xref: /linux/fs/ceph/addr.c (revision f6154d8babbb8a98f0d3ea325aafae2e33bfd8be)
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 = mapping->host;
83 	struct ceph_client *cl = ceph_inode_to_client(inode);
84 	struct ceph_inode_info *ci;
85 	struct ceph_snap_context *snapc;
86 
87 	if (folio_test_dirty(folio)) {
88 		doutc(cl, "%llx.%llx %p idx %lu -- already dirty\n",
89 		      ceph_vinop(inode), folio, folio->index);
90 		VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
91 		return false;
92 	}
93 
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 	doutc(cl, "%llx.%llx %p idx %lu head %d/%d -> %d/%d "
115 	      "snapc %p seq %lld (%d snaps)\n",
116 	      ceph_vinop(inode), 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 = folio->mapping->host;
141 	struct ceph_client *cl = ceph_inode_to_client(inode);
142 	struct ceph_inode_info *ci = ceph_inode(inode);
143 	struct ceph_snap_context *snapc;
144 
145 
146 	if (offset != 0 || length != folio_size(folio)) {
147 		doutc(cl, "%llx.%llx idx %lu partial dirty page %zu~%zu\n",
148 		      ceph_vinop(inode), folio->index, offset, length);
149 		return;
150 	}
151 
152 	WARN_ON(!folio_test_locked(folio));
153 	if (folio_test_private(folio)) {
154 		doutc(cl, "%llx.%llx idx %lu full dirty page\n",
155 		      ceph_vinop(inode), folio->index);
156 
157 		snapc = folio_detach_private(folio);
158 		ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
159 		ceph_put_snap_context(snapc);
160 	}
161 
162 	folio_wait_fscache(folio);
163 }
164 
165 static bool ceph_release_folio(struct folio *folio, gfp_t gfp)
166 {
167 	struct inode *inode = folio->mapping->host;
168 	struct ceph_client *cl = ceph_inode_to_client(inode);
169 
170 	doutc(cl, "%llx.%llx idx %lu (%sdirty)\n", ceph_vinop(inode),
171 	      folio->index, folio_test_dirty(folio) ? "" : "not ");
172 
173 	if (folio_test_private(folio))
174 		return false;
175 
176 	if (folio_test_fscache(folio)) {
177 		if (current_is_kswapd() || !(gfp & __GFP_FS))
178 			return false;
179 		folio_wait_fscache(folio);
180 	}
181 	ceph_fscache_note_page_release(inode);
182 	return true;
183 }
184 
185 static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
186 {
187 	struct inode *inode = rreq->inode;
188 	struct ceph_inode_info *ci = ceph_inode(inode);
189 	struct ceph_file_layout *lo = &ci->i_layout;
190 	unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
191 	loff_t end = rreq->start + rreq->len, new_end;
192 	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
193 	unsigned long max_len;
194 	u32 blockoff;
195 
196 	if (priv) {
197 		/* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
198 		if (priv->file_ra_disabled)
199 			max_pages = 0;
200 		else
201 			max_pages = priv->file_ra_pages;
202 
203 	}
204 
205 	/* Readahead is disabled */
206 	if (!max_pages)
207 		return;
208 
209 	max_len = max_pages << PAGE_SHIFT;
210 
211 	/*
212 	 * Try to expand the length forward by rounding up it to the next
213 	 * block, but do not exceed the file size, unless the original
214 	 * request already exceeds it.
