xref: /openbmc/linux/fs/ceph/addr.c (revision dbfb5232)
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/swap.h>
8 #include <linux/pagemap.h>
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 #include <linux/ktime.h>
15 #include <linux/netfs.h>
16 
17 #include "super.h"
18 #include "mds_client.h"
19 #include "cache.h"
20 #include "metric.h"
21 #include "crypto.h"
22 #include <linux/ceph/osd_client.h>
23 #include <linux/ceph/striper.h>
24 
25 /*
26  * Ceph address space ops.
27  *
28  * There are a few funny things going on here.
29  *
30  * The page->private field is used to reference a struct
31  * ceph_snap_context for _every_ dirty page.  This indicates which
32  * snapshot the page was logically dirtied in, and thus which snap
33  * context needs to be associated with the osd write during writeback.
34  *
35  * Similarly, struct ceph_inode_info maintains a set of counters to
36  * count dirty pages on the inode.  In the absence of snapshots,
37  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
38  *
39  * When a snapshot is taken (that is, when the client receives
40  * notification that a snapshot was taken), each inode with caps and
41  * with dirty pages (dirty pages implies there is a cap) gets a new
42  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
43  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
44  * moved to capsnap->dirty. (Unless a sync write is currently in
45  * progress.  In that case, the capsnap is said to be "pending", new
46  * writes cannot start, and the capsnap isn't "finalized" until the
47  * write completes (or fails) and a final size/mtime for the inode for
48  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
49  *
50  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
51  * we look for the first capsnap in i_cap_snaps and write out pages in
52  * that snap context _only_.  Then we move on to the next capsnap,
53  * eventually reaching the "live" or "head" context (i.e., pages that
54  * are not yet snapped) and are writing the most recently dirtied
55  * pages.
56  *
57  * Invalidate and so forth must take care to ensure the dirty page
58  * accounting is preserved.
59  */
60 
61 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
62 #define CONGESTION_OFF_THRESH(congestion_kb)				\
63 	(CONGESTION_ON_THRESH(congestion_kb) -				\
64 	 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
65 
66 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
67 					struct folio **foliop, void **_fsdata);
68 
69 static inline struct ceph_snap_context *page_snap_context(struct page *page)
70 {
71 	if (PagePrivate(page))
72 		return (void *)page->private;
73 	return NULL;
74 }
75 
76 /*
77  * Dirty a page.  Optimistically adjust accounting, on the assumption
78  * that we won't race with invalidate.  If we do, readjust.
79  */
80 static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
81 {
82 	struct inode *inode;
83 	struct ceph_inode_info *ci;
84 	struct ceph_snap_context *snapc;
85 
86 	if (folio_test_dirty(folio)) {
87 		dout("%p dirty_folio %p idx %lu -- already dirty\n",
88 		     mapping->host, folio, folio->index);
89 		VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
90 		return false;
91 	}
92 
93 	inode = mapping->host;
94 	ci = ceph_inode(inode);
95 
96 	/* dirty the head */
97 	spin_lock(&ci->i_ceph_lock);
98 	BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
99 	if (__ceph_have_pending_cap_snap(ci)) {
100 		struct ceph_cap_snap *capsnap =
101 				list_last_entry(&ci->i_cap_snaps,
102 						struct ceph_cap_snap,
103 						ci_item);
104 		snapc = ceph_get_snap_context(capsnap->context);
105 		capsnap->dirty_pages++;
106 	} else {
107 		BUG_ON(!ci->i_head_snapc);
108 		snapc = ceph_get_snap_context(ci->i_head_snapc);
109 		++ci->i_wrbuffer_ref_head;
110 	}
111 	if (ci->i_wrbuffer_ref == 0)
112 		ihold(inode);
113 	++ci->i_wrbuffer_ref;
114 	dout("%p dirty_folio %p idx %lu head %d/%d -> %d/%d "
115 	     "snapc %p seq %lld (%d snaps)\n",
116 	     mapping->host, folio, folio->index,
117 	     ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
118 	     ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
119 	     snapc, snapc->seq, snapc->num_snaps);
120 	spin_unlock(&ci->i_ceph_lock);
121 
122 	/*
123 	 * Reference snap context in folio->private.  Also set
124 	 * PagePrivate so that we get invalidate_folio callback.
125 	 */
126 	VM_WARN_ON_FOLIO(folio->private, folio);
127 	folio_attach_private(folio, snapc);
128 
129 	return ceph_fscache_dirty_folio(mapping, folio);
130 }
131 
132 /*
133  * If we are truncating the full folio (i.e. offset == 0), adjust the
134  * dirty folio counters appropriately.  Only called if there is private
135  * data on the folio.
136  */
137 static void ceph_invalidate_folio(struct folio *folio, size_t offset,
138 				size_t length)
139 {
140 	struct inode *inode;
141 	struct ceph_inode_info *ci;
142 	struct ceph_snap_context *snapc;
143 
144 	inode = folio->mapping->host;
145 	ci = ceph_inode(inode);
146 
147 	if (offset != 0 || length != folio_size(folio)) {
148 		dout("%p invalidate_folio idx %lu partial dirty page %zu~%zu\n",
149 		     inode, folio->index, offset, length);
150 		return;
151 	}
152 
153 	WARN_ON(!folio_test_locked(folio));
154 	if (folio_test_private(folio)) {
155 		dout("%p invalidate_folio idx %lu full dirty page\n",
156 		     inode, folio->index);
157 
158 		snapc = folio_detach_private(folio);
159 		ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
160 		ceph_put_snap_context(snapc);
161 	}
162 
163 	folio_wait_fscache(folio);
164 }
165 
166 static bool ceph_release_folio(struct folio *folio, gfp_t gfp)
167 {
168 	struct inode *inode = folio->mapping->host;
169 
170 	dout("%llx:%llx release_folio idx %lu (%sdirty)\n",
171 	     ceph_vinop(inode),
172 	     folio->index, folio_test_dirty(folio) ? "" : "not ");
173 
174 	if (folio_test_private(folio))
175 		return false;
176 
177 	if (folio_test_fscache(folio)) {
178 		if (current_is_kswapd() || !(gfp & __GFP_FS))
179 			return false;
180 		folio_wait_fscache(folio);
181 	}
182 	ceph_fscache_note_page_release(inode);
183 	return true;
184 }
185 
186 static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
187 {
188 	struct inode *inode = rreq->inode;
189 	struct ceph_inode_info *ci = ceph_inode(inode);
190 	struct ceph_file_layout *lo = &ci->i_layout;
191 	unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
192 	loff_t end = rreq->start + rreq->len, new_end;
193 	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
194 	unsigned long max_len;
195 	u32 blockoff;
196 
197 	if (priv) {
198 		/* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
199 		if (priv->file_ra_disabled)
200 			max_pages = 0;
201 		else
202 			max_pages = priv->file_ra_pages;
203 
204 	}
205 
206 	/* Readahead is disabled */
207 	if (!max_pages)
208 		return;
209 
210 	max_len = max_pages << PAGE_SHIFT;
211 
212 	/*
213 	 * Try to expand the length forward by rounding up it to the next
214 	 * block, but do not exceed the file size, unless the original
215 	 * request already exceeds it.
