xref: /linux/fs/ceph/addr.c (revision f8e17c17b81070f38062dce79ca7f4541851dadd)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3 
4 #include <linux/backing-dev.h>
5 #include <linux/fs.h>
6 #include <linux/mm.h>
7 #include <linux/pagemap.h>
8 #include <linux/writeback.h>	/* generic_writepages */
9 #include <linux/slab.h>
10 #include <linux/pagevec.h>
11 #include <linux/task_io_accounting_ops.h>
12 #include <linux/signal.h>
13 #include <linux/iversion.h>
14 
15 #include "super.h"
16 #include "mds_client.h"
17 #include "cache.h"
18 #include <linux/ceph/osd_client.h>
19 #include <linux/ceph/striper.h>
20 
21 /*
22  * Ceph address space ops.
23  *
24  * There are a few funny things going on here.
25  *
26  * The page->private field is used to reference a struct
27  * ceph_snap_context for _every_ dirty page.  This indicates which
28  * snapshot the page was logically dirtied in, and thus which snap
29  * context needs to be associated with the osd write during writeback.
30  *
31  * Similarly, struct ceph_inode_info maintains a set of counters to
32  * count dirty pages on the inode.  In the absence of snapshots,
33  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
34  *
35  * When a snapshot is taken (that is, when the client receives
36  * notification that a snapshot was taken), each inode with caps and
37  * with dirty pages (dirty pages implies there is a cap) gets a new
38  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
39  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
40  * moved to capsnap->dirty. (Unless a sync write is currently in
41  * progress.  In that case, the capsnap is said to be "pending", new
42  * writes cannot start, and the capsnap isn't "finalized" until the
43  * write completes (or fails) and a final size/mtime for the inode for
44  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
45  *
46  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
47  * we look for the first capsnap in i_cap_snaps and write out pages in
48  * that snap context _only_.  Then we move on to the next capsnap,
49  * eventually reaching the "live" or "head" context (i.e., pages that
50  * are not yet snapped) and are writing the most recently dirtied
51  * pages.
52  *
53  * Invalidate and so forth must take care to ensure the dirty page
54  * accounting is preserved.
55  */
56 
57 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
58 #define CONGESTION_OFF_THRESH(congestion_kb)				\
59 	(CONGESTION_ON_THRESH(congestion_kb) -				\
60 	 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
61 
62 static inline struct ceph_snap_context *page_snap_context(struct page *page)
63 {
64 	if (PagePrivate(page))
65 		return (void *)page->private;
66 	return NULL;
67 }
68 
69 /*
70  * Dirty a page.  Optimistically adjust accounting, on the assumption
71  * that we won't race with invalidate.  If we do, readjust.
72  */
73 static int ceph_set_page_dirty(struct page *page)
74 {
75 	struct address_space *mapping = page->mapping;
76 	struct inode *inode;
77 	struct ceph_inode_info *ci;
78 	struct ceph_snap_context *snapc;
79 	int ret;
80 
81 	if (unlikely(!mapping))
82 		return !TestSetPageDirty(page);
83 
84 	if (PageDirty(page)) {
85 		dout("%p set_page_dirty %p idx %lu -- already dirty\n",
86 		     mapping->host, page, page->index);
87 		BUG_ON(!PagePrivate(page));
88 		return 0;
89 	}
90 
91 	inode = mapping->host;
92 	ci = ceph_inode(inode);
93 
94 	/* dirty the head */
95 	spin_lock(&ci->i_ceph_lock);
96 	BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
97 	if (__ceph_have_pending_cap_snap(ci)) {
98 		struct ceph_cap_snap *capsnap =
99 				list_last_entry(&ci->i_cap_snaps,
100 						struct ceph_cap_snap,
101 						ci_item);
102 		snapc = ceph_get_snap_context(capsnap->context);
103 		capsnap->dirty_pages++;
104 	} else {
105 		BUG_ON(!ci->i_head_snapc);
106 		snapc = ceph_get_snap_context(ci->i_head_snapc);
107 		++ci->i_wrbuffer_ref_head;
108 	}
109 	if (ci->i_wrbuffer_ref == 0)
110 		ihold(inode);
111 	++ci->i_wrbuffer_ref;
112 	dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
113 	     "snapc %p seq %lld (%d snaps)\n",
114 	     mapping->host, page, page->index,
115 	     ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
116 	     ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
117 	     snapc, snapc->seq, snapc->num_snaps);
118 	spin_unlock(&ci->i_ceph_lock);
119 
120 	/*
121 	 * Reference snap context in page->private.  Also set
122 	 * PagePrivate so that we get invalidatepage callback.
123 	 */
124 	BUG_ON(PagePrivate(page));
125 	page->private = (unsigned long)snapc;
126 	SetPagePrivate(page);
127 
128 	ret = __set_page_dirty_nobuffers(page);
129 	WARN_ON(!PageLocked(page));
130 	WARN_ON(!page->mapping);
131 
132 	return ret;
133 }
134 
135 /*
136  * If we are truncating the full page (i.e. offset == 0), adjust the
137  * dirty page counters appropriately.  Only called if there is private
138  * data on the page.
139  */
140 static void ceph_invalidatepage(struct page *page, unsigned int offset,
141 				unsigned int length)
142 {
143 	struct inode *inode;
144 	struct ceph_inode_info *ci;
145 	struct ceph_snap_context *snapc = page_snap_context(page);
146 
147 	inode = page->mapping->host;
148 	ci = ceph_inode(inode);
149 
150 	if (offset != 0 || length != PAGE_SIZE) {
151 		dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
152 		     inode, page, page->index, offset, length);
153 		return;
154 	}
155 
156 	ceph_invalidate_fscache_page(inode, page);
157 
158 	WARN_ON(!PageLocked(page));
159 	if (!PagePrivate(page))
160 		return;
161 
162 	ClearPageChecked(page);
163 
164 	dout("%p invalidatepage %p idx %lu full dirty page\n",
165 	     inode, page, page->index);
166 
167 	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
168 	ceph_put_snap_context(snapc);
169 	page->private = 0;
170 	ClearPagePrivate(page);
171 }
172 
173 static int ceph_releasepage(struct page *page, gfp_t g)
174 {
175 	dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
176 	     page, page->index, PageDirty(page) ? "" : "not ");
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 ceph_do_readpage(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_fs_client *fsc = ceph_inode_to_client(inode);
193 	int err = 0;
194 	u64 off = page_offset(page);
195 	u64 len = PAGE_SIZE;
196 
197 	if (off >= i_size_read(inode)) {
198 		zero_user_segment(page, 0, PAGE_SIZE);
199 		SetPageUptodate(page);
200 		return 0;
201 	}
202 
203 	if (ci->i_inline_version != CEPH_INLINE_NONE) {
204 		/*
205 		 * Uptodate inline data should have been added
206 		 * into page cache while getting Fcr caps.
207 		 */
208 		if (off == 0)
209 			return -EINVAL;
210 		zero_user_segment(page, 0, PAGE_SIZE);
211 		SetPageUptodate(page);
212 		return 0;
213 	}
214 
215 	err = ceph_readpage_from_fscache(inode, page);
216 	if (err == 0)
217 		return -EINPROGRESS;
218 
219 	dout("readpage inode %p file %p page %p index %lu\n",
220 	     inode, filp, page, page->index);
221 	err = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
222 				  &ci->i_layout, off, &len,
223 				  ci->i_truncate_seq, ci->i_truncate_size,
224 				  &page, 1, 0);
225 	if (err == -ENOENT)
226 		err = 0;
227 	if (err < 0) {
228 		SetPageError(page);
229 		ceph_fscache_readpage_cancel(inode, page);
230 		if (err == -EBLACKLISTED)
231 			fsc->blacklisted = true;
232 		goto out;
233 	}
234 	if (err < PAGE_SIZE)
235 		/* zero fill remainder of page */
236 		zero_user_segment(page, err, PAGE_SIZE);
237 	else
238 		flush_dcache_page(page);
239 
240 	SetPageUptodate(page);
241 	ceph_readpage_to_fscache(inode, page);
242 
243 out:
244 	return err < 0 ? err : 0;
245 }
246 
247 static int ceph_readpage(struct file *filp, struct page *page)
248 {
249 	int r = ceph_do_readpage(filp, page);
250 	if (r != -EINPROGRESS)
251 		unlock_page(page);
252 	else
253 		r = 0;
254 	return r;
255 }
256 
257 /*
258  * Finish an async read(ahead) op.