215 	 */
216 	new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
217 	if (new_end > end && new_end <= rreq->start + max_len)
218 		rreq->len = new_end - rreq->start;
219 
220 	/* Try to expand the start downward */
221 	div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
222 	if (rreq->len + blockoff <= max_len) {
223 		rreq->start -= blockoff;
224 		rreq->len += blockoff;
225 	}
226 }
227 
228 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
229 {
230 	struct inode *inode = subreq->rreq->inode;
231 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
232 	struct ceph_inode_info *ci = ceph_inode(inode);
233 	u64 objno, objoff;
234 	u32 xlen;
235 
236 	/* Truncate the extent at the end of the current block */
237 	ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
238 				      &objno, &objoff, &xlen);
239 	subreq->len = min(xlen, fsc->mount_options->rsize);
240 	return true;
241 }
242 
243 static void finish_netfs_read(struct ceph_osd_request *req)
244 {
245 	struct inode *inode = req->r_inode;
246 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
247 	struct ceph_client *cl = fsc->client;
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 	doutc(cl, "result %d subreq->len=%zu i_size=%lld\n", 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_client *cl = fsc->client;
353 	struct ceph_osd_request *req = NULL;
354 	struct ceph_vino vino = ceph_vino(inode);
355 	struct iov_iter iter;
356 	int err = 0;
357 	u64 len = subreq->len;
358 	bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
359 	u64 off = subreq->start;
360 
361 	if (ceph_inode_is_shutdown(inode)) {
362 		err = -EIO;
363 		goto out;
364 	}
365 
366 	if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
367 		return;
368 
369 	ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
370 
371 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
372 			off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
373 			CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
374 			NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
375 	if (IS_ERR(req)) {
376 		err = PTR_ERR(req);
377 		req = NULL;
378 		goto out;
379 	}
380 
381 	if (sparse) {
382 		err = ceph_alloc_sparse_ext_map(&req->r_ops[0]);
383 		if (err)
384 			goto out;
385 	}
386 
387 	doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
388 	      ceph_vinop(inode), subreq->start, subreq->len, len);
389 
390 	iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
391 
392 	/*
393 	 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
394 	 * encrypted inodes. We'd need infrastructure that handles an iov_iter
395 	 * instead of page arrays, and we don't have that as of yet. Once the
396 	 * dust settles on the write helpers and encrypt/decrypt routines for
397 	 * netfs, we should be able to rework this.
398 	 */
399 	if (IS_ENCRYPTED(inode)) {
400 		struct page **pages;
401 		size_t page_off;
402 
403 		err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
404 		if (err < 0) {
405 			doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
406 			      ceph_vinop(inode), err);
407 			goto out;
408 		}
409 
410 		/* should always give us a page-aligned read */
411 		WARN_ON_ONCE(page_off);
412 		len = err;
413 		err = 0;
414 
415 		osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
416 						 false);
417 	} else {
418 		osd_req_op_extent_osd_iter(req, 0, &iter);
419 	}
420 	if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
421 		err = -EIO;
422 		goto out;
423 	}
424 	req->r_callback = finish_netfs_read;
425 	req->r_priv = subreq;
426 	req->r_inode = inode;
427 	ihold(inode);
428 
429 	ceph_osdc_start_request(req->r_osdc, req);
430 out:
431 	ceph_osdc_put_request(req);
432 	if (err)
433 		netfs_subreq_terminated(subreq, err, false);
434 	doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
435 }
436 
437 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
438 {
439 	struct inode *inode = rreq->inode;
440 	struct ceph_client *cl = ceph_inode_to_client(inode);
441 	int got = 0, want = CEPH_CAP_FILE_CACHE;
442 	struct ceph_netfs_request_data *priv;
443 	int ret = 0;
444 
445 	if (rreq->origin != NETFS_READAHEAD)
446 		return 0;
447 
448 	priv = kzalloc(sizeof(*priv), GFP_NOFS);
449 	if (!priv)
450 		return -ENOMEM;
451 
452 	if (file) {
453 		struct ceph_rw_context *rw_ctx;
454 		struct ceph_file_info *fi = file->private_data;
455 
456 		priv->file_ra_pages = file->f_ra.ra_pages;
457 		priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
458 
459 		rw_ctx = ceph_find_rw_context(fi);
460 		if (rw_ctx) {
461 			rreq->netfs_priv = priv;
462 			return 0;
463 		}
464 	}
465 
466 	/*
467 	 * readahead callers do not necessarily hold Fcb caps
468 	 * (e.g. fadvise, madvise).
