xref: /openbmc/linux/fs/ceph/addr.c (revision c819e2cf)
1 #include <linux/ceph/ceph_debug.h>
2 
3 #include <linux/backing-dev.h>
4 #include <linux/fs.h>
5 #include <linux/mm.h>
6 #include <linux/pagemap.h>
7 #include <linux/writeback.h>	/* generic_writepages */
8 #include <linux/slab.h>
9 #include <linux/pagevec.h>
10 #include <linux/task_io_accounting_ops.h>
11 
12 #include "super.h"
13 #include "mds_client.h"
14 #include "cache.h"
15 #include <linux/ceph/osd_client.h>
16 
17 /*
18  * Ceph address space ops.
19  *
20  * There are a few funny things going on here.
21  *
22  * The page->private field is used to reference a struct
23  * ceph_snap_context for _every_ dirty page.  This indicates which
24  * snapshot the page was logically dirtied in, and thus which snap
25  * context needs to be associated with the osd write during writeback.
26  *
27  * Similarly, struct ceph_inode_info maintains a set of counters to
28  * count dirty pages on the inode.  In the absence of snapshots,
29  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
30  *
31  * When a snapshot is taken (that is, when the client receives
32  * notification that a snapshot was taken), each inode with caps and
33  * with dirty pages (dirty pages implies there is a cap) gets a new
34  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
35  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
36  * moved to capsnap->dirty. (Unless a sync write is currently in
37  * progress.  In that case, the capsnap is said to be "pending", new
38  * writes cannot start, and the capsnap isn't "finalized" until the
39  * write completes (or fails) and a final size/mtime for the inode for
40  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
41  *
42  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
43  * we look for the first capsnap in i_cap_snaps and write out pages in
44  * that snap context _only_.  Then we move on to the next capsnap,
45  * eventually reaching the "live" or "head" context (i.e., pages that
46  * are not yet snapped) and are writing the most recently dirtied
47  * pages.
48  *
49  * Invalidate and so forth must take care to ensure the dirty page
50  * accounting is preserved.
51  */
52 
53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
54 #define CONGESTION_OFF_THRESH(congestion_kb)				\
55 	(CONGESTION_ON_THRESH(congestion_kb) -				\
56 	 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
57 
58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
59 {
60 	if (PagePrivate(page))
61 		return (void *)page->private;
62 	return NULL;
63 }
64 
65 /*
66  * Dirty a page.  Optimistically adjust accounting, on the assumption
67  * that we won't race with invalidate.  If we do, readjust.
68  */
69 static int ceph_set_page_dirty(struct page *page)
70 {
71 	struct address_space *mapping = page->mapping;
72 	struct inode *inode;
73 	struct ceph_inode_info *ci;
74 	struct ceph_snap_context *snapc;
75 	int ret;
76 
77 	if (unlikely(!mapping))
78 		return !TestSetPageDirty(page);
79 
80 	if (PageDirty(page)) {
81 		dout("%p set_page_dirty %p idx %lu -- already dirty\n",
82 		     mapping->host, page, page->index);
83 		BUG_ON(!PagePrivate(page));
84 		return 0;
85 	}
86 
87 	inode = mapping->host;
88 	ci = ceph_inode(inode);
89 
90 	/*
91 	 * Note that we're grabbing a snapc ref here without holding
92 	 * any locks!
93 	 */
94 	snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
95 
96 	/* dirty the head */
97 	spin_lock(&ci->i_ceph_lock);
98 	if (ci->i_head_snapc == NULL)
99 		ci->i_head_snapc = ceph_get_snap_context(snapc);
100 	++ci->i_wrbuffer_ref_head;
101 	if (ci->i_wrbuffer_ref == 0)
102 		ihold(inode);
103 	++ci->i_wrbuffer_ref;
104 	dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
105 	     "snapc %p seq %lld (%d snaps)\n",
106 	     mapping->host, page, page->index,
107 	     ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
108 	     ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
109 	     snapc, snapc->seq, snapc->num_snaps);
110 	spin_unlock(&ci->i_ceph_lock);
111 
112 	/*
113 	 * Reference snap context in page->private.  Also set
114 	 * PagePrivate so that we get invalidatepage callback.
115 	 */
116 	BUG_ON(PagePrivate(page));
117 	page->private = (unsigned long)snapc;
118 	SetPagePrivate(page);
119 
120 	ret = __set_page_dirty_nobuffers(page);
121 	WARN_ON(!PageLocked(page));
122 	WARN_ON(!page->mapping);
123 
124 	return ret;
125 }
126 
127 /*
128  * If we are truncating the full page (i.e. offset == 0), adjust the
129  * dirty page counters appropriately.  Only called if there is private
130  * data on the page.
131  */
132 static void ceph_invalidatepage(struct page *page, unsigned int offset,
133 				unsigned int length)
134 {
135 	struct inode *inode;
136 	struct ceph_inode_info *ci;
137 	struct ceph_snap_context *snapc = page_snap_context(page);
138 
139 	inode = page->mapping->host;
140 	ci = ceph_inode(inode);
141 
142 	if (offset != 0 || length != PAGE_CACHE_SIZE) {
143 		dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
144 		     inode, page, page->index, offset, length);
145 		return;
146 	}
147 
148 	ceph_invalidate_fscache_page(inode, page);
149 
150 	if (!PagePrivate(page))
151 		return;
152 
153 	/*
154 	 * We can get non-dirty pages here due to races between
155 	 * set_page_dirty and truncate_complete_page; just spit out a
156 	 * warning, in case we end up with accounting problems later.
157 	 */
158 	if (!PageDirty(page))
159 		pr_err("%p invalidatepage %p page not dirty\n", inode, page);
160 
161 	ClearPageChecked(page);
162 
163 	dout("%p invalidatepage %p idx %lu full dirty page\n",
164 	     inode, page, page->index);
165 
166 	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
167 	ceph_put_snap_context(snapc);
168 	page->private = 0;
169 	ClearPagePrivate(page);
170 }
171 
172 static int ceph_releasepage(struct page *page, gfp_t g)
173 {
174 	struct inode *inode = page->mapping ? page->mapping->host : NULL;
175 	dout("%p releasepage %p idx %lu\n", inode, page, page->index);
176 	WARN_ON(PageDirty(page));
177 
178 	/* Can we release the page from the cache? */
179 	if (!ceph_release_fscache_page(page, g))
180 		return 0;
181 
182 	return !PagePrivate(page);
183 }
184 
185 /*
186  * read a single page, without unlocking it.
