Loading...
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#include "pnfs.h"
24
25static struct kmem_cache *nfs_page_cachep;
26
27static inline struct nfs_page *
28nfs_page_alloc(void)
29{
30 struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_KERNEL);
31 if (p)
32 INIT_LIST_HEAD(&p->wb_list);
33 return p;
34}
35
36static inline void
37nfs_page_free(struct nfs_page *p)
38{
39 kmem_cache_free(nfs_page_cachep, p);
40}
41
42/**
43 * nfs_create_request - Create an NFS read/write request.
44 * @file: file descriptor to use
45 * @inode: inode to which the request is attached
46 * @page: page to write
47 * @offset: starting offset within the page for the write
48 * @count: number of bytes to read/write
49 *
50 * The page must be locked by the caller. This makes sure we never
51 * create two different requests for the same page.
52 * User should ensure it is safe to sleep in this function.
53 */
54struct nfs_page *
55nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
56 struct page *page,
57 unsigned int offset, unsigned int count)
58{
59 struct nfs_page *req;
60
61 /* try to allocate the request struct */
62 req = nfs_page_alloc();
63 if (req == NULL)
64 return ERR_PTR(-ENOMEM);
65
66 /* get lock context early so we can deal with alloc failures */
67 req->wb_lock_context = nfs_get_lock_context(ctx);
68 if (req->wb_lock_context == NULL) {
69 nfs_page_free(req);
70 return ERR_PTR(-ENOMEM);
71 }
72
73 /* Initialize the request struct. Initially, we assume a
74 * long write-back delay. This will be adjusted in
75 * update_nfs_request below if the region is not locked. */
76 req->wb_page = page;
77 atomic_set(&req->wb_complete, 0);
78 req->wb_index = page->index;
79 page_cache_get(page);
80 BUG_ON(PagePrivate(page));
81 BUG_ON(!PageLocked(page));
82 BUG_ON(page->mapping->host != inode);
83 req->wb_offset = offset;
84 req->wb_pgbase = offset;
85 req->wb_bytes = count;
86 req->wb_context = get_nfs_open_context(ctx);
87 kref_init(&req->wb_kref);
88 return req;
89}
90
91/**
92 * nfs_unlock_request - Unlock request and wake up sleepers.
93 * @req:
94 */
95void nfs_unlock_request(struct nfs_page *req)
96{
97 if (!NFS_WBACK_BUSY(req)) {
98 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
99 BUG();
100 }
101 smp_mb__before_clear_bit();
102 clear_bit(PG_BUSY, &req->wb_flags);
103 smp_mb__after_clear_bit();
104 wake_up_bit(&req->wb_flags, PG_BUSY);
105 nfs_release_request(req);
106}
107
108/**
109 * nfs_set_page_tag_locked - Tag a request as locked
110 * @req:
111 */
112int nfs_set_page_tag_locked(struct nfs_page *req)
113{
114 if (!nfs_lock_request_dontget(req))
115 return 0;
116 if (test_bit(PG_MAPPED, &req->wb_flags))
117 radix_tree_tag_set(&NFS_I(req->wb_context->dentry->d_inode)->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
118 return 1;
119}
120
121/**
122 * nfs_clear_page_tag_locked - Clear request tag and wake up sleepers
123 */
124void nfs_clear_page_tag_locked(struct nfs_page *req)
125{
126 if (test_bit(PG_MAPPED, &req->wb_flags)) {
127 struct inode *inode = req->wb_context->dentry->d_inode;
128 struct nfs_inode *nfsi = NFS_I(inode);
129
130 spin_lock(&inode->i_lock);
131 radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
132 nfs_unlock_request(req);
133 spin_unlock(&inode->i_lock);
134 } else
135 nfs_unlock_request(req);
136}
137
138/*
139 * nfs_clear_request - Free up all resources allocated to the request
140 * @req:
141 *
142 * Release page and open context resources associated with a read/write
143 * request after it has completed.
144 */
145static void nfs_clear_request(struct nfs_page *req)
146{
147 struct page *page = req->wb_page;
148 struct nfs_open_context *ctx = req->wb_context;
149 struct nfs_lock_context *l_ctx = req->wb_lock_context;
150
151 if (page != NULL) {
152 page_cache_release(page);
153 req->wb_page = NULL;
154 }
155 if (l_ctx != NULL) {
156 nfs_put_lock_context(l_ctx);
157 req->wb_lock_context = NULL;
158 }
159 if (ctx != NULL) {
160 put_nfs_open_context(ctx);
161 req->wb_context = NULL;
162 }
163}
164
165
166/**
167 * nfs_release_request - Release the count on an NFS read/write request
168 * @req: request to release
169 *
170 * Note: Should never be called with the spinlock held!
171 */
172static void nfs_free_request(struct kref *kref)
173{
174 struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
175
176 /* Release struct file and open context */
177 nfs_clear_request(req);
178 nfs_page_free(req);
179}
180
181void nfs_release_request(struct nfs_page *req)
182{
183 kref_put(&req->wb_kref, nfs_free_request);
184}
185
186static int nfs_wait_bit_uninterruptible(void *word)
187{
188 io_schedule();
189 return 0;
190}
191
192/**
193 * nfs_wait_on_request - Wait for a request to complete.
194 * @req: request to wait upon.
195 *
196 * Interruptible by fatal signals only.
197 * The user is responsible for holding a count on the request.
198 */
199int
200nfs_wait_on_request(struct nfs_page *req)
201{
202 return wait_on_bit(&req->wb_flags, PG_BUSY,
203 nfs_wait_bit_uninterruptible,
204 TASK_UNINTERRUPTIBLE);
205}
206
207bool nfs_generic_pg_test(struct nfs_pageio_descriptor *desc, struct nfs_page *prev, struct nfs_page *req)
208{
209 /*
210 * FIXME: ideally we should be able to coalesce all requests
211 * that are not block boundary aligned, but currently this
212 * is problematic for the case of bsize < PAGE_CACHE_SIZE,
213 * since nfs_flush_multi and nfs_pagein_multi assume you
214 * can have only one struct nfs_page.
215 */
216 if (desc->pg_bsize < PAGE_SIZE)
217 return 0;
218
219 return desc->pg_count + req->wb_bytes <= desc->pg_bsize;
220}
221EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
222
223/**
224 * nfs_pageio_init - initialise a page io descriptor
225 * @desc: pointer to descriptor
226 * @inode: pointer to inode
227 * @doio: pointer to io function
228 * @bsize: io block size
229 * @io_flags: extra parameters for the io function
230 */
231void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
232 struct inode *inode,
233 const struct nfs_pageio_ops *pg_ops,
234 size_t bsize,
235 int io_flags)
236{
237 INIT_LIST_HEAD(&desc->pg_list);
238 desc->pg_bytes_written = 0;
239 desc->pg_count = 0;
240 desc->pg_bsize = bsize;
241 desc->pg_base = 0;
242 desc->pg_moreio = 0;
243 desc->pg_recoalesce = 0;
244 desc->pg_inode = inode;
245 desc->pg_ops = pg_ops;
246 desc->pg_ioflags = io_flags;
247 desc->pg_error = 0;
248 desc->pg_lseg = NULL;
249}
250
251/**
252 * nfs_can_coalesce_requests - test two requests for compatibility
253 * @prev: pointer to nfs_page
254 * @req: pointer to nfs_page
255 *
256 * The nfs_page structures 'prev' and 'req' are compared to ensure that the
257 * page data area they describe is contiguous, and that their RPC
258 * credentials, NFSv4 open state, and lockowners are the same.
259 *
260 * Return 'true' if this is the case, else return 'false'.
261 */
262static bool nfs_can_coalesce_requests(struct nfs_page *prev,
263 struct nfs_page *req,
264 struct nfs_pageio_descriptor *pgio)
265{
266 if (req->wb_context->cred != prev->wb_context->cred)
267 return false;
268 if (req->wb_lock_context->lockowner != prev->wb_lock_context->lockowner)
269 return false;
270 if (req->wb_context->state != prev->wb_context->state)
271 return false;
272 if (req->wb_index != (prev->wb_index + 1))
273 return false;
274 if (req->wb_pgbase != 0)
275 return false;
276 if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
277 return false;
278 return pgio->pg_ops->pg_test(pgio, prev, req);
279}
280
281/**
282 * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
283 * @desc: destination io descriptor
284 * @req: request
285 *
286 * Returns true if the request 'req' was successfully coalesced into the
287 * existing list of pages 'desc'.