469 	 */
470 	ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
471 	if (ret < 0) {
472 		doutc(cl, "%llx.%llx, error getting cap\n", ceph_vinop(inode));
473 		goto out;
474 	}
475 
476 	if (!(got & want)) {
477 		doutc(cl, "%llx.%llx, no cache cap\n", ceph_vinop(inode));
478 		ret = -EACCES;
479 		goto out;
480 	}
481 	if (ret == 0) {
482 		ret = -EACCES;
483 		goto out;
484 	}
485 
486 	priv->caps = got;
487 	rreq->netfs_priv = priv;
488 
489 out:
490 	if (ret < 0)
491 		kfree(priv);
492 
493 	return ret;
494 }
495 
496 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
497 {
498 	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
499 
500 	if (!priv)
501 		return;
502 
503 	if (priv->caps)
504 		ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
505 	kfree(priv);
506 	rreq->netfs_priv = NULL;
507 }
508 
509 const struct netfs_request_ops ceph_netfs_ops = {
510 	.init_request		= ceph_init_request,
511 	.free_request		= ceph_netfs_free_request,
512 	.begin_cache_operation	= ceph_begin_cache_operation,
513 	.issue_read		= ceph_netfs_issue_read,
514 	.expand_readahead	= ceph_netfs_expand_readahead,
515 	.clamp_length		= ceph_netfs_clamp_length,
516 	.check_write_begin	= ceph_netfs_check_write_begin,
517 };
518 
519 #ifdef CONFIG_CEPH_FSCACHE
520 static void ceph_set_page_fscache(struct page *page)
521 {
522 	set_page_fscache(page);
523 }
524 
525 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
526 {
527 	struct inode *inode = priv;
528 
529 	if (IS_ERR_VALUE(error) && error != -ENOBUFS)
530 		ceph_fscache_invalidate(inode, false);
531 }
532 
533 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
534 {
535 	struct ceph_inode_info *ci = ceph_inode(inode);
536 	struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
537 
538 	fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
539 			       ceph_fscache_write_terminated, inode, caching);
540 }
541 #else
542 static inline void ceph_set_page_fscache(struct page *page)
543 {
544 }
545 
546 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
547 {
548 }
549 #endif /* CONFIG_CEPH_FSCACHE */
550 
551 struct ceph_writeback_ctl
552 {
553 	loff_t i_size;
554 	u64 truncate_size;
555 	u32 truncate_seq;
556 	bool size_stable;
557 	bool head_snapc;
558 };
559 
560 /*
561  * Get ref for the oldest snapc for an inode with dirty data... that is, the
562  * only snap context we are allowed to write back.
563  */
564 static struct ceph_snap_context *
565 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
566 		   struct ceph_snap_context *page_snapc)
567 {
568 	struct ceph_inode_info *ci = ceph_inode(inode);
569 	struct ceph_client *cl = ceph_inode_to_client(inode);
570 	struct ceph_snap_context *snapc = NULL;
571 	struct ceph_cap_snap *capsnap = NULL;
572 
573 	spin_lock(&ci->i_ceph_lock);
574 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
575 		doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
576 		      capsnap, capsnap->context, capsnap->dirty_pages);
577 		if (!capsnap->dirty_pages)
578 			continue;
579 
580 		/* get i_size, truncate_{seq,size} for page_snapc? */
581 		if (snapc && capsnap->context != page_snapc)
582 			continue;
583 
584 		if (ctl) {
585 			if (capsnap->writing) {
586 				ctl->i_size = i_size_read(inode);
587 				ctl->size_stable = false;
588 			} else {
589 				ctl->i_size = capsnap->size;
590 				ctl->size_stable = true;
591 			}
592 			ctl->truncate_size = capsnap->truncate_size;
593 			ctl->truncate_seq = capsnap->truncate_seq;
594 			ctl->head_snapc = false;
595 		}
596 
597 		if (snapc)
598 			break;
599 
600 		snapc = ceph_get_snap_context(capsnap->context);
601 		if (!page_snapc ||
602 		    page_snapc == snapc ||
603 		    page_snapc->seq > snapc->seq)
604 			break;
605 	}
606 	if (!snapc && ci->i_wrbuffer_ref_head) {
607 		snapc = ceph_get_snap_context(ci->i_head_snapc);
608 		doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
609 		      ci->i_wrbuffer_ref_head);
610 		if (ctl) {
611 			ctl->i_size = i_size_read(inode);
612 			ctl->truncate_size = ci->i_truncate_size;
613 			ctl->truncate_seq = ci->i_truncate_seq;
614 			ctl->size_stable = false;
615 			ctl->head_snapc = true;
616 		}
617 	}
618 	spin_unlock(&ci->i_ceph_lock);
619 	return snapc;
620 }
621 
622 static u64 get_writepages_data_length(struct inode *inode,
623 				      struct page *page, u64 start)
624 {
625 	struct ceph_inode_info *ci = ceph_inode(inode);
626 	struct ceph_snap_context *snapc;
627 	struct ceph_cap_snap *capsnap = NULL;
628 	u64 end = i_size_read(inode);
629 	u64 ret;
630 
631 	snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
632 	if (snapc != ci->i_head_snapc) {
633 		bool found = false;
634 		spin_lock(&ci->i_ceph_lock);
635 		list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
636 			if (capsnap->context == snapc) {
637 				if (!capsnap->writing)
638 					end = capsnap->size;
639 				found = true;
640 				break;
641 			}
642 		}
643 		spin_unlock(&ci->i_ceph_lock);
644 		WARN_ON(!found);
645 	}
646 	if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
647 		end = ceph_fscrypt_page_offset(page) + thp_size(page);
648 	ret = end > start ? end - start : 0;
649 	if (ret && fscrypt_is_bounce_page(page))
650 		ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
651 	return ret;
652 }
653 
654 /*
655  * Write a single page, but leave the page locked.
