xref: /openbmc/linux/fs/nfs/pagelist.c (revision b6dcefde)
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_page		*req;
62 
63 	for (;;) {
64 		/* try to allocate the request struct */
65 		req = nfs_page_alloc();
66 		if (req != NULL)
67 			break;
68 
69 		if (fatal_signal_pending(current))
70 			return ERR_PTR(-ERESTARTSYS);
71 		yield();
72 	}
73 
74 	/* Initialize the request struct. Initially, we assume a
75 	 * long write-back delay. This will be adjusted in
76 	 * update_nfs_request below if the region is not locked. */
77 	req->wb_page    = page;
78 	atomic_set(&req->wb_complete, 0);
79 	req->wb_index	= page->index;
80 	page_cache_get(page);
81 	BUG_ON(PagePrivate(page));
82 	BUG_ON(!PageLocked(page));
83 	BUG_ON(page->mapping->host != inode);
84 	req->wb_offset  = offset;
85 	req->wb_pgbase	= offset;
86 	req->wb_bytes   = count;
87 	req->wb_context = get_nfs_open_context(ctx);
88 	kref_init(&req->wb_kref);
89 	return req;
90 }
91 
92 /**
93  * nfs_unlock_request - Unlock request and wake up sleepers.
94  * @req:
95  */
96 void nfs_unlock_request(struct nfs_page *req)
97 {
98 	if (!NFS_WBACK_BUSY(req)) {
99 		printk(KERN_ERR "NFS: Invalid unlock attempted\n");
100 		BUG();
101 	}
102 	smp_mb__before_clear_bit();
103 	clear_bit(PG_BUSY, &req->wb_flags);
104 	smp_mb__after_clear_bit();
105 	wake_up_bit(&req->wb_flags, PG_BUSY);
106 	nfs_release_request(req);
107 }
108 
109 /**
110  * nfs_set_page_tag_locked - Tag a request as locked
111  * @req:
112  */
113 int nfs_set_page_tag_locked(struct nfs_page *req)
114 {
115 	struct nfs_inode *nfsi = NFS_I(req->wb_context->path.dentry->d_inode);
116 
117 	if (!nfs_lock_request_dontget(req))
118 		return 0;
119 	if (req->wb_page != NULL)
120 		radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
121 	return 1;
122 }
123 
124 /**
125  * nfs_clear_page_tag_locked - Clear request tag and wake up sleepers
126  */
127 void nfs_clear_page_tag_locked(struct nfs_page *req)
128 {
129 	struct inode *inode = req->wb_context->path.dentry->d_inode;
130 	struct nfs_inode *nfsi = NFS_I(inode);
131 
132 	if (req->wb_page != NULL) {
133 		spin_lock(&inode->i_lock);
134 		radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
135 		nfs_unlock_request(req);
136 		spin_unlock(&inode->i_lock);
137 	} else
138 		nfs_unlock_request(req);
139 }
140 
141 /**
142  * nfs_clear_request - Free up all resources allocated to the request
143  * @req:
144  *
145  * Release page resources associated with a write request after it
146  * has completed.
147  */
148 void nfs_clear_request(struct nfs_page *req)
149 {
150 	struct page *page = req->wb_page;
151 	if (page != NULL) {
152 		page_cache_release(page);
153 		req->wb_page = NULL;
154 	}
155 }
156 
157 
158 /**
159  * nfs_release_request - Release the count on an NFS read/write request
160  * @req: request to release
161  *
162  * Note: Should never be called with the spinlock held!
163  */
164 static void nfs_free_request(struct kref *kref)
165 {
166 	struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
167 
168 	/* Release struct file or cached credential */
169 	nfs_clear_request(req);
170 	put_nfs_open_context(req->wb_context);
171 	nfs_page_free(req);
172 }
173 
174 void nfs_release_request(struct nfs_page *req)
175 {
176 	kref_put(&req->wb_kref, nfs_free_request);
177 }
178 
179 static int nfs_wait_bit_uninterruptible(void *word)
180 {
181 	io_schedule();
182 	return 0;
183 }
184 
185 /**
186  * nfs_wait_on_request - Wait for a request to complete.
187  * @req: request to wait upon.