216 	 */
217 	new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
218 	if (new_end > end && new_end <= rreq->start + max_len)
219 		rreq->len = new_end - rreq->start;
220 
221 	/* Try to expand the start downward */
222 	div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
223 	if (rreq->len + blockoff <= max_len) {
224 		rreq->start -= blockoff;
225 		rreq->len += blockoff;
226 	}
227 }
228 
229 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
230 {
231 	struct inode *inode = subreq->rreq->inode;
232 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
233 	struct ceph_inode_info *ci = ceph_inode(inode);
234 	u64 objno, objoff;
235 	u32 xlen;
236 
237 	/* Truncate the extent at the end of the current block */
238 	ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
239 				      &objno, &objoff, &xlen);
240 	subreq->len = min(xlen, fsc->mount_options->rsize);
241 	return true;
242 }
243 
244 static void finish_netfs_read(struct ceph_osd_request *req)
245 {
246 	struct inode *inode = req->r_inode;
247 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
248 	struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
249 	struct netfs_io_subrequest *subreq = req->r_priv;
250 	struct ceph_osd_req_op *op = &req->r_ops[0];
251 	int err = req->r_result;
252 	bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
253 
254 	ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
255 				 req->r_end_latency, osd_data->length, err);
256 
257 	dout("%s: result %d subreq->len=%zu i_size=%lld\n", __func__, req->r_result,
258 	     subreq->len, i_size_read(req->r_inode));
259 
260 	/* no object means success but no data */
261 	if (err == -ENOENT)
262 		err = 0;
263 	else if (err == -EBLOCKLISTED)
264 		fsc->blocklisted = true;
265 
266 	if (err >= 0) {
267 		if (sparse && err > 0)
268 			err = ceph_sparse_ext_map_end(op);
269 		if (err < subreq->len)
270 			__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
271 		if (IS_ENCRYPTED(inode) && err > 0) {
272 			err = ceph_fscrypt_decrypt_extents(inode,
273 					osd_data->pages, subreq->start,
274 					op->extent.sparse_ext,
275 					op->extent.sparse_ext_cnt);
276 			if (err > subreq->len)
277 				err = subreq->len;
278 		}
279 	}
280 
281 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
282 		ceph_put_page_vector(osd_data->pages,
283 				     calc_pages_for(osd_data->alignment,
284 					osd_data->length), false);
285 	}
286 	netfs_subreq_terminated(subreq, err, false);
287 	iput(req->r_inode);
288 	ceph_dec_osd_stopping_blocker(fsc->mdsc);
289 }
290 
291 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
292 {
293 	struct netfs_io_request *rreq = subreq->rreq;
294 	struct inode *inode = rreq->inode;
295 	struct ceph_mds_reply_info_parsed *rinfo;
296 	struct ceph_mds_reply_info_in *iinfo;
297 	struct ceph_mds_request *req;
298 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
299 	struct ceph_inode_info *ci = ceph_inode(inode);
300 	struct iov_iter iter;
301 	ssize_t err = 0;
302 	size_t len;
303 	int mode;
304 
305 	__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
306 	__clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
307 
308 	if (subreq->start >= inode->i_size)
309 		goto out;
310 
311 	/* We need to fetch the inline data. */
312 	mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
313 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
314 	if (IS_ERR(req)) {
315 		err = PTR_ERR(req);
316 		goto out;
317 	}
318 	req->r_ino1 = ci->i_vino;
319 	req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
320 	req->r_num_caps = 2;
321 
322 	err = ceph_mdsc_do_request(mdsc, NULL, req);
323 	if (err < 0)
324 		goto out;
325 
326 	rinfo = &req->r_reply_info;
327 	iinfo = &rinfo->targeti;
328 	if (iinfo->inline_version == CEPH_INLINE_NONE) {
329 		/* The data got uninlined */
330 		ceph_mdsc_put_request(req);
331 		return false;
332 	}
333 
334 	len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
335 	iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
336 	err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
337 	if (err == 0)
338 		err = -EFAULT;
339 
340 	ceph_mdsc_put_request(req);
341 out:
342 	netfs_subreq_terminated(subreq, err, false);
343 	return true;
344 }
345 
346 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
347 {
348 	struct netfs_io_request *rreq = subreq->rreq;
349 	struct inode *inode = rreq->inode;
350 	struct ceph_inode_info *ci = ceph_inode(inode);
351 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
352 	struct ceph_osd_request *req = NULL;
353 	struct ceph_vino vino = ceph_vino(inode);
354 	struct iov_iter iter;
355 	int err = 0;
356 	u64 len = subreq->len;
357 	bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
358 	u64 off = subreq->start;
359 
360 	if (ceph_inode_is_shutdown(inode)) {
361 		err = -EIO;
362 		goto out;
363 	}
364 
365 	if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
366 		return;
367 
368 	ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
369 
370 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
371 			off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
372 			CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
373 			NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
374 	if (IS_ERR(req)) {
375 		err = PTR_ERR(req);
376 		req = NULL;
377 		goto out;
378 	}
379 
380 	if (sparse) {
381 		err = ceph_alloc_sparse_ext_map(&req->r_ops[0]);
382 		if (err)
383 			goto out;
384 	}
385 
386 	dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
387 
388 	iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
389 
390 	/*
391 	 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
392 	 * encrypted inodes. We'd need infrastructure that handles an iov_iter
393 	 * instead of page arrays, and we don't have that as of yet. Once the
394 	 * dust settles on the write helpers and encrypt/decrypt routines for
395 	 * netfs, we should be able to rework this.
396 	 */
397 	if (IS_ENCRYPTED(inode)) {
398 		struct page **pages;
399 		size_t page_off;
400 
401 		err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
402 		if (err < 0) {
403 			dout("%s: iov_ter_get_pages_alloc returned %d\n",
404 			     __func__, err);
405 			goto out;
406 		}
407 
408 		/* should always give us a page-aligned read */
409 		WARN_ON_ONCE(page_off);
410 		len = err;
411 		err = 0;
412 
413 		osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
414 						 false);
415 	} else {
416 		osd_req_op_extent_osd_iter(req, 0, &iter);
417 	}
418 	if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
419 		err = -EIO;
420 		goto out;
421 	}
422 	req->r_callback = finish_netfs_read;
423 	req->r_priv = subreq;
424 	req->r_inode = inode;
425 	ihold(inode);
426 
427 	ceph_osdc_start_request(req->r_osdc, req);
428 out:
429 	ceph_osdc_put_request(req);
430 	if (err)
431 		netfs_subreq_terminated(subreq, err, false);
432 	dout("%s: result %d\n", __func__, err);
433 }
434 
435 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
436 {
437 	struct inode *inode = rreq->inode;
438 	int got = 0, want = CEPH_CAP_FILE_CACHE;
439 	struct ceph_netfs_request_data *priv;
440 	int ret = 0;
441 
442 	if (rreq->origin != NETFS_READAHEAD)
443 		return 0;
444 
445 	priv = kzalloc(sizeof(*priv), GFP_NOFS);
446 	if (!priv)
447 		return -ENOMEM;
448 
449 	if (file) {
450 		struct ceph_rw_context *rw_ctx;
451 		struct ceph_file_info *fi = file->private_data;
452 
453 		priv->file_ra_pages = file->f_ra.ra_pages;
454 		priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
455 
456 		rw_ctx = ceph_find_rw_context(fi);
457 		if (rw_ctx) {
458 			rreq->netfs_priv = priv;
459 			return 0;
460 		}
461 	}
462 
463 	/*
464 	 * readahead callers do not necessarily hold Fcb caps
465 	 * (e.g. fadvise, madvise).
466 	 */
467 	ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
468 	if (ret < 0) {
469 		dout("start_read %p, error getting cap\n", inode);
470 		goto out;
471 	}
472 
473 	if (!(got & want)) {
474 		dout("start_read %p, no cache cap\n", inode);
475 		ret = -EACCES;
476 		goto out;
477 	}
478 	if (ret == 0) {
479 		ret = -EACCES;
480 		goto out;
481 	}
482 
483 	priv->caps = got;
484 	rreq->netfs_priv = priv;
485 
486 out:
487 	if (ret < 0)
488 		kfree(priv);
489 
490 	return ret;
491 }
492 
493 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
494 {
495 	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
496 
497 	if (!priv)
498 		return;
499 
500 	if (priv->caps)
501 		ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
502 	kfree(priv);
503 	rreq->netfs_priv = NULL;
504 }
505 
506 const struct netfs_request_ops ceph_netfs_ops = {
507 	.init_request		= ceph_init_request,
508 	.free_request		= ceph_netfs_free_request,
509 	.begin_cache_operation	= ceph_begin_cache_operation,
510 	.issue_read		= ceph_netfs_issue_read,
511 	.expand_readahead	= ceph_netfs_expand_readahead,
512 	.clamp_length		= ceph_netfs_clamp_length,
513 	.check_write_begin	= ceph_netfs_check_write_begin,
514 };
515 
516 #ifdef CONFIG_CEPH_FSCACHE
517 static void ceph_set_page_fscache(struct page *page)
518 {
519 	set_page_fscache(page);
520 }
521 
522 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
523 {
524 	struct inode *inode = priv;
525 
526 	if (IS_ERR_VALUE(error) && error != -ENOBUFS)
527 		ceph_fscache_invalidate(inode, false);
528 }
529 
530 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
531 {
532 	struct ceph_inode_info *ci = ceph_inode(inode);
533 	struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
534 
535 	fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
536 			       ceph_fscache_write_terminated, inode, caching);
537 }
538 #else
539 static inline void ceph_set_page_fscache(struct page *page)
540 {
541 }
542 
543 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
544 {
545 }
546 #endif /* CONFIG_CEPH_FSCACHE */
547 
548 struct ceph_writeback_ctl
549 {
550 	loff_t i_size;
551 	u64 truncate_size;
552 	u32 truncate_seq;
553 	bool size_stable;
554 	bool head_snapc;
555 };
556 
557 /*
558  * Get ref for the oldest snapc for an inode with dirty data... that is, the
559  * only snap context we are allowed to write back.
560  */
561 static struct ceph_snap_context *
562 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
563 		   struct ceph_snap_context *page_snapc)
564 {
565 	struct ceph_inode_info *ci = ceph_inode(inode);
566 	struct ceph_snap_context *snapc = NULL;
567 	struct ceph_cap_snap *capsnap = NULL;
568 
569 	spin_lock(&ci->i_ceph_lock);
570 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
571 		dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
572 		     capsnap->context, capsnap->dirty_pages);
573 		if (!capsnap->dirty_pages)
574 			continue;
575 
576 		/* get i_size, truncate_{seq,size} for page_snapc? */
577 		if (snapc && capsnap->context != page_snapc)
578 			continue;
579 
580 		if (ctl) {
581 			if (capsnap->writing) {
582 				ctl->i_size = i_size_read(inode);
583 				ctl->size_stable = false;
584 			} else {
585 				ctl->i_size = capsnap->size;
586 				ctl->size_stable = true;
587 			}
588 			ctl->truncate_size = capsnap->truncate_size;
589 			ctl->truncate_seq = capsnap->truncate_seq;
590 			ctl->head_snapc = false;
591 		}
592 
593 		if (snapc)
594 			break;
595 
596 		snapc = ceph_get_snap_context(capsnap->context);
597 		if (!page_snapc ||
598 		    page_snapc == snapc ||
599 		    page_snapc->seq > snapc->seq)
600 			break;
601 	}
602 	if (!snapc && ci->i_wrbuffer_ref_head) {
603 		snapc = ceph_get_snap_context(ci->i_head_snapc);
604 		dout(" head snapc %p has %d dirty pages\n",
605 		     snapc, ci->i_wrbuffer_ref_head);
606 		if (ctl) {
607 			ctl->i_size = i_size_read(inode);
608 			ctl->truncate_size = ci->i_truncate_size;
609 			ctl->truncate_seq = ci->i_truncate_seq;
610 			ctl->size_stable = false;
611 			ctl->head_snapc = true;
612 		}
613 	}
614 	spin_unlock(&ci->i_ceph_lock);
615 	return snapc;
616 }
617 
618 static u64 get_writepages_data_length(struct inode *inode,
619 				      struct page *page, u64 start)
620 {
621 	struct ceph_inode_info *ci = ceph_inode(inode);
622 	struct ceph_snap_context *snapc;
623 	struct ceph_cap_snap *capsnap = NULL;
624 	u64 end = i_size_read(inode);
625 	u64 ret;
626 
627 	snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
628 	if (snapc != ci->i_head_snapc) {
629 		bool found = false;
630 		spin_lock(&ci->i_ceph_lock);
631 		list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
632 			if (capsnap->context == snapc) {
633 				if (!capsnap->writing)
634 					end = capsnap->size;
635 				found = true;
636 				break;
637 			}
638 		}
639 		spin_unlock(&ci->i_ceph_lock);
640 		WARN_ON(!found);
641 	}
642 	if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
643 		end = ceph_fscrypt_page_offset(page) + thp_size(page);
644 	ret = end > start ? end - start : 0;
645 	if (ret && fscrypt_is_bounce_page(page))
646 		ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
647 	return ret;
648 }
649 
650 /*
651  * Write a single page, but leave the page locked.