259  */
260 static void finish_read(struct ceph_osd_request *req)
261 {
262 	struct inode *inode = req->r_inode;
263 	struct ceph_osd_data *osd_data;
264 	int rc = req->r_result <= 0 ? req->r_result : 0;
265 	int bytes = req->r_result >= 0 ? req->r_result : 0;
266 	int num_pages;
267 	int i;
268 
269 	dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
270 	if (rc == -EBLACKLISTED)
271 		ceph_inode_to_client(inode)->blacklisted = true;
272 
273 	/* unlock all pages, zeroing any data we didn't read */
274 	osd_data = osd_req_op_extent_osd_data(req, 0);
275 	BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
276 	num_pages = calc_pages_for((u64)osd_data->alignment,
277 					(u64)osd_data->length);
278 	for (i = 0; i < num_pages; i++) {
279 		struct page *page = osd_data->pages[i];
280 
281 		if (rc < 0 && rc != -ENOENT) {
282 			ceph_fscache_readpage_cancel(inode, page);
283 			goto unlock;
284 		}
285 		if (bytes < (int)PAGE_SIZE) {
286 			/* zero (remainder of) page */
287 			int s = bytes < 0 ? 0 : bytes;
288 			zero_user_segment(page, s, PAGE_SIZE);
289 		}
290  		dout("finish_read %p uptodate %p idx %lu\n", inode, page,
291 		     page->index);
292 		flush_dcache_page(page);
293 		SetPageUptodate(page);
294 		ceph_readpage_to_fscache(inode, page);
295 unlock:
296 		unlock_page(page);
297 		put_page(page);
298 		bytes -= PAGE_SIZE;
299 	}
300 	kfree(osd_data->pages);
301 }
302 
303 /*
304  * start an async read(ahead) operation.  return nr_pages we submitted
305  * a read for on success, or negative error code.
306  */
307 static int start_read(struct inode *inode, struct ceph_rw_context *rw_ctx,
308 		      struct list_head *page_list, int max)
309 {
310 	struct ceph_osd_client *osdc =
311 		&ceph_inode_to_client(inode)->client->osdc;
312 	struct ceph_inode_info *ci = ceph_inode(inode);
313 	struct page *page = lru_to_page(page_list);
314 	struct ceph_vino vino;
315 	struct ceph_osd_request *req;
316 	u64 off;
317 	u64 len;
318 	int i;
319 	struct page **pages;
320 	pgoff_t next_index;
321 	int nr_pages = 0;
322 	int got = 0;
323 	int ret = 0;
324 
325 	if (!rw_ctx) {
326 		/* caller of readpages does not hold buffer and read caps
327 		 * (fadvise, madvise and readahead cases) */
328 		int want = CEPH_CAP_FILE_CACHE;
329 		ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want,
330 					true, &got);
331 		if (ret < 0) {
332 			dout("start_read %p, error getting cap\n", inode);
333 		} else if (!(got & want)) {
334 			dout("start_read %p, no cache cap\n", inode);
335 			ret = 0;
336 		}
337 		if (ret <= 0) {
338 			if (got)
339 				ceph_put_cap_refs(ci, got);
340 			while (!list_empty(page_list)) {
341 				page = lru_to_page(page_list);
342 				list_del(&page->lru);
343 				put_page(page);
344 			}
345 			return ret;
346 		}
347 	}
348 
349 	off = (u64) page_offset(page);
350 
351 	/* count pages */
352 	next_index = page->index;
353 	list_for_each_entry_reverse(page, page_list, lru) {
354 		if (page->index != next_index)
355 			break;
356 		nr_pages++;
357 		next_index++;
358 		if (max && nr_pages == max)
359 			break;
360 	}
361 	len = nr_pages << PAGE_SHIFT;
362 	dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
363 	     off, len);
364 	vino = ceph_vino(inode);
365 	req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
366 				    0, 1, CEPH_OSD_OP_READ,
367 				    CEPH_OSD_FLAG_READ, NULL,
368 				    ci->i_truncate_seq, ci->i_truncate_size,
369 				    false);
370 	if (IS_ERR(req)) {
371 		ret = PTR_ERR(req);
372 		goto out;
373 	}
374 
375 	/* build page vector */
376 	nr_pages = calc_pages_for(0, len);
377 	pages = kmalloc_array(nr_pages, sizeof(*pages), GFP_KERNEL);
378 	if (!pages) {
379 		ret = -ENOMEM;
380 		goto out_put;
381 	}
382 	for (i = 0; i < nr_pages; ++i) {
383 		page = list_entry(page_list->prev, struct page, lru);
384 		BUG_ON(PageLocked(page));
385 		list_del(&page->lru);
386 
387  		dout("start_read %p adding %p idx %lu\n", inode, page,
388 		     page->index);
389 		if (add_to_page_cache_lru(page, &inode->i_data, page->index,
390 					  GFP_KERNEL)) {
391 			ceph_fscache_uncache_page(inode, page);
392 			put_page(page);
393 			dout("start_read %p add_to_page_cache failed %p\n",
394 			     inode, page);
395 			nr_pages = i;
396 			if (nr_pages > 0) {
397 				len = nr_pages << PAGE_SHIFT;
398 				osd_req_op_extent_update(req, 0, len);
399 				break;
400 			}
401 			goto out_pages;
402 		}
403 		pages[i] = page;
404 	}
405 	osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
406 	req->r_callback = finish_read;
407 	req->r_inode = inode;
408 
409 	dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
410 	ret = ceph_osdc_start_request(osdc, req, false);
411 	if (ret < 0)
412 		goto out_pages;
413 	ceph_osdc_put_request(req);
414 
415 	/* After adding locked pages to page cache, the inode holds cache cap.
416 	 * So we can drop our cap refs. */
417 	if (got)
418 		ceph_put_cap_refs(ci, got);
419 
420 	return nr_pages;
421 
422 out_pages:
423 	for (i = 0; i < nr_pages; ++i) {
424 		ceph_fscache_readpage_cancel(inode, pages[i]);
425 		unlock_page(pages[i]);
426 	}
427 	ceph_put_page_vector(pages, nr_pages, false);
428 out_put:
429 	ceph_osdc_put_request(req);
430 out:
431 	if (got)
432 		ceph_put_cap_refs(ci, got);
433 	return ret;
434 }
435 
436 
437 /*
438  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
439  * the caller (VM) cleans them up.
440  */
441 static int ceph_readpages(struct file *file, struct address_space *mapping,
442 			  struct list_head *page_list, unsigned nr_pages)
443 {
444 	struct inode *inode = file_inode(file);
445 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
446 	struct ceph_file_info *fi = file->private_data;
447 	struct ceph_rw_context *rw_ctx;
448 	int rc = 0;
449 	int max = 0;
450 
451 	if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
452 		return -EINVAL;
453 
454 	rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
455 					 &nr_pages);
456 
457 	if (rc == 0)
458 		goto out;
459 
460 	rw_ctx = ceph_find_rw_context(fi);
461 	max = fsc->mount_options->rsize >> PAGE_SHIFT;
462 	dout("readpages %p file %p ctx %p nr_pages %d max %d\n",
463 	     inode, file, rw_ctx, nr_pages, max);
464 	while (!list_empty(page_list)) {
465 		rc = start_read(inode, rw_ctx, page_list, max);
466 		if (rc < 0)
467 			goto out;
468 	}
469 out:
470 	ceph_fscache_readpages_cancel(inode, page_list);
471 
472 	dout("readpages %p file %p ret %d\n", inode, file, rc);
473 	return rc;
474 }
475 
476 struct ceph_writeback_ctl
477 {
478 	loff_t i_size;
479 	u64 truncate_size;
480 	u32 truncate_seq;
481 	bool size_stable;
482 	bool head_snapc;
483 };
484 
485 /*
486  * Get ref for the oldest snapc for an inode with dirty data... that is, the
487  * only snap context we are allowed to write back.