187  */
188 static int readpage_nounlock(struct file *filp, struct page *page)
189 {
190 	struct inode *inode = file_inode(filp);
191 	struct ceph_inode_info *ci = ceph_inode(inode);
192 	struct ceph_osd_client *osdc =
193 		&ceph_inode_to_client(inode)->client->osdc;
194 	int err = 0;
195 	u64 off = page_offset(page);
196 	u64 len = PAGE_CACHE_SIZE;
197 
198 	if (off >= i_size_read(inode)) {
199 		zero_user_segment(page, err, PAGE_CACHE_SIZE);
200 		SetPageUptodate(page);
201 		return 0;
202 	}
203 
204 	/*
205 	 * Uptodate inline data should have been added into page cache
206 	 * while getting Fcr caps.
207 	 */
208 	if (ci->i_inline_version != CEPH_INLINE_NONE)
209 		return -EINVAL;
210 
211 	err = ceph_readpage_from_fscache(inode, page);
212 	if (err == 0)
213 		goto out;
214 
215 	dout("readpage inode %p file %p page %p index %lu\n",
216 	     inode, filp, page, page->index);
217 	err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
218 				  off, &len,
219 				  ci->i_truncate_seq, ci->i_truncate_size,
220 				  &page, 1, 0);
221 	if (err == -ENOENT)
222 		err = 0;
223 	if (err < 0) {
224 		SetPageError(page);
225 		ceph_fscache_readpage_cancel(inode, page);
226 		goto out;
227 	}
228 	if (err < PAGE_CACHE_SIZE)
229 		/* zero fill remainder of page */
230 		zero_user_segment(page, err, PAGE_CACHE_SIZE);
231 	else
232 		flush_dcache_page(page);
233 
234 	SetPageUptodate(page);
235 	ceph_readpage_to_fscache(inode, page);
236 
237 out:
238 	return err < 0 ? err : 0;
239 }
240 
241 static int ceph_readpage(struct file *filp, struct page *page)
242 {
243 	int r = readpage_nounlock(filp, page);
244 	unlock_page(page);
245 	return r;
246 }
247 
248 /*
249  * Finish an async read(ahead) op.
250  */
251 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
252 {
253 	struct inode *inode = req->r_inode;
254 	struct ceph_osd_data *osd_data;
255 	int rc = req->r_result;
256 	int bytes = le32_to_cpu(msg->hdr.data_len);
257 	int num_pages;
258 	int i;
259 
260 	dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
261 
262 	/* unlock all pages, zeroing any data we didn't read */
263 	osd_data = osd_req_op_extent_osd_data(req, 0);
264 	BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
265 	num_pages = calc_pages_for((u64)osd_data->alignment,
266 					(u64)osd_data->length);
267 	for (i = 0; i < num_pages; i++) {
268 		struct page *page = osd_data->pages[i];
269 
270 		if (rc < 0)
271 			goto unlock;
272 		if (bytes < (int)PAGE_CACHE_SIZE) {
273 			/* zero (remainder of) page */
274 			int s = bytes < 0 ? 0 : bytes;
275 			zero_user_segment(page, s, PAGE_CACHE_SIZE);
276 		}
277  		dout("finish_read %p uptodate %p idx %lu\n", inode, page,
278 		     page->index);
279 		flush_dcache_page(page);
280 		SetPageUptodate(page);
281 		ceph_readpage_to_fscache(inode, page);
282 unlock:
283 		unlock_page(page);
284 		page_cache_release(page);
285 		bytes -= PAGE_CACHE_SIZE;
286 	}
287 	kfree(osd_data->pages);
288 }
289 
290 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
291 {
292 	int i;
293 
294 	for (i = 0; i < num_pages; i++)
295 		unlock_page(pages[i]);
296 }
297 
298 /*
299  * start an async read(ahead) operation.  return nr_pages we submitted
300  * a read for on success, or negative error code.
301  */
302 static int start_read(struct inode *inode, struct list_head *page_list, int max)
303 {
304 	struct ceph_osd_client *osdc =
305 		&ceph_inode_to_client(inode)->client->osdc;
306 	struct ceph_inode_info *ci = ceph_inode(inode);
307 	struct page *page = list_entry(page_list->prev, struct page, lru);
308 	struct ceph_vino vino;
309 	struct ceph_osd_request *req;
310 	u64 off;
311 	u64 len;
312 	int i;
313 	struct page **pages;
314 	pgoff_t next_index;
315 	int nr_pages = 0;
316 	int ret;
317 
318 	off = (u64) page_offset(page);
319 
320 	/* count pages */
321 	next_index = page->index;
322 	list_for_each_entry_reverse(page, page_list, lru) {
323 		if (page->index != next_index)
324 			break;
325 		nr_pages++;
326 		next_index++;
327 		if (max && nr_pages == max)
328 			break;
329 	}
330 	len = nr_pages << PAGE_CACHE_SHIFT;
331 	dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
332 	     off, len);
333 	vino = ceph_vino(inode);
334 	req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
335 				    0, 1, CEPH_OSD_OP_READ,
336 				    CEPH_OSD_FLAG_READ, NULL,
337 				    ci->i_truncate_seq, ci->i_truncate_size,
338 				    false);
339 	if (IS_ERR(req))
340 		return PTR_ERR(req);
341 
342 	/* build page vector */
343 	nr_pages = calc_pages_for(0, len);
344 	pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
345 	ret = -ENOMEM;
346 	if (!pages)
347 		goto out;
348 	for (i = 0; i < nr_pages; ++i) {
349 		page = list_entry(page_list->prev, struct page, lru);
350 		BUG_ON(PageLocked(page));
351 		list_del(&page->lru);
352 
353  		dout("start_read %p adding %p idx %lu\n", inode, page,
354 		     page->index);
355 		if (add_to_page_cache_lru(page, &inode->i_data, page->index,
356 					  GFP_NOFS)) {
357 			ceph_fscache_uncache_page(inode, page);
358 			page_cache_release(page);
359 			dout("start_read %p add_to_page_cache failed %p\n",
360 			     inode, page);
361 			nr_pages = i;
362 			goto out_pages;
363 		}
364 		pages[i] = page;
365 	}
366 	osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
367 	req->r_callback = finish_read;
368 	req->r_inode = inode;
369 
370 	ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
371 
372 	dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
373 	ret = ceph_osdc_start_request(osdc, req, false);
374 	if (ret < 0)
375 		goto out_pages;
376 	ceph_osdc_put_request(req);
377 	return nr_pages;
378 
379 out_pages:
380 	ceph_unlock_page_vector(pages, nr_pages);
381 	ceph_release_page_vector(pages, nr_pages);
382 out:
383 	ceph_osdc_put_request(req);
384 	return ret;
385 }
386 
387 
388 /*
389  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
390  * the caller (VM) cleans them up.
