xref: /openbmc/linux/fs/nfs/pagelist.c (revision 64c70b1c)
1 /*
2  * linux/fs/nfs/pagelist.c
3  *
4  * A set of helper functions for managing NFS read and write requests.
5  * The main purpose of these routines is to provide support for the
6  * coalescing of several requests into a single RPC call.
7  *
8  * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
9  *
10  */
11 
12 #include <linux/slab.h>
13 #include <linux/file.h>
14 #include <linux/sched.h>
15 #include <linux/sunrpc/clnt.h>
16 #include <linux/nfs3.h>
17 #include <linux/nfs4.h>
18 #include <linux/nfs_page.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 
22 #include "internal.h"
23 
24 static struct kmem_cache *nfs_page_cachep;
25 
26 static inline struct nfs_page *
27 nfs_page_alloc(void)
28 {
29 	struct nfs_page	*p;
30 	p = kmem_cache_alloc(nfs_page_cachep, GFP_KERNEL);
31 	if (p) {
32 		memset(p, 0, sizeof(*p));
33 		INIT_LIST_HEAD(&p->wb_list);
34 	}
35 	return p;
36 }
37 
38 static inline void
39 nfs_page_free(struct nfs_page *p)
40 {
41 	kmem_cache_free(nfs_page_cachep, p);
42 }
43 
44 /**
45  * nfs_create_request - Create an NFS read/write request.
46  * @file: file descriptor to use
47  * @inode: inode to which the request is attached
48  * @page: page to write
49  * @offset: starting offset within the page for the write
50  * @count: number of bytes to read/write
51  *
52  * The page must be locked by the caller. This makes sure we never
53  * create two different requests for the same page.
54  * User should ensure it is safe to sleep in this function.
55  */
56 struct nfs_page *
57 nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
58 		   struct page *page,
59 		   unsigned int offset, unsigned int count)
60 {
61 	struct nfs_server *server = NFS_SERVER(inode);
62 	struct nfs_page		*req;
63 
64 	for (;;) {
65 		/* try to allocate the request struct */
66 		req = nfs_page_alloc();
67 		if (req != NULL)
68 			break;
69 
70 		if (signalled() && (server->flags & NFS_MOUNT_INTR))
71 			return ERR_PTR(-ERESTARTSYS);
72 		yield();
73 	}
74 
75 	/* Initialize the request struct. Initially, we assume a
76 	 * long write-back delay. This will be adjusted in
77 	 * update_nfs_request below if the region is not locked. */
78 	req->wb_page    = page;
79 	atomic_set(&req->wb_complete, 0);
80 	req->wb_index	= page->index;
81 	page_cache_get(page);
82 	BUG_ON(PagePrivate(page));
83 	BUG_ON(!PageLocked(page));
84 	BUG_ON(page->mapping->host != inode);
85 	req->wb_offset  = offset;
86 	req->wb_pgbase	= offset;
87 	req->wb_bytes   = count;
88 	atomic_set(&req->wb_count, 1);
89 	req->wb_context = get_nfs_open_context(ctx);
90 
91 	return req;
92 }
93 
94 /**
95  * nfs_unlock_request - Unlock request and wake up sleepers.
96  * @req:
97  */
98 void nfs_unlock_request(struct nfs_page *req)
99 {
100 	if (!NFS_WBACK_BUSY(req)) {
101 		printk(KERN_ERR "NFS: Invalid unlock attempted\n");
102 		BUG();
103 	}
104 	smp_mb__before_clear_bit();
105 	clear_bit(PG_BUSY, &req->wb_flags);
106 	smp_mb__after_clear_bit();
107 	wake_up_bit(&req->wb_flags, PG_BUSY);
108 	nfs_release_request(req);
109 }
110 
111 /**
112  * nfs_set_page_writeback_locked - Lock a request for writeback
113  * @req:
114  */
115 int nfs_set_page_writeback_locked(struct nfs_page *req)
116 {
117 	struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
118 
119 	if (!nfs_lock_request(req))
120 		return 0;
121 	radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_WRITEBACK);
122 	return 1;
123 }
124 
125 /**
126  * nfs_clear_page_writeback - Unlock request and wake up sleepers
127  */
128 void nfs_clear_page_writeback(struct nfs_page *req)
129 {
130 	struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
131 
132 	if (req->wb_page != NULL) {
133 		spin_lock(&nfsi->req_lock);
134 		radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_WRITEBACK);
135 		spin_unlock(&nfsi->req_lock);
136 	}
137 	nfs_unlock_request(req);
138 }
139 
140 /**
141  * nfs_clear_request - Free up all resources allocated to the request
142  * @req:
143  *
144  * Release page resources associated with a write request after it
145  * has completed.