488  */
489 static struct ceph_snap_context *
490 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
491 		   struct ceph_snap_context *page_snapc)
492 {
493 	struct ceph_inode_info *ci = ceph_inode(inode);
494 	struct ceph_snap_context *snapc = NULL;
495 	struct ceph_cap_snap *capsnap = NULL;
496 
497 	spin_lock(&ci->i_ceph_lock);
498 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
499 		dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
500 		     capsnap->context, capsnap->dirty_pages);
501 		if (!capsnap->dirty_pages)
502 			continue;
503 
504 		/* get i_size, truncate_{seq,size} for page_snapc? */
505 		if (snapc && capsnap->context != page_snapc)
506 			continue;
507 
508 		if (ctl) {
509 			if (capsnap->writing) {
510 				ctl->i_size = i_size_read(inode);
511 				ctl->size_stable = false;
512 			} else {
513 				ctl->i_size = capsnap->size;
514 				ctl->size_stable = true;
515 			}
516 			ctl->truncate_size = capsnap->truncate_size;
517 			ctl->truncate_seq = capsnap->truncate_seq;
518 			ctl->head_snapc = false;
519 		}
520 
521 		if (snapc)
522 			break;
523 
524 		snapc = ceph_get_snap_context(capsnap->context);
525 		if (!page_snapc ||
526 		    page_snapc == snapc ||
527 		    page_snapc->seq > snapc->seq)
528 			break;
529 	}
530 	if (!snapc && ci->i_wrbuffer_ref_head) {
531 		snapc = ceph_get_snap_context(ci->i_head_snapc);
532 		dout(" head snapc %p has %d dirty pages\n",
533 		     snapc, ci->i_wrbuffer_ref_head);
534 		if (ctl) {
535 			ctl->i_size = i_size_read(inode);
536 			ctl->truncate_size = ci->i_truncate_size;
537 			ctl->truncate_seq = ci->i_truncate_seq;
538 			ctl->size_stable = false;
539 			ctl->head_snapc = true;
540 		}
541 	}
542 	spin_unlock(&ci->i_ceph_lock);
543 	return snapc;
544 }
545 
546 static u64 get_writepages_data_length(struct inode *inode,
547 				      struct page *page, u64 start)
548 {
549 	struct ceph_inode_info *ci = ceph_inode(inode);
550 	struct ceph_snap_context *snapc = page_snap_context(page);
551 	struct ceph_cap_snap *capsnap = NULL;
552 	u64 end = i_size_read(inode);
553 
554 	if (snapc != ci->i_head_snapc) {
555 		bool found = false;
556 		spin_lock(&ci->i_ceph_lock);
557 		list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
558 			if (capsnap->context == snapc) {
559 				if (!capsnap->writing)
560 					end = capsnap->size;
561 				found = true;
562 				break;
563 			}
564 		}
565 		spin_unlock(&ci->i_ceph_lock);
566 		WARN_ON(!found);
567 	}
568 	if (end > page_offset(page) + PAGE_SIZE)
569 		end = page_offset(page) + PAGE_SIZE;
570 	return end > start ? end - start : 0;
571 }
572 
573 /*
574  * Write a single page, but leave the page locked.
575  *
576  * If we get a write error, mark the mapping for error, but still adjust the
577  * dirty page accounting (i.e., page is no longer dirty).
578  */
579 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
580 {
581 	struct inode *inode;
582 	struct ceph_inode_info *ci;
583 	struct ceph_fs_client *fsc;
584 	struct ceph_snap_context *snapc, *oldest;
585 	loff_t page_off = page_offset(page);
586 	int err, len = PAGE_SIZE;
587 	struct ceph_writeback_ctl ceph_wbc;
588 
589 	dout("writepage %p idx %lu\n", page, page->index);
590 
591 	inode = page->mapping->host;
592 	ci = ceph_inode(inode);
593 	fsc = ceph_inode_to_client(inode);
594 
595 	/* verify this is a writeable snap context */
596 	snapc = page_snap_context(page);
597 	if (!snapc) {
598 		dout("writepage %p page %p not dirty?\n", inode, page);
599 		return 0;
600 	}
601 	oldest = get_oldest_context(inode, &ceph_wbc, snapc);
602 	if (snapc->seq > oldest->seq) {
603 		dout("writepage %p page %p snapc %p not writeable - noop\n",
604 		     inode, page, snapc);
605 		/* we should only noop if called by kswapd */
606 		WARN_ON(!(current->flags & PF_MEMALLOC));
607 		ceph_put_snap_context(oldest);
608 		redirty_page_for_writepage(wbc, page);
609 		return 0;
610 	}
611 	ceph_put_snap_context(oldest);
612 
613 	/* is this a partial page at end of file? */
614 	if (page_off >= ceph_wbc.i_size) {
615 		dout("%p page eof %llu\n", page, ceph_wbc.i_size);
616 		page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
617 		return 0;
618 	}
619 
620 	if (ceph_wbc.i_size < page_off + len)
621 		len = ceph_wbc.i_size - page_off;
622 
623 	dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
624 	     inode, page, page->index, page_off, len, snapc, snapc->seq);
625 
626 	if (atomic_long_inc_return(&fsc->writeback_count) >
627 	    CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
628 		set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
629 
630 	set_page_writeback(page);
631 	err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
632 				   &ci->i_layout, snapc, page_off, len,
633 				   ceph_wbc.truncate_seq,
634 				   ceph_wbc.truncate_size,
635 				   &inode->i_mtime, &page, 1);
636 	if (err < 0) {
637 		struct writeback_control tmp_wbc;
638 		if (!wbc)
639 			wbc = &tmp_wbc;
640 		if (err == -ERESTARTSYS) {
641 			/* killed by SIGKILL */
642 			dout("writepage interrupted page %p\n", page);
643 			redirty_page_for_writepage(wbc, page);
644 			end_page_writeback(page);
645 			return err;
646 		}
647 		if (err == -EBLACKLISTED)
648 			fsc->blacklisted = true;
649 		dout("writepage setting page/mapping error %d %p\n",
650 		     err, page);
651 		mapping_set_error(&inode->i_data, err);
652 		wbc->pages_skipped++;
653 	} else {
654 		dout("writepage cleaned page %p\n", page);
655 		err = 0;  /* vfs expects us to return 0 */
656 	}
657 	page->private = 0;
658 	ClearPagePrivate(page);
659 	end_page_writeback(page);
660 	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
661 	ceph_put_snap_context(snapc);  /* page's reference */
662 
663 	if (atomic_long_dec_return(&fsc->writeback_count) <
664 	    CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
665 		clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
666 
667 	return err;
668 }
669 
670 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
671 {
672 	int err;
673 	struct inode *inode = page->mapping->host;
674 	BUG_ON(!inode);
675 	ihold(inode);
676 	err = writepage_nounlock(page, wbc);
677 	if (err == -ERESTARTSYS) {
678 		/* direct memory reclaimer was killed by SIGKILL. return 0
679 		 * to prevent caller from setting mapping/page error */
680 		err = 0;
681 	}
682 	unlock_page(page);
683 	iput(inode);
684 	return err;
685 }
686 
687 /*
688  * async writeback completion handler.
689  *
690  * If we get an error, set the mapping error bit, but not the individual
691  * page error bits.
692  */
693 static void writepages_finish(struct ceph_osd_request *req)
694 {
695 	struct inode *inode = req->r_inode;
696 	struct ceph_inode_info *ci = ceph_inode(inode);
697 	struct ceph_osd_data *osd_data;
698 	struct page *page;
699 	int num_pages, total_pages = 0;
700 	int i, j;
701 	int rc = req->r_result;
702 	struct ceph_snap_context *snapc = req->r_snapc;
703 	struct address_space *mapping = inode->i_mapping;
704 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
705 	bool remove_page;
706 
707 	dout("writepages_finish %p rc %d\n", inode, rc);
708 	if (rc < 0) {
709 		mapping_set_error(mapping, rc);
710 		ceph_set_error_write(ci);
711 		if (rc == -EBLACKLISTED)
712 			fsc->blacklisted = true;
713 	} else {
714 		ceph_clear_error_write(ci);
715 	}
716 
717 	/*
718 	 * We lost the cache cap, need to truncate the page before
719 	 * it is unlocked, otherwise we'd truncate it later in the
720 	 * page truncation thread, possibly losing some data that
721 	 * raced its way in
722 	 */
723 	remove_page = !(ceph_caps_issued(ci) &
724 			(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
725 
726 	/* clean all pages */
727 	for (i = 0; i < req->r_num_ops; i++) {
728 		if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
729 			break;
730 
731 		osd_data = osd_req_op_extent_osd_data(req, i);
732 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
733 		num_pages = calc_pages_for((u64)osd_data->alignment,
734 					   (u64)osd_data->length);
735 		total_pages += num_pages;
736 		for (j = 0; j < num_pages; j++) {
737 			page = osd_data->pages[j];
738 			BUG_ON(!page);
739 			WARN_ON(!PageUptodate(page));
740 
741 			if (atomic_long_dec_return(&fsc->writeback_count) <
742 			     CONGESTION_OFF_THRESH(
743 					fsc->mount_options->congestion_kb))
744 				clear_bdi_congested(inode_to_bdi(inode),
745 						    BLK_RW_ASYNC);
746 
747 			ceph_put_snap_context(page_snap_context(page));
748 			page->private = 0;
749 			ClearPagePrivate(page);
750 			dout("unlocking %p\n", page);
751 			end_page_writeback(page);
752 
753 			if (remove_page)
754 				generic_error_remove_page(inode->i_mapping,
755 							  page);
756 
757 			unlock_page(page);
758 		}
759 		dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
760 		     inode, osd_data->length, rc >= 0 ? num_pages : 0);
761 
762 		release_pages(osd_data->pages, num_pages);
763 	}
764 
765 	ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
766 
767 	osd_data = osd_req_op_extent_osd_data(req, 0);
768 	if (osd_data->pages_from_pool)
769 		mempool_free(osd_data->pages,
770 			     ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
771 	else
772 		kfree(osd_data->pages);
773 	ceph_osdc_put_request(req);
774 }
775 
776 /*
777  * initiate async writeback
778  */
779 static int ceph_writepages_start(struct address_space *mapping,
780 				 struct writeback_control *wbc)
781 {
782 	struct inode *inode = mapping->host;
783 	struct ceph_inode_info *ci = ceph_inode(inode);
784 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
785 	struct ceph_vino vino = ceph_vino(inode);
786 	pgoff_t index, start_index, end = -1;
787 	struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
788 	struct pagevec pvec;
789 	int rc = 0;
790 	unsigned int wsize = i_blocksize(inode);
791 	struct ceph_osd_request *req = NULL;
792 	struct ceph_writeback_ctl ceph_wbc;
793 	bool should_loop, range_whole = false;
794 	bool done = false;
795 
796 	dout("writepages_start %p (mode=%s)\n", inode,
797 	     wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
798 	     (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
799 
800 	if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
801 		if (ci->i_wrbuffer_ref > 0) {
802 			pr_warn_ratelimited(
803 				"writepage_start %p %lld forced umount\n",
804 				inode, ceph_ino(inode));
805 		}
806 		mapping_set_error(mapping, -EIO);
807 		return -EIO; /* we're in a forced umount, don't write! */
808 	}
809 	if (fsc->mount_options->wsize < wsize)
810 		wsize = fsc->mount_options->wsize;
811 
812 	pagevec_init(&pvec);
813 
814 	start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
815 	index = start_index;
816 
817 retry:
818 	/* find oldest snap context with dirty data */
819 	snapc = get_oldest_context(inode, &ceph_wbc, NULL);
820 	if (!snapc) {
821 		/* hmm, why does writepages get called when there
822 		   is no dirty data? */
823 		dout(" no snap context with dirty data?\n");
824 		goto out;
825 	}
826 	dout(" oldest snapc is %p seq %lld (%d snaps)\n",
827 	     snapc, snapc->seq, snapc->num_snaps);
828 
829 	should_loop = false;
830 	if (ceph_wbc.head_snapc && snapc != last_snapc) {
831 		/* where to start/end? */
832 		if (wbc->range_cyclic) {
833 			index = start_index;
834 			end = -1;
835 			if (index > 0)
836 				should_loop = true;
837 			dout(" cyclic, start at %lu\n", index);
838 		} else {
839 			index = wbc->range_start >> PAGE_SHIFT;
840 			end = wbc->range_end >> PAGE_SHIFT;
841 			if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
842 				range_whole = true;
843 			dout(" not cyclic, %lu to %lu\n", index, end);
844 		}
845 	} else if (!ceph_wbc.head_snapc) {
846 		/* Do not respect wbc->range_{start,end}. Dirty pages
847 		 * in that range can be associated with newer snapc.