652  *
653  * If we get a write error, mark the mapping for error, but still adjust the
654  * dirty page accounting (i.e., page is no longer dirty).
655  */
656 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
657 {
658 	struct folio *folio = page_folio(page);
659 	struct inode *inode = page->mapping->host;
660 	struct ceph_inode_info *ci = ceph_inode(inode);
661 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
662 	struct ceph_snap_context *snapc, *oldest;
663 	loff_t page_off = page_offset(page);
664 	int err;
665 	loff_t len = thp_size(page);
666 	loff_t wlen;
667 	struct ceph_writeback_ctl ceph_wbc;
668 	struct ceph_osd_client *osdc = &fsc->client->osdc;
669 	struct ceph_osd_request *req;
670 	bool caching = ceph_is_cache_enabled(inode);
671 	struct page *bounce_page = NULL;
672 
673 	dout("writepage %p idx %lu\n", page, page->index);
674 
675 	if (ceph_inode_is_shutdown(inode))
676 		return -EIO;
677 
678 	/* verify this is a writeable snap context */
679 	snapc = page_snap_context(page);
680 	if (!snapc) {
681 		dout("writepage %p page %p not dirty?\n", inode, page);
682 		return 0;
683 	}
684 	oldest = get_oldest_context(inode, &ceph_wbc, snapc);
685 	if (snapc->seq > oldest->seq) {
686 		dout("writepage %p page %p snapc %p not writeable - noop\n",
687 		     inode, page, snapc);
688 		/* we should only noop if called by kswapd */
689 		WARN_ON(!(current->flags & PF_MEMALLOC));
690 		ceph_put_snap_context(oldest);
691 		redirty_page_for_writepage(wbc, page);
692 		return 0;
693 	}
694 	ceph_put_snap_context(oldest);
695 
696 	/* is this a partial page at end of file? */
697 	if (page_off >= ceph_wbc.i_size) {
698 		dout("folio at %lu beyond eof %llu\n", folio->index,
699 				ceph_wbc.i_size);
700 		folio_invalidate(folio, 0, folio_size(folio));
701 		return 0;
702 	}
703 
704 	if (ceph_wbc.i_size < page_off + len)
705 		len = ceph_wbc.i_size - page_off;
706 
707 	wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
708 	dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
709 	     inode, page, page->index, page_off, wlen, snapc, snapc->seq);
710 
711 	if (atomic_long_inc_return(&fsc->writeback_count) >
712 	    CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
713 		fsc->write_congested = true;
714 
715 	req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
716 				    page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
717 				    CEPH_OSD_FLAG_WRITE, snapc,
718 				    ceph_wbc.truncate_seq,
719 				    ceph_wbc.truncate_size, true);
720 	if (IS_ERR(req)) {
721 		redirty_page_for_writepage(wbc, page);
722 		return PTR_ERR(req);
723 	}
724 
725 	if (wlen < len)
726 		len = wlen;
727 
728 	set_page_writeback(page);
729 	if (caching)
730 		ceph_set_page_fscache(page);
731 	ceph_fscache_write_to_cache(inode, page_off, len, caching);
732 
733 	if (IS_ENCRYPTED(inode)) {
734 		bounce_page = fscrypt_encrypt_pagecache_blocks(page,
735 						    CEPH_FSCRYPT_BLOCK_SIZE, 0,
736 						    GFP_NOFS);
737 		if (IS_ERR(bounce_page)) {
738 			redirty_page_for_writepage(wbc, page);
739 			end_page_writeback(page);
740 			ceph_osdc_put_request(req);
741 			return PTR_ERR(bounce_page);
742 		}
743 	}
744 
745 	/* it may be a short write due to an object boundary */
746 	WARN_ON_ONCE(len > thp_size(page));
747 	osd_req_op_extent_osd_data_pages(req, 0,
748 			bounce_page ? &bounce_page : &page, wlen, 0,
749 			false, false);
750 	dout("writepage %llu~%llu (%llu bytes, %sencrypted)\n",
751 	     page_off, len, wlen, IS_ENCRYPTED(inode) ? "" : "not ");
752 
753 	req->r_mtime = inode->i_mtime;
754 	ceph_osdc_start_request(osdc, req);
755 	err = ceph_osdc_wait_request(osdc, req);
756 
757 	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
758 				  req->r_end_latency, len, err);
759 	fscrypt_free_bounce_page(bounce_page);
760 	ceph_osdc_put_request(req);
761 	if (err == 0)
762 		err = len;
763 
764 	if (err < 0) {
765 		struct writeback_control tmp_wbc;
766 		if (!wbc)
767 			wbc = &tmp_wbc;
768 		if (err == -ERESTARTSYS) {
769 			/* killed by SIGKILL */
770 			dout("writepage interrupted page %p\n", page);
771 			redirty_page_for_writepage(wbc, page);
772 			end_page_writeback(page);
773 			return err;
774 		}
775 		if (err == -EBLOCKLISTED)
776 			fsc->blocklisted = true;
777 		dout("writepage setting page/mapping error %d %p\n",
778 		     err, page);
779 		mapping_set_error(&inode->i_data, err);
780 		wbc->pages_skipped++;
781 	} else {
782 		dout("writepage cleaned page %p\n", page);
783 		err = 0;  /* vfs expects us to return 0 */
784 	}
785 	oldest = detach_page_private(page);
786 	WARN_ON_ONCE(oldest != snapc);
787 	end_page_writeback(page);
788 	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
789 	ceph_put_snap_context(snapc);  /* page's reference */
790 
791 	if (atomic_long_dec_return(&fsc->writeback_count) <
792 	    CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
793 		fsc->write_congested = false;
794 
795 	return err;
796 }
797 
798 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
799 {
800 	int err;
801 	struct inode *inode = page->mapping->host;
802 	BUG_ON(!inode);
803 	ihold(inode);
804 
805 	if (wbc->sync_mode == WB_SYNC_NONE &&
806 	    ceph_inode_to_fs_client(inode)->write_congested) {
807 		redirty_page_for_writepage(wbc, page);
808 		return AOP_WRITEPAGE_ACTIVATE;
809 	}
810 
811 	wait_on_page_fscache(page);
812 
813 	err = writepage_nounlock(page, wbc);
814 	if (err == -ERESTARTSYS) {
815 		/* direct memory reclaimer was killed by SIGKILL. return 0
816 		 * to prevent caller from setting mapping/page error */
817 		err = 0;
818 	}
819 	unlock_page(page);
820 	iput(inode);
821 	return err;
822 }
823 
824 /*
825  * async writeback completion handler.
826  *
827  * If we get an error, set the mapping error bit, but not the individual
828  * page error bits.
