -	The error flag is set for this page. This happens only when a
296  *	previous async read operation failed.
297  */
298 int nfs_readpage(struct file *file, struct page *page)
299 {
300 	struct nfs_open_context *ctx;
301 	struct inode *inode = page_file_mapping(page)->host;
302 	int		error;
303 
304 	dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
305 		page, PAGE_SIZE, page_index(page));
306 	nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
307 	nfs_add_stats(inode, NFSIOS_READPAGES, 1);
308 
309 	/*
310 	 * Try to flush any pending writes to the file..
311 	 *
312 	 * NOTE! Because we own the page lock, there cannot
313 	 * be any new pending writes generated at this point
314 	 * for this page (other pages can be written to).
315 	 */
316 	error = nfs_wb_page(inode, page);
317 	if (error)
318 		goto out_unlock;
319 	if (PageUptodate(page))
320 		goto out_unlock;
321 
322 	error = -ESTALE;
323 	if (NFS_STALE(inode))
324 		goto out_unlock;
325 
326 	if (file == NULL) {
327 		error = -EBADF;
328 		ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
329 		if (ctx == NULL)
330 			goto out_unlock;
331 	} else
332 		ctx = get_nfs_open_context(nfs_file_open_context(file));
333 
334 	if (!IS_SYNC(inode)) {
335 		error = nfs_readpage_from_fscache(ctx, inode, page);
336 		if (error == 0)
337 			goto out;
338 	}
339 
340 	error = nfs_readpage_async(ctx, inode, page);
341 
342 out:
343 	put_nfs_open_context(ctx);
344 	return error;
345 out_unlock:
346 	unlock_page(page);
347 	return error;
348 }
349 
350 struct nfs_readdesc {
351 	struct nfs_pageio_descriptor *pgio;
352 	struct nfs_open_context *ctx;
353 };
354 
355 static int
356 readpage_async_filler(void *data, struct page *page)
357 {
358 	struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
359 	struct nfs_page *new;
360 	unsigned int len;
361 	int error;
362 
363 	len = nfs_page_length(page);
364 	if (len == 0)
365 		return nfs_return_empty_page(page);
366 
367 	new = nfs_create_request(desc->ctx, page, 0, len);
368 	if (IS_ERR(new))
369 		goto out_error;
370 
371 	if (len < PAGE_SIZE)
372 		zero_user_segment(page, len, PAGE_SIZE);
373 	if (!nfs_pageio_add_request(desc->pgio, new)) {
374 		nfs_list_remove_request(new);
375 		nfs_readpage_release(new);
376 		error = desc->pgio->pg_error;
377 		goto out;
378 	}
379 	return 0;
380 out_error:
381 	error = PTR_ERR(new);
382 	unlock_page(page);
383 out:
384 	return error;
385 }
386 
387 int nfs_readpages(struct file *filp, struct address_space *mapping,
388 		struct list_head *pages, unsigned nr_pages)
389 {
390 	struct nfs_pageio_descriptor pgio;
391 	struct nfs_pgio_mirror *pgm;
392 	struct nfs_readdesc desc = {
393 		.pgio = &pgio,
394 	};
395 	struct inode *inode = mapping->host;
396 	unsigned long npages;
397 	int ret = -ESTALE;
398 
399 	dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
400 			inode->i_sb->s_id,
401 			(unsigned long long)NFS_FILEID(inode),
402 			nr_pages);
403 	nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
404 
405 	if (NFS_STALE(inode))
406 		goto out;
407 
408 	if (filp == NULL) {
409 		desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
410 		if (desc.ctx == NULL)
411 			return -EBADF;
412 	} else
413 		desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
414 
415 	/* attempt to read as many of the pages as possible from the cache
416 	 * - this returns -ENOBUFS immediately if the cookie is negative
417 	 */
418 	ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
419 					 pages, &nr_pages);
420 	if (ret == 0)
421 		goto read_complete; /* all pages were read */
422 
423 	nfs_pageio_init_read(&pgio, inode, false,
424 			     &nfs_async_read_completion_ops);
425 
426 	ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
427 	nfs_pageio_complete(&pgio);
428 
429 	/* It doesn't make sense to do mirrored reads! */
430 	WARN_ON_ONCE(pgio.pg_mirror_count != 1);
431 
432 	pgm = &pgio.pg_mirrors[0];
433 	NFS_I(inode)->read_io += pgm->pg_bytes_written;
434 	npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >>
435 		 PAGE_SHIFT;
436 	nfs_add_stats(inode, NFSIOS_READPAGES, npages);
437 read_complete:
438 	put_nfs_open_context(desc.ctx);
439 out:
440 	return ret;
441 }
442 
443 int __init nfs_init_readpagecache(void)
444 {
445 	nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
446 					     sizeof(struct nfs_pgio_header),
447 					     0, SLAB_HWCACHE_ALIGN,
448 					     NULL);
449 	if (nfs_rdata_cachep == NULL)
450 		return -ENOMEM;
451 
452 	return 0;
453 }
454 
455 void nfs_destroy_readpagecache(void)
456 {
457 	kmem_cache_destroy(nfs_rdata_cachep);
458 }
459 
460 static const struct nfs_rw_ops nfs_rw_read_ops = {
461 	.rw_alloc_header	= nfs_readhdr_alloc,
462 	.rw_free_header		= nfs_readhdr_free,
463 	.rw_done		= nfs_readpage_done,
464 	.rw_result		= nfs_readpage_result,
465 	.rw_initiate		= nfs_initiate_read,
466 };
467