288 */
289static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
290 struct nfs_page *req)
291{
292 if (desc->pg_count != 0) {
293 struct nfs_page *prev;
294
295 prev = nfs_list_entry(desc->pg_list.prev);
296 if (!nfs_can_coalesce_requests(prev, req, desc))
297 return 0;
298 } else {
299 if (desc->pg_ops->pg_init)
300 desc->pg_ops->pg_init(desc, req);
301 desc->pg_base = req->wb_pgbase;
302 }
303 nfs_list_remove_request(req);
304 nfs_list_add_request(req, &desc->pg_list);
305 desc->pg_count += req->wb_bytes;
306 return 1;
307}
308
309/*
310 * Helper for nfs_pageio_add_request and nfs_pageio_complete
311 */
312static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
313{
314 if (!list_empty(&desc->pg_list)) {
315 int error = desc->pg_ops->pg_doio(desc);
316 if (error < 0)
317 desc->pg_error = error;
318 else
319 desc->pg_bytes_written += desc->pg_count;
320 }
321 if (list_empty(&desc->pg_list)) {
322 desc->pg_count = 0;
323 desc->pg_base = 0;
324 }
325}
326
327/**
328 * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
329 * @desc: destination io descriptor
330 * @req: request
331 *
332 * Returns true if the request 'req' was successfully coalesced into the
333 * existing list of pages 'desc'.
334 */
335static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
336 struct nfs_page *req)
337{
338 while (!nfs_pageio_do_add_request(desc, req)) {
339 desc->pg_moreio = 1;
340 nfs_pageio_doio(desc);
341 if (desc->pg_error < 0)
342 return 0;
343 desc->pg_moreio = 0;
344 if (desc->pg_recoalesce)
345 return 0;
346 }
347 return 1;
348}
349
350static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
351{
352 LIST_HEAD(head);
353
354 do {
355 list_splice_init(&desc->pg_list, &head);
356 desc->pg_bytes_written -= desc->pg_count;
357 desc->pg_count = 0;
358 desc->pg_base = 0;
359 desc->pg_recoalesce = 0;
360
361 while (!list_empty(&head)) {
362 struct nfs_page *req;
363
364 req = list_first_entry(&head, struct nfs_page, wb_list);
365 nfs_list_remove_request(req);
366 if (__nfs_pageio_add_request(desc, req))
367 continue;
368 if (desc->pg_error < 0)
369 return 0;
370 break;
371 }
372 } while (desc->pg_recoalesce);
373 return 1;
374}
375
376int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
377 struct nfs_page *req)
378{
379 int ret;
380
381 do {
382 ret = __nfs_pageio_add_request(desc, req);
383 if (ret)
384 break;
385 if (desc->pg_error < 0)
386 break;
387 ret = nfs_do_recoalesce(desc);
388 } while (ret);
389 return ret;
390}
391
392/**
393 * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
394 * @desc: pointer to io descriptor
395 */
396void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
397{
398 for (;;) {
399 nfs_pageio_doio(desc);
400 if (!desc->pg_recoalesce)
401 break;
402 if (!nfs_do_recoalesce(desc))
403 break;
404 }
405}
406
407/**
408 * nfs_pageio_cond_complete - Conditional I/O completion
409 * @desc: pointer to io descriptor
410 * @index: page index
411 *
412 * It is important to ensure that processes don't try to take locks
413 * on non-contiguous ranges of pages as that might deadlock. This
414 * function should be called before attempting to wait on a locked
415 * nfs_page. It will complete the I/O if the page index 'index'
416 * is not contiguous with the existing list of pages in 'desc'.
417 */
418void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
419{
420 if (!list_empty(&desc->pg_list)) {
421 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
422 if (index != prev->wb_index + 1)
423 nfs_pageio_complete(desc);
424 }
425}
426
427#define NFS_SCAN_MAXENTRIES 16
428/**
429 * nfs_scan_list - Scan a list for matching requests
430 * @nfsi: NFS inode
431 * @dst: Destination list
432 * @idx_start: lower bound of page->index to scan
433 * @npages: idx_start + npages sets the upper bound to scan.
434 * @tag: tag to scan for
435 *
436 * Moves elements from one of the inode request lists.
437 * If the number of requests is set to 0, the entire address_space
438 * starting at index idx_start, is scanned.
439 * The requests are *not* checked to ensure that they form a contiguous set.
440 * You must be holding the inode's i_lock when calling this function
441 */
442int nfs_scan_list(struct nfs_inode *nfsi,
443 struct list_head *dst, pgoff_t idx_start,
444 unsigned int npages, int tag)
445{
446 struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
447 struct nfs_page *req;
448 pgoff_t idx_end;
449 int found, i;
450 int res;
451 struct list_head *list;
452
453 res = 0;
454 if (npages == 0)
455 idx_end = ~0;
456 else
457 idx_end = idx_start + npages - 1;
458
459 for (;;) {
460 found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
461 (void **)&pgvec[0], idx_start,
462 NFS_SCAN_MAXENTRIES, tag);
463 if (found <= 0)
464 break;
465 for (i = 0; i < found; i++) {
466 req = pgvec[i];
467 if (req->wb_index > idx_end)
468 goto out;
469 idx_start = req->wb_index + 1;
470 if (nfs_set_page_tag_locked(req)) {
471 kref_get(&req->wb_kref);
472 radix_tree_tag_clear(&nfsi->nfs_page_tree,
473 req->wb_index, tag);
474 list = pnfs_choose_commit_list(req, dst);
475 nfs_list_add_request(req, list);
476 res++;
477 if (res == INT_MAX)
478 goto out;
479 }
480 }
481 /* for latency reduction */
482 cond_resched_lock(&nfsi->vfs_inode.i_lock);
483 }
484out:
485 return res;
486}
487
488int __init nfs_init_nfspagecache(void)
489{
490 nfs_page_cachep = kmem_cache_create("nfs_page",
491 sizeof(struct nfs_page),
492 0, SLAB_HWCACHE_ALIGN,
493 NULL);
494 if (nfs_page_cachep == NULL)
495 return -ENOMEM;
496
497 return 0;
498}
499
500void nfs_destroy_nfspagecache(void)
501{
502 kmem_cache_destroy(nfs_page_cachep);
503}
504
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/fs/nfs/pagelist.c
4 *
5 * A set of helper functions for managing NFS read and write requests.
6 * The main purpose of these routines is to provide support for the
7 * coalescing of several requests into a single RPC call.