391  */
392 static int ceph_readpages(struct file *file, struct address_space *mapping,
393 			  struct list_head *page_list, unsigned nr_pages)
394 {
395 	struct inode *inode = file_inode(file);
396 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
397 	int rc = 0;
398 	int max = 0;
399 
400 	if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
401 		return -EINVAL;
402 
403 	rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
404 					 &nr_pages);
405 
406 	if (rc == 0)
407 		goto out;
408 
409 	if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
410 		max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
411 			>> PAGE_SHIFT;
412 
413 	dout("readpages %p file %p nr_pages %d max %d\n", inode,
414 		file, nr_pages,
415 	     max);
416 	while (!list_empty(page_list)) {
417 		rc = start_read(inode, page_list, max);
418 		if (rc < 0)
419 			goto out;
420 		BUG_ON(rc == 0);
421 	}
422 out:
423 	ceph_fscache_readpages_cancel(inode, page_list);
424 
425 	dout("readpages %p file %p ret %d\n", inode, file, rc);
426 	return rc;
427 }
428 
429 /*
430  * Get ref for the oldest snapc for an inode with dirty data... that is, the
431  * only snap context we are allowed to write back.
432  */
433 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
434 						    u64 *snap_size)
435 {
436 	struct ceph_inode_info *ci = ceph_inode(inode);
437 	struct ceph_snap_context *snapc = NULL;
438 	struct ceph_cap_snap *capsnap = NULL;
439 
440 	spin_lock(&ci->i_ceph_lock);
441 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
442 		dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
443 		     capsnap->context, capsnap->dirty_pages);
444 		if (capsnap->dirty_pages) {
445 			snapc = ceph_get_snap_context(capsnap->context);
446 			if (snap_size)
447 				*snap_size = capsnap->size;
448 			break;
449 		}
450 	}
451 	if (!snapc && ci->i_wrbuffer_ref_head) {
452 		snapc = ceph_get_snap_context(ci->i_head_snapc);
453 		dout(" head snapc %p has %d dirty pages\n",
454 		     snapc, ci->i_wrbuffer_ref_head);
455 	}
456 	spin_unlock(&ci->i_ceph_lock);
457 	return snapc;
458 }
459 
460 /*
461  * Write a single page, but leave the page locked.
462  *
463  * If we get a write error, set the page error bit, but still adjust the
464  * dirty page accounting (i.e., page is no longer dirty).
465  */
466 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
467 {
468 	struct inode *inode;
469 	struct ceph_inode_info *ci;
470 	struct ceph_fs_client *fsc;
471 	struct ceph_osd_client *osdc;
472 	struct ceph_snap_context *snapc, *oldest;
473 	loff_t page_off = page_offset(page);
474 	long writeback_stat;
475 	u64 truncate_size, snap_size = 0;
476 	u32 truncate_seq;
477 	int err = 0, len = PAGE_CACHE_SIZE;
478 
479 	dout("writepage %p idx %lu\n", page, page->index);
480 
481 	if (!page->mapping || !page->mapping->host) {
482 		dout("writepage %p - no mapping\n", page);
483 		return -EFAULT;
484 	}
485 	inode = page->mapping->host;
486 	ci = ceph_inode(inode);
487 	fsc = ceph_inode_to_client(inode);
488 	osdc = &fsc->client->osdc;
489 
490 	/* verify this is a writeable snap context */
491 	snapc = page_snap_context(page);
492 	if (snapc == NULL) {
493 		dout("writepage %p page %p not dirty?\n", inode, page);
494 		goto out;
495 	}
496 	oldest = get_oldest_context(inode, &snap_size);
497 	if (snapc->seq > oldest->seq) {
498 		dout("writepage %p page %p snapc %p not writeable - noop\n",
499 		     inode, page, snapc);
500 		/* we should only noop if called by kswapd */
501 		WARN_ON((current->flags & PF_MEMALLOC) == 0);
502 		ceph_put_snap_context(oldest);
503 		goto out;
504 	}
505 	ceph_put_snap_context(oldest);
506 
507 	spin_lock(&ci->i_ceph_lock);
508 	truncate_seq = ci->i_truncate_seq;
509 	truncate_size = ci->i_truncate_size;
510 	if (!snap_size)
511 		snap_size = i_size_read(inode);
512 	spin_unlock(&ci->i_ceph_lock);
513 
514 	/* is this a partial page at end of file? */
515 	if (page_off >= snap_size) {
516 		dout("%p page eof %llu\n", page, snap_size);
517 		goto out;
518 	}
519 	if (snap_size < page_off + len)
520 		len = snap_size - page_off;
521 
522 	dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
523 	     inode, page, page->index, page_off, len, snapc);
524 
525 	writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
526 	if (writeback_stat >
527 	    CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
528 		set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
529 
530 	ceph_readpage_to_fscache(inode, page);
531 
532 	set_page_writeback(page);
533 	err = ceph_osdc_writepages(osdc, ceph_vino(inode),
534 				   &ci->i_layout, snapc,
535 				   page_off, len,
536 				   truncate_seq, truncate_size,
537 				   &inode->i_mtime, &page, 1);
538 	if (err < 0) {
539 		dout("writepage setting page/mapping error %d %p\n", err, page);
540 		SetPageError(page);
541 		mapping_set_error(&inode->i_data, err);
542 		if (wbc)
543 			wbc->pages_skipped++;
544 	} else {
545 		dout("writepage cleaned page %p\n", page);
546 		err = 0;  /* vfs expects us to return 0 */
547 	}
548 	page->private = 0;
549 	ClearPagePrivate(page);
550 	end_page_writeback(page);
551 	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
552 	ceph_put_snap_context(snapc);  /* page's reference */
553 out:
554 	return err;
555 }
556 
557 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
558 {
559 	int err;
560 	struct inode *inode = page->mapping->host;
561 	BUG_ON(!inode);
562 	ihold(inode);
563 	err = writepage_nounlock(page, wbc);
564 	unlock_page(page);
565 	iput(inode);
566 	return err;
567 }
568 
569 
570 /*
571  * lame release_pages helper.  release_pages() isn't exported to
572  * modules.
573  */
574 static void ceph_release_pages(struct page **pages, int num)
575 {
576 	struct pagevec pvec;
577 	int i;
578 
579 	pagevec_init(&pvec, 0);
580 	for (i = 0; i < num; i++) {
581 		if (pagevec_add(&pvec, pages[i]) == 0)
582 			pagevec_release(&pvec);
583 	}
584 	pagevec_release(&pvec);
585 }
586 
587 /*
588  * async writeback completion handler.
589  *
590  * If we get an error, set the mapping error bit, but not the individual
591  * page error bits.