829  */
830 static void writepages_finish(struct ceph_osd_request *req)
831 {
832 	struct inode *inode = req->r_inode;
833 	struct ceph_inode_info *ci = ceph_inode(inode);
834 	struct ceph_osd_data *osd_data;
835 	struct page *page;
836 	int num_pages, total_pages = 0;
837 	int i, j;
838 	int rc = req->r_result;
839 	struct ceph_snap_context *snapc = req->r_snapc;
840 	struct address_space *mapping = inode->i_mapping;
841 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
842 	unsigned int len = 0;
843 	bool remove_page;
844 
845 	dout("writepages_finish %p rc %d\n", inode, rc);
846 	if (rc < 0) {
847 		mapping_set_error(mapping, rc);
848 		ceph_set_error_write(ci);
849 		if (rc == -EBLOCKLISTED)
850 			fsc->blocklisted = true;
851 	} else {
852 		ceph_clear_error_write(ci);
853 	}
854 
855 	/*
856 	 * We lost the cache cap, need to truncate the page before
857 	 * it is unlocked, otherwise we'd truncate it later in the
858 	 * page truncation thread, possibly losing some data that
859 	 * raced its way in
860 	 */
861 	remove_page = !(ceph_caps_issued(ci) &
862 			(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
863 
864 	/* clean all pages */
865 	for (i = 0; i < req->r_num_ops; i++) {
866 		if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
867 			pr_warn("%s incorrect op %d req %p index %d tid %llu\n",
868 				__func__, req->r_ops[i].op, req, i, req->r_tid);
869 			break;
870 		}
871 
872 		osd_data = osd_req_op_extent_osd_data(req, i);
873 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
874 		len += osd_data->length;
875 		num_pages = calc_pages_for((u64)osd_data->alignment,
876 					   (u64)osd_data->length);
877 		total_pages += num_pages;
878 		for (j = 0; j < num_pages; j++) {
879 			page = osd_data->pages[j];
880 			if (fscrypt_is_bounce_page(page)) {
881 				page = fscrypt_pagecache_page(page);
882 				fscrypt_free_bounce_page(osd_data->pages[j]);
883 				osd_data->pages[j] = page;
884 			}
885 			BUG_ON(!page);
886 			WARN_ON(!PageUptodate(page));
887 
888 			if (atomic_long_dec_return(&fsc->writeback_count) <
889 			     CONGESTION_OFF_THRESH(
890 					fsc->mount_options->congestion_kb))
891 				fsc->write_congested = false;
892 
893 			ceph_put_snap_context(detach_page_private(page));
894 			end_page_writeback(page);
895 			dout("unlocking %p\n", page);
896 
897 			if (remove_page)
898 				generic_error_remove_page(inode->i_mapping,
899 							  page);
900 
901 			unlock_page(page);
902 		}
903 		dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
904 		     inode, osd_data->length, rc >= 0 ? num_pages : 0);
905 
906 		release_pages(osd_data->pages, num_pages);
907 	}
908 
909 	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
910 				  req->r_end_latency, len, rc);
911 
912 	ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
913 
914 	osd_data = osd_req_op_extent_osd_data(req, 0);
915 	if (osd_data->pages_from_pool)
916 		mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
917 	else
918 		kfree(osd_data->pages);
919 	ceph_osdc_put_request(req);
920 	ceph_dec_osd_stopping_blocker(fsc->mdsc);
921 }
922 
923 /*
924  * initiate async writeback
925  */
926 static int ceph_writepages_start(struct address_space *mapping,
927 				 struct writeback_control *wbc)
928 {
929 	struct inode *inode = mapping->host;
930 	struct ceph_inode_info *ci = ceph_inode(inode);
931 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
932 	struct ceph_vino vino = ceph_vino(inode);
933 	pgoff_t index, start_index, end = -1;
934 	struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
935 	struct folio_batch fbatch;
936 	int rc = 0;
937 	unsigned int wsize = i_blocksize(inode);
938 	struct ceph_osd_request *req = NULL;
939 	struct ceph_writeback_ctl ceph_wbc;
940 	bool should_loop, range_whole = false;
941 	bool done = false;
942 	bool caching = ceph_is_cache_enabled(inode);
943 	xa_mark_t tag;
944 
945 	if (wbc->sync_mode == WB_SYNC_NONE &&
946 	    fsc->write_congested)
947 		return 0;
948 
949 	dout("writepages_start %p (mode=%s)\n", inode,
950 	     wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
951 	     (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
952 
953 	if (ceph_inode_is_shutdown(inode)) {
954 		if (ci->i_wrbuffer_ref > 0) {
955 			pr_warn_ratelimited(
956 				"writepage_start %p %lld forced umount\n",
957 				inode, ceph_ino(inode));
958 		}
959 		mapping_set_error(mapping, -EIO);
960 		return -EIO; /* we're in a forced umount, don't write! */
961 	}
962 	if (fsc->mount_options->wsize < wsize)
963 		wsize = fsc->mount_options->wsize;
964 
965 	folio_batch_init(&fbatch);
966 
967 	start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
968 	index = start_index;
969 
970 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
971 		tag = PAGECACHE_TAG_TOWRITE;
972 	} else {
973 		tag = PAGECACHE_TAG_DIRTY;
974 	}
975 retry:
976 	/* find oldest snap context with dirty data */
977 	snapc = get_oldest_context(inode, &ceph_wbc, NULL);
978 	if (!snapc) {
979 		/* hmm, why does writepages get called when there
980 		   is no dirty data? */
981 		dout(" no snap context with dirty data?\n");
982 		goto out;
983 	}
984 	dout(" oldest snapc is %p seq %lld (%d snaps)\n",
985 	     snapc, snapc->seq, snapc->num_snaps);
986 
987 	should_loop = false;
988 	if (ceph_wbc.head_snapc && snapc != last_snapc) {
989 		/* where to start/end? */
990 		if (wbc->range_cyclic) {
991 			index = start_index;
992 			end = -1;
993 			if (index > 0)
994 				should_loop = true;
995 			dout(" cyclic, start at %lu\n", index);
996 		} else {
997 			index = wbc->range_start >> PAGE_SHIFT;
998 			end = wbc->range_end >> PAGE_SHIFT;
999 			if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1000 				range_whole = true;
1001 			dout(" not cyclic, %lu to %lu\n", index, end);
1002 		}
1003 	} else if (!ceph_wbc.head_snapc) {
1004 		/* Do not respect wbc->range_{start,end}. Dirty pages
1005 		 * in that range can be associated with newer snapc.
1006 		 * They are not writeable until we write all dirty pages
1007 		 * associated with 'snapc' get written */
1008 		if (index > 0)
1009 			should_loop = true;
1010 		dout(" non-head snapc, range whole\n");
1011 	}
1012 
1013 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1014 		tag_pages_for_writeback(mapping, index, end);
1015 
1016 	ceph_put_snap_context(last_snapc);
1017 	last_snapc = snapc;
1018 
1019 	while (!done && index <= end) {
1020 		int num_ops = 0, op_idx;
1021 		unsigned i, nr_folios, max_pages, locked_pages = 0;
1022 		struct page **pages = NULL, **data_pages;
1023 		struct page *page;
1024 		pgoff_t strip_unit_end = 0;
1025 		u64 offset = 0, len = 0;
1026 		bool from_pool = false;
1027 
1028 		max_pages = wsize >> PAGE_SHIFT;
1029 
1030 get_more_pages:
1031 		nr_folios = filemap_get_folios_tag(mapping, &index,
1032 						   end, tag, &fbatch);
1033 		dout("pagevec_lookup_range_tag got %d\n", nr_folios);
1034 		if (!nr_folios && !locked_pages)
1035 			break;
1036 		for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1037 			page = &fbatch.folios[i]->page;
1038 			dout("? %p idx %lu\n", page, page->index);
1039 			if (locked_pages == 0)
1040 				lock_page(page);  /* first page */
1041 			else if (!trylock_page(page))
1042 				break;
1043 
1044 			/* only dirty pages, or our accounting breaks */
1045 			if (unlikely(!PageDirty(page)) ||
1046 			    unlikely(page->mapping != mapping)) {
1047 				dout("!dirty or !mapping %p\n", page);
1048 				unlock_page(page);
1049 				continue;
1050 			}
1051 			/* only if matching snap context */
1052 			pgsnapc = page_snap_context(page);
1053 			if (pgsnapc != snapc) {
1054 				dout("page snapc %p %lld != oldest %p %lld\n",
1055 				     pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1056 				if (!should_loop &&
1057 				    !ceph_wbc.head_snapc &&
1058 				    wbc->sync_mode != WB_SYNC_NONE)
1059 					should_loop = true;
1060 				unlock_page(page);
1061 				continue;
1062 			}
1063 			if (page_offset(page) >= ceph_wbc.i_size) {
1064 				struct folio *folio = page_folio(page);
1065 
1066 				dout("folio at %lu beyond eof %llu\n",
1067 				     folio->index, ceph_wbc.i_size);
1068 				if ((ceph_wbc.size_stable ||
1069 				    folio_pos(folio) >= i_size_read(inode)) &&
1070 				    folio_clear_dirty_for_io(folio))
1071 					folio_invalidate(folio, 0,
1072 							folio_size(folio));
1073 				folio_unlock(folio);
1074 				continue;
1075 			}
1076 			if (strip_unit_end && (page->index > strip_unit_end)) {
1077 				dout("end of strip unit %p\n", page);
1078 				unlock_page(page);
1079 				break;
1080 			}
1081 			if (PageWriteback(page) || PageFsCache(page)) {
1082 				if (wbc->sync_mode == WB_SYNC_NONE) {
1083 					dout("%p under writeback\n", page);
1084 					unlock_page(page);
1085 					continue;
1086 				}
1087 				dout("waiting on writeback %p\n", page);
1088 				wait_on_page_writeback(page);
1089 				wait_on_page_fscache(page);
1090 			}
1091 
1092 			if (!clear_page_dirty_for_io(page)) {
1093 				dout("%p !clear_page_dirty_for_io\n", page);
1094 				unlock_page(page);
1095 				continue;
1096 			}
1097 
1098 			/*
1099 			 * We have something to write.  If this is
1100 			 * the first locked page this time through,
1101 			 * calculate max possinle write size and
1102 			 * allocate a page array
1103 			 */
1104 			if (locked_pages == 0) {
1105 				u64 objnum;
1106 				u64 objoff;
1107 				u32 xlen;
1108 