8 *
9 * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
10 *
11 */
12
13#include <linux/slab.h>
14#include <linux/file.h>
15#include <linux/sched.h>
16#include <linux/sunrpc/clnt.h>
17#include <linux/nfs.h>
18#include <linux/nfs3.h>
19#include <linux/nfs4.h>
20#include <linux/nfs_fs.h>
21#include <linux/nfs_page.h>
22#include <linux/nfs_mount.h>
23#include <linux/export.h>
24#include <linux/filelock.h>
25
26#include "internal.h"
27#include "pnfs.h"
28#include "nfstrace.h"
29#include "fscache.h"
30
31#define NFSDBG_FACILITY NFSDBG_PAGECACHE
32
33static struct kmem_cache *nfs_page_cachep;
34static const struct rpc_call_ops nfs_pgio_common_ops;
35
36struct nfs_page_iter_page {
37 const struct nfs_page *req;
38 size_t count;
39};
40
41static void nfs_page_iter_page_init(struct nfs_page_iter_page *i,
42 const struct nfs_page *req)
43{
44 i->req = req;
45 i->count = 0;
46}
47
48static void nfs_page_iter_page_advance(struct nfs_page_iter_page *i, size_t sz)
49{
50 const struct nfs_page *req = i->req;
51 size_t tmp = i->count + sz;
52
53 i->count = (tmp < req->wb_bytes) ? tmp : req->wb_bytes;
54}
55
56static struct page *nfs_page_iter_page_get(struct nfs_page_iter_page *i)
57{
58 const struct nfs_page *req = i->req;
59 struct page *page;
60
61 if (i->count != req->wb_bytes) {
62 size_t base = i->count + req->wb_pgbase;
63 size_t len = PAGE_SIZE - offset_in_page(base);
64
65 page = nfs_page_to_page(req, base);
66 nfs_page_iter_page_advance(i, len);
67 return page;
68 }
69 return NULL;
70}
71
72static struct nfs_pgio_mirror *
73nfs_pgio_get_mirror(struct nfs_pageio_descriptor *desc, u32 idx)
74{
75 if (desc->pg_ops->pg_get_mirror)
76 return desc->pg_ops->pg_get_mirror(desc, idx);
77 return &desc->pg_mirrors[0];
78}
79
80struct nfs_pgio_mirror *
81nfs_pgio_current_mirror(struct nfs_pageio_descriptor *desc)
82{
83 return nfs_pgio_get_mirror(desc, desc->pg_mirror_idx);
84}
85EXPORT_SYMBOL_GPL(nfs_pgio_current_mirror);
86
87static u32
88nfs_pgio_set_current_mirror(struct nfs_pageio_descriptor *desc, u32 idx)
89{
90 if (desc->pg_ops->pg_set_mirror)
91 return desc->pg_ops->pg_set_mirror(desc, idx);
92 return desc->pg_mirror_idx;
93}
94
95void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
96 struct nfs_pgio_header *hdr,
97 void (*release)(struct nfs_pgio_header *hdr))
98{
99 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
100
101
102 hdr->req = nfs_list_entry(mirror->pg_list.next);
103 hdr->inode = desc->pg_inode;
104 hdr->cred = nfs_req_openctx(hdr->req)->cred;
105 hdr->io_start = req_offset(hdr->req);
106 hdr->good_bytes = mirror->pg_count;
107 hdr->io_completion = desc->pg_io_completion;
108 hdr->dreq = desc->pg_dreq;
109 nfs_netfs_set_pgio_header(hdr, desc);
110 hdr->release = release;
111 hdr->completion_ops = desc->pg_completion_ops;
112 if (hdr->completion_ops->init_hdr)
113 hdr->completion_ops->init_hdr(hdr);
114
115 hdr->pgio_mirror_idx = desc->pg_mirror_idx;
116}
117EXPORT_SYMBOL_GPL(nfs_pgheader_init);
118
119void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
120{
121 unsigned int new = pos - hdr->io_start;
122
123 trace_nfs_pgio_error(hdr, error, pos);
124 if (hdr->good_bytes > new) {
125 hdr->good_bytes = new;
126 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
127 if (!test_and_set_bit(NFS_IOHDR_ERROR, &hdr->flags))
128 hdr->error = error;
129 }
130}
131
132static inline struct nfs_page *nfs_page_alloc(void)
133{
134 struct nfs_page *p =
135 kmem_cache_zalloc(nfs_page_cachep, nfs_io_gfp_mask());
136 if (p)
137 INIT_LIST_HEAD(&p->wb_list);
138 return p;
139}
140
141static inline void
142nfs_page_free(struct nfs_page *p)
143{
144 kmem_cache_free(nfs_page_cachep, p);
145}
146
147/**
148 * nfs_iocounter_wait - wait for i/o to complete
149 * @l_ctx: nfs_lock_context with io_counter to use
150 *
151 * returns -ERESTARTSYS if interrupted by a fatal signal.
152 * Otherwise returns 0 once the io_count hits 0.
153 */
154int
155nfs_iocounter_wait(struct nfs_lock_context *l_ctx)
156{
157 return wait_var_event_killable(&l_ctx->io_count,
158 !atomic_read(&l_ctx->io_count));
159}
160
161/**
162 * nfs_async_iocounter_wait - wait on a rpc_waitqueue for I/O
163 * to complete
164 * @task: the rpc_task that should wait
165 * @l_ctx: nfs_lock_context with io_counter to check
166 *
167 * Returns true if there is outstanding I/O to wait on and the
168 * task has been put to sleep.
169 */
170bool
171nfs_async_iocounter_wait(struct rpc_task *task, struct nfs_lock_context *l_ctx)
172{
173 struct inode *inode = d_inode(l_ctx->open_context->dentry);
174 bool ret = false;
175
176 if (atomic_read(&l_ctx->io_count) > 0) {
177 rpc_sleep_on(&NFS_SERVER(inode)->uoc_rpcwaitq, task, NULL);
178 ret = true;
179 }
180
181 if (atomic_read(&l_ctx->io_count) == 0) {
182 rpc_wake_up_queued_task(&NFS_SERVER(inode)->uoc_rpcwaitq, task);
183 ret = false;
184 }
185
186 return ret;
187}
188EXPORT_SYMBOL_GPL(nfs_async_iocounter_wait);
189
190/*
191 * nfs_page_lock_head_request - page lock the head of the page group
192 * @req: any member of the page group
193 */
194struct nfs_page *
195nfs_page_group_lock_head(struct nfs_page *req)
196{
197 struct nfs_page *head = req->wb_head;
198
199 while (!nfs_lock_request(head)) {
200 int ret = nfs_wait_on_request(head);
201 if (ret < 0)
202 return ERR_PTR(ret);
203 }
204 if (head != req)
205 kref_get(&head->wb_kref);
206 return head;
207}
208
209/*
210 * nfs_unroll_locks - unlock all newly locked reqs and wait on @req
211 * @head: head request of page group, must be holding head lock
212 * @req: request that couldn't lock and needs to wait on the req bit lock
213 *
214 * This is a helper function for nfs_lock_and_join_requests
215 * returns 0 on success, < 0 on error.
216 */
217static void
218nfs_unroll_locks(struct nfs_page *head, struct nfs_page *req)
219{
220 struct nfs_page *tmp;
221
222 /* relinquish all the locks successfully grabbed this run */
223 for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
224 if (!kref_read(&tmp->wb_kref))
225 continue;
226 nfs_unlock_and_release_request(tmp);
227 }
228}
229
230/*
231 * nfs_page_group_lock_subreq - try to lock a subrequest
232 * @head: head request of page group
233 * @subreq: request to lock
234 *
235 * This is a helper function for nfs_lock_and_join_requests which
236 * must be called with the head request and page group both locked.
237 * On error, it returns with the page group unlocked.
238 */
239static int
240nfs_page_group_lock_subreq(struct nfs_page *head, struct nfs_page *subreq)
241{
242 int ret;
243
244 if (!kref_get_unless_zero(&subreq->wb_kref))
245 return 0;
246 while (!nfs_lock_request(subreq)) {
247 nfs_page_group_unlock(head);
248 ret = nfs_wait_on_request(subreq);
249 if (!ret)
250 ret = nfs_page_group_lock(head);
251 if (ret < 0) {
252 nfs_unroll_locks(head, subreq);
253 nfs_release_request(subreq);
254 return ret;
255 }
256 }
257 return 0;
258}
259
260/*
261 * nfs_page_group_lock_subrequests - try to lock the subrequests
262 * @head: head request of page group
263 *
264 * This is a helper function for nfs_lock_and_join_requests which
265 * must be called with the head request locked.
266 */
267int nfs_page_group_lock_subrequests(struct nfs_page *head)
268{
269 struct nfs_page *subreq;
270 int ret;
271
272 ret = nfs_page_group_lock(head);
273 if (ret < 0)
274 return ret;
275 /* lock each request in the page group */
276 for (subreq = head->wb_this_page; subreq != head;
277 subreq = subreq->wb_this_page) {
278 ret = nfs_page_group_lock_subreq(head, subreq);
279 if (ret < 0)
280 return ret;
281 }
282 nfs_page_group_unlock(head);
283 return 0;
284}
285
286/*
287 * nfs_page_set_headlock - set the request PG_HEADLOCK
288 * @req: request that is to be locked
289 *
290 * this lock must be held when modifying req->wb_head
291 *
292 * return 0 on success, < 0 on error
293 */
294int
295nfs_page_set_headlock(struct nfs_page *req)
296{
297 if (!test_and_set_bit(PG_HEADLOCK, &req->wb_flags))
298 return 0;
299
300 set_bit(PG_CONTENDED1, &req->wb_flags);
301 smp_mb__after_atomic();
302 return wait_on_bit_lock(&req->wb_flags, PG_HEADLOCK,
303 TASK_UNINTERRUPTIBLE);
304}
305
306/*
307 * nfs_page_clear_headlock - clear the request PG_HEADLOCK
308 * @req: request that is to be locked
309 */
310void
311nfs_page_clear_headlock(struct nfs_page *req)
312{
313 clear_bit_unlock(PG_HEADLOCK, &req->wb_flags);
314 smp_mb__after_atomic();
315 if (!test_bit(PG_CONTENDED1, &req->wb_flags))
316 return;
317 wake_up_bit(&req->wb_flags, PG_HEADLOCK);
318}
319
320/*
321 * nfs_page_group_lock - lock the head of the page group
322 * @req: request in group that is to be locked
323 *
324 * this lock must be held when traversing or modifying the page
325 * group list
326 *
327 * return 0 on success, < 0 on error
328 */
329int
330nfs_page_group_lock(struct nfs_page *req)
331{
332 int ret;
333
334 ret = nfs_page_set_headlock(req);
335 if (ret || req->wb_head == req)
336 return ret;
337 return nfs_page_set_headlock(req->wb_head);
338}
339
340/*
341 * nfs_page_group_unlock - unlock the head of the page group
342 * @req: request in group that is to be unlocked
343 */
344void
345nfs_page_group_unlock(struct nfs_page *req)
346{
347 if (req != req->wb_head)
348 nfs_page_clear_headlock(req->wb_head);
349 nfs_page_clear_headlock(req);
350}
351
352/*
353 * nfs_page_group_sync_on_bit_locked
354 *
355 * must be called with page group lock held
356 */
357static bool
358nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
359{
360 struct nfs_page *head = req->wb_head;
361 struct nfs_page *tmp;
362
363 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags));
364 WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags));
365
366 tmp = req->wb_this_page;
367 while (tmp != req) {
368 if (!test_bit(bit, &tmp->wb_flags))
369 return false;
370 tmp = tmp->wb_this_page;
371 }
372
373 /* true! reset all bits */
374 tmp = req;
375 do {
376 clear_bit(bit, &tmp->wb_flags);
377 tmp = tmp->wb_this_page;
378 } while (tmp != req);
379
380 return true;
381}
382
383/*
384 * nfs_page_group_sync_on_bit - set bit on current request, but only
385 * return true if the bit is set for all requests in page group
386 * @req - request in page group
387 * @bit - PG_* bit that is used to sync page group
388 */
389bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit)
390{
391 bool ret;
392
393 nfs_page_group_lock(req);
394 ret = nfs_page_group_sync_on_bit_locked(req, bit);
395 nfs_page_group_unlock(req);
396
397 return ret;
398}
399
400/*
401 * nfs_page_group_init - Initialize the page group linkage for @req
402 * @req - a new nfs request
403 * @prev - the previous request in page group, or NULL if @req is the first
404 * or only request in the group (the head).