592  */
593 static void writepages_finish(struct ceph_osd_request *req,
594 			      struct ceph_msg *msg)
595 {
596 	struct inode *inode = req->r_inode;
597 	struct ceph_inode_info *ci = ceph_inode(inode);
598 	struct ceph_osd_data *osd_data;
599 	unsigned wrote;
600 	struct page *page;
601 	int num_pages;
602 	int i;
603 	struct ceph_snap_context *snapc = req->r_snapc;
604 	struct address_space *mapping = inode->i_mapping;
605 	int rc = req->r_result;
606 	u64 bytes = req->r_ops[0].extent.length;
607 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
608 	long writeback_stat;
609 	unsigned issued = ceph_caps_issued(ci);
610 
611 	osd_data = osd_req_op_extent_osd_data(req, 0);
612 	BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
613 	num_pages = calc_pages_for((u64)osd_data->alignment,
614 					(u64)osd_data->length);
615 	if (rc >= 0) {
616 		/*
617 		 * Assume we wrote the pages we originally sent.  The
618 		 * osd might reply with fewer pages if our writeback
619 		 * raced with a truncation and was adjusted at the osd,
620 		 * so don't believe the reply.
621 		 */
622 		wrote = num_pages;
623 	} else {
624 		wrote = 0;
625 		mapping_set_error(mapping, rc);
626 	}
627 	dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
628 	     inode, rc, bytes, wrote);
629 
630 	/* clean all pages */
631 	for (i = 0; i < num_pages; i++) {
632 		page = osd_data->pages[i];
633 		BUG_ON(!page);
634 		WARN_ON(!PageUptodate(page));
635 
636 		writeback_stat =
637 			atomic_long_dec_return(&fsc->writeback_count);
638 		if (writeback_stat <
639 		    CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
640 			clear_bdi_congested(&fsc->backing_dev_info,
641 					    BLK_RW_ASYNC);
642 
643 		ceph_put_snap_context(page_snap_context(page));
644 		page->private = 0;
645 		ClearPagePrivate(page);
646 		dout("unlocking %d %p\n", i, page);
647 		end_page_writeback(page);
648 
649 		/*
650 		 * We lost the cache cap, need to truncate the page before
651 		 * it is unlocked, otherwise we'd truncate it later in the
652 		 * page truncation thread, possibly losing some data that
653 		 * raced its way in
654 		 */
655 		if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
656 			generic_error_remove_page(inode->i_mapping, page);
657 
658 		unlock_page(page);
659 	}
660 	dout("%p wrote+cleaned %d pages\n", inode, wrote);
661 	ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
662 
663 	ceph_release_pages(osd_data->pages, num_pages);
664 	if (osd_data->pages_from_pool)
665 		mempool_free(osd_data->pages,
666 			     ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
667 	else
668 		kfree(osd_data->pages);
669 	ceph_osdc_put_request(req);
670 }
671 
672 /*
673  * initiate async writeback
674  */
675 static int ceph_writepages_start(struct address_space *mapping,
676 				 struct writeback_control *wbc)
677 {
678 	struct inode *inode = mapping->host;
679 	struct ceph_inode_info *ci = ceph_inode(inode);
680 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
681 	struct ceph_vino vino = ceph_vino(inode);
682 	pgoff_t index, start, end;
683 	int range_whole = 0;
684 	int should_loop = 1;
685 	pgoff_t max_pages = 0, max_pages_ever = 0;
686 	struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
687 	struct pagevec pvec;
688 	int done = 0;
689 	int rc = 0;
690 	unsigned wsize = 1 << inode->i_blkbits;
691 	struct ceph_osd_request *req = NULL;
692 	int do_sync = 0;
693 	u64 truncate_size, snap_size;
694 	u32 truncate_seq;
695 
696 	/*
697 	 * Include a 'sync' in the OSD request if this is a data
698 	 * integrity write (e.g., O_SYNC write or fsync()), or if our
699 	 * cap is being revoked.
700 	 */
701 	if ((wbc->sync_mode == WB_SYNC_ALL) ||
702 		ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
703 		do_sync = 1;
704 	dout("writepages_start %p dosync=%d (mode=%s)\n",
705 	     inode, do_sync,
706 	     wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
707 	     (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
708 
709 	if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
710 		pr_warn("writepage_start %p on forced umount\n", inode);
711 		return -EIO; /* we're in a forced umount, don't write! */
712 	}
713 	if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
714 		wsize = fsc->mount_options->wsize;
715 	if (wsize < PAGE_CACHE_SIZE)
716 		wsize = PAGE_CACHE_SIZE;
717 	max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
718 
719 	pagevec_init(&pvec, 0);
720 
721 	/* where to start/end? */
722 	if (wbc->range_cyclic) {
723 		start = mapping->writeback_index; /* Start from prev offset */
724 		end = -1;
725 		dout(" cyclic, start at %lu\n", start);
726 	} else {
727 		start = wbc->range_start >> PAGE_CACHE_SHIFT;
728 		end = wbc->range_end >> PAGE_CACHE_SHIFT;
729 		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
730 			range_whole = 1;
731 		should_loop = 0;
732 		dout(" not cyclic, %lu to %lu\n", start, end);
733 	}
734 	index = start;
735 
736 retry:
737 	/* find oldest snap context with dirty data */
738 	ceph_put_snap_context(snapc);
739 	snap_size = 0;
740 	snapc = get_oldest_context(inode, &snap_size);
741 	if (!snapc) {
742 		/* hmm, why does writepages get called when there
743 		   is no dirty data? */
744 		dout(" no snap context with dirty data?\n");
745 		goto out;
746 	}
747 	if (snap_size == 0)
748 		snap_size = i_size_read(inode);
749 	dout(" oldest snapc is %p seq %lld (%d snaps)\n",
750 	     snapc, snapc->seq, snapc->num_snaps);
751 
752 	spin_lock(&ci->i_ceph_lock);
753 	truncate_seq = ci->i_truncate_seq;
754 	truncate_size = ci->i_truncate_size;
755 	if (!snap_size)
756 		snap_size = i_size_read(inode);
757 	spin_unlock(&ci->i_ceph_lock);
758 
759 	if (last_snapc && snapc != last_snapc) {
760 		/* if we switched to a newer snapc, restart our scan at the
761 		 * start of the original file range. */
762 		dout("  snapc differs from last pass, restarting at %lu\n",
763 		     index);
764 		index = start;
765 	}
766 	last_snapc = snapc;
767 
768 	while (!done && index <= end) {
769 		unsigned i;
770 		int first;
771 		pgoff_t next;
772 		int pvec_pages, locked_pages;
773 		struct page **pages = NULL;
774 		mempool_t *pool = NULL;	/* Becomes non-null if mempool used */
775 		struct page *page;
776 		int want;
777 		u64 offset, len;
778 		long writeback_stat;
779 
780 		next = 0;
781 		locked_pages = 0;
782 		max_pages = max_pages_ever;
783 
784 get_more_pages:
785 		first = -1;
786 		want = min(end - index,
787 			   min((pgoff_t)PAGEVEC_SIZE,
788 			       max_pages - (pgoff_t)locked_pages) - 1)
789 			+ 1;
790 		pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
791 						PAGECACHE_TAG_DIRTY,
792 						want);
793 		dout("pagevec_lookup_tag got %d\n", pvec_pages);
794 		if (!pvec_pages && !locked_pages)
795 			break;
796 		for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
797 			page = pvec.pages[i];
798 			dout("? %p idx %lu\n", page, page->index);
799 			if (locked_pages == 0)
800 				lock_page(page);  /* first page */
801 			else if (!trylock_page(page))
802 				break;
803 
804 			/* only dirty pages, or our accounting breaks */
805 			if (unlikely(!PageDirty(page)) ||
806 			    unlikely(page->mapping != mapping)) {
807 				dout("!dirty or !mapping %p\n", page);
808 				unlock_page(page);
809 				break;
810 			}
811 			if (!wbc->range_cyclic && page->index > end) {
812 				dout("end of range %p\n", page);
813 				done = 1;
814 				unlock_page(page);
815 				break;
816 			}
817 			if (next && (page->index != next)) {
818 				dout("not consecutive %p\n", page);
819 				unlock_page(page);
820 				break;
821 			}
822 			if (wbc->sync_mode != WB_SYNC_NONE) {
823 				dout("waiting on writeback %p\n", page);
824 				wait_on_page_writeback(page);
825 			}
826 			if (page_offset(page) >= snap_size) {
827 				dout("%p page eof %llu\n", page, snap_size);
828 				done = 1;
829 				unlock_page(page);
830 				break;
831 			}
832 			if (PageWriteback(page)) {
833 				dout("%p under writeback\n", page);
834 				unlock_page(page);
835 				break;
836 			}
837 
838 			/* only if matching snap context */
839 			pgsnapc = page_snap_context(page);
840 			if (pgsnapc->seq > snapc->seq) {
841 				dout("page snapc %p %lld > oldest %p %lld\n",
842 				     pgsnapc, pgsnapc->seq, snapc, snapc->seq);
843 				unlock_page(page);
844 				if (!locked_pages)
845 					continue; /* keep looking for snap */
846 				break;
847 			}
848 
849 			if (!clear_page_dirty_for_io(page)) {
850 				dout("%p !clear_page_dirty_for_io\n", page);
851 				unlock_page(page);
852 				break;
853 			}
854 
855 			/*
856 			 * We have something to write.  If this is
857 			 * the first locked page this time through,
858 			 * allocate an osd request and a page array
859 			 * that it will use.
860 			 */
861 			if (locked_pages == 0) {
862 				BUG_ON(pages);
863 				/* prepare async write request */
864 				offset = (u64)page_offset(page);
865 				len = wsize;
866 				req = ceph_osdc_new_request(&fsc->client->osdc,
867 							&ci->i_layout, vino,
868 							offset, &len, 0,
869 							do_sync ? 2 : 1,
870 							CEPH_OSD_OP_WRITE,
871 							CEPH_OSD_FLAG_WRITE |
872 							CEPH_OSD_FLAG_ONDISK,
873 							snapc, truncate_seq,
874 							truncate_size, true);
875 				if (IS_ERR(req)) {
876 					rc = PTR_ERR(req);
877 					unlock_page(page);
878 					break;
879 				}
880 
881 				if (do_sync)
882 					osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
883 
884 				req->r_callback = writepages_finish;
885 				req->r_inode = inode;
886 
887 				max_pages = calc_pages_for(0, (u64)len);
888 				pages = kmalloc(max_pages * sizeof (*pages),
889 						GFP_NOFS);
890 				if (!pages) {
891 					pool = fsc->wb_pagevec_pool;
892 					pages = mempool_alloc(pool, GFP_NOFS);
893 					BUG_ON(!pages);
894 				}
895 			}
896 
897 			/* note position of first page in pvec */
898 			if (first < 0)
899 				first = i;
900 			dout("%p will write page %p idx %lu\n",
901 			     inode, page, page->index);
902 
903 			writeback_stat =
904 			       atomic_long_inc_return(&fsc->writeback_count);
905 			if (writeback_stat > CONGESTION_ON_THRESH(
906 				    fsc->mount_options->congestion_kb)) {
907 				set_bdi_congested(&fsc->backing_dev_info,
908 						  BLK_RW_ASYNC);
909 			}
910 
911 			set_page_writeback(page);
912 			pages[locked_pages] = page;
913 			locked_pages++;
914 			next = page->index + 1;
915 		}
916 
917 		/* did we get anything? */
918 		if (!locked_pages)
919 			goto release_pvec_pages;
920 		if (i) {
921 			int j;
922 			BUG_ON(!locked_pages || first < 0);
923 
924 			if (pvec_pages && i == pvec_pages &&
925 			    locked_pages < max_pages) {
926 				dout("reached end pvec, trying for more\n");
927 				pagevec_reinit(&pvec);
928 				goto get_more_pages;
929 			}
930 
931 			/* shift unused pages over in the pvec...  we
932 			 * will need to release them below. */
933 			for (j = i; j < pvec_pages; j++) {
934 				dout(" pvec leftover page %p\n",
935 				     pvec.pages[j]);
936 				pvec.pages[j-i+first] = pvec.pages[j];
937 			}
938 			pvec.nr -= i-first;
939 		}
940 
941 		/* Format the osd request message and submit the write */
942 
943 		offset = page_offset(pages[0]);
944 		len = min(snap_size - offset,
945 			  (u64)locked_pages << PAGE_CACHE_SHIFT);
946 		dout("writepages got %d pages at %llu~%llu\n",
947 		     locked_pages, offset, len);
948 
949 		osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
950 							!!pool, false);
951 
952 		pages = NULL;	/* request message now owns the pages array */
953 		pool = NULL;
954 
955 		/* Update the write op length in case we changed it */
956 
957 		osd_req_op_extent_update(req, 0, len);
958 
959 		vino = ceph_vino(inode);
960 		ceph_osdc_build_request(req, offset, snapc, vino.snap,
961 					&inode->i_mtime);
962 
963 		rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
964 		BUG_ON(rc);
965 		req = NULL;
966 
967 		/* continue? */
968 		index = next;
969 		wbc->nr_to_write -= locked_pages;
970 		if (wbc->nr_to_write <= 0)
971 			done = 1;
972 
973 release_pvec_pages:
974 		dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
975 		     pvec.nr ? pvec.pages[0] : NULL);
976 		pagevec_release(&pvec);
977 
978 		if (locked_pages && !done)
979 			goto retry;
980 	}
981 
982 	if (should_loop && !done) {
983 		/* more to do; loop back to beginning of file */
984 		dout("writepages looping back to beginning of file\n");
985 		should_loop = 0;
986 		index = 0;
987 		goto retry;
988 	}
989 
990 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
991 		mapping->writeback_index = index;
992 
993 out:
994 	if (req)
995 		ceph_osdc_put_request(req);
996 	ceph_put_snap_context(snapc);
997 	dout("writepages done, rc = %d\n", rc);
998 	return rc;
999 }
1000 
1001 
1002 
1003 /*
1004  * See if a given @snapc is either writeable, or already written.