1109 				/* prepare async write request */
1110 				offset = (u64)page_offset(page);
1111 				ceph_calc_file_object_mapping(&ci->i_layout,
1112 							      offset, wsize,
1113 							      &objnum, &objoff,
1114 							      &xlen);
1115 				len = xlen;
1116 
1117 				num_ops = 1;
1118 				strip_unit_end = page->index +
1119 					((len - 1) >> PAGE_SHIFT);
1120 
1121 				BUG_ON(pages);
1122 				max_pages = calc_pages_for(0, (u64)len);
1123 				pages = kmalloc_array(max_pages,
1124 						      sizeof(*pages),
1125 						      GFP_NOFS);
1126 				if (!pages) {
1127 					from_pool = true;
1128 					pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1129 					BUG_ON(!pages);
1130 				}
1131 
1132 				len = 0;
1133 			} else if (page->index !=
1134 				   (offset + len) >> PAGE_SHIFT) {
1135 				if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1136 							     CEPH_OSD_MAX_OPS)) {
1137 					redirty_page_for_writepage(wbc, page);
1138 					unlock_page(page);
1139 					break;
1140 				}
1141 
1142 				num_ops++;
1143 				offset = (u64)page_offset(page);
1144 				len = 0;
1145 			}
1146 
1147 			/* note position of first page in fbatch */
1148 			dout("%p will write page %p idx %lu\n",
1149 			     inode, page, page->index);
1150 
1151 			if (atomic_long_inc_return(&fsc->writeback_count) >
1152 			    CONGESTION_ON_THRESH(
1153 				    fsc->mount_options->congestion_kb))
1154 				fsc->write_congested = true;
1155 
1156 			if (IS_ENCRYPTED(inode)) {
1157 				pages[locked_pages] =
1158 					fscrypt_encrypt_pagecache_blocks(page,
1159 						PAGE_SIZE, 0,
1160 						locked_pages ? GFP_NOWAIT : GFP_NOFS);
1161 				if (IS_ERR(pages[locked_pages])) {
1162 					if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1163 						pr_err("%s: inode->i_blkbits=%hhu\n",
1164 							__func__, inode->i_blkbits);
1165 					/* better not fail on first page! */
1166 					BUG_ON(locked_pages == 0);
1167 					pages[locked_pages] = NULL;
1168 					redirty_page_for_writepage(wbc, page);
1169 					unlock_page(page);
1170 					break;
1171 				}
1172 				++locked_pages;
1173 			} else {
1174 				pages[locked_pages++] = page;
1175 			}
1176 
1177 			fbatch.folios[i] = NULL;
1178 			len += thp_size(page);
1179 		}
1180 
1181 		/* did we get anything? */
1182 		if (!locked_pages)
1183 			goto release_folios;
1184 		if (i) {
1185 			unsigned j, n = 0;
1186 			/* shift unused page to beginning of fbatch */
1187 			for (j = 0; j < nr_folios; j++) {
1188 				if (!fbatch.folios[j])
1189 					continue;
1190 				if (n < j)
1191 					fbatch.folios[n] = fbatch.folios[j];
1192 				n++;
1193 			}
1194 			fbatch.nr = n;
1195 
1196 			if (nr_folios && i == nr_folios &&
1197 			    locked_pages < max_pages) {
1198 				dout("reached end fbatch, trying for more\n");
1199 				folio_batch_release(&fbatch);
1200 				goto get_more_pages;
1201 			}
1202 		}
1203 
1204 new_request:
1205 		offset = ceph_fscrypt_page_offset(pages[0]);
1206 		len = wsize;
1207 
1208 		req = ceph_osdc_new_request(&fsc->client->osdc,
1209 					&ci->i_layout, vino,
1210 					offset, &len, 0, num_ops,
1211 					CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1212 					snapc, ceph_wbc.truncate_seq,
1213 					ceph_wbc.truncate_size, false);
1214 		if (IS_ERR(req)) {
1215 			req = ceph_osdc_new_request(&fsc->client->osdc,
1216 						&ci->i_layout, vino,
1217 						offset, &len, 0,
1218 						min(num_ops,
1219 						    CEPH_OSD_SLAB_OPS),
1220 						CEPH_OSD_OP_WRITE,
1221 						CEPH_OSD_FLAG_WRITE,
1222 						snapc, ceph_wbc.truncate_seq,
1223 						ceph_wbc.truncate_size, true);
1224 			BUG_ON(IS_ERR(req));
1225 		}
1226 		BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1227 			     thp_size(pages[locked_pages - 1]) - offset);
1228 
1229 		if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1230 			rc = -EIO;
1231 			goto release_folios;
1232 		}
1233 		req->r_callback = writepages_finish;
1234 		req->r_inode = inode;
1235 
1236 		/* Format the osd request message and submit the write */
1237 		len = 0;
1238 		data_pages = pages;
1239 		op_idx = 0;
1240 		for (i = 0; i < locked_pages; i++) {
1241 			struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1242 
1243 			u64 cur_offset = page_offset(page);
1244 			/*
1245 			 * Discontinuity in page range? Ceph can handle that by just passing
1246 			 * multiple extents in the write op.
1247 			 */
1248 			if (offset + len != cur_offset) {
1249 				/* If it's full, stop here */
1250 				if (op_idx + 1 == req->r_num_ops)
1251 					break;
1252 
1253 				/* Kick off an fscache write with what we have so far. */
1254 				ceph_fscache_write_to_cache(inode, offset, len, caching);
1255 
1256 				/* Start a new extent */
1257 				osd_req_op_extent_dup_last(req, op_idx,
1258 							   cur_offset - offset);
1259 				dout("writepages got pages at %llu~%llu\n",
1260 				     offset, len);
1261 				osd_req_op_extent_osd_data_pages(req, op_idx,
1262 							data_pages, len, 0,
1263 							from_pool, false);
1264 				osd_req_op_extent_update(req, op_idx, len);
1265 
1266 				len = 0;
1267 				offset = cur_offset;
1268 				data_pages = pages + i;
1269 				op_idx++;
1270 			}
1271 
1272 			set_page_writeback(page);
1273 			if (caching)
1274 				ceph_set_page_fscache(page);
1275 			len += thp_size(page);
1276 		}
1277 		ceph_fscache_write_to_cache(inode, offset, len, caching);
1278 
1279 		if (ceph_wbc.size_stable) {
1280 			len = min(len, ceph_wbc.i_size - offset);
1281 		} else if (i == locked_pages) {
1282 			/* writepages_finish() clears writeback pages
1283 			 * according to the data length, so make sure
1284 			 * data length covers all locked pages */
1285 			u64 min_len = len + 1 - thp_size(page);
1286 			len = get_writepages_data_length(inode, pages[i - 1],
1287 							 offset);
1288 			len = max(len, min_len);
1289 		}
1290 		if (IS_ENCRYPTED(inode))
1291 			len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1292 
1293 		dout("writepages got pages at %llu~%llu\n", offset, len);
1294 
1295 		if (IS_ENCRYPTED(inode) &&
1296 		    ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1297 			pr_warn("%s: bad encrypted write offset=%lld len=%llu\n",
1298 				__func__, offset, len);
1299 
1300 		osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1301 						 0, from_pool, false);
1302 		osd_req_op_extent_update(req, op_idx, len);
1303 
1304 		BUG_ON(op_idx + 1 != req->r_num_ops);
1305 
1306 		from_pool = false;
1307 		if (i < locked_pages) {
1308 			BUG_ON(num_ops <= req->r_num_ops);
1309 			num_ops -= req->r_num_ops;
1310 			locked_pages -= i;
1311 
1312 			/* allocate new pages array for next request */
1313 			data_pages = pages;
1314 			pages = kmalloc_array(locked_pages, sizeof(*pages),
1315 					      GFP_NOFS);
1316 			if (!pages) {
1317 				from_pool = true;
1318 				pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1319 				BUG_ON(!pages);
1320 			}
1321 			memcpy(pages, data_pages + i,
1322 			       locked_pages * sizeof(*pages));
1323 			memset(data_pages + i, 0,
1324 			       locked_pages * sizeof(*pages));
1325 		} else {
1326 			BUG_ON(num_ops != req->r_num_ops);
1327 			index = pages[i - 1]->index + 1;
1328 			/* request message now owns the pages array */
1329 			pages = NULL;
1330 		}
1331 
1332 		req->r_mtime = inode->i_mtime;
1333 		ceph_osdc_start_request(&fsc->client->osdc, req);
1334 		req = NULL;
1335 
1336 		wbc->nr_to_write -= i;
1337 		if (pages)
1338 			goto new_request;
1339 
1340 		/*
1341 		 * We stop writing back only if we are not doing
1342 		 * integrity sync. In case of integrity sync we have to
1343 		 * keep going until we have written all the pages
1344 		 * we tagged for writeback prior to entering this loop.
1345 		 */
1346 		if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1347 			done = true;
1348 
1349 release_folios:
1350 		dout("folio_batch release on %d folios (%p)\n", (int)fbatch.nr,
1351 		     fbatch.nr ? fbatch.folios[0] : NULL);
1352 		folio_batch_release(&fbatch);
1353 	}
1354 
1355 	if (should_loop && !done) {
1356 		/* more to do; loop back to beginning of file */
1357 		dout("writepages looping back to beginning of file\n");
1358 		end = start_index - 1; /* OK even when start_index == 0 */
1359 
1360 		/* to write dirty pages associated with next snapc,
1361 		 * we need to wait until current writes complete */
1362 		if (wbc->sync_mode != WB_SYNC_NONE &&
1363 		    start_index == 0 && /* all dirty pages were checked */
1364 		    !ceph_wbc.head_snapc) {
1365 			struct page *page;
1366 			unsigned i, nr;
1367 			index = 0;
1368 			while ((index <= end) &&
1369 			       (nr = filemap_get_folios_tag(mapping, &index,
1370 						(pgoff_t)-1,
1371 						PAGECACHE_TAG_WRITEBACK,
1372 						&fbatch))) {
1373 				for (i = 0; i < nr; i++) {
1374 					page = &fbatch.folios[i]->page;
1375 					if (page_snap_context(page) != snapc)
1376 						continue;
1377 					wait_on_page_writeback(page);
1378 				}
1379 				folio_batch_release(&fbatch);
1380 				cond_resched();
1381 			}
1382 		}
1383 
1384 		start_index = 0;
1385 		index = 0;
1386 		goto retry;
1387 	}
1388 
1389 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1390 		mapping->writeback_index = index;
1391 
1392 out:
1393 	ceph_osdc_put_request(req);
1394 	ceph_put_snap_context(last_snapc);
1395 	dout("writepages dend - startone, rc = %d\n", rc);
1396 	return rc;
1397 }
1398 
1399 
1400 
1401 /*
1402  * See if a given @snapc is either writeable, or already written.