405 */
406static inline void
407nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev)
408{
409 struct inode *inode;
410 WARN_ON_ONCE(prev == req);
411
412 if (!prev) {
413 /* a head request */
414 req->wb_head = req;
415 req->wb_this_page = req;
416 } else {
417 /* a subrequest */
418 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head);
419 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags));
420 req->wb_head = prev->wb_head;
421 req->wb_this_page = prev->wb_this_page;
422 prev->wb_this_page = req;
423
424 /* All subrequests take a ref on the head request until
425 * nfs_page_group_destroy is called */
426 kref_get(&req->wb_head->wb_kref);
427
428 /* grab extra ref and bump the request count if head request
429 * has extra ref from the write/commit path to handle handoff
430 * between write and commit lists. */
431 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) {
432 inode = nfs_page_to_inode(req);
433 set_bit(PG_INODE_REF, &req->wb_flags);
434 kref_get(&req->wb_kref);
435 atomic_long_inc(&NFS_I(inode)->nrequests);
436 }
437 }
438}
439
440/*
441 * nfs_page_group_destroy - sync the destruction of page groups
442 * @req - request that no longer needs the page group
443 *
444 * releases the page group reference from each member once all
445 * members have called this function.
446 */
447static void
448nfs_page_group_destroy(struct kref *kref)
449{
450 struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
451 struct nfs_page *head = req->wb_head;
452 struct nfs_page *tmp, *next;
453
454 if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN))
455 goto out;
456
457 tmp = req;
458 do {
459 next = tmp->wb_this_page;
460 /* unlink and free */
461 tmp->wb_this_page = tmp;
462 tmp->wb_head = tmp;
463 nfs_free_request(tmp);
464 tmp = next;
465 } while (tmp != req);
466out:
467 /* subrequests must release the ref on the head request */
468 if (head != req)
469 nfs_release_request(head);
470}
471
472static struct nfs_page *nfs_page_create(struct nfs_lock_context *l_ctx,
473 unsigned int pgbase, pgoff_t index,
474 unsigned int offset, unsigned int count)
475{
476 struct nfs_page *req;
477 struct nfs_open_context *ctx = l_ctx->open_context;
478
479 if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
480 return ERR_PTR(-EBADF);
481 /* try to allocate the request struct */
482 req = nfs_page_alloc();
483 if (req == NULL)
484 return ERR_PTR(-ENOMEM);
485
486 req->wb_lock_context = l_ctx;
487 refcount_inc(&l_ctx->count);
488 atomic_inc(&l_ctx->io_count);
489
490 /* Initialize the request struct. Initially, we assume a
491 * long write-back delay. This will be adjusted in
492 * update_nfs_request below if the region is not locked. */
493 req->wb_pgbase = pgbase;
494 req->wb_index = index;
495 req->wb_offset = offset;
496 req->wb_bytes = count;
497 kref_init(&req->wb_kref);
498 req->wb_nio = 0;
499 return req;
500}
501
502static void nfs_page_assign_folio(struct nfs_page *req, struct folio *folio)
503{
504 if (folio != NULL) {
505 req->wb_folio = folio;
506 folio_get(folio);
507 set_bit(PG_FOLIO, &req->wb_flags);
508 }
509}
510
511static void nfs_page_assign_page(struct nfs_page *req, struct page *page)
512{
513 if (page != NULL) {
514 req->wb_page = page;
515 get_page(page);
516 }
517}
518
519/**
520 * nfs_page_create_from_page - Create an NFS read/write request.
521 * @ctx: open context to use
522 * @page: page to write
523 * @pgbase: starting offset within the page for the write
524 * @offset: file offset for the write
525 * @count: number of bytes to read/write
526 *
527 * The page must be locked by the caller. This makes sure we never
528 * create two different requests for the same page.
529 * User should ensure it is safe to sleep in this function.
530 */
531struct nfs_page *nfs_page_create_from_page(struct nfs_open_context *ctx,
532 struct page *page,
533 unsigned int pgbase, loff_t offset,
534 unsigned int count)
535{
536 struct nfs_lock_context *l_ctx = nfs_get_lock_context(ctx);
537 struct nfs_page *ret;
538
539 if (IS_ERR(l_ctx))
540 return ERR_CAST(l_ctx);
541 ret = nfs_page_create(l_ctx, pgbase, offset >> PAGE_SHIFT,
542 offset_in_page(offset), count);
543 if (!IS_ERR(ret)) {
544 nfs_page_assign_page(ret, page);
545 nfs_page_group_init(ret, NULL);
546 }
547 nfs_put_lock_context(l_ctx);
548 return ret;
549}
550
551/**
552 * nfs_page_create_from_folio - Create an NFS read/write request.
553 * @ctx: open context to use
554 * @folio: folio to write
555 * @offset: starting offset within the folio for the write
556 * @count: number of bytes to read/write
557 *
558 * The page must be locked by the caller. This makes sure we never
559 * create two different requests for the same page.
560 * User should ensure it is safe to sleep in this function.
561 */
562struct nfs_page *nfs_page_create_from_folio(struct nfs_open_context *ctx,
563 struct folio *folio,
564 unsigned int offset,
565 unsigned int count)
566{
567 struct nfs_lock_context *l_ctx = nfs_get_lock_context(ctx);
568 struct nfs_page *ret;
569
570 if (IS_ERR(l_ctx))
571 return ERR_CAST(l_ctx);
572 ret = nfs_page_create(l_ctx, offset, folio_index(folio), offset, count);
573 if (!IS_ERR(ret)) {
574 nfs_page_assign_folio(ret, folio);
575 nfs_page_group_init(ret, NULL);
576 }
577 nfs_put_lock_context(l_ctx);
578 return ret;
579}
580
581static struct nfs_page *
582nfs_create_subreq(struct nfs_page *req,
583 unsigned int pgbase,
584 unsigned int offset,
585 unsigned int count)
586{
587 struct nfs_page *last;
588 struct nfs_page *ret;
589 struct folio *folio = nfs_page_to_folio(req);
590 struct page *page = nfs_page_to_page(req, pgbase);
591
592 ret = nfs_page_create(req->wb_lock_context, pgbase, req->wb_index,
593 offset, count);
594 if (!IS_ERR(ret)) {
595 if (folio)
596 nfs_page_assign_folio(ret, folio);
597 else
598 nfs_page_assign_page(ret, page);
599 /* find the last request */
600 for (last = req->wb_head;
601 last->wb_this_page != req->wb_head;
602 last = last->wb_this_page)
603 ;
604
605 nfs_lock_request(ret);
606 nfs_page_group_init(ret, last);
607 ret->wb_nio = req->wb_nio;
608 }
609 return ret;
610}
611
612/**
613 * nfs_unlock_request - Unlock request and wake up sleepers.