188  *
189  * Interruptible by fatal signals only.
190  * The user is responsible for holding a count on the request.
191  */
192 int
193 nfs_wait_on_request(struct nfs_page *req)
194 {
195 	return wait_on_bit(&req->wb_flags, PG_BUSY,
196 			nfs_wait_bit_uninterruptible,
197 			TASK_UNINTERRUPTIBLE);
198 }
199 
200 /**
201  * nfs_pageio_init - initialise a page io descriptor
202  * @desc: pointer to descriptor
203  * @inode: pointer to inode
204  * @doio: pointer to io function
205  * @bsize: io block size
206  * @io_flags: extra parameters for the io function
207  */
208 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
209 		     struct inode *inode,
210 		     int (*doio)(struct inode *, struct list_head *, unsigned int, size_t, int),
211 		     size_t bsize,
212 		     int io_flags)
213 {
214 	INIT_LIST_HEAD(&desc->pg_list);
215 	desc->pg_bytes_written = 0;
216 	desc->pg_count = 0;
217 	desc->pg_bsize = bsize;
218 	desc->pg_base = 0;
219 	desc->pg_inode = inode;
220 	desc->pg_doio = doio;
221 	desc->pg_ioflags = io_flags;
222 	desc->pg_error = 0;
223 }
224 
225 /**
226  * nfs_can_coalesce_requests - test two requests for compatibility
227  * @prev: pointer to nfs_page
228  * @req: pointer to nfs_page
229  *
230  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
231  * page data area they describe is contiguous, and that their RPC
232  * credentials, NFSv4 open state, and lockowners are the same.
233  *
234  * Return 'true' if this is the case, else return 'false'.
235  */
236 static int nfs_can_coalesce_requests(struct nfs_page *prev,
237 				     struct nfs_page *req)
238 {
239 	if (req->wb_context->cred != prev->wb_context->cred)
240 		return 0;
241 	if (req->wb_context->lockowner != prev->wb_context->lockowner)
242 		return 0;
243 	if (req->wb_context->state != prev->wb_context->state)
244 		return 0;
245 	if (req->wb_index != (prev->wb_index + 1))
246 		return 0;
247 	if (req->wb_pgbase != 0)
248 		return 0;
249 	if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
250 		return 0;
251 	return 1;
252 }
253 
254 /**
255  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
256  * @desc: destination io descriptor
257  * @req: request
258  *
259  * Returns true if the request 'req' was successfully coalesced into the
260  * existing list of pages 'desc'.
261  */
262 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
263 				     struct nfs_page *req)
264 {
265 	size_t newlen = req->wb_bytes;
266 
267 	if (desc->pg_count != 0) {
268 		struct nfs_page *prev;
269 
270 		/*
271 		 * FIXME: ideally we should be able to coalesce all requests
272 		 * that are not block boundary aligned, but currently this
273 		 * is problematic for the case of bsize < PAGE_CACHE_SIZE,
274 		 * since nfs_flush_multi and nfs_pagein_multi assume you
275 		 * can have only one struct nfs_page.
276 		 */
277 		if (desc->pg_bsize < PAGE_SIZE)
278 			return 0;
279 		newlen += desc->pg_count;
280 		if (newlen > desc->pg_bsize)
281 			return 0;
282 		prev = nfs_list_entry(desc->pg_list.prev);
283 		if (!nfs_can_coalesce_requests(prev, req))
284 			return 0;
285 	} else
286 		desc->pg_base = req->wb_pgbase;
287 	nfs_list_remove_request(req);
288 	nfs_list_add_request(req, &desc->pg_list);
289 	desc->pg_count = newlen;
290 	return 1;
291 }
292 
293 /*
294  * Helper for nfs_pageio_add_request and nfs_pageio_complete
295  */
296 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
297 {
298 	if (!list_empty(&desc->pg_list)) {
299 		int error = desc->pg_doio(desc->pg_inode,
300 					  &desc->pg_list,
301 					  nfs_page_array_len(desc->pg_base,
302 							     desc->pg_count),
303 					  desc->pg_count,
304 					  desc->pg_ioflags);
305 		if (error < 0)
306 			desc->pg_error = error;
307 		else
308 			desc->pg_bytes_written += desc->pg_count;
309 	}
310 	if (list_empty(&desc->pg_list)) {
311 		desc->pg_count = 0;
312 		desc->pg_base = 0;
313 	}
314 }
315 
316 /**
317  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
318  * @desc: destination io descriptor
319  * @req: request
320  *
321  * Returns true if the request 'req' was successfully coalesced into the
322  * existing list of pages 'desc'.