848 		 * They are not writeable until we write all dirty pages
849 		 * associated with 'snapc' get written */
850 		if (index > 0)
851 			should_loop = true;
852 		dout(" non-head snapc, range whole\n");
853 	}
854 
855 	ceph_put_snap_context(last_snapc);
856 	last_snapc = snapc;
857 
858 	while (!done && index <= end) {
859 		int num_ops = 0, op_idx;
860 		unsigned i, pvec_pages, max_pages, locked_pages = 0;
861 		struct page **pages = NULL, **data_pages;
862 		mempool_t *pool = NULL;	/* Becomes non-null if mempool used */
863 		struct page *page;
864 		pgoff_t strip_unit_end = 0;
865 		u64 offset = 0, len = 0;
866 
867 		max_pages = wsize >> PAGE_SHIFT;
868 
869 get_more_pages:
870 		pvec_pages = pagevec_lookup_range_nr_tag(&pvec, mapping, &index,
871 						end, PAGECACHE_TAG_DIRTY,
872 						max_pages - locked_pages);
873 		dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
874 		if (!pvec_pages && !locked_pages)
875 			break;
876 		for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
877 			page = pvec.pages[i];
878 			dout("? %p idx %lu\n", page, page->index);
879 			if (locked_pages == 0)
880 				lock_page(page);  /* first page */
881 			else if (!trylock_page(page))
882 				break;
883 
884 			/* only dirty pages, or our accounting breaks */
885 			if (unlikely(!PageDirty(page)) ||
886 			    unlikely(page->mapping != mapping)) {
887 				dout("!dirty or !mapping %p\n", page);
888 				unlock_page(page);
889 				continue;
890 			}
891 			/* only if matching snap context */
892 			pgsnapc = page_snap_context(page);
893 			if (pgsnapc != snapc) {
894 				dout("page snapc %p %lld != oldest %p %lld\n",
895 				     pgsnapc, pgsnapc->seq, snapc, snapc->seq);
896 				if (!should_loop &&
897 				    !ceph_wbc.head_snapc &&
898 				    wbc->sync_mode != WB_SYNC_NONE)
899 					should_loop = true;
900 				unlock_page(page);
901 				continue;
902 			}
903 			if (page_offset(page) >= ceph_wbc.i_size) {
904 				dout("%p page eof %llu\n",
905 				     page, ceph_wbc.i_size);
906 				if ((ceph_wbc.size_stable ||
907 				    page_offset(page) >= i_size_read(inode)) &&
908 				    clear_page_dirty_for_io(page))
909 					mapping->a_ops->invalidatepage(page,
910 								0, PAGE_SIZE);
911 				unlock_page(page);
912 				continue;
913 			}
914 			if (strip_unit_end && (page->index > strip_unit_end)) {
915 				dout("end of strip unit %p\n", page);
916 				unlock_page(page);
917 				break;
918 			}
919 			if (PageWriteback(page)) {
920 				if (wbc->sync_mode == WB_SYNC_NONE) {
921 					dout("%p under writeback\n", page);
922 					unlock_page(page);
923 					continue;
924 				}
925 				dout("waiting on writeback %p\n", page);
926 				wait_on_page_writeback(page);
927 			}
928 
929 			if (!clear_page_dirty_for_io(page)) {
930 				dout("%p !clear_page_dirty_for_io\n", page);
931 				unlock_page(page);
932 				continue;
933 			}
934 
935 			/*
936 			 * We have something to write.  If this is
937 			 * the first locked page this time through,
938 			 * calculate max possinle write size and
939 			 * allocate a page array
940 			 */
941 			if (locked_pages == 0) {
942 				u64 objnum;
943 				u64 objoff;
944 				u32 xlen;
945 
946 				/* prepare async write request */
947 				offset = (u64)page_offset(page);
948 				ceph_calc_file_object_mapping(&ci->i_layout,
949 							      offset, wsize,
950 							      &objnum, &objoff,
951 							      &xlen);
952 				len = xlen;
953 
954 				num_ops = 1;
955 				strip_unit_end = page->index +
956 					((len - 1) >> PAGE_SHIFT);
957 
958 				BUG_ON(pages);
959 				max_pages = calc_pages_for(0, (u64)len);
960 				pages = kmalloc_array(max_pages,
961 						      sizeof(*pages),
962 						      GFP_NOFS);
963 				if (!pages) {
964 					pool = fsc->wb_pagevec_pool;
965 					pages = mempool_alloc(pool, GFP_NOFS);
966 					BUG_ON(!pages);
967 				}
968 
969 				len = 0;
970 			} else if (page->index !=
971 				   (offset + len) >> PAGE_SHIFT) {
972 				if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
973 							CEPH_OSD_MAX_OPS)) {
974 					redirty_page_for_writepage(wbc, page);
975 					unlock_page(page);
976 					break;
977 				}
978 
979 				num_ops++;
980 				offset = (u64)page_offset(page);
981 				len = 0;
982 			}
983 
984 			/* note position of first page in pvec */
985 			dout("%p will write page %p idx %lu\n",
986 			     inode, page, page->index);
987 
988 			if (atomic_long_inc_return(&fsc->writeback_count) >
989 			    CONGESTION_ON_THRESH(
990 				    fsc->mount_options->congestion_kb)) {
991 				set_bdi_congested(inode_to_bdi(inode),
992 						  BLK_RW_ASYNC);
993 			}
994 
995 
996 			pages[locked_pages++] = page;
997 			pvec.pages[i] = NULL;
998 
999 			len += PAGE_SIZE;
1000 		}
1001 
1002 		/* did we get anything? */
1003 		if (!locked_pages)
1004 			goto release_pvec_pages;
1005 		if (i) {
1006 			unsigned j, n = 0;
1007 			/* shift unused page to beginning of pvec */
1008 			for (j = 0; j < pvec_pages; j++) {
1009 				if (!pvec.pages[j])
1010 					continue;
1011 				if (n < j)
1012 					pvec.pages[n] = pvec.pages[j];
1013 				n++;
1014 			}
1015 			pvec.nr = n;
1016 
1017 			if (pvec_pages && i == pvec_pages &&
1018 			    locked_pages < max_pages) {
1019 				dout("reached end pvec, trying for more\n");
1020 				pagevec_release(&pvec);
1021 				goto get_more_pages;
1022 			}
1023 		}
1024 
1025 new_request:
1026 		offset = page_offset(pages[0]);
1027 		len = wsize;
1028 
1029 		req = ceph_osdc_new_request(&fsc->client->osdc,
1030 					&ci->i_layout, vino,
1031 					offset, &len, 0, num_ops,
1032 					CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1033 					snapc, ceph_wbc.truncate_seq,
1034 					ceph_wbc.truncate_size, false);
1035 		if (IS_ERR(req)) {
1036 			req = ceph_osdc_new_request(&fsc->client->osdc,
1037 						&ci->i_layout, vino,
1038 						offset, &len, 0,
1039 						min(num_ops,
1040 						    CEPH_OSD_SLAB_OPS),
1041 						CEPH_OSD_OP_WRITE,
1042 						CEPH_OSD_FLAG_WRITE,
1043 						snapc, ceph_wbc.truncate_seq,
1044 						ceph_wbc.truncate_size, true);
1045 			BUG_ON(IS_ERR(req));
1046 		}
1047 		BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1048 			     PAGE_SIZE - offset);
1049 
1050 		req->r_callback = writepages_finish;