656  *
657  * If we get a write error, mark the mapping for error, but still adjust the
658  * dirty page accounting (i.e., page is no longer dirty).
659  */
660 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
661 {
662 	struct folio *folio = page_folio(page);
663 	struct inode *inode = page->mapping->host;
664 	struct ceph_inode_info *ci = ceph_inode(inode);
665 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
666 	struct ceph_client *cl = fsc->client;
667 	struct ceph_snap_context *snapc, *oldest;
668 	loff_t page_off = page_offset(page);
669 	int err;
670 	loff_t len = thp_size(page);
671 	loff_t wlen;
672 	struct ceph_writeback_ctl ceph_wbc;
673 	struct ceph_osd_client *osdc = &fsc->client->osdc;
674 	struct ceph_osd_request *req;
675 	bool caching = ceph_is_cache_enabled(inode);
676 	struct page *bounce_page = NULL;
677 
678 	doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
679 	      page->index);
680 
681 	if (ceph_inode_is_shutdown(inode))
682 		return -EIO;
683 
684 	/* verify this is a writeable snap context */
685 	snapc = page_snap_context(page);
686 	if (!snapc) {
687 		doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
688 		      page);
689 		return 0;
690 	}
691 	oldest = get_oldest_context(inode, &ceph_wbc, snapc);
692 	if (snapc->seq > oldest->seq) {
693 		doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
694 		      ceph_vinop(inode), page, snapc);
695 		/* we should only noop if called by kswapd */
696 		WARN_ON(!(current->flags & PF_MEMALLOC));
697 		ceph_put_snap_context(oldest);
698 		redirty_page_for_writepage(wbc, page);
699 		return 0;
700 	}
701 	ceph_put_snap_context(oldest);
702 
703 	/* is this a partial page at end of file? */
704 	if (page_off >= ceph_wbc.i_size) {
705 		doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
706 		      ceph_vinop(inode), folio->index, ceph_wbc.i_size);
707 		folio_invalidate(folio, 0, folio_size(folio));
708 		return 0;
709 	}
710 
711 	if (ceph_wbc.i_size < page_off + len)
712 		len = ceph_wbc.i_size - page_off;
713 
714 	wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
715 	doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
716 	      ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
717 	      snapc->seq);
718 
719 	if (atomic_long_inc_return(&fsc->writeback_count) >
720 	    CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
721 		fsc->write_congested = true;
722 
723 	req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
724 				    page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
725 				    CEPH_OSD_FLAG_WRITE, snapc,
726 				    ceph_wbc.truncate_seq,
727 				    ceph_wbc.truncate_size, true);
728 	if (IS_ERR(req)) {
729 		redirty_page_for_writepage(wbc, page);
730 		return PTR_ERR(req);
731 	}
732 
733 	if (wlen < len)
734 		len = wlen;
735 
736 	set_page_writeback(page);
737 	if (caching)
738 		ceph_set_page_fscache(page);
739 	ceph_fscache_write_to_cache(inode, page_off, len, caching);
740 
741 	if (IS_ENCRYPTED(inode)) {
742 		bounce_page = fscrypt_encrypt_pagecache_blocks(page,
743 						    CEPH_FSCRYPT_BLOCK_SIZE, 0,
744 						    GFP_NOFS);
745 		if (IS_ERR(bounce_page)) {
746 			redirty_page_for_writepage(wbc, page);
747 			end_page_writeback(page);
748 			ceph_osdc_put_request(req);
749 			return PTR_ERR(bounce_page);
750 		}
751 	}
752 
753 	/* it may be a short write due to an object boundary */
754 	WARN_ON_ONCE(len > thp_size(page));
755 	osd_req_op_extent_osd_data_pages(req, 0,
756 			bounce_page ? &bounce_page : &page, wlen, 0,
757 			false, false);
758 	doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
759 	      ceph_vinop(inode), page_off, len, wlen,
760 	      IS_ENCRYPTED(inode) ? "" : "not ");
761 
762 	req->r_mtime = inode_get_mtime(inode);
763 	ceph_osdc_start_request(osdc, req);
764 	err = ceph_osdc_wait_request(osdc, req);
765 
766 	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
767 				  req->r_end_latency, len, err);
768 	fscrypt_free_bounce_page(bounce_page);
769 	ceph_osdc_put_request(req);
770 	if (err == 0)
771 		err = len;
772 
773 	if (err < 0) {