146  */
147 void nfs_clear_request(struct nfs_page *req)
148 {
149 	struct page *page = req->wb_page;
150 	if (page != NULL) {
151 		page_cache_release(page);
152 		req->wb_page = NULL;
153 	}
154 }
155 
156 
157 /**
158  * nfs_release_request - Release the count on an NFS read/write request
159  * @req: request to release
160  *
161  * Note: Should never be called with the spinlock held!
162  */
163 void
164 nfs_release_request(struct nfs_page *req)
165 {
166 	if (!atomic_dec_and_test(&req->wb_count))
167 		return;
168 
169 	/* Release struct file or cached credential */
170 	nfs_clear_request(req);
171 	put_nfs_open_context(req->wb_context);
172 	nfs_page_free(req);
173 }
174 
175 static int nfs_wait_bit_interruptible(void *word)
176 {
177 	int ret = 0;
178 
179 	if (signal_pending(current))
180 		ret = -ERESTARTSYS;
181 	else
182 		schedule();
183 	return ret;
184 }
185 
186 /**
187  * nfs_wait_on_request - Wait for a request to complete.
188  * @req: request to wait upon.
189  *
190  * Interruptible by signals only if mounted with intr flag.
191  * The user is responsible for holding a count on the request.
192  */
193 int
194 nfs_wait_on_request(struct nfs_page *req)
195 {
196         struct rpc_clnt	*clnt = NFS_CLIENT(req->wb_context->dentry->d_inode);
197 	sigset_t oldmask;
198 	int ret = 0;
199 
200 	if (!test_bit(PG_BUSY, &req->wb_flags))
201 		goto out;
202 	/*
203 	 * Note: the call to rpc_clnt_sigmask() suffices to ensure that we
204 	 *	 are not interrupted if intr flag is not set
205 	 */
206 	rpc_clnt_sigmask(clnt, &oldmask);
207 	ret = out_of_line_wait_on_bit(&req->wb_flags, PG_BUSY,
208 			nfs_wait_bit_interruptible, TASK_INTERRUPTIBLE);
209 	rpc_clnt_sigunmask(clnt, &oldmask);
210 out:
211 	return ret;
212 }
213 
214 /**
215  * nfs_pageio_init - initialise a page io descriptor
216  * @desc: pointer to descriptor
217  * @inode: pointer to inode
218  * @doio: pointer to io function
219  * @bsize: io block size
220  * @io_flags: extra parameters for the io function
221  */
222 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
223 		     struct inode *inode,
224 		     int (*doio)(struct inode *, struct list_head *, unsigned int, size_t, int),
225 		     size_t bsize,
226 		     int io_flags)
227 {
228 	INIT_LIST_HEAD(&desc->pg_list);
229 	desc->pg_bytes_written = 0;
230 	desc->pg_count = 0;
231 	desc->pg_bsize = bsize;
232 	desc->pg_base = 0;
233 	desc->pg_inode = inode;
234 	desc->pg_doio = doio;
235 	desc->pg_ioflags = io_flags;
236 	desc->pg_error = 0;
237 }
238 
239 /**
240  * nfs_can_coalesce_requests - test two requests for compatibility
241  * @prev: pointer to nfs_page
242  * @req: pointer to nfs_page
243  *
244  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
245  * page data area they describe is contiguous, and that their RPC
246  * credentials, NFSv4 open state, and lockowners are the same.
247  *
248  * Return 'true' if this is the case, else return 'false'.
249  */
250 static int nfs_can_coalesce_requests(struct nfs_page *prev,
251 				     struct nfs_page *req)
252 {
253 	if (req->wb_context->cred != prev->wb_context->cred)
254 		return 0;
255 	if (req->wb_context->lockowner != prev->wb_context->lockowner)
256 		return 0;
257 	if (req->wb_context->state != prev->wb_context->state)
258 		return 0;
259 	if (req->wb_index != (prev->wb_index + 1))
260 		return 0;
261 	if (req->wb_pgbase != 0)
262 		return 0;
263 	if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
264 		return 0;
265 	return 1;
266 }
267 
268 /**
269  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
270  * @desc: destination io descriptor
271  * @req: request
272  *
273  * Returns true if the request 'req' was successfully coalesced into the
274  * existing list of pages 'desc'.
275  */
276 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
277 				     struct nfs_page *req)
278 {
279 	size_t newlen = req->wb_bytes;
280 
281 	if (desc->pg_count != 0) {
282 		struct nfs_page *prev;
283 
284 		/*
285 		 * FIXME: ideally we should be able to coalesce all requests
286 		 * that are not block boundary aligned, but currently this
287 		 * is problematic for the case of bsize < PAGE_CACHE_SIZE,
288 		 * since nfs_flush_multi and nfs_pagein_multi assume you
289 		 * can have only one struct nfs_page.