1005  */
1006 static int context_is_writeable_or_written(struct inode *inode,
1007 					   struct ceph_snap_context *snapc)
1008 {
1009 	struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1010 	int ret = !oldest || snapc->seq <= oldest->seq;
1011 
1012 	ceph_put_snap_context(oldest);
1013 	return ret;
1014 }
1015 
1016 /*
1017  * We are only allowed to write into/dirty the page if the page is
1018  * clean, or already dirty within the same snap context.
1019  *
1020  * called with page locked.
1021  * return success with page locked,
1022  * or any failure (incl -EAGAIN) with page unlocked.
1023  */
1024 static int ceph_update_writeable_page(struct file *file,
1025 			    loff_t pos, unsigned len,
1026 			    struct page *page)
1027 {
1028 	struct inode *inode = file_inode(file);
1029 	struct ceph_inode_info *ci = ceph_inode(inode);
1030 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1031 	loff_t page_off = pos & PAGE_CACHE_MASK;
1032 	int pos_in_page = pos & ~PAGE_CACHE_MASK;
1033 	int end_in_page = pos_in_page + len;
1034 	loff_t i_size;
1035 	int r;
1036 	struct ceph_snap_context *snapc, *oldest;
1037 
1038 retry_locked:
1039 	/* writepages currently holds page lock, but if we change that later, */
1040 	wait_on_page_writeback(page);
1041 
1042 	/* check snap context */
1043 	BUG_ON(!ci->i_snap_realm);
1044 	down_read(&mdsc->snap_rwsem);
1045 	BUG_ON(!ci->i_snap_realm->cached_context);
1046 	snapc = page_snap_context(page);
1047 	if (snapc && snapc != ci->i_head_snapc) {
1048 		/*
1049 		 * this page is already dirty in another (older) snap
1050 		 * context!  is it writeable now?
1051 		 */
1052 		oldest = get_oldest_context(inode, NULL);
1053 		up_read(&mdsc->snap_rwsem);
1054 
1055 		if (snapc->seq > oldest->seq) {
1056 			ceph_put_snap_context(oldest);
1057 			dout(" page %p snapc %p not current or oldest\n",
1058 			     page, snapc);
1059 			/*
1060 			 * queue for writeback, and wait for snapc to
1061 			 * be writeable or written
1062 			 */
1063 			snapc = ceph_get_snap_context(snapc);
1064 			unlock_page(page);
1065 			ceph_queue_writeback(inode);
1066 			r = wait_event_interruptible(ci->i_cap_wq,
1067 			       context_is_writeable_or_written(inode, snapc));
1068 			ceph_put_snap_context(snapc);
1069 			if (r == -ERESTARTSYS)
1070 				return r;
1071 			return -EAGAIN;
1072 		}
1073 		ceph_put_snap_context(oldest);
1074 
1075 		/* yay, writeable, do it now (without dropping page lock) */
1076 		dout(" page %p snapc %p not current, but oldest\n",
1077 		     page, snapc);
1078 		if (!clear_page_dirty_for_io(page))
1079 			goto retry_locked;
1080 		r = writepage_nounlock(page, NULL);
1081 		if (r < 0)
1082 			goto fail_nosnap;
1083 		goto retry_locked;
1084 	}
1085 
1086 	if (PageUptodate(page)) {
1087 		dout(" page %p already uptodate\n", page);
1088 		return 0;
1089 	}
1090 
1091 	/* full page? */
1092 	if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1093 		return 0;
1094 
1095 	/* past end of file? */
1096 	i_size = inode->i_size;   /* caller holds i_mutex */
1097 
1098 	if (page_off >= i_size ||
1099 	    (pos_in_page == 0 && (pos+len) >= i_size &&
1100 	     end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1101 		dout(" zeroing %p 0 - %d and %d - %d\n",
1102 		     page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1103 		zero_user_segments(page,
1104 				   0, pos_in_page,
1105 				   end_in_page, PAGE_CACHE_SIZE);
1106 		return 0;
1107 	}
1108 
1109 	/* we need to read it. */
1110 	up_read(&mdsc->snap_rwsem);
1111 	r = readpage_nounlock(file, page);
1112 	if (r < 0)
1113 		goto fail_nosnap;
1114 	goto retry_locked;
1115 fail_nosnap:
1116 	unlock_page(page);
1117 	return r;
1118 }
1119 
1120 /*
1121  * We are only allowed to write into/dirty the page if the page is
1122  * clean, or already dirty within the same snap context.
1123  */
1124 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1125 			    loff_t pos, unsigned len, unsigned flags,
1126 			    struct page **pagep, void **fsdata)
1127 {
1128 	struct inode *inode = file_inode(file);
1129 	struct page *page;
1130 	pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1131 	int r;
1132 
1133 	do {
1134 		/* get a page */
1135 		page = grab_cache_page_write_begin(mapping, index, 0);
1136 		if (!page)
1137 			return -ENOMEM;
1138 		*pagep = page;
1139 
1140 		dout("write_begin file %p inode %p page %p %d~%d\n", file,
1141 		     inode, page, (int)pos, (int)len);
1142 
1143 		r = ceph_update_writeable_page(file, pos, len, page);
1144 	} while (r == -EAGAIN);
1145 
1146 	return r;
1147 }
1148 
1149 /*
1150  * we don't do anything in here that simple_write_end doesn't do
1151  * except adjust dirty page accounting and drop read lock on
1152  * mdsc->snap_rwsem.
1153  */
1154 static int ceph_write_end(struct file *file, struct address_space *mapping,
1155 			  loff_t pos, unsigned len, unsigned copied,
1156 			  struct page *page, void *fsdata)
1157 {
1158 	struct inode *inode = file_inode(file);
1159 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1160 	struct ceph_mds_client *mdsc = fsc->mdsc;
1161 	unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1162 	int check_cap = 0;
1163 
1164 	dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1165 	     inode, page, (int)pos, (int)copied, (int)len);
1166 
1167 	/* zero the stale part of the page if we did a short copy */
1168 	if (copied < len)
1169 		zero_user_segment(page, from+copied, len);
1170 
1171 	/* did file size increase? */
1172 	/* (no need for i_size_read(); we caller holds i_mutex */
1173 	if (pos+copied > inode->i_size)
1174 		check_cap = ceph_inode_set_size(inode, pos+copied);
1175 
1176 	if (!PageUptodate(page))
1177 		SetPageUptodate(page);
1178 
1179 	set_page_dirty(page);
1180 
1181 	unlock_page(page);
1182 	up_read(&mdsc->snap_rwsem);
1183 	page_cache_release(page);
1184 
1185 	if (check_cap)
1186 		ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1187 
1188 	return copied;
1189 }
1190 
1191 /*
1192  * we set .direct_IO to indicate direct io is supported, but since we
1193  * intercept O_DIRECT reads and writes early, this function should
1194  * never get called.