614 * @req: pointer to request
615 */
616void nfs_unlock_request(struct nfs_page *req)
617{
618 clear_bit_unlock(PG_BUSY, &req->wb_flags);
619 smp_mb__after_atomic();
620 if (!test_bit(PG_CONTENDED2, &req->wb_flags))
621 return;
622 wake_up_bit(&req->wb_flags, PG_BUSY);
623}
624
625/**
626 * nfs_unlock_and_release_request - Unlock request and release the nfs_page
627 * @req: pointer to request
628 */
629void nfs_unlock_and_release_request(struct nfs_page *req)
630{
631 nfs_unlock_request(req);
632 nfs_release_request(req);
633}
634
635/*
636 * nfs_clear_request - Free up all resources allocated to the request
637 * @req:
638 *
639 * Release page and open context resources associated with a read/write
640 * request after it has completed.
641 */
642static void nfs_clear_request(struct nfs_page *req)
643{
644 struct folio *folio = nfs_page_to_folio(req);
645 struct page *page = req->wb_page;
646 struct nfs_lock_context *l_ctx = req->wb_lock_context;
647 struct nfs_open_context *ctx;
648
649 if (folio != NULL) {
650 folio_put(folio);
651 req->wb_folio = NULL;
652 clear_bit(PG_FOLIO, &req->wb_flags);
653 } else if (page != NULL) {
654 put_page(page);
655 req->wb_page = NULL;
656 }
657 if (l_ctx != NULL) {
658 if (atomic_dec_and_test(&l_ctx->io_count)) {
659 wake_up_var(&l_ctx->io_count);
660 ctx = l_ctx->open_context;
661 if (test_bit(NFS_CONTEXT_UNLOCK, &ctx->flags))
662 rpc_wake_up(&NFS_SERVER(d_inode(ctx->dentry))->uoc_rpcwaitq);
663 }
664 nfs_put_lock_context(l_ctx);
665 req->wb_lock_context = NULL;
666 }
667}
668
669/**
670 * nfs_free_request - Release the count on an NFS read/write request
671 * @req: request to release
672 *
673 * Note: Should never be called with the spinlock held!
674 */
675void nfs_free_request(struct nfs_page *req)
676{
677 WARN_ON_ONCE(req->wb_this_page != req);
678
679 /* extra debug: make sure no sync bits are still set */
680 WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
681 WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
682 WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
683 WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
684 WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
685
686 /* Release struct file and open context */
687 nfs_clear_request(req);
688 nfs_page_free(req);
689}
690
691void nfs_release_request(struct nfs_page *req)
692{
693 kref_put(&req->wb_kref, nfs_page_group_destroy);
694}
695EXPORT_SYMBOL_GPL(nfs_release_request);
696
697/**
698 * nfs_wait_on_request - Wait for a request to complete.
699 * @req: request to wait upon.
700 *
701 * Interruptible by fatal signals only.
702 * The user is responsible for holding a count on the request.
703 */
704int
705nfs_wait_on_request(struct nfs_page *req)
706{
707 if (!test_bit(PG_BUSY, &req->wb_flags))
708 return 0;
709 set_bit(PG_CONTENDED2, &req->wb_flags);
710 smp_mb__after_atomic();
711 return wait_on_bit_io(&req->wb_flags, PG_BUSY,
712 TASK_UNINTERRUPTIBLE);
713}
714EXPORT_SYMBOL_GPL(nfs_wait_on_request);
715
716/*
717 * nfs_generic_pg_test - determine if requests can be coalesced
718 * @desc: pointer to descriptor
719 * @prev: previous request in desc, or NULL
720 * @req: this request
721 *
722 * Returns zero if @req cannot be coalesced into @desc, otherwise it returns
723 * the size of the request.
724 */
725size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc,
726 struct nfs_page *prev, struct nfs_page *req)
727{
728 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
729
730
731 if (mirror->pg_count > mirror->pg_bsize) {
732 /* should never happen */
733 WARN_ON_ONCE(1);
734 return 0;
735 }
736
737 /*
738 * Limit the request size so that we can still allocate a page array
739 * for it without upsetting the slab allocator.
740 */
741 if (((mirror->pg_count + req->wb_bytes) >> PAGE_SHIFT) *
742 sizeof(struct page *) > PAGE_SIZE)
743 return 0;
744
745 return min(mirror->pg_bsize - mirror->pg_count, (size_t)req->wb_bytes);
746}
747EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
748
749struct nfs_pgio_header *nfs_pgio_header_alloc(const struct nfs_rw_ops *ops)
750{
751 struct nfs_pgio_header *hdr = ops->rw_alloc_header();
752
753 if (hdr) {
754 INIT_LIST_HEAD(&hdr->pages);
755 hdr->rw_ops = ops;
756 }
757 return hdr;
758}
759EXPORT_SYMBOL_GPL(nfs_pgio_header_alloc);
760
761/**
762 * nfs_pgio_data_destroy - make @hdr suitable for reuse
763 *
764 * Frees memory and releases refs from nfs_generic_pgio, so that it may
765 * be called again.
766 *
767 * @hdr: A header that has had nfs_generic_pgio called
768 */
769static void nfs_pgio_data_destroy(struct nfs_pgio_header *hdr)
770{
771 if (hdr->args.context)
772 put_nfs_open_context(hdr->args.context);
773 if (hdr->page_array.pagevec != hdr->page_array.page_array)
774 kfree(hdr->page_array.pagevec);
775}
776
777/*
778 * nfs_pgio_header_free - Free a read or write header
779 * @hdr: The header to free
780 */
781void nfs_pgio_header_free(struct nfs_pgio_header *hdr)
782{
783 nfs_pgio_data_destroy(hdr);
784 hdr->rw_ops->rw_free_header(hdr);
785}
786EXPORT_SYMBOL_GPL(nfs_pgio_header_free);
787
788/**
789 * nfs_pgio_rpcsetup - Set up arguments for a pageio call
790 * @hdr: The pageio hdr
791 * @pgbase: base
792 * @count: Number of bytes to read
793 * @how: How to commit data (writes only)
794 * @cinfo: Commit information for the call (writes only)
795 */
796static void nfs_pgio_rpcsetup(struct nfs_pgio_header *hdr, unsigned int pgbase,
797 unsigned int count, int how,
798 struct nfs_commit_info *cinfo)
799{
800 struct nfs_page *req = hdr->req;
801
802 /* Set up the RPC argument and reply structs
803 * NB: take care not to mess about with hdr->commit et al. */
804
805 hdr->args.fh = NFS_FH(hdr->inode);
806 hdr->args.offset = req_offset(req);
807 /* pnfs_set_layoutcommit needs this */
808 hdr->mds_offset = hdr->args.offset;
809 hdr->args.pgbase = pgbase;
810 hdr->args.pages = hdr->page_array.pagevec;
811 hdr->args.count = count;
812 hdr->args.context = get_nfs_open_context(nfs_req_openctx(req));
813 hdr->args.lock_context = req->wb_lock_context;
814 hdr->args.stable = NFS_UNSTABLE;
815 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
816 case 0:
817 break;
818 case FLUSH_COND_STABLE:
819 if (nfs_reqs_to_commit(cinfo))
820 break;
821 fallthrough;
822 default:
823 hdr->args.stable = NFS_FILE_SYNC;
824 }
825
826 hdr->res.fattr = &hdr->fattr;
827 hdr->res.count = 0;
828 hdr->res.eof = 0;
829 hdr->res.verf = &hdr->verf;
830 nfs_fattr_init(&hdr->fattr);
831}
832
833/**
834 * nfs_pgio_prepare - Prepare pageio hdr to go over the wire
835 * @task: The current task
836 * @calldata: pageio header to prepare
837 */
838static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
839{