1051 		req->r_inode = inode;
1052 
1053 		/* Format the osd request message and submit the write */
1054 		len = 0;
1055 		data_pages = pages;
1056 		op_idx = 0;
1057 		for (i = 0; i < locked_pages; i++) {
1058 			u64 cur_offset = page_offset(pages[i]);
1059 			if (offset + len != cur_offset) {
1060 				if (op_idx + 1 == req->r_num_ops)
1061 					break;
1062 				osd_req_op_extent_dup_last(req, op_idx,
1063 							   cur_offset - offset);
1064 				dout("writepages got pages at %llu~%llu\n",
1065 				     offset, len);
1066 				osd_req_op_extent_osd_data_pages(req, op_idx,
1067 							data_pages, len, 0,
1068 							!!pool, false);
1069 				osd_req_op_extent_update(req, op_idx, len);
1070 
1071 				len = 0;
1072 				offset = cur_offset;
1073 				data_pages = pages + i;
1074 				op_idx++;
1075 			}
1076 
1077 			set_page_writeback(pages[i]);
1078 			len += PAGE_SIZE;
1079 		}
1080 
1081 		if (ceph_wbc.size_stable) {
1082 			len = min(len, ceph_wbc.i_size - offset);
1083 		} else if (i == locked_pages) {
1084 			/* writepages_finish() clears writeback pages
1085 			 * according to the data length, so make sure
1086 			 * data length covers all locked pages */
1087 			u64 min_len = len + 1 - PAGE_SIZE;
1088 			len = get_writepages_data_length(inode, pages[i - 1],
1089 							 offset);
1090 			len = max(len, min_len);
1091 		}
1092 		dout("writepages got pages at %llu~%llu\n", offset, len);
1093 
1094 		osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1095 						 0, !!pool, false);
1096 		osd_req_op_extent_update(req, op_idx, len);
1097 
1098 		BUG_ON(op_idx + 1 != req->r_num_ops);
1099 
1100 		pool = NULL;
1101 		if (i < locked_pages) {
1102 			BUG_ON(num_ops <= req->r_num_ops);
1103 			num_ops -= req->r_num_ops;
1104 			locked_pages -= i;
1105 
1106 			/* allocate new pages array for next request */
1107 			data_pages = pages;
1108 			pages = kmalloc_array(locked_pages, sizeof(*pages),
1109 					      GFP_NOFS);
1110 			if (!pages) {
1111 				pool = fsc->wb_pagevec_pool;
1112 				pages = mempool_alloc(pool, GFP_NOFS);
1113 				BUG_ON(!pages);
1114 			}
1115 			memcpy(pages, data_pages + i,
1116 			       locked_pages * sizeof(*pages));
1117 			memset(data_pages + i, 0,
1118 			       locked_pages * sizeof(*pages));
1119 		} else {
1120 			BUG_ON(num_ops != req->r_num_ops);
1121 			index = pages[i - 1]->index + 1;
1122 			/* request message now owns the pages array */
1123 			pages = NULL;
1124 		}
1125 
1126 		req->r_mtime = inode->i_mtime;
1127 		rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1128 		BUG_ON(rc);
1129 		req = NULL;
1130 
1131 		wbc->nr_to_write -= i;
1132 		if (pages)
1133 			goto new_request;
1134 
1135 		/*
1136 		 * We stop writing back only if we are not doing
1137 		 * integrity sync. In case of integrity sync we have to
1138 		 * keep going until we have written all the pages
1139 		 * we tagged for writeback prior to entering this loop.
1140 		 */
1141 		if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1142 			done = true;
1143 
1144 release_pvec_pages:
1145 		dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1146 		     pvec.nr ? pvec.pages[0] : NULL);
1147 		pagevec_release(&pvec);
1148 	}
1149 
1150 	if (should_loop && !done) {
1151 		/* more to do; loop back to beginning of file */
1152 		dout("writepages looping back to beginning of file\n");
1153 		end = start_index - 1; /* OK even when start_index == 0 */
1154 
1155 		/* to write dirty pages associated with next snapc,
1156 		 * we need to wait until current writes complete */
1157 		if (wbc->sync_mode != WB_SYNC_NONE &&
1158 		    start_index == 0 && /* all dirty pages were checked */
1159 		    !ceph_wbc.head_snapc) {
1160 			struct page *page;
1161 			unsigned i, nr;
1162 			index = 0;
1163 			while ((index <= end) &&
1164 			       (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1165 						PAGECACHE_TAG_WRITEBACK))) {
1166 				for (i = 0; i < nr; i++) {
1167 					page = pvec.pages[i];
1168 					if (page_snap_context(page) != snapc)
1169 						continue;
1170 					wait_on_page_writeback(page);
1171 				}
1172 				pagevec_release(&pvec);
1173 				cond_resched();
1174 			}
1175 		}
1176 
1177 		start_index = 0;
1178 		index = 0;
1179 		goto retry;
1180 	}
1181 
1182 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1183 		mapping->writeback_index = index;
1184 
1185 out:
1186 	ceph_osdc_put_request(req);
1187 	ceph_put_snap_context(last_snapc);
1188 	dout("writepages dend - startone, rc = %d\n", rc);
1189 	return rc;
1190 }
1191 
1192 
1193 
1194 /*
1195  * See if a given @snapc is either writeable, or already written.
1196  */
1197 static int context_is_writeable_or_written(struct inode *inode,
1198 					   struct ceph_snap_context *snapc)
1199 {
1200 	struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1201 	int ret = !oldest || snapc->seq <= oldest->seq;
1202 
1203 	ceph_put_snap_context(oldest);
1204 	return ret;
1205 }
1206 
1207 /*
1208  * We are only allowed to write into/dirty the page if the page is
1209  * clean, or already dirty within the same snap context.
1210  *
1211  * called with page locked.
1212  * return success with page locked,
1213  * or any failure (incl -EAGAIN) with page unlocked.
1214  */
1215 static int ceph_update_writeable_page(struct file *file,
1216 			    loff_t pos, unsigned len,
1217 			    struct page *page)
1218 {
1219 	struct inode *inode = file_inode(file);
1220 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1221 	struct ceph_inode_info *ci = ceph_inode(inode);
1222 	loff_t page_off = pos & PAGE_MASK;
1223 	int pos_in_page = pos & ~PAGE_MASK;
1224 	int end_in_page = pos_in_page + len;
1225 	loff_t i_size;
1226 	int r;
1227 	struct ceph_snap_context *snapc, *oldest;
1228 
1229 	if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1230 		dout(" page %p forced umount\n", page);
1231 		unlock_page(page);
1232 		return -EIO;
1233 	}
1234 
1235 retry_locked:
1236 	/* writepages currently holds page lock, but if we change that later, */
1237 	wait_on_page_writeback(page);
1238 
1239 	snapc = page_snap_context(page);
1240 	if (snapc && snapc != ci->i_head_snapc) {
1241 		/*
1242 		 * this page is already dirty in another (older) snap
1243 		 * context!  is it writeable now?