1195  */
1196 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1197 			      struct iov_iter *iter,
1198 			      loff_t pos)
1199 {
1200 	WARN_ON(1);
1201 	return -EINVAL;
1202 }
1203 
1204 const struct address_space_operations ceph_aops = {
1205 	.readpage = ceph_readpage,
1206 	.readpages = ceph_readpages,
1207 	.writepage = ceph_writepage,
1208 	.writepages = ceph_writepages_start,
1209 	.write_begin = ceph_write_begin,
1210 	.write_end = ceph_write_end,
1211 	.set_page_dirty = ceph_set_page_dirty,
1212 	.invalidatepage = ceph_invalidatepage,
1213 	.releasepage = ceph_releasepage,
1214 	.direct_IO = ceph_direct_io,
1215 };
1216 
1217 
1218 /*
1219  * vm ops
1220  */
1221 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1222 {
1223 	struct inode *inode = file_inode(vma->vm_file);
1224 	struct ceph_inode_info *ci = ceph_inode(inode);
1225 	struct ceph_file_info *fi = vma->vm_file->private_data;
1226 	struct page *pinned_page = NULL;
1227 	loff_t off = vmf->pgoff << PAGE_CACHE_SHIFT;
1228 	int want, got, ret;
1229 
1230 	dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1231 	     inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE);
1232 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1233 		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1234 	else
1235 		want = CEPH_CAP_FILE_CACHE;
1236 	while (1) {
1237 		got = 0;
1238 		ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want,
1239 				    -1, &got, &pinned_page);
1240 		if (ret == 0)
1241 			break;
1242 		if (ret != -ERESTARTSYS) {
1243 			WARN_ON(1);
1244 			return VM_FAULT_SIGBUS;
1245 		}
1246 	}
1247 	dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1248 	     inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got));
1249 
1250 	if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1251 	    ci->i_inline_version == CEPH_INLINE_NONE)
1252 		ret = filemap_fault(vma, vmf);
1253 	else
1254 		ret = -EAGAIN;
1255 
1256 	dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1257 	     inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret);
1258 	if (pinned_page)
1259 		page_cache_release(pinned_page);
1260 	ceph_put_cap_refs(ci, got);
1261 
1262 	if (ret != -EAGAIN)
1263 		return ret;
1264 
1265 	/* read inline data */
1266 	if (off >= PAGE_CACHE_SIZE) {
1267 		/* does not support inline data > PAGE_SIZE */
1268 		ret = VM_FAULT_SIGBUS;
1269 	} else {
1270 		int ret1;
1271 		struct address_space *mapping = inode->i_mapping;
1272 		struct page *page = find_or_create_page(mapping, 0,
1273 						mapping_gfp_mask(mapping) &
1274 						~__GFP_FS);
1275 		if (!page) {
1276 			ret = VM_FAULT_OOM;
1277 			goto out;
1278 		}
1279 		ret1 = __ceph_do_getattr(inode, page,
1280 					 CEPH_STAT_CAP_INLINE_DATA, true);
1281 		if (ret1 < 0 || off >= i_size_read(inode)) {
1282 			unlock_page(page);
1283 			page_cache_release(page);
1284 			ret = VM_FAULT_SIGBUS;
1285 			goto out;
1286 		}
1287 		if (ret1 < PAGE_CACHE_SIZE)
1288 			zero_user_segment(page, ret1, PAGE_CACHE_SIZE);
1289 		else
1290 			flush_dcache_page(page);
1291 		SetPageUptodate(page);
1292 		vmf->page = page;
1293 		ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1294 	}
1295 out:
1296 	dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1297 	     inode, off, (size_t)PAGE_CACHE_SIZE, ret);
1298 	return ret;
1299 }
1300 
1301 /*
1302  * Reuse write_begin here for simplicity.
1303  */
1304 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1305 {
1306 	struct inode *inode = file_inode(vma->vm_file);
1307 	struct ceph_inode_info *ci = ceph_inode(inode);
1308 	struct ceph_file_info *fi = vma->vm_file->private_data;
1309 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1310 	struct page *page = vmf->page;
1311 	loff_t off = page_offset(page);
1312 	loff_t size = i_size_read(inode);
1313 	size_t len;
1314 	int want, got, ret;
1315 
1316 	if (ci->i_inline_version != CEPH_INLINE_NONE) {
1317 		struct page *locked_page = NULL;
1318 		if (off == 0) {
1319 			lock_page(page);
1320 			locked_page = page;
1321 		}
1322 		ret = ceph_uninline_data(vma->vm_file, locked_page);
1323 		if (locked_page)
1324 			unlock_page(locked_page);
1325 		if (ret < 0)
1326 			return VM_FAULT_SIGBUS;
1327 	}
1328 
1329 	if (off + PAGE_CACHE_SIZE <= size)
1330 		len = PAGE_CACHE_SIZE;
1331 	else
1332 		len = size & ~PAGE_CACHE_MASK;
1333 
1334 	dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1335 	     inode, ceph_vinop(inode), off, len, size);
1336 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1337 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1338 	else
1339 		want = CEPH_CAP_FILE_BUFFER;
1340 	while (1) {
1341 		got = 0;
1342 		ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1343 				    &got, NULL);
1344 		if (ret == 0)
1345 			break;
1346 		if (ret != -ERESTARTSYS) {
1347 			WARN_ON(1);
1348 			return VM_FAULT_SIGBUS;
1349 		}
1350 	}
1351 	dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1352 	     inode, off, len, ceph_cap_string(got));
1353 
1354 	/* Update time before taking page lock */
1355 	file_update_time(vma->vm_file);
1356 
1357 	lock_page(page);
1358 
1359 	ret = VM_FAULT_NOPAGE;
1360 	if ((off > size) ||
1361 	    (page->mapping != inode->i_mapping))
1362 		goto out;
1363 
1364 	ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1365 	if (ret == 0) {
1366 		/* success.  we'll keep the page locked. */
1367 		set_page_dirty(page);
1368 		up_read(&mdsc->snap_rwsem);
1369 		ret = VM_FAULT_LOCKED;
1370 	} else {
1371 		if (ret == -ENOMEM)
1372 			ret = VM_FAULT_OOM;
1373 		else
1374 			ret = VM_FAULT_SIGBUS;
1375 	}
1376 out:
1377 	if (ret != VM_FAULT_LOCKED)
1378 		unlock_page(page);
1379 	if (ret == VM_FAULT_LOCKED ||
1380 	    ci->i_inline_version != CEPH_INLINE_NONE) {
1381 		int dirty;
1382 		spin_lock(&ci->i_ceph_lock);
1383 		ci->i_inline_version = CEPH_INLINE_NONE;