840 struct nfs_pgio_header *hdr = calldata;
841 int err;
842 err = NFS_PROTO(hdr->inode)->pgio_rpc_prepare(task, hdr);
843 if (err)
844 rpc_exit(task, err);
845}
846
847int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_header *hdr,
848 const struct cred *cred, const struct nfs_rpc_ops *rpc_ops,
849 const struct rpc_call_ops *call_ops, int how, int flags)
850{
851 struct rpc_task *task;
852 struct rpc_message msg = {
853 .rpc_argp = &hdr->args,
854 .rpc_resp = &hdr->res,
855 .rpc_cred = cred,
856 };
857 struct rpc_task_setup task_setup_data = {
858 .rpc_client = clnt,
859 .task = &hdr->task,
860 .rpc_message = &msg,
861 .callback_ops = call_ops,
862 .callback_data = hdr,
863 .workqueue = nfsiod_workqueue,
864 .flags = RPC_TASK_ASYNC | flags,
865 };
866
867 if (nfs_server_capable(hdr->inode, NFS_CAP_MOVEABLE))
868 task_setup_data.flags |= RPC_TASK_MOVEABLE;
869
870 hdr->rw_ops->rw_initiate(hdr, &msg, rpc_ops, &task_setup_data, how);
871
872 dprintk("NFS: initiated pgio call "
873 "(req %s/%llu, %u bytes @ offset %llu)\n",
874 hdr->inode->i_sb->s_id,
875 (unsigned long long)NFS_FILEID(hdr->inode),
876 hdr->args.count,
877 (unsigned long long)hdr->args.offset);
878
879 task = rpc_run_task(&task_setup_data);
880 if (IS_ERR(task))
881 return PTR_ERR(task);
882 rpc_put_task(task);
883 return 0;
884}
885EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
886
887/**
888 * nfs_pgio_error - Clean up from a pageio error
889 * @hdr: pageio header
890 */
891static void nfs_pgio_error(struct nfs_pgio_header *hdr)
892{
893 set_bit(NFS_IOHDR_REDO, &hdr->flags);
894 hdr->completion_ops->completion(hdr);
895}
896
897/**
898 * nfs_pgio_release - Release pageio data
899 * @calldata: The pageio header to release
900 */
901static void nfs_pgio_release(void *calldata)
902{
903 struct nfs_pgio_header *hdr = calldata;
904 hdr->completion_ops->completion(hdr);
905}
906
907static void nfs_pageio_mirror_init(struct nfs_pgio_mirror *mirror,
908 unsigned int bsize)
909{
910 INIT_LIST_HEAD(&mirror->pg_list);
911 mirror->pg_bytes_written = 0;
912 mirror->pg_count = 0;
913 mirror->pg_bsize = bsize;
914 mirror->pg_base = 0;
915 mirror->pg_recoalesce = 0;
916}
917
918/**
919 * nfs_pageio_init - initialise a page io descriptor
920 * @desc: pointer to descriptor
921 * @inode: pointer to inode
922 * @pg_ops: pointer to pageio operations
923 * @compl_ops: pointer to pageio completion operations
924 * @rw_ops: pointer to nfs read/write operations
925 * @bsize: io block size
926 * @io_flags: extra parameters for the io function
927 */
928void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
929 struct inode *inode,
930 const struct nfs_pageio_ops *pg_ops,
931 const struct nfs_pgio_completion_ops *compl_ops,
932 const struct nfs_rw_ops *rw_ops,
933 size_t bsize,
934 int io_flags)
935{
936 desc->pg_moreio = 0;
937 desc->pg_inode = inode;
938 desc->pg_ops = pg_ops;
939 desc->pg_completion_ops = compl_ops;
940 desc->pg_rw_ops = rw_ops;
941 desc->pg_ioflags = io_flags;
942 desc->pg_error = 0;
943 desc->pg_lseg = NULL;
944 desc->pg_io_completion = NULL;
945 desc->pg_dreq = NULL;
946 nfs_netfs_reset_pageio_descriptor(desc);
947 desc->pg_bsize = bsize;
948
949 desc->pg_mirror_count = 1;
950 desc->pg_mirror_idx = 0;
951
952 desc->pg_mirrors_dynamic = NULL;
953 desc->pg_mirrors = desc->pg_mirrors_static;
954 nfs_pageio_mirror_init(&desc->pg_mirrors[0], bsize);
955 desc->pg_maxretrans = 0;
956}
957
958/**
959 * nfs_pgio_result - Basic pageio error handling
960 * @task: The task that ran
961 * @calldata: Pageio header to check
962 */
963static void nfs_pgio_result(struct rpc_task *task, void *calldata)
964{
965 struct nfs_pgio_header *hdr = calldata;
966 struct inode *inode = hdr->inode;
967
968 if (hdr->rw_ops->rw_done(task, hdr, inode) != 0)
969 return;
970 if (task->tk_status < 0)
971 nfs_set_pgio_error(hdr, task->tk_status, hdr->args.offset);
972 else
973 hdr->rw_ops->rw_result(task, hdr);
974}
975
976/*
977 * Create an RPC task for the given read or write request and kick it.
978 * The page must have been locked by the caller.
979 *
980 * It may happen that the page we're passed is not marked dirty.
981 * This is the case if nfs_updatepage detects a conflicting request
982 * that has been written but not committed.
983 */
984int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
985 struct nfs_pgio_header *hdr)
986{
987 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
988
989 struct nfs_page *req;
990 struct page **pages,
991 *last_page;
992 struct list_head *head = &mirror->pg_list;
993 struct nfs_commit_info cinfo;
994 struct nfs_page_array *pg_array = &hdr->page_array;
995 unsigned int pagecount, pageused;
996 unsigned int pg_base = offset_in_page(mirror->pg_base);
997 gfp_t gfp_flags = nfs_io_gfp_mask();
998
999 pagecount = nfs_page_array_len(pg_base, mirror->pg_count);
1000 pg_array->npages = pagecount;
1001
1002 if (pagecount <= ARRAY_SIZE(pg_array->page_array))
1003 pg_array->pagevec = pg_array->page_array;
1004 else {
1005 pg_array->pagevec = kcalloc(pagecount, sizeof(struct page *), gfp_flags);
1006 if (!pg_array->pagevec) {
1007 pg_array->npages = 0;
1008 nfs_pgio_error(hdr);
1009 desc->pg_error = -ENOMEM;
1010 return desc->pg_error;
1011 }
1012 }
1013
1014 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1015 pages = hdr->page_array.pagevec;
1016 last_page = NULL;
1017 pageused = 0;
1018 while (!list_empty(head)) {
1019 struct nfs_page_iter_page i;
1020 struct page *page;
1021
1022 req = nfs_list_entry(head->next);
1023 nfs_list_move_request(req, &hdr->pages);
1024
1025 if (req->wb_pgbase == 0)
1026 last_page = NULL;
1027
1028 nfs_page_iter_page_init(&i, req);
1029 while ((page = nfs_page_iter_page_get(&i)) != NULL) {
1030 if (last_page != page) {
1031 pageused++;
1032 if (pageused > pagecount)
1033 goto full;
1034 *pages++ = last_page = page;
1035 }
1036 }
1037 }
1038full:
1039 if (WARN_ON_ONCE(pageused != pagecount)) {
1040 nfs_pgio_error(hdr);
1041 desc->pg_error = -EINVAL;
1042 return desc->pg_error;
1043 }
1044
1045 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1046 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1047 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1048
1049 /* Set up the argument struct */
1050 nfs_pgio_rpcsetup(hdr, pg_base, mirror->pg_count, desc->pg_ioflags,
1051 &cinfo);
1052 desc->pg_rpc_callops = &nfs_pgio_common_ops;
1053 return 0;
1054}
1055EXPORT_SYMBOL_GPL(nfs_generic_pgio);
1056
1057static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
1058{
1059 struct nfs_pgio_header *hdr;
1060 int ret;
1061 unsigned short task_flags = 0;
1062
1063 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1064 if (!hdr) {
1065 desc->pg_error = -ENOMEM;