1244 		 */
1245 		oldest = get_oldest_context(inode, NULL, NULL);
1246 		if (snapc->seq > oldest->seq) {
1247 			ceph_put_snap_context(oldest);
1248 			dout(" page %p snapc %p not current or oldest\n",
1249 			     page, snapc);
1250 			/*
1251 			 * queue for writeback, and wait for snapc to
1252 			 * be writeable or written
1253 			 */
1254 			snapc = ceph_get_snap_context(snapc);
1255 			unlock_page(page);
1256 			ceph_queue_writeback(inode);
1257 			r = wait_event_killable(ci->i_cap_wq,
1258 			       context_is_writeable_or_written(inode, snapc));
1259 			ceph_put_snap_context(snapc);
1260 			if (r == -ERESTARTSYS)
1261 				return r;
1262 			return -EAGAIN;
1263 		}
1264 		ceph_put_snap_context(oldest);
1265 
1266 		/* yay, writeable, do it now (without dropping page lock) */
1267 		dout(" page %p snapc %p not current, but oldest\n",
1268 		     page, snapc);
1269 		if (!clear_page_dirty_for_io(page))
1270 			goto retry_locked;
1271 		r = writepage_nounlock(page, NULL);
1272 		if (r < 0)
1273 			goto fail_unlock;
1274 		goto retry_locked;
1275 	}
1276 
1277 	if (PageUptodate(page)) {
1278 		dout(" page %p already uptodate\n", page);
1279 		return 0;
1280 	}
1281 
1282 	/* full page? */
1283 	if (pos_in_page == 0 && len == PAGE_SIZE)
1284 		return 0;
1285 
1286 	/* past end of file? */
1287 	i_size = i_size_read(inode);
1288 
1289 	if (page_off >= i_size ||
1290 	    (pos_in_page == 0 && (pos+len) >= i_size &&
1291 	     end_in_page - pos_in_page != PAGE_SIZE)) {
1292 		dout(" zeroing %p 0 - %d and %d - %d\n",
1293 		     page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1294 		zero_user_segments(page,
1295 				   0, pos_in_page,
1296 				   end_in_page, PAGE_SIZE);
1297 		return 0;
1298 	}
1299 
1300 	/* we need to read it. */
1301 	r = ceph_do_readpage(file, page);
1302 	if (r < 0) {
1303 		if (r == -EINPROGRESS)
1304 			return -EAGAIN;
1305 		goto fail_unlock;
1306 	}
1307 	goto retry_locked;
1308 fail_unlock:
1309 	unlock_page(page);
1310 	return r;
1311 }
1312 
1313 /*
1314  * We are only allowed to write into/dirty the page if the page is
1315  * clean, or already dirty within the same snap context.
1316  */
1317 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1318 			    loff_t pos, unsigned len, unsigned flags,
1319 			    struct page **pagep, void **fsdata)
1320 {
1321 	struct inode *inode = file_inode(file);
1322 	struct page *page;
1323 	pgoff_t index = pos >> PAGE_SHIFT;
1324 	int r;
1325 
1326 	do {
1327 		/* get a page */
1328 		page = grab_cache_page_write_begin(mapping, index, 0);
1329 		if (!page)
1330 			return -ENOMEM;
1331 
1332 		dout("write_begin file %p inode %p page %p %d~%d\n", file,
1333 		     inode, page, (int)pos, (int)len);
1334 
1335 		r = ceph_update_writeable_page(file, pos, len, page);
1336 		if (r < 0)
1337 			put_page(page);
1338 		else
1339 			*pagep = page;
1340 	} while (r == -EAGAIN);
1341 
1342 	return r;
1343 }
1344 
1345 /*
1346  * we don't do anything in here that simple_write_end doesn't do
1347  * except adjust dirty page accounting
1348  */
1349 static int ceph_write_end(struct file *file, struct address_space *mapping,
1350 			  loff_t pos, unsigned len, unsigned copied,
1351 			  struct page *page, void *fsdata)
1352 {
1353 	struct inode *inode = file_inode(file);
1354 	bool check_cap = false;
1355 
1356 	dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1357 	     inode, page, (int)pos, (int)copied, (int)len);
1358 
1359 	/* zero the stale part of the page if we did a short copy */
1360 	if (!PageUptodate(page)) {
1361 		if (copied < len) {
1362 			copied = 0;
1363 			goto out;
1364 		}
1365 		SetPageUptodate(page);
1366 	}
1367 
1368 	/* did file size increase? */
1369 	if (pos+copied > i_size_read(inode))
1370 		check_cap = ceph_inode_set_size(inode, pos+copied);
1371 
1372 	set_page_dirty(page);
1373 
1374 out:
1375 	unlock_page(page);
1376 	put_page(page);
1377 
1378 	if (check_cap)
1379 		ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1380 
1381 	return copied;
1382 }
1383 
1384 /*
1385  * we set .direct_IO to indicate direct io is supported, but since we
1386  * intercept O_DIRECT reads and writes early, this function should
1387  * never get called.
1388  */
1389 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1390 {
1391 	WARN_ON(1);
1392 	return -EINVAL;
1393 }
1394 
1395 const struct address_space_operations ceph_aops = {
1396 	.readpage = ceph_readpage,
1397 	.readpages = ceph_readpages,
1398 	.writepage = ceph_writepage,
1399 	.writepages = ceph_writepages_start,
1400 	.write_begin = ceph_write_begin,
1401 	.write_end = ceph_write_end,
1402 	.set_page_dirty = ceph_set_page_dirty,
1403 	.invalidatepage = ceph_invalidatepage,
1404 	.releasepage = ceph_releasepage,
1405 	.direct_IO = ceph_direct_io,
1406 };
1407 
1408 static void ceph_block_sigs(sigset_t *oldset)
1409 {
1410 	sigset_t mask;
1411 	siginitsetinv(&mask, sigmask(SIGKILL));
1412 	sigprocmask(SIG_BLOCK, &mask, oldset);
1413 }
1414 
1415 static void ceph_restore_sigs(sigset_t *oldset)
1416 {
1417 	sigprocmask(SIG_SETMASK, oldset, NULL);
1418 }
1419 
1420 /*
1421  * vm ops
1422  */
1423 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1424 {
1425 	struct vm_area_struct *vma = vmf->vma;
1426 	struct inode *inode = file_inode(vma->vm_file);
1427 	struct ceph_inode_info *ci = ceph_inode(inode);
1428 	struct ceph_file_info *fi = vma->vm_file->private_data;
1429 	struct page *pinned_page = NULL;
1430 	loff_t off = vmf->pgoff << PAGE_SHIFT;
1431 	int want, got, err;
1432 	sigset_t oldset;
1433 	vm_fault_t ret = VM_FAULT_SIGBUS;
1434 
1435 	ceph_block_sigs(&oldset);
1436 
1437 	dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1438 	     inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1439 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1440 		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1441 	else
1442 		want = CEPH_CAP_FILE_CACHE;
1443 
1444 	got = 0;
1445 	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1,
1446 			    &got, &pinned_page);
1447 	if (err < 0)
1448 		goto out_restore;
1449 
1450 	dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1451 	     inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1452 
1453 	if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1454 	    ci->i_inline_version == CEPH_INLINE_NONE) {
1455 		CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1456 		ceph_add_rw_context(fi, &rw_ctx);
1457 		ret = filemap_fault(vmf);
1458 		ceph_del_rw_context(fi, &rw_ctx);
1459 		dout("filemap_fault %p %llu~%zd drop cap refs %s ret %x\n",
1460 			inode, off, (size_t)PAGE_SIZE,
1461 				ceph_cap_string(got), ret);
1462 	} else
1463 		err = -EAGAIN;
1464 
1465 	if (pinned_page)
1466 		put_page(pinned_page);
1467 	ceph_put_cap_refs(ci, got);
1468 
1469 	if (err != -EAGAIN)
1470 		goto out_restore;
1471 
1472 	/* read inline data */
1473 	if (off >= PAGE_SIZE) {
1474 		/* does not support inline data > PAGE_SIZE */
1475 		ret = VM_FAULT_SIGBUS;
1476 	} else {
1477 		struct address_space *mapping = inode->i_mapping;
1478 		struct page *page = find_or_create_page(mapping, 0,
1479 						mapping_gfp_constraint(mapping,
1480 						~__GFP_FS));
1481 		if (!page) {
1482 			ret = VM_FAULT_OOM;
1483 			goto out_inline;
1484 		}
1485 		err = __ceph_do_getattr(inode, page,
1486 					 CEPH_STAT_CAP_INLINE_DATA, true);
1487 		if (err < 0 || off >= i_size_read(inode)) {
1488 			unlock_page(page);
1489 			put_page(page);
1490 			ret = vmf_error(err);
1491 			goto out_inline;
1492 		}
1493 		if (err < PAGE_SIZE)
1494 			zero_user_segment(page, err, PAGE_SIZE);
1495 		else
1496 			flush_dcache_page(page);
1497 		SetPageUptodate(page);
1498 		vmf->page = page;
1499 		ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1500 out_inline:
1501 		dout("filemap_fault %p %llu~%zd read inline data ret %x\n",
1502 		     inode, off, (size_t)PAGE_SIZE, ret);
1503 	}
1504 out_restore:
1505 	ceph_restore_sigs(&oldset);
1506 	if (err < 0)
1507 		ret = vmf_error(err);
1508 
1509 	return ret;
1510 }
1511 
1512 /*
1513  * Reuse write_begin here for simplicity.