290 		 */
291 		if (desc->pg_bsize < PAGE_SIZE)
292 			return 0;
293 		newlen += desc->pg_count;
294 		if (newlen > desc->pg_bsize)
295 			return 0;
296 		prev = nfs_list_entry(desc->pg_list.prev);
297 		if (!nfs_can_coalesce_requests(prev, req))
298 			return 0;
299 	} else
300 		desc->pg_base = req->wb_pgbase;
301 	nfs_list_remove_request(req);
302 	nfs_list_add_request(req, &desc->pg_list);
303 	desc->pg_count = newlen;
304 	return 1;
305 }
306 
307 /*
308  * Helper for nfs_pageio_add_request and nfs_pageio_complete
309  */
310 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
311 {
312 	if (!list_empty(&desc->pg_list)) {
313 		int error = desc->pg_doio(desc->pg_inode,
314 					  &desc->pg_list,
315 					  nfs_page_array_len(desc->pg_base,
316 							     desc->pg_count),
317 					  desc->pg_count,
318 					  desc->pg_ioflags);
319 		if (error < 0)
320 			desc->pg_error = error;
321 		else
322 			desc->pg_bytes_written += desc->pg_count;
323 	}
324 	if (list_empty(&desc->pg_list)) {
325 		desc->pg_count = 0;
326 		desc->pg_base = 0;
327 	}
328 }
329 
330 /**
331  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
332  * @desc: destination io descriptor
333  * @req: request
334  *
335  * Returns true if the request 'req' was successfully coalesced into the
336  * existing list of pages 'desc'.
337  */
338 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
339 			   struct nfs_page *req)
340 {
341 	while (!nfs_pageio_do_add_request(desc, req)) {
342 		nfs_pageio_doio(desc);
343 		if (desc->pg_error < 0)
344 			return 0;
345 	}
346 	return 1;
347 }
348 
349 /**
350  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
351  * @desc: pointer to io descriptor
352  */
353 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
354 {
355 	nfs_pageio_doio(desc);
356 }
357 
358 /**
359  * nfs_pageio_cond_complete - Conditional I/O completion
360  * @desc: pointer to io descriptor
361  * @index: page index
362  *
363  * It is important to ensure that processes don't try to take locks
364  * on non-contiguous ranges of pages as that might deadlock. This
365  * function should be called before attempting to wait on a locked
366  * nfs_page. It will complete the I/O if the page index 'index'
367  * is not contiguous with the existing list of pages in 'desc'.
368  */
369 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
370 {
371 	if (!list_empty(&desc->pg_list)) {
372 		struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
373 		if (index != prev->wb_index + 1)
374 			nfs_pageio_doio(desc);
375 	}
376 }
377 
378 #define NFS_SCAN_MAXENTRIES 16
379 /**
380  * nfs_scan_list - Scan a list for matching requests
381  * @nfsi: NFS inode
382  * @head: One of the NFS inode request lists
383  * @dst: Destination list
384  * @idx_start: lower bound of page->index to scan
385  * @npages: idx_start + npages sets the upper bound to scan.
386  *
387  * Moves elements from one of the inode request lists.
388  * If the number of requests is set to 0, the entire address_space
389  * starting at index idx_start, is scanned.
390  * The requests are *not* checked to ensure that they form a contiguous set.
391  * You must be holding the inode's req_lock when calling this function
392  */
393 int nfs_scan_list(struct nfs_inode *nfsi, struct list_head *head,
394 		struct list_head *dst, pgoff_t idx_start,
395 		unsigned int npages)
396 {
397 	struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
398 	struct nfs_page *req;
399 	pgoff_t idx_end;
400 	int found, i;
401 	int res;
402 
403 	res = 0;
404 	if (npages == 0)
405 		idx_end = ~0;
406 	else
407 		idx_end = idx_start + npages - 1;
408 
409 	for (;;) {
410 		found = radix_tree_gang_lookup(&nfsi->nfs_page_tree,
411 				(void **)&pgvec[0], idx_start,
412 				NFS_SCAN_MAXENTRIES);
413 		if (found <= 0)
414 			break;
415 		for (i = 0; i < found; i++) {
416 			req = pgvec[i];
417 			if (req->wb_index > idx_end)
418 				goto out;
419 			idx_start = req->wb_index + 1;
420 			if (req->wb_list_head != head)
421 				continue;
422 			if (nfs_set_page_writeback_locked(req)) {
423 				nfs_list_remove_request(req);
424 				nfs_list_add_request(req, dst);
425 				res++;
426 			}
427 		}
428 
429 	}
430 out:
431 	return res;
432 }
433 
434 int __init nfs_init_nfspagecache(void)
435 {
436 	nfs_page_cachep = kmem_cache_create("nfs_page",
437 					    sizeof(struct nfs_page),
438 					    0, SLAB_HWCACHE_ALIGN,
439 					    NULL, NULL);
440 	if (nfs_page_cachep == NULL)
441 		return -ENOMEM;
442 
443 	return 0;
444 }
445 
446 void nfs_destroy_nfspagecache(void)
447 {
448 	kmem_cache_destroy(nfs_page_cachep);
449 }
450 
451