1403  */
1404 static int context_is_writeable_or_written(struct inode *inode,
1405 					   struct ceph_snap_context *snapc)
1406 {
1407 	struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1408 	int ret = !oldest || snapc->seq <= oldest->seq;
1409 
1410 	ceph_put_snap_context(oldest);
1411 	return ret;
1412 }
1413 
1414 /**
1415  * ceph_find_incompatible - find an incompatible context and return it
1416  * @page: page being dirtied
1417  *
1418  * We are only allowed to write into/dirty a page if the page is
1419  * clean, or already dirty within the same snap context. Returns a
1420  * conflicting context if there is one, NULL if there isn't, or a
1421  * negative error code on other errors.
1422  *
1423  * Must be called with page lock held.
1424  */
1425 static struct ceph_snap_context *
1426 ceph_find_incompatible(struct page *page)
1427 {
1428 	struct inode *inode = page->mapping->host;
1429 	struct ceph_inode_info *ci = ceph_inode(inode);
1430 
1431 	if (ceph_inode_is_shutdown(inode)) {
1432 		dout(" page %p %llx:%llx is shutdown\n", page,
1433 		     ceph_vinop(inode));
1434 		return ERR_PTR(-ESTALE);
1435 	}
1436 
1437 	for (;;) {
1438 		struct ceph_snap_context *snapc, *oldest;
1439 
1440 		wait_on_page_writeback(page);
1441 
1442 		snapc = page_snap_context(page);
1443 		if (!snapc || snapc == ci->i_head_snapc)
1444 			break;
1445 
1446 		/*
1447 		 * this page is already dirty in another (older) snap
1448 		 * context!  is it writeable now?
1449 		 */
1450 		oldest = get_oldest_context(inode, NULL, NULL);
1451 		if (snapc->seq > oldest->seq) {
1452 			/* not writeable -- return it for the caller to deal with */
1453 			ceph_put_snap_context(oldest);
1454 			dout(" page %p snapc %p not current or oldest\n", page, snapc);
1455 			return ceph_get_snap_context(snapc);
1456 		}
1457 		ceph_put_snap_context(oldest);
1458 
1459 		/* yay, writeable, do it now (without dropping page lock) */
1460 		dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1461 		if (clear_page_dirty_for_io(page)) {
1462 			int r = writepage_nounlock(page, NULL);
1463 			if (r < 0)
1464 				return ERR_PTR(r);
1465 		}
1466 	}
1467 	return NULL;
1468 }
1469 
1470 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1471 					struct folio **foliop, void **_fsdata)
1472 {
1473 	struct inode *inode = file_inode(file);
1474 	struct ceph_inode_info *ci = ceph_inode(inode);
1475 	struct ceph_snap_context *snapc;
1476 
1477 	snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1478 	if (snapc) {
1479 		int r;
1480 
1481 		folio_unlock(*foliop);
1482 		folio_put(*foliop);
1483 		*foliop = NULL;
1484 		if (IS_ERR(snapc))
1485 			return PTR_ERR(snapc);
1486 
1487 		ceph_queue_writeback(inode);
1488 		r = wait_event_killable(ci->i_cap_wq,
1489 					context_is_writeable_or_written(inode, snapc));
1490 		ceph_put_snap_context(snapc);
1491 		return r == 0 ? -EAGAIN : r;
1492 	}
1493 	return 0;
1494 }
1495 
1496 /*
1497  * We are only allowed to write into/dirty the page if the page is
1498  * clean, or already dirty within the same snap context.
1499  */
1500 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1501 			    loff_t pos, unsigned len,
1502 			    struct page **pagep, void **fsdata)
1503 {
1504 	struct inode *inode = file_inode(file);
1505 	struct ceph_inode_info *ci = ceph_inode(inode);
1506 	struct folio *folio = NULL;
1507 	int r;
1508 
1509 	r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1510 	if (r < 0)
1511 		return r;
1512 
1513 	folio_wait_fscache(folio);
1514 	WARN_ON_ONCE(!folio_test_locked(folio));
1515 	*pagep = &folio->page;
1516 	return 0;
1517 }
1518 
1519 /*
1520  * we don't do anything in here that simple_write_end doesn't do
1521  * except adjust dirty page accounting
1522  */
1523 static int ceph_write_end(struct file *file, struct address_space *mapping,
1524 			  loff_t pos, unsigned len, unsigned copied,
1525 			  struct page *subpage, void *fsdata)
1526 {
1527 	struct folio *folio = page_folio(subpage);
1528 	struct inode *inode = file_inode(file);
1529 	bool check_cap = false;
1530 
1531 	dout("write_end file %p inode %p folio %p %d~%d (%d)\n", file,
1532 	     inode, folio, (int)pos, (int)copied, (int)len);
1533 
1534 	if (!folio_test_uptodate(folio)) {
1535 		/* just return that nothing was copied on a short copy */
1536 		if (copied < len) {
1537 			copied = 0;
1538 			goto out;
1539 		}
1540 		folio_mark_uptodate(folio);
1541 	}
1542 
1543 	/* did file size increase? */
1544 	if (pos+copied > i_size_read(inode))
1545 		check_cap = ceph_inode_set_size(inode, pos+copied);
1546 
1547 	folio_mark_dirty(folio);
1548 
1549 out:
1550 	folio_unlock(folio);
1551 	folio_put(folio);
1552 
1553 	if (check_cap)
1554 		ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1555 
1556 	return copied;
1557 }
1558 
1559 const struct address_space_operations ceph_aops = {
1560 	.read_folio = netfs_read_folio,
1561 	.readahead = netfs_readahead,
1562 	.writepage = ceph_writepage,
1563 	.writepages = ceph_writepages_start,
1564 	.write_begin = ceph_write_begin,
1565 	.write_end = ceph_write_end,
1566 	.dirty_folio = ceph_dirty_folio,
1567 	.invalidate_folio = ceph_invalidate_folio,
1568 	.release_folio = ceph_release_folio,
1569 	.direct_IO = noop_direct_IO,
1570 };
1571 
1572 static void ceph_block_sigs(sigset_t *oldset)
1573 {
1574 	sigset_t mask;
1575 	siginitsetinv(&mask, sigmask(SIGKILL));
1576 	sigprocmask(SIG_BLOCK, &mask, oldset);
1577 }
1578 
1579 static void ceph_restore_sigs(sigset_t *oldset)
1580 {
1581 	sigprocmask(SIG_SETMASK, oldset, NULL);
1582 }
1583 
1584 /*
1585  * vm ops
1586  */
1587 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1588 {
1589 	struct vm_area_struct *vma = vmf->vma;
1590 	struct inode *inode = file_inode(vma->vm_file);
1591 	struct ceph_inode_info *ci = ceph_inode(inode);
1592 	struct ceph_file_info *fi = vma->vm_file->private_data;
1593 	loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1594 	int want, got, err;
1595 	sigset_t oldset;
1596 	vm_fault_t ret = VM_FAULT_SIGBUS;
1597 
1598 	if (ceph_inode_is_shutdown(inode))
1599 		return ret;
1600 
1601 	ceph_block_sigs(&oldset);
1602 
1603 	dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1604 	     inode, ceph_vinop(inode), off);
1605 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1606 		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1607 	else
1608 		want = CEPH_CAP_FILE_CACHE;
1609 
1610 	got = 0;
1611 	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1612 	if (err < 0)
1613 		goto out_restore;
1614 
1615 	dout("filemap_fault %p %llu got cap refs on %s\n",
1616 	     inode, off, ceph_cap_string(got));
1617 
1618 	if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1619 	    !ceph_has_inline_data(ci)) {
1620 		CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1621 		ceph_add_rw_context(fi, &rw_ctx);
1622 		ret = filemap_fault(vmf);
1623 		ceph_del_rw_context(fi, &rw_ctx);
1624 		dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1625 		     inode, off, ceph_cap_string(got), ret);
1626 	} else
1627 		err = -EAGAIN;
1628 
1629 	ceph_put_cap_refs(ci, got);
1630 
1631 	if (err != -EAGAIN)
1632 		goto out_restore;
1633 
1634 	/* read inline data */
1635 	if (off >= PAGE_SIZE) {
1636 		/* does not support inline data > PAGE_SIZE */
1637 		ret = VM_FAULT_SIGBUS;
1638 	} else {
1639 		struct address_space *mapping = inode->i_mapping;
1640 		struct page *page;
1641 
1642 		filemap_invalidate_lock_shared(mapping);
1643 		page = find_or_create_page(mapping, 0,
1644 				mapping_gfp_constraint(mapping, ~__GFP_FS));
1645 		if (!page) {
1646 			ret = VM_FAULT_OOM;
1647 			goto out_inline;
1648 		}
1649 		err = __ceph_do_getattr(inode, page,
1650 					 CEPH_STAT_CAP_INLINE_DATA, true);
1651 		if (err < 0 || off >= i_size_read(inode)) {
1652 			unlock_page(page);
1653 			put_page(page);
1654 			ret = vmf_error(err);
1655 			goto out_inline;
1656 		}
1657 		if (err < PAGE_SIZE)
1658 			zero_user_segment(page, err, PAGE_SIZE);
1659 		else
1660 			flush_dcache_page(page);
1661 		SetPageUptodate(page);
1662 		vmf->page = page;
1663 		ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1664 out_inline:
1665 		filemap_invalidate_unlock_shared(mapping);