1384 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
1385 		spin_unlock(&ci->i_ceph_lock);
1386 		if (dirty)
1387 			__mark_inode_dirty(inode, dirty);
1388 	}
1389 
1390 	dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1391 	     inode, off, len, ceph_cap_string(got), ret);
1392 	ceph_put_cap_refs(ci, got);
1393 
1394 	return ret;
1395 }
1396 
1397 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1398 			   char	*data, size_t len)
1399 {
1400 	struct address_space *mapping = inode->i_mapping;
1401 	struct page *page;
1402 
1403 	if (locked_page) {
1404 		page = locked_page;
1405 	} else {
1406 		if (i_size_read(inode) == 0)
1407 			return;
1408 		page = find_or_create_page(mapping, 0,
1409 					   mapping_gfp_mask(mapping) & ~__GFP_FS);
1410 		if (!page)
1411 			return;
1412 		if (PageUptodate(page)) {
1413 			unlock_page(page);
1414 			page_cache_release(page);
1415 			return;
1416 		}
1417 	}
1418 
1419 	dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1420 	     inode, ceph_vinop(inode), len, locked_page);
1421 
1422 	if (len > 0) {
1423 		void *kaddr = kmap_atomic(page);
1424 		memcpy(kaddr, data, len);
1425 		kunmap_atomic(kaddr);
1426 	}
1427 
1428 	if (page != locked_page) {
1429 		if (len < PAGE_CACHE_SIZE)
1430 			zero_user_segment(page, len, PAGE_CACHE_SIZE);
1431 		else
1432 			flush_dcache_page(page);
1433 
1434 		SetPageUptodate(page);
1435 		unlock_page(page);
1436 		page_cache_release(page);
1437 	}
1438 }
1439 
1440 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1441 {
1442 	struct inode *inode = file_inode(filp);
1443 	struct ceph_inode_info *ci = ceph_inode(inode);
1444 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1445 	struct ceph_osd_request *req;
1446 	struct page *page = NULL;
1447 	u64 len, inline_version;
1448 	int err = 0;
1449 	bool from_pagecache = false;
1450 
1451 	spin_lock(&ci->i_ceph_lock);
1452 	inline_version = ci->i_inline_version;
1453 	spin_unlock(&ci->i_ceph_lock);
1454 
1455 	dout("uninline_data %p %llx.%llx inline_version %llu\n",
1456 	     inode, ceph_vinop(inode), inline_version);
1457 
1458 	if (inline_version == 1 || /* initial version, no data */
1459 	    inline_version == CEPH_INLINE_NONE)
1460 		goto out;
1461 
1462 	if (locked_page) {
1463 		page = locked_page;
1464 		WARN_ON(!PageUptodate(page));
1465 	} else if (ceph_caps_issued(ci) &
1466 		   (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1467 		page = find_get_page(inode->i_mapping, 0);
1468 		if (page) {
1469 			if (PageUptodate(page)) {
1470 				from_pagecache = true;
1471 				lock_page(page);
1472 			} else {
1473 				page_cache_release(page);
1474 				page = NULL;
1475 			}
1476 		}
1477 	}
1478 
1479 	if (page) {
1480 		len = i_size_read(inode);
1481 		if (len > PAGE_CACHE_SIZE)
1482 			len = PAGE_CACHE_SIZE;
1483 	} else {
1484 		page = __page_cache_alloc(GFP_NOFS);
1485 		if (!page) {
1486 			err = -ENOMEM;
1487 			goto out;
1488 		}
1489 		err = __ceph_do_getattr(inode, page,
1490 					CEPH_STAT_CAP_INLINE_DATA, true);
1491 		if (err < 0) {
1492 			/* no inline data */
1493 			if (err == -ENODATA)
1494 				err = 0;
1495 			goto out;
1496 		}
1497 		len = err;
1498 	}
1499 
1500 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1501 				    ceph_vino(inode), 0, &len, 0, 1,
1502 				    CEPH_OSD_OP_CREATE,
1503 				    CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1504 				    ci->i_snap_realm->cached_context,
1505 				    0, 0, false);
1506 	if (IS_ERR(req)) {
1507 		err = PTR_ERR(req);
1508 		goto out;
1509 	}
1510 
1511 	ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1512 	err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1513 	if (!err)
1514 		err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1515 	ceph_osdc_put_request(req);
1516 	if (err < 0)
1517 		goto out;
1518 
1519 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1520 				    ceph_vino(inode), 0, &len, 1, 3,
1521 				    CEPH_OSD_OP_WRITE,
1522 				    CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1523 				    ci->i_snap_realm->cached_context,
1524 				    ci->i_truncate_seq, ci->i_truncate_size,
1525 				    false);
1526 	if (IS_ERR(req)) {
1527 		err = PTR_ERR(req);
1528 		goto out;
1529 	}
1530 
1531 	osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1532 
1533 	err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1534 				    "inline_version", &inline_version,
1535 				    sizeof(inline_version),
1536 				    CEPH_OSD_CMPXATTR_OP_GT,
1537 				    CEPH_OSD_CMPXATTR_MODE_U64);
1538 	if (err)
1539 		goto out_put;
1540 
1541 	err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1542 				    "inline_version", &inline_version,
1543 				    sizeof(inline_version), 0, 0);
1544 	if (err)
1545 		goto out_put;
1546 
1547 	ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1548 	err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1549 	if (!err)
1550 		err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1551 out_put:
1552 	ceph_osdc_put_request(req);
1553 	if (err == -ECANCELED)
1554 		err = 0;
1555 out:
1556 	if (page && page != locked_page) {
1557 		if (from_pagecache) {
1558 			unlock_page(page);
1559 			page_cache_release(page);
1560 		} else
1561 			__free_pages(page, 0);
1562 	}
1563 
1564 	dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1565 	     inode, ceph_vinop(inode), inline_version, err);
1566 	return err;
1567 }
1568 
1569 static struct vm_operations_struct ceph_vmops = {
1570 	.fault		= ceph_filemap_fault,
1571 	.page_mkwrite	= ceph_page_mkwrite,
1572 	.remap_pages	= generic_file_remap_pages,
1573 };
1574 
1575 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1576 {
1577 	struct address_space *mapping = file->f_mapping;
1578 
1579 	if (!mapping->a_ops->readpage)
1580 		return -ENOEXEC;
1581 	file_accessed(file);
1582 	vma->vm_ops = &ceph_vmops;
1583 	return 0;
1584 }
1585