1066 return desc->pg_error;
1067 }
1068 nfs_pgheader_init(desc, hdr, nfs_pgio_header_free);
1069 ret = nfs_generic_pgio(desc, hdr);
1070 if (ret == 0) {
1071 if (NFS_SERVER(hdr->inode)->nfs_client->cl_minorversion)
1072 task_flags = RPC_TASK_MOVEABLE;
1073 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
1074 hdr,
1075 hdr->cred,
1076 NFS_PROTO(hdr->inode),
1077 desc->pg_rpc_callops,
1078 desc->pg_ioflags,
1079 RPC_TASK_CRED_NOREF | task_flags);
1080 }
1081 return ret;
1082}
1083
1084static struct nfs_pgio_mirror *
1085nfs_pageio_alloc_mirrors(struct nfs_pageio_descriptor *desc,
1086 unsigned int mirror_count)
1087{
1088 struct nfs_pgio_mirror *ret;
1089 unsigned int i;
1090
1091 kfree(desc->pg_mirrors_dynamic);
1092 desc->pg_mirrors_dynamic = NULL;
1093 if (mirror_count == 1)
1094 return desc->pg_mirrors_static;
1095 ret = kmalloc_array(mirror_count, sizeof(*ret), nfs_io_gfp_mask());
1096 if (ret != NULL) {
1097 for (i = 0; i < mirror_count; i++)
1098 nfs_pageio_mirror_init(&ret[i], desc->pg_bsize);
1099 desc->pg_mirrors_dynamic = ret;
1100 }
1101 return ret;
1102}
1103
1104/*
1105 * nfs_pageio_setup_mirroring - determine if mirroring is to be used
1106 * by calling the pg_get_mirror_count op
1107 */
1108static void nfs_pageio_setup_mirroring(struct nfs_pageio_descriptor *pgio,
1109 struct nfs_page *req)
1110{
1111 unsigned int mirror_count = 1;
1112
1113 if (pgio->pg_ops->pg_get_mirror_count)
1114 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
1115 if (mirror_count == pgio->pg_mirror_count || pgio->pg_error < 0)
1116 return;
1117
1118 if (!mirror_count || mirror_count > NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX) {
1119 pgio->pg_error = -EINVAL;
1120 return;
1121 }
1122
1123 pgio->pg_mirrors = nfs_pageio_alloc_mirrors(pgio, mirror_count);
1124 if (pgio->pg_mirrors == NULL) {
1125 pgio->pg_error = -ENOMEM;
1126 pgio->pg_mirrors = pgio->pg_mirrors_static;
1127 mirror_count = 1;
1128 }
1129 pgio->pg_mirror_count = mirror_count;
1130}
1131
1132static void nfs_pageio_cleanup_mirroring(struct nfs_pageio_descriptor *pgio)
1133{
1134 pgio->pg_mirror_count = 1;
1135 pgio->pg_mirror_idx = 0;
1136 pgio->pg_mirrors = pgio->pg_mirrors_static;
1137 kfree(pgio->pg_mirrors_dynamic);
1138 pgio->pg_mirrors_dynamic = NULL;
1139}
1140
1141static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
1142 const struct nfs_lock_context *l2)
1143{
1144 return l1->lockowner == l2->lockowner;
1145}
1146
1147static bool nfs_page_is_contiguous(const struct nfs_page *prev,
1148 const struct nfs_page *req)
1149{
1150 size_t prev_end = prev->wb_pgbase + prev->wb_bytes;
1151
1152 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
1153 return false;
1154 if (req->wb_pgbase == 0)
1155 return prev_end == nfs_page_max_length(prev);
1156 if (req->wb_pgbase == prev_end) {
1157 struct folio *folio = nfs_page_to_folio(req);
1158 if (folio)
1159 return folio == nfs_page_to_folio(prev);
1160 return req->wb_page == prev->wb_page;
1161 }
1162 return false;
1163}
1164
1165/**
1166 * nfs_coalesce_size - test two requests for compatibility
1167 * @prev: pointer to nfs_page
1168 * @req: pointer to nfs_page
1169 * @pgio: pointer to nfs_pagio_descriptor
1170 *
1171 * The nfs_page structures 'prev' and 'req' are compared to ensure that the
1172 * page data area they describe is contiguous, and that their RPC
1173 * credentials, NFSv4 open state, and lockowners are the same.
1174 *
1175 * Returns size of the request that can be coalesced
1176 */
1177static unsigned int nfs_coalesce_size(struct nfs_page *prev,
1178 struct nfs_page *req,
1179 struct nfs_pageio_descriptor *pgio)
1180{
1181 struct file_lock_context *flctx;
1182
1183 if (prev) {
1184 if (!nfs_match_open_context(nfs_req_openctx(req), nfs_req_openctx(prev)))
1185 return 0;
1186 flctx = locks_inode_context(d_inode(nfs_req_openctx(req)->dentry));
1187 if (flctx != NULL &&
1188 !(list_empty_careful(&flctx->flc_posix) &&
1189 list_empty_careful(&flctx->flc_flock)) &&
1190 !nfs_match_lock_context(req->wb_lock_context,
1191 prev->wb_lock_context))
1192 return 0;
1193 if (!nfs_page_is_contiguous(prev, req))
1194 return 0;
1195 }
1196 return pgio->pg_ops->pg_test(pgio, prev, req);
1197}
1198
1199/**
1200 * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
1201 * @desc: destination io descriptor
1202 * @req: request
1203 *
1204 * If the request 'req' was successfully coalesced into the existing list
1205 * of pages 'desc', it returns the size of req.
1206 */
1207static unsigned int
1208nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
1209 struct nfs_page *req)
1210{
1211 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1212 struct nfs_page *prev = NULL;
1213 unsigned int size;
1214
1215 if (list_empty(&mirror->pg_list)) {
1216 if (desc->pg_ops->pg_init)
1217 desc->pg_ops->pg_init(desc, req);
1218 if (desc->pg_error < 0)
1219 return 0;
1220 mirror->pg_base = req->wb_pgbase;
1221 mirror->pg_count = 0;
1222 mirror->pg_recoalesce = 0;
1223 } else
1224 prev = nfs_list_entry(mirror->pg_list.prev);
1225
1226 if (desc->pg_maxretrans && req->wb_nio > desc->pg_maxretrans) {
1227 if (NFS_SERVER(desc->pg_inode)->flags & NFS_MOUNT_SOFTERR)
1228 desc->pg_error = -ETIMEDOUT;
1229 else
1230 desc->pg_error = -EIO;
1231 return 0;
1232 }
1233
1234 size = nfs_coalesce_size(prev, req, desc);
1235 if (size < req->wb_bytes)
1236 return size;
1237 nfs_list_move_request(req, &mirror->pg_list);
1238 mirror->pg_count += req->wb_bytes;
1239 return req->wb_bytes;
1240}
1241
1242/*
1243 * Helper for nfs_pageio_add_request and nfs_pageio_complete
1244 */
1245static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
1246{
1247 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1248
1249 if (!list_empty(&mirror->pg_list)) {
1250 int error = desc->pg_ops->pg_doio(desc);
1251 if (error < 0)
1252 desc->pg_error = error;
1253 if (list_empty(&mirror->pg_list))
1254 mirror->pg_bytes_written += mirror->pg_count;
1255 }
1256}
1257
1258static void
1259nfs_pageio_cleanup_request(struct nfs_pageio_descriptor *desc,
1260 struct nfs_page *req)
1261{
1262 LIST_HEAD(head);
1263
1264 nfs_list_move_request(req, &head);
1265 desc->pg_completion_ops->error_cleanup(&head, desc->pg_error);
1266}
1267
1268/**
1269 * __nfs_pageio_add_request - Attempt to coalesce a request into a page list.
1270 * @desc: destination io descriptor
1271 * @req: request
1272 *
1273 * This may split a request into subrequests which are all part of the
1274 * same page group. If so, it will submit @req as the last one, to ensure
1275 * the pointer to @req is still valid in case of failure.
1276 *
1277 * Returns true if the request 'req' was successfully coalesced into the
1278 * existing list of pages 'desc'.