323  */
324 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
325 			   struct nfs_page *req)
326 {
327 	while (!nfs_pageio_do_add_request(desc, req)) {
328 		nfs_pageio_doio(desc);
329 		if (desc->pg_error < 0)
330 			return 0;
331 	}
332 	return 1;
333 }
334 
335 /**
336  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
337  * @desc: pointer to io descriptor
338  */
339 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
340 {
341 	nfs_pageio_doio(desc);
342 }
343 
344 /**
345  * nfs_pageio_cond_complete - Conditional I/O completion
346  * @desc: pointer to io descriptor
347  * @index: page index
348  *
349  * It is important to ensure that processes don't try to take locks
350  * on non-contiguous ranges of pages as that might deadlock. This
351  * function should be called before attempting to wait on a locked
352  * nfs_page. It will complete the I/O if the page index 'index'
353  * is not contiguous with the existing list of pages in 'desc'.
354  */
355 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
356 {
357 	if (!list_empty(&desc->pg_list)) {
358 		struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
359 		if (index != prev->wb_index + 1)
360 			nfs_pageio_doio(desc);
361 	}
362 }
363 
364 #define NFS_SCAN_MAXENTRIES 16
365 /**
366  * nfs_scan_list - Scan a list for matching requests
367  * @nfsi: NFS inode
368  * @dst: Destination list
369  * @idx_start: lower bound of page->index to scan
370  * @npages: idx_start + npages sets the upper bound to scan.
371  * @tag: tag to scan for
372  *
373  * Moves elements from one of the inode request lists.
374  * If the number of requests is set to 0, the entire address_space
375  * starting at index idx_start, is scanned.
376  * The requests are *not* checked to ensure that they form a contiguous set.
377  * You must be holding the inode's i_lock when calling this function
378  */
379 int nfs_scan_list(struct nfs_inode *nfsi,
380 		struct list_head *dst, pgoff_t idx_start,
381 		unsigned int npages, int tag)
382 {
383 	struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
384 	struct nfs_page *req;
385 	pgoff_t idx_end;
386 	int found, i;
387 	int res;
388 
389 	res = 0;
390 	if (npages == 0)
391 		idx_end = ~0;
392 	else
393 		idx_end = idx_start + npages - 1;
394 
395 	for (;;) {
396 		found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
397 				(void **)&pgvec[0], idx_start,
398 				NFS_SCAN_MAXENTRIES, tag);
399 		if (found <= 0)
400 			break;
401 		for (i = 0; i < found; i++) {
402 			req = pgvec[i];
403 			if (req->wb_index > idx_end)
404 				goto out;
405 			idx_start = req->wb_index + 1;
406 			if (nfs_set_page_tag_locked(req)) {
407 				kref_get(&req->wb_kref);
408 				nfs_list_remove_request(req);
409 				radix_tree_tag_clear(&nfsi->nfs_page_tree,
410 						req->wb_index, tag);
411 				nfs_list_add_request(req, dst);
412 				res++;
413 				if (res == INT_MAX)
414 					goto out;
415 			}
416 		}
417 		/* for latency reduction */
418 		cond_resched_lock(&nfsi->vfs_inode.i_lock);
419 	}
420 out:
421 	return res;
422 }
423 
424 int __init nfs_init_nfspagecache(void)
425 {
426 	nfs_page_cachep = kmem_cache_create("nfs_page",
427 					    sizeof(struct nfs_page),
428 					    0, SLAB_HWCACHE_ALIGN,
429 					    NULL);
430 	if (nfs_page_cachep == NULL)
431 		return -ENOMEM;
432 
433 	return 0;
434 }
435 
436 void nfs_destroy_nfspagecache(void)
437 {
438 	kmem_cache_destroy(nfs_page_cachep);
439 }
440 
441