1666 		dout("filemap_fault %p %llu read inline data ret %x\n",
1667 		     inode, off, ret);
1668 	}
1669 out_restore:
1670 	ceph_restore_sigs(&oldset);
1671 	if (err < 0)
1672 		ret = vmf_error(err);
1673 
1674 	return ret;
1675 }
1676 
1677 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1678 {
1679 	struct vm_area_struct *vma = vmf->vma;
1680 	struct inode *inode = file_inode(vma->vm_file);
1681 	struct ceph_inode_info *ci = ceph_inode(inode);
1682 	struct ceph_file_info *fi = vma->vm_file->private_data;
1683 	struct ceph_cap_flush *prealloc_cf;
1684 	struct page *page = vmf->page;
1685 	loff_t off = page_offset(page);
1686 	loff_t size = i_size_read(inode);
1687 	size_t len;
1688 	int want, got, err;
1689 	sigset_t oldset;
1690 	vm_fault_t ret = VM_FAULT_SIGBUS;
1691 
1692 	if (ceph_inode_is_shutdown(inode))
1693 		return ret;
1694 
1695 	prealloc_cf = ceph_alloc_cap_flush();
1696 	if (!prealloc_cf)
1697 		return VM_FAULT_OOM;
1698 
1699 	sb_start_pagefault(inode->i_sb);
1700 	ceph_block_sigs(&oldset);
1701 
1702 	if (off + thp_size(page) <= size)
1703 		len = thp_size(page);
1704 	else
1705 		len = offset_in_thp(page, size);
1706 
1707 	dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1708 	     inode, ceph_vinop(inode), off, len, size);
1709 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1710 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1711 	else
1712 		want = CEPH_CAP_FILE_BUFFER;
1713 
1714 	got = 0;
1715 	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1716 	if (err < 0)
1717 		goto out_free;
1718 
1719 	dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1720 	     inode, off, len, ceph_cap_string(got));
1721 
1722 	/* Update time before taking page lock */
1723 	file_update_time(vma->vm_file);
1724 	inode_inc_iversion_raw(inode);
1725 
1726 	do {
1727 		struct ceph_snap_context *snapc;
1728 
1729 		lock_page(page);
1730 
1731 		if (page_mkwrite_check_truncate(page, inode) < 0) {
1732 			unlock_page(page);
1733 			ret = VM_FAULT_NOPAGE;
1734 			break;
1735 		}
1736 
1737 		snapc = ceph_find_incompatible(page);
1738 		if (!snapc) {
1739 			/* success.  we'll keep the page locked. */
1740 			set_page_dirty(page);
1741 			ret = VM_FAULT_LOCKED;
1742 			break;
1743 		}
1744 
1745 		unlock_page(page);
1746 
1747 		if (IS_ERR(snapc)) {
1748 			ret = VM_FAULT_SIGBUS;
1749 			break;
1750 		}
1751 
1752 		ceph_queue_writeback(inode);
1753 		err = wait_event_killable(ci->i_cap_wq,
1754 				context_is_writeable_or_written(inode, snapc));
1755 		ceph_put_snap_context(snapc);
1756 	} while (err == 0);
1757 
1758 	if (ret == VM_FAULT_LOCKED) {
1759 		int dirty;
1760 		spin_lock(&ci->i_ceph_lock);
1761 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1762 					       &prealloc_cf);
1763 		spin_unlock(&ci->i_ceph_lock);
1764 		if (dirty)
1765 			__mark_inode_dirty(inode, dirty);
1766 	}
1767 
1768 	dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1769 	     inode, off, len, ceph_cap_string(got), ret);
1770 	ceph_put_cap_refs_async(ci, got);
1771 out_free:
1772 	ceph_restore_sigs(&oldset);
1773 	sb_end_pagefault(inode->i_sb);
1774 	ceph_free_cap_flush(prealloc_cf);
1775 	if (err < 0)
1776 		ret = vmf_error(err);
1777 	return ret;
1778 }
1779 
1780 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1781 			   char	*data, size_t len)
1782 {
1783 	struct address_space *mapping = inode->i_mapping;
1784 	struct page *page;
1785 
1786 	if (locked_page) {
1787 		page = locked_page;
1788 	} else {
1789 		if (i_size_read(inode) == 0)
1790 			return;
1791 		page = find_or_create_page(mapping, 0,
1792 					   mapping_gfp_constraint(mapping,
1793 					   ~__GFP_FS));
1794 		if (!page)
1795 			return;
1796 		if (PageUptodate(page)) {
1797 			unlock_page(page);
1798 			put_page(page);
1799 			return;
1800 		}
1801 	}
1802 
1803 	dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1804 	     inode, ceph_vinop(inode), len, locked_page);
1805 
1806 	if (len > 0) {
1807 		void *kaddr = kmap_atomic(page);
1808 		memcpy(kaddr, data, len);
1809 		kunmap_atomic(kaddr);
1810 	}
1811 
1812 	if (page != locked_page) {
1813 		if (len < PAGE_SIZE)
1814 			zero_user_segment(page, len, PAGE_SIZE);
1815 		else
1816 			flush_dcache_page(page);
1817 
1818 		SetPageUptodate(page);
1819 		unlock_page(page);
1820 		put_page(page);
1821 	}
1822 }
1823 
1824 int ceph_uninline_data(struct file *file)
1825 {
1826 	struct inode *inode = file_inode(file);
1827 	struct ceph_inode_info *ci = ceph_inode(inode);
1828 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1829 	struct ceph_osd_request *req = NULL;
1830 	struct ceph_cap_flush *prealloc_cf = NULL;
1831 	struct folio *folio = NULL;
1832 	u64 inline_version = CEPH_INLINE_NONE;
1833 	struct page *pages[1];
1834 	int err = 0;
1835 	u64 len;
1836 
1837 	spin_lock(&ci->i_ceph_lock);
1838 	inline_version = ci->i_inline_version;
1839 	spin_unlock(&ci->i_ceph_lock);
1840 
1841 	dout("uninline_data %p %llx.%llx inline_version %llu\n",
1842 	     inode, ceph_vinop(inode), inline_version);
1843 
1844 	if (ceph_inode_is_shutdown(inode)) {
1845 		err = -EIO;
1846 		goto out;
1847 	}
1848 
1849 	if (inline_version == CEPH_INLINE_NONE)
1850 		return 0;
1851 
1852 	prealloc_cf = ceph_alloc_cap_flush();
1853 	if (!prealloc_cf)
1854 		return -ENOMEM;
1855 
1856 	if (inline_version == 1) /* initial version, no data */
1857 		goto out_uninline;
1858 
1859 	folio = read_mapping_folio(inode->i_mapping, 0, file);
1860 	if (IS_ERR(folio)) {
1861 		err = PTR_ERR(folio);
1862 		goto out;
1863 	}
1864 
1865 	folio_lock(folio);
1866 
1867 	len = i_size_read(inode);
1868 	if (len > folio_size(folio))
1869 		len = folio_size(folio);
1870 
1871 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1872 				    ceph_vino(inode), 0, &len, 0, 1,
1873 				    CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1874 				    NULL, 0, 0, false);
1875 	if (IS_ERR(req)) {
1876 		err = PTR_ERR(req);
1877 		goto out_unlock;
1878 	}
1879 
1880 	req->r_mtime = inode->i_mtime;
1881 	ceph_osdc_start_request(&fsc->client->osdc, req);
1882 	err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1883 	ceph_osdc_put_request(req);
1884 	if (err < 0)
1885 		goto out_unlock;
1886 
1887 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1888 				    ceph_vino(inode), 0, &len, 1, 3,
1889 				    CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1890 				    NULL, ci->i_truncate_seq,
1891 				    ci->i_truncate_size, false);
1892 	if (IS_ERR(req)) {
1893 		err = PTR_ERR(req);
1894 		goto out_unlock;
1895 	}
1896 
1897 	pages[0] = folio_page(folio, 0);
1898 	osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1899 
1900 	{
1901 		__le64 xattr_buf = cpu_to_le64(inline_version);
1902 		err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1903 					    "inline_version", &xattr_buf,
1904 					    sizeof(xattr_buf),
1905 					    CEPH_OSD_CMPXATTR_OP_GT,
1906 					    CEPH_OSD_CMPXATTR_MODE_U64);
1907 		if (err)
1908 			goto out_put_req;
1909 	}
1910 
1911 	{
1912 		char xattr_buf[32];
1913 		int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1914 					 "%llu", inline_version);
1915 		err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1916 					    "inline_version",
1917 					    xattr_buf, xattr_len, 0, 0);
1918 		if (err)
1919 			goto out_put_req;
1920 	}
1921 
1922 	req->r_mtime = inode->i_mtime;
1923 	ceph_osdc_start_request(&fsc->client->osdc, req);
1924 	err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1925 
1926 	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1927 				  req->r_end_latency, len, err);
1928 
1929 out_uninline:
1930 	if (!err) {
1931 		int dirty;
1932 
1933 		/* Set to CAP_INLINE_NONE and dirty the caps */
1934 		down_read(&fsc->mdsc->snap_rwsem);
1935 		spin_lock(&ci->i_ceph_lock);
1936 		ci->i_inline_version = CEPH_INLINE_NONE;
1937 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1938 		spin_unlock(&ci->i_ceph_lock);
1939 		up_read(&fsc->mdsc->snap_rwsem);
1940 		if (dirty)
1941 			__mark_inode_dirty(inode, dirty);
1942 	}
1943 out_put_req:
1944 	ceph_osdc_put_request(req);
1945 	if (err == -ECANCELED)
1946 		err = 0;
1947 out_unlock:
1948 	if (folio) {
1949 		folio_unlock(folio);