1279 */
1280static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1281 struct nfs_page *req)
1282{
1283 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1284 struct nfs_page *subreq;
1285 unsigned int size, subreq_size;
1286
1287 nfs_page_group_lock(req);
1288
1289 subreq = req;
1290 subreq_size = subreq->wb_bytes;
1291 for(;;) {
1292 size = nfs_pageio_do_add_request(desc, subreq);
1293 if (size == subreq_size) {
1294 /* We successfully submitted a request */
1295 if (subreq == req)
1296 break;
1297 req->wb_pgbase += size;
1298 req->wb_bytes -= size;
1299 req->wb_offset += size;
1300 subreq_size = req->wb_bytes;
1301 subreq = req;
1302 continue;
1303 }
1304 if (WARN_ON_ONCE(subreq != req)) {
1305 nfs_page_group_unlock(req);
1306 nfs_pageio_cleanup_request(desc, subreq);
1307 subreq = req;
1308 subreq_size = req->wb_bytes;
1309 nfs_page_group_lock(req);
1310 }
1311 if (!size) {
1312 /* Can't coalesce any more, so do I/O */
1313 nfs_page_group_unlock(req);
1314 desc->pg_moreio = 1;
1315 nfs_pageio_doio(desc);
1316 if (desc->pg_error < 0 || mirror->pg_recoalesce)
1317 return 0;
1318 /* retry add_request for this subreq */
1319 nfs_page_group_lock(req);
1320 continue;
1321 }
1322 subreq = nfs_create_subreq(req, req->wb_pgbase,
1323 req->wb_offset, size);
1324 if (IS_ERR(subreq))
1325 goto err_ptr;
1326 subreq_size = size;
1327 }
1328
1329 nfs_page_group_unlock(req);
1330 return 1;
1331err_ptr:
1332 desc->pg_error = PTR_ERR(subreq);
1333 nfs_page_group_unlock(req);
1334 return 0;
1335}
1336
1337static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
1338{
1339 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1340 LIST_HEAD(head);
1341
1342 do {
1343 list_splice_init(&mirror->pg_list, &head);
1344 mirror->pg_recoalesce = 0;
1345
1346 while (!list_empty(&head)) {
1347 struct nfs_page *req;
1348
1349 req = list_first_entry(&head, struct nfs_page, wb_list);
1350 if (__nfs_pageio_add_request(desc, req))
1351 continue;
1352 if (desc->pg_error < 0) {
1353 list_splice_tail(&head, &mirror->pg_list);
1354 mirror->pg_recoalesce = 1;
1355 return 0;
1356 }
1357 break;
1358 }
1359 } while (mirror->pg_recoalesce);
1360 return 1;
1361}
1362
1363static int nfs_pageio_add_request_mirror(struct nfs_pageio_descriptor *desc,
1364 struct nfs_page *req)
1365{
1366 int ret;
1367
1368 do {
1369 ret = __nfs_pageio_add_request(desc, req);
1370 if (ret)
1371 break;
1372 if (desc->pg_error < 0)
1373 break;
1374 ret = nfs_do_recoalesce(desc);
1375 } while (ret);
1376
1377 return ret;
1378}
1379
1380static void nfs_pageio_error_cleanup(struct nfs_pageio_descriptor *desc)
1381{
1382 u32 midx;
1383 struct nfs_pgio_mirror *mirror;
1384
1385 if (!desc->pg_error)
1386 return;
1387
1388 for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1389 mirror = nfs_pgio_get_mirror(desc, midx);
1390 desc->pg_completion_ops->error_cleanup(&mirror->pg_list,
1391 desc->pg_error);
1392 }
1393}
1394
1395int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1396 struct nfs_page *req)
1397{
1398 u32 midx;
1399 unsigned int pgbase, offset, bytes;
1400 struct nfs_page *dupreq;
1401
1402 pgbase = req->wb_pgbase;
1403 offset = req->wb_offset;
1404 bytes = req->wb_bytes;
1405
1406 nfs_pageio_setup_mirroring(desc, req);
1407 if (desc->pg_error < 0)
1408 goto out_failed;
1409
1410 /* Create the mirror instances first, and fire them off */
1411 for (midx = 1; midx < desc->pg_mirror_count; midx++) {
1412 nfs_page_group_lock(req);
1413
1414 dupreq = nfs_create_subreq(req,
1415 pgbase, offset, bytes);
1416
1417 nfs_page_group_unlock(req);
1418 if (IS_ERR(dupreq)) {
1419 desc->pg_error = PTR_ERR(dupreq);
1420 goto out_failed;
1421 }
1422
1423 nfs_pgio_set_current_mirror(desc, midx);
1424 if (!nfs_pageio_add_request_mirror(desc, dupreq))
1425 goto out_cleanup_subreq;
1426 }
1427
1428 nfs_pgio_set_current_mirror(desc, 0);
1429 if (!nfs_pageio_add_request_mirror(desc, req))
1430 goto out_failed;
1431
1432 return 1;
1433
1434out_cleanup_subreq:
1435 nfs_pageio_cleanup_request(desc, dupreq);
1436out_failed:
1437 nfs_pageio_error_cleanup(desc);
1438 return 0;
1439}
1440
1441/*
1442 * nfs_pageio_complete_mirror - Complete I/O on the current mirror of an
1443 * nfs_pageio_descriptor
1444 * @desc: pointer to io descriptor
1445 * @mirror_idx: pointer to mirror index
1446 */
1447static void nfs_pageio_complete_mirror(struct nfs_pageio_descriptor *desc,
1448 u32 mirror_idx)
1449{
1450 struct nfs_pgio_mirror *mirror;
1451 u32 restore_idx;
1452
1453 restore_idx = nfs_pgio_set_current_mirror(desc, mirror_idx);
1454 mirror = nfs_pgio_current_mirror(desc);
1455
1456 for (;;) {
1457 nfs_pageio_doio(desc);
1458 if (desc->pg_error < 0 || !mirror->pg_recoalesce)
1459 break;
1460 if (!nfs_do_recoalesce(desc))
1461 break;
1462 }
1463 nfs_pgio_set_current_mirror(desc, restore_idx);
1464}
1465
1466/*
1467 * nfs_pageio_resend - Transfer requests to new descriptor and resend
1468 * @hdr - the pgio header to move request from
1469 * @desc - the pageio descriptor to add requests to
1470 *
1471 * Try to move each request (nfs_page) from @hdr to @desc then attempt
1472 * to send them.
1473 *
1474 * Returns 0 on success and < 0 on error.
1475 */
1476int nfs_pageio_resend(struct nfs_pageio_descriptor *desc,
1477 struct nfs_pgio_header *hdr)
1478{
1479 LIST_HEAD(pages);
1480
1481 desc->pg_io_completion = hdr->io_completion;
1482 desc->pg_dreq = hdr->dreq;
1483 nfs_netfs_set_pageio_descriptor(desc, hdr);
1484 list_splice_init(&hdr->pages, &pages);
1485 while (!list_empty(&pages)) {
1486 struct nfs_page *req = nfs_list_entry(pages.next);
1487
1488 if (!nfs_pageio_add_request(desc, req))
1489 break;
1490 }
1491 nfs_pageio_complete(desc);
1492 if (!list_empty(&pages)) {
1493 int err = desc->pg_error < 0 ? desc->pg_error : -EIO;
1494 hdr->completion_ops->error_cleanup(&pages, err);
1495 nfs_set_pgio_error(hdr, err, hdr->io_start);
1496 return err;
1497 }
1498 return 0;
1499}
1500EXPORT_SYMBOL_GPL(nfs_pageio_resend);
1501
1502/**
1503 * nfs_pageio_complete - Complete I/O then cleanup an nfs_pageio_descriptor
1504 * @desc: pointer to io descriptor
1505 */
1506void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1507{
1508 u32 midx;
1509
1510 for (midx = 0; midx < desc->pg_mirror_count; midx++)
1511 nfs_pageio_complete_mirror(desc, midx);
1512
1513 if (desc->pg_error < 0)
1514 nfs_pageio_error_cleanup(desc);
1515 if (desc->pg_ops->pg_cleanup)
1516 desc->pg_ops->pg_cleanup(desc);
1517 nfs_pageio_cleanup_mirroring(desc);
1518}
1519
1520/**
1521 * nfs_pageio_cond_complete - Conditional I/O completion
1522 * @desc: pointer to io descriptor
1523 * @index: page index
1524 *
1525 * It is important to ensure that processes don't try to take locks
1526 * on non-contiguous ranges of pages as that might deadlock. This
1527 * function should be called before attempting to wait on a locked
1528 * nfs_page. It will complete the I/O if the page index 'index'
1529 * is not contiguous with the existing list of pages in 'desc'.
1530 */
1531void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1532{
1533 struct nfs_pgio_mirror *mirror;
1534 struct nfs_page *prev;
1535 struct folio *folio;
1536 u32 midx;
1537
1538 for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1539 mirror = nfs_pgio_get_mirror(desc, midx);
1540 if (!list_empty(&mirror->pg_list)) {
1541 prev = nfs_list_entry(mirror->pg_list.prev);
1542 folio = nfs_page_to_folio(prev);
1543 if (folio) {
1544 if (index == folio_next_index(folio))
1545 continue;
1546 } else if (index == prev->wb_index + 1)
1547 continue;
1548 nfs_pageio_complete(desc);
1549 break;
1550 }
1551 }
1552}
1553
1554/*
1555 * nfs_pageio_stop_mirroring - stop using mirroring (set mirror count to 1)
1556 */
1557void nfs_pageio_stop_mirroring(struct nfs_pageio_descriptor *pgio)
1558{
1559 nfs_pageio_complete(pgio);
1560}
1561
1562int __init nfs_init_nfspagecache(void)
1563{
1564 nfs_page_cachep = kmem_cache_create("nfs_page",
1565 sizeof(struct nfs_page),
1566 0, SLAB_HWCACHE_ALIGN,
1567 NULL);
1568 if (nfs_page_cachep == NULL)
1569 return -ENOMEM;
1570
1571 return 0;
1572}
1573
1574void nfs_destroy_nfspagecache(void)
1575{
1576 kmem_cache_destroy(nfs_page_cachep);
1577}
1578
1579static const struct rpc_call_ops nfs_pgio_common_ops = {
1580 .rpc_call_prepare = nfs_pgio_prepare,
1581 .rpc_call_done = nfs_pgio_result,
1582 .rpc_release = nfs_pgio_release,
1583};
1584
1585const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1586 .pg_test = nfs_generic_pg_test,
1587 .pg_doio = nfs_generic_pg_pgios,
1588};