774 		struct writeback_control tmp_wbc;
775 		if (!wbc)
776 			wbc = &tmp_wbc;
777 		if (err == -ERESTARTSYS) {
778 			/* killed by SIGKILL */
779 			doutc(cl, "%llx.%llx interrupted page %p\n",
780 			      ceph_vinop(inode), page);
781 			redirty_page_for_writepage(wbc, page);
782 			end_page_writeback(page);
783 			return err;
784 		}
785 		if (err == -EBLOCKLISTED)
786 			fsc->blocklisted = true;
787 		doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
788 		      ceph_vinop(inode), err, page);
789 		mapping_set_error(&inode->i_data, err);
790 		wbc->pages_skipped++;
791 	} else {
792 		doutc(cl, "%llx.%llx cleaned page %p\n",
793 		      ceph_vinop(inode), page);
794 		err = 0;  /* vfs expects us to return 0 */
795 	}
796 	oldest = detach_page_private(page);
797 	WARN_ON_ONCE(oldest != snapc);
798 	end_page_writeback(page);
799 	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
800 	ceph_put_snap_context(snapc);  /* page's reference */
801 
802 	if (atomic_long_dec_return(&fsc->writeback_count) <
803 	    CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
804 		fsc->write_congested = false;
805 
806 	return err;
807 }
808 
809 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
810 {
811 	int err;
812 	struct inode *inode = page->mapping->host;
813 	BUG_ON(!inode);
814 	ihold(inode);
815 
816 	if (wbc->sync_mode == WB_SYNC_NONE &&
817 	    ceph_inode_to_fs_client(inode)->write_congested)
818 		return AOP_WRITEPAGE_ACTIVATE;
819 
820 	wait_on_page_fscache(page);
821 
822 	err = writepage_nounlock(page, wbc);
823 	if (err == -ERESTARTSYS) {
824 		/* direct memory reclaimer was killed by SIGKILL. return 0
825 		 * to prevent caller from setting mapping/page error */
826 		err = 0;
827 	}
828 	unlock_page(page);
829 	iput(inode);
830 	return err;
831 }
832 
833 /*
834  * async writeback completion handler.
835  *
836  * If we get an error, set the mapping error bit, but not the individual
837  * page error bits.
838  */
839 static void writepages_finish(struct ceph_osd_request *req)
840 {
841 	struct inode *inode = req->r_inode;
842 	struct ceph_inode_info *ci = ceph_inode(inode);
843 	struct ceph_client *cl = ceph_inode_to_client(inode);
844 	struct ceph_osd_data *osd_data;
845 	struct page *page;
846 	int num_pages, total_pages = 0;
847 	int i, j;
848 	int rc = req->r_result;
849 	struct ceph_snap_context *snapc = req->r_snapc;
850 	struct address_space *mapping = inode->i_mapping;
851 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
852 	unsigned int len = 0;
853 	bool remove_page;
854 
855 	doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
856 	if (rc < 0) {
857 		mapping_set_error(mapping, rc);
858 		ceph_set_error_write(ci);
859 		if (rc == -EBLOCKLISTED)
860 			fsc->blocklisted = true;
861 	} else {
862 		ceph_clear_error_write(ci);
863 	}
864 
865 	/*
866 	 * We lost the cache cap, need to truncate the page before
867 	 * it is unlocked, otherwise we'd truncate it later in the
868 	 * page truncation thread, possibly losing some data that
869 	 * raced its way in
870 	 */
871 	remove_page = !(ceph_caps_issued(ci) &
872 			(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
873 
874 	/* clean all pages */
875 	for (i = 0; i < req->r_num_ops; i++) {
876 		if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
877 			pr_warn_client(cl,
878 				"%llx.%llx incorrect op %d req %p index %d tid %llu\n",
879 				ceph_vinop(inode), req->r_ops[i].op, req, i,
880 				req->r_tid);
881 			break;
882 		}
883 
884 		osd_data = osd_req_op_extent_osd_data(req, i);
885 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
886 		len += osd_data->length;
887 		num_pages = calc_pages_for((u64)osd_data->alignment,
888 					   (u64)osd_data->length);
889 		total_pages += num_pages;
890 		for (j = 0; j < num_pages; j++) {
891 			page = osd_data->pages[j];
892 			if (fscrypt_is_bounce_page(page)) {
893 				page = fscrypt_pagecache_page(page);
894 				fscrypt_free_bounce_page(osd_data->pages[j]);
895 				osd_data->pages[j] = page;
896 			}
897 			BUG_ON(!page);
898 			WARN_ON(!PageUptodate(page));
899 
900 			if (atomic_long_dec