1514  */
1515 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1516 {
1517 	struct vm_area_struct *vma = vmf->vma;
1518 	struct inode *inode = file_inode(vma->vm_file);
1519 	struct ceph_inode_info *ci = ceph_inode(inode);
1520 	struct ceph_file_info *fi = vma->vm_file->private_data;
1521 	struct ceph_cap_flush *prealloc_cf;
1522 	struct page *page = vmf->page;
1523 	loff_t off = page_offset(page);
1524 	loff_t size = i_size_read(inode);
1525 	size_t len;
1526 	int want, got, err;
1527 	sigset_t oldset;
1528 	vm_fault_t ret = VM_FAULT_SIGBUS;
1529 
1530 	prealloc_cf = ceph_alloc_cap_flush();
1531 	if (!prealloc_cf)
1532 		return VM_FAULT_OOM;
1533 
1534 	sb_start_pagefault(inode->i_sb);
1535 	ceph_block_sigs(&oldset);
1536 
1537 	if (ci->i_inline_version != CEPH_INLINE_NONE) {
1538 		struct page *locked_page = NULL;
1539 		if (off == 0) {
1540 			lock_page(page);
1541 			locked_page = page;
1542 		}
1543 		err = ceph_uninline_data(vma->vm_file, locked_page);
1544 		if (locked_page)
1545 			unlock_page(locked_page);
1546 		if (err < 0)
1547 			goto out_free;
1548 	}
1549 
1550 	if (off + PAGE_SIZE <= size)
1551 		len = PAGE_SIZE;
1552 	else
1553 		len = size & ~PAGE_MASK;
1554 
1555 	dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1556 	     inode, ceph_vinop(inode), off, len, size);
1557 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1558 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1559 	else
1560 		want = CEPH_CAP_FILE_BUFFER;
1561 
1562 	got = 0;
1563 	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len,
1564 			    &got, NULL);
1565 	if (err < 0)
1566 		goto out_free;
1567 
1568 	dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1569 	     inode, off, len, ceph_cap_string(got));
1570 
1571 	/* Update time before taking page lock */
1572 	file_update_time(vma->vm_file);
1573 	inode_inc_iversion_raw(inode);
1574 
1575 	do {
1576 		lock_page(page);
1577 
1578 		if ((off > size) || (page->mapping != inode->i_mapping)) {
1579 			unlock_page(page);
1580 			ret = VM_FAULT_NOPAGE;
1581 			break;
1582 		}
1583 
1584 		err = ceph_update_writeable_page(vma->vm_file, off, len, page);
1585 		if (err >= 0) {
1586 			/* success.  we'll keep the page locked. */
1587 			set_page_dirty(page);
1588 			ret = VM_FAULT_LOCKED;
1589 		}
1590 	} while (err == -EAGAIN);
1591 
1592 	if (ret == VM_FAULT_LOCKED ||
1593 	    ci->i_inline_version != CEPH_INLINE_NONE) {
1594 		int dirty;
1595 		spin_lock(&ci->i_ceph_lock);
1596 		ci->i_inline_version = CEPH_INLINE_NONE;
1597 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1598 					       &prealloc_cf);
1599 		spin_unlock(&ci->i_ceph_lock);
1600 		if (dirty)
1601 			__mark_inode_dirty(inode, dirty);
1602 	}
1603 
1604 	dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1605 	     inode, off, len, ceph_cap_string(got), ret);
1606 	ceph_put_cap_refs(ci, got);
1607 out_free:
1608 	ceph_restore_sigs(&oldset);
1609 	sb_end_pagefault(inode->i_sb);
1610 	ceph_free_cap_flush(prealloc_cf);
1611 	if (err < 0)
1612 		ret = vmf_error(err);
1613 	return ret;
1614 }
1615 
1616 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1617 			   char	*data, size_t len)
1618 {
1619 	struct address_space *mapping = inode->i_mapping;
1620 	struct page *page;
1621 
1622 	if (locked_page) {
1623 		page = locked_page;
1624 	} else {
1625 		if (i_size_read(inode) == 0)
1626 			return;
1627 		page = find_or_create_page(mapping, 0,
1628 					   mapping_gfp_constraint(mapping,
1629 					   ~__GFP_FS));
1630 		if (!page)
1631 			return;
1632 		if (PageUptodate(page)) {
1633 			unlock_page(page);
1634 			put_page(page);
1635 			return;
1636 		}
1637 	}
1638 
1639 	dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1640 	     inode, ceph_vinop(inode), len, locked_page);
1641 
1642 	if (len > 0) {
1643 		void *kaddr = kmap_atomic(page);
1644 		memcpy(kaddr, data, len);
1645 		kunmap_atomic(kaddr);
1646 	}
1647 
1648 	if (page != locked_page) {
1649 		if (len < PAGE_SIZE)
1650 			zero_user_segment(page, len, PAGE_SIZE);
1651 		else
1652 			flush_dcache_page(page);
1653 
1654 		SetPageUptodate(page);
1655 		unlock_page(page);
1656 		put_page(page);
1657 	}
1658 }
1659 
1660 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1661 {
1662 	struct inode *inode = file_inode(filp);
1663 	struct ceph_inode_info *ci = ceph_inode(inode);
1664 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1665 	struct ceph_osd_request *req;
1666 	struct page *page = NULL;
1667 	u64 len, inline_version;
1668 	int err = 0;
1669 	bool from_pagecache = false;
1670 
1671 	spin_lock(&ci->i_ceph_lock);
1672 	inline_version = ci->i_inline_version;
1673 	spin_unlock(&ci->i_ceph_lock);
1674 
1675 	dout("uninline_data %p %llx.%llx inline_version %llu\n",
1676 	     inode, ceph_vinop(inode), inline_version);
1677 
1678 	if (inline_version == 1 || /* initial version, no data */
1679 	    inline_version == CEPH_INLINE_NONE)
1680 		goto out;
1681 
1682 	if (locked_page) {
1683 		page = locked_page;
1684 		WARN_ON(!PageUptodate(page));
1685 	} else if (ceph_caps_issued(ci) &
1686 		   (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1687 		page = find_get_page(inode->i_mapping, 0);
1688 		if (page) {
1689 			if (PageUptodate(page)) {
1690 				from_pagecache = true;
1691 				lock_page(page);
1692 			} else {
1693 				put_page(page);
1694 				page = NULL;
1695 			}
1696 		}
1697 	}
1698 
1699 	if (page) {
1700 		len = i_size_read(inode);
1701 		if (len > PAGE_SIZE)
1702 			len = PAGE_SIZE;
1703 	} else {
1704 		page = __page_cache_alloc(GFP_NOFS);
1705 		if (!page) {
1706 			err = -ENOMEM;
1707 			goto out;
1708 		}
1709 		err = __ceph_do_getattr(inode, page,
1710 					CEPH_STAT_CAP_INLINE_DATA, true);
1711 		if (err < 0) {
1712 			/* no inline data */
1713 			if (err == -ENODATA)
1714 				err = 0;
1715 			goto out;
1716 		}
1717 		len = err;
1718 	}
1719 
1720 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1721 				    ceph_vino(inode), 0, &len, 0, 1,
1722 				    CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1723 				    NULL, 0, 0, false);
1724 	if (IS_ERR(req)) {
1725 		err = PTR_ERR(req);
1726 		goto out;
1727 	}
1728 
1729 	req->r_mtime = inode->i_mtime;
1730 	err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1731 	if (!err)
1732 		err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1733 	ceph_osdc_put_request(req);
1734 	if (err < 0)
1735 		goto out;
1736 
1737 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1738 				    ceph_vino(inode), 0, &len, 1, 3,
1739 				    CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1740 				    NULL, ci->i_truncate_seq,
1741 				    ci->i_truncate_size, false);
1742 	if (IS_ERR(req)) {
1743 		err = PTR_ERR(req);
1744 		goto out;
1745 	}
1746 
1747 	osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1748 
1749 	{
1750 		__le64 xattr_buf = cpu_to_le64(inline_version);
1751 		err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1752 					    "inline_version", &xattr_buf,
1753 					    sizeof(xattr_buf),
1754 					    CEPH_OSD_CMPXATTR_OP_GT,
1755 					    CEPH_OSD_CMPXATTR_MODE_U64);
1756 		if (err)
1757 			goto out_put;
1758 	}
1759 
1760 	{
1761 		char xattr_buf[32];
1762 		int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1763 					 "%llu", inline_version);
1764 		err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1765 					    "inline_version",
1766 					    xattr_buf, xattr_len, 0, 0);
1767 		if (err)
1768 			goto out_put;
1769 	}
1770 
1771 	req->r_mtime = inode->i_mtime;
1772 	err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1773 	if (!err)
1774 		err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1775 out_put:
1776 	ceph_osdc_put_request(req);
1777 	if (err == -ECANCELED)
1778 		err = 0;
1779 out:
1780 	if (page && page != locked_page) {
1781 		if (from_pagecache) {
1782 			unlock_page(page);
1783 			put_page(page);
1784 		} else
1785 			__free_pages(page, 0);
1786 	}
1787 
1788 	dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1789 	     inode, ceph_vinop(inode), inline_version, err);