1950 		folio_put(folio);
1951 	}
1952 out:
1953 	ceph_free_cap_flush(prealloc_cf);
1954 	dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1955 	     inode, ceph_vinop(inode), inline_version, err);
1956 	return err;
1957 }
1958 
1959 static const struct vm_operations_struct ceph_vmops = {
1960 	.fault		= ceph_filemap_fault,
1961 	.page_mkwrite	= ceph_page_mkwrite,
1962 };
1963 
1964 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1965 {
1966 	struct address_space *mapping = file->f_mapping;
1967 
1968 	if (!mapping->a_ops->read_folio)
1969 		return -ENOEXEC;
1970 	vma->vm_ops = &ceph_vmops;
1971 	return 0;
1972 }
1973 
1974 enum {
1975 	POOL_READ	= 1,
1976 	POOL_WRITE	= 2,
1977 };
1978 
1979 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1980 				s64 pool, struct ceph_string *pool_ns)
1981 {
1982 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
1983 	struct ceph_mds_client *mdsc = fsc->mdsc;
1984 	struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1985 	struct rb_node **p, *parent;
1986 	struct ceph_pool_perm *perm;
1987 	struct page **pages;
1988 	size_t pool_ns_len;
1989 	int err = 0, err2 = 0, have = 0;
1990 
1991 	down_read(&mdsc->pool_perm_rwsem);
1992 	p = &mdsc->pool_perm_tree.rb_node;
1993 	while (*p) {
1994 		perm = rb_entry(*p, struct ceph_pool_perm, node);
1995 		if (pool < perm->pool)
1996 			p = &(*p)->rb_left;
1997 		else if (pool > perm->pool)
1998 			p = &(*p)->rb_right;
1999 		else {
2000 			int ret = ceph_compare_string(pool_ns,
2001 						perm->pool_ns,
2002 						perm->pool_ns_len);
2003 			if (ret < 0)
2004 				p = &(*p)->rb_left;
2005 			else if (ret > 0)
2006 				p = &(*p)->rb_right;
2007 			else {
2008 				have = perm->perm;
2009 				break;
2010 			}
2011 		}
2012 	}
2013 	up_read(&mdsc->pool_perm_rwsem);
2014 	if (*p)
2015 		goto out;
2016 
2017 	if (pool_ns)
2018 		dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
2019 		     pool, (int)pool_ns->len, pool_ns->str);
2020 	else
2021 		dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
2022 
2023 	down_write(&mdsc->pool_perm_rwsem);
2024 	p = &mdsc->pool_perm_tree.rb_node;
2025 	parent = NULL;
2026 	while (*p) {
2027 		parent = *p;
2028 		perm = rb_entry(parent, struct ceph_pool_perm, node);
2029 		if (pool < perm->pool)
2030 			p = &(*p)->rb_left;
2031 		else if (pool > perm->pool)
2032 			p = &(*p)->rb_right;
2033 		else {
2034 			int ret = ceph_compare_string(pool_ns,
2035 						perm->pool_ns,
2036 						perm->pool_ns_len);
2037 			if (ret < 0)
2038 				p = &(*p)->rb_left;
2039 			else if (ret > 0)
2040 				p = &(*p)->rb_right;
2041 			else {
2042 				have = perm->perm;
2043 				break;
2044 			}
2045 		}
2046 	}
2047 	if (*p) {
2048 		up_write(&mdsc->pool_perm_rwsem);
2049 		goto out;
2050 	}
2051 
2052 	rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2053 					 1, false, GFP_NOFS);
2054 	if (!rd_req) {
2055 		err = -ENOMEM;
2056 		goto out_unlock;
2057 	}
2058 
2059 	rd_req->r_flags = CEPH_OSD_FLAG_READ;
2060 	osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2061 	rd_req->r_base_oloc.pool = pool;
2062 	if (pool_ns)
2063 		rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2064 	ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2065 
2066 	err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2067 	if (err)
2068 		goto out_unlock;
2069 
2070 	wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2071 					 1, false, GFP_NOFS);
2072 	if (!wr_req) {
2073 		err = -ENOMEM;
2074 		goto out_unlock;
2075 	}
2076 
2077 	wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2078 	osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2079 	ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2080 	ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2081 
2082 	err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2083 	if (err)
2084 		goto out_unlock;
2085 
2086 	/* one page should be large enough for STAT data */
2087 	pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2088 	if (IS_ERR(pages)) {
2089 		err = PTR_ERR(pages);
2090 		goto out_unlock;
2091 	}
2092 
2093 	osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2094 				     0, false, true);
2095 	ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2096 
2097 	wr_req->r_mtime = ci->netfs.inode.i_mtime;
2098 	ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2099 
2100 	err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2101 	err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2102 
2103 	if (err >= 0 || err == -ENOENT)
2104 		have |= POOL_READ;
2105 	else if (err != -EPERM) {
2106 		if (err == -EBLOCKLISTED)
2107 			fsc->blocklisted = true;
2108 		goto out_unlock;
2109 	}
2110 
2111 	if (err2 == 0 || err2 == -EEXIST)
2112 		have |= POOL_WRITE;
2113 	else if (err2 != -EPERM) {
2114 		if (err2 == -EBLOCKLISTED)
2115 			fsc->blocklisted = true;
2116 		err = err2;
2117 		goto out_unlock;
2118 	}
2119 
2120 	pool_ns_len = pool_ns ? pool_ns->len : 0;
2121 	perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2122 	if (!perm) {
2123 		err = -ENOMEM;
2124 		goto out_unlock;
2125 	}
2126 
2127 	perm->pool = pool;
2128 	perm->perm = have;
2129 	perm->pool_ns_len = pool_ns_len;
2130 	if (pool_ns_len > 0)
2131 		memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2132 	perm->pool_ns[pool_ns_len] = 0;
2133 
2134 	rb_link_node(&perm->node, parent, p);
2135 	rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2136 	err = 0;
2137 out_unlock:
2138 	up_write(&mdsc->pool_perm_rwsem);
2139 
2140 	ceph_osdc_put_request(rd_req);
2141 	ceph_osdc_put_request(wr_req);
2142 out:
2143 	if (!err)
2144 		err = have;
2145 	if (pool_ns)
2146 		dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
2147 		     pool, (int)pool_ns->len, pool_ns->str, err);
2148 	else
2149 		dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
2150 	return err;
2151 }
2152 
2153 int ceph_pool_perm_check(struct inode *inode, int need)
2154 {
2155 	struct ceph_inode_info *ci = ceph_inode(inode);
2156 	struct ceph_string *pool_ns;
2157 	s64 pool;
2158 	int ret, flags;
2159 
2160 	/* Only need to do this for regular files */
2161 	if (!S_ISREG(inode->i_mode))
2162 		return 0;
2163 
2164 	if (ci->i_vino.snap != CEPH_NOSNAP) {
2165 		/*
2166 		 * Pool permission check needs to write to the first object.
2167 		 * But for snapshot, head of the first object may have alread
2168 		 * been deleted. Skip check to avoid creating orphan object.
2169 		 */
2170 		return 0;
2171 	}
2172 
2173 	if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2174 				NOPOOLPERM))
2175 		return 0;
2176 
2177 	spin_lock(&ci->i_ceph_lock);
2178 	flags = ci->i_ceph_flags;
2179 	pool = ci->i_layout.pool_id;
2180 	spin_unlock(&ci->i_ceph_lock);
2181 check:
2182 	if (flags & CEPH_I_POOL_PERM) {
2183 		if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2184 			dout("ceph_pool_perm_check pool %lld no read perm\n",
2185 			     pool);
2186 			return -EPERM;
2187 		}
2188 		if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2189 			dout("ceph_pool_perm_check pool %lld no write perm\n",
2190 			     pool);
2191 			return -EPERM;
2192 		}
2193 		return 0;
2194 	}
2195 
2196 	pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2197 	ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2198 	ceph_put_string(pool_ns);
2199 	if (ret < 0)
2200 		return ret;
2201 
2202 	flags = CEPH_I_POOL_PERM;
2203 	if (ret & POOL_READ)
2204 		flags |= CEPH_I_POOL_RD;
2205 	if (ret & POOL_WRITE)
2206 		flags |= CEPH_I_POOL_WR;
2207 
2208 	spin_lock(&ci->i_ceph_lock);
2209 	if (pool == ci->i_layout.pool_id &&
2210 	    pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2211 		ci->i_ceph_flags |= flags;
2212         } else {
2213 		pool = ci->i_layout.pool_id;
2214 		flags = ci->i_ceph_flags;
2215 	}
2216 	spin_unlock(&ci->i_ceph_lock);
2217 	goto check;
2218 }
2219 
2220 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2221 {
2222 	struct ceph_pool_perm *perm;
2223 	struct rb_node *n;
2224 
2225 	while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2226 		n = rb_first(&mdsc->pool_perm_tree);
2227 		perm = rb_entry(n, struct ceph_pool_perm, node);
2228 		rb_erase(n, &mdsc->pool_perm_tree);
2229 		kfree(perm);
2230 	}
2231 }
2232