1790 	return err;
1791 }
1792 
1793 static const struct vm_operations_struct ceph_vmops = {
1794 	.fault		= ceph_filemap_fault,
1795 	.page_mkwrite	= ceph_page_mkwrite,
1796 };
1797 
1798 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1799 {
1800 	struct address_space *mapping = file->f_mapping;
1801 
1802 	if (!mapping->a_ops->readpage)
1803 		return -ENOEXEC;
1804 	file_accessed(file);
1805 	vma->vm_ops = &ceph_vmops;
1806 	return 0;
1807 }
1808 
1809 enum {
1810 	POOL_READ	= 1,
1811 	POOL_WRITE	= 2,
1812 };
1813 
1814 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1815 				s64 pool, struct ceph_string *pool_ns)
1816 {
1817 	struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1818 	struct ceph_mds_client *mdsc = fsc->mdsc;
1819 	struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1820 	struct rb_node **p, *parent;
1821 	struct ceph_pool_perm *perm;
1822 	struct page **pages;
1823 	size_t pool_ns_len;
1824 	int err = 0, err2 = 0, have = 0;
1825 
1826 	down_read(&mdsc->pool_perm_rwsem);
1827 	p = &mdsc->pool_perm_tree.rb_node;
1828 	while (*p) {
1829 		perm = rb_entry(*p, struct ceph_pool_perm, node);
1830 		if (pool < perm->pool)
1831 			p = &(*p)->rb_left;
1832 		else if (pool > perm->pool)
1833 			p = &(*p)->rb_right;
1834 		else {
1835 			int ret = ceph_compare_string(pool_ns,
1836 						perm->pool_ns,
1837 						perm->pool_ns_len);
1838 			if (ret < 0)
1839 				p = &(*p)->rb_left;
1840 			else if (ret > 0)
1841 				p = &(*p)->rb_right;
1842 			else {
1843 				have = perm->perm;
1844 				break;
1845 			}
1846 		}
1847 	}
1848 	up_read(&mdsc->pool_perm_rwsem);
1849 	if (*p)
1850 		goto out;
1851 
1852 	if (pool_ns)
1853 		dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1854 		     pool, (int)pool_ns->len, pool_ns->str);
1855 	else
1856 		dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1857 
1858 	down_write(&mdsc->pool_perm_rwsem);
1859 	p = &mdsc->pool_perm_tree.rb_node;
1860 	parent = NULL;
1861 	while (*p) {
1862 		parent = *p;
1863 		perm = rb_entry(parent, struct ceph_pool_perm, node);
1864 		if (pool < perm->pool)
1865 			p = &(*p)->rb_left;
1866 		else if (pool > perm->pool)
1867 			p = &(*p)->rb_right;
1868 		else {
1869 			int ret = ceph_compare_string(pool_ns,
1870 						perm->pool_ns,
1871 						perm->pool_ns_len);
1872 			if (ret < 0)
1873 				p = &(*p)->rb_left;
1874 			else if (ret > 0)
1875 				p = &(*p)->rb_right;
1876 			else {
1877 				have = perm->perm;
1878 				break;
1879 			}
1880 		}
1881 	}
1882 	if (*p) {
1883 		up_write(&mdsc->pool_perm_rwsem);
1884 		goto out;
1885 	}
1886 
1887 	rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1888 					 1, false, GFP_NOFS);
1889 	if (!rd_req) {
1890 		err = -ENOMEM;
1891 		goto out_unlock;
1892 	}
1893 
1894 	rd_req->r_flags = CEPH_OSD_FLAG_READ;
1895 	osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1896 	rd_req->r_base_oloc.pool = pool;
1897 	if (pool_ns)
1898 		rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1899 	ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1900 
1901 	err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1902 	if (err)
1903 		goto out_unlock;
1904 
1905 	wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1906 					 1, false, GFP_NOFS);
1907 	if (!wr_req) {
1908 		err = -ENOMEM;
1909 		goto out_unlock;
1910 	}
1911 
1912 	wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1913 	osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1914 	ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1915 	ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1916 
1917 	err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1918 	if (err)
1919 		goto out_unlock;
1920 
1921 	/* one page should be large enough for STAT data */
1922 	pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1923 	if (IS_ERR(pages)) {
1924 		err = PTR_ERR(pages);
1925 		goto out_unlock;
1926 	}
1927 
1928 	osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1929 				     0, false, true);
1930 	err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1931 
1932 	wr_req->r_mtime = ci->vfs_inode.i_mtime;
1933 	err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1934 
1935 	if (!err)
1936 		err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1937 	if (!err2)
1938 		err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1939 
1940 	if (err >= 0 || err == -ENOENT)
1941 		have |= POOL_READ;
1942 	else if (err != -EPERM) {
1943 		if (err == -EBLACKLISTED)
1944 			fsc->blacklisted = true;
1945 		goto out_unlock;
1946 	}
1947 
1948 	if (err2 == 0 || err2 == -EEXIST)
1949 		have |= POOL_WRITE;
1950 	else if (err2 != -EPERM) {
1951 		if (err2 == -EBLACKLISTED)
1952 			fsc->blacklisted = true;
1953 		err = err2;
1954 		goto out_unlock;
1955 	}
1956 
1957 	pool_ns_len = pool_ns ? pool_ns->len : 0;
1958 	perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1959 	if (!perm) {
1960 		err = -ENOMEM;
1961 		goto out_unlock;
1962 	}
1963 
1964 	perm->pool = pool;
1965 	perm->perm = have;
1966 	perm->pool_ns_len = pool_ns_len;
1967 	if (pool_ns_len > 0)
1968 		memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1969 	perm->pool_ns[pool_ns_len] = 0;
1970 
1971 	rb_link_node(&perm->node, parent, p);
1972 	rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1973 	err = 0;
1974 out_unlock:
1975 	up_write(&mdsc->pool_perm_rwsem);
1976 
1977 	ceph_osdc_put_request(rd_req);
1978 	ceph_osdc_put_request(wr_req);
1979 out:
1980 	if (!err)
1981 		err = have;
1982 	if (pool_ns)
1983 		dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1984 		     pool, (int)pool_ns->len, pool_ns->str, err);
1985 	else
1986 		dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1987 	return err;
1988 }
1989 
1990 int ceph_pool_perm_check(struct inode *inode, int need)
1991 {
1992 	struct ceph_inode_info *ci = ceph_inode(inode);
1993 	struct ceph_string *pool_ns;
1994 	s64 pool;
1995 	int ret, flags;
1996 
1997 	if (ci->i_vino.snap != CEPH_NOSNAP) {
1998 		/*
1999 		 * Pool permission check needs to write to the first object.
2000 		 * But for snapshot, head of the first object may have alread
2001 		 * been deleted. Skip check to avoid creating orphan object.
2002 		 */
2003 		return 0;
2004 	}
2005 
2006 	if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2007 				NOPOOLPERM))
2008 		return 0;
2009 
2010 	spin_lock(&ci->i_ceph_lock);
2011 	flags = ci->i_ceph_flags;
2012 	pool = ci->i_layout.pool_id;
2013 	spin_unlock(&ci->i_ceph_lock);
2014 check:
2015 	if (flags & CEPH_I_POOL_PERM) {
2016 		if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2017 			dout("ceph_pool_perm_check pool %lld no read perm\n",
2018 			     pool);
2019 			return -EPERM;
2020 		}
2021 		if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2022 			dout("ceph_pool_perm_check pool %lld no write perm\n",
2023 			     pool);
2024 			return -EPERM;
2025 		}
2026 		return 0;
2027 	}
2028 
2029 	pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2030 	ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2031 	ceph_put_string(pool_ns);
2032 	if (ret < 0)
2033 		return ret;
2034 
2035 	flags = CEPH_I_POOL_PERM;
2036 	if (ret & POOL_READ)
2037 		flags |= CEPH_I_POOL_RD;
2038 	if (ret & POOL_WRITE)
2039 		flags |= CEPH_I_POOL_WR;
2040 
2041 	spin_lock(&ci->i_ceph_lock);
2042 	if (pool == ci->i_layout.pool_id &&
2043 	    pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2044 		ci->i_ceph_flags |= flags;
2045         } else {
2046 		pool = ci->i_layout.pool_id;
2047 		flags = ci->i_ceph_flags;
2048 	}
2049 	spin_unlock(&ci->i_ceph_lock);
2050 	goto check;
2051 }
2052 
2053 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2054 {
2055 	struct ceph_pool_perm *perm;
2056 	struct rb_node *n;
2057 
2058 	while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2059 		n = rb_first(&mdsc->pool_perm_tree);
2060 		perm = rb_entry(n, struct ceph_pool_perm, node);
2061 		rb_erase(n, &mdsc->pool_perm_tree);
2062 		kfree(perm);
2063 	}
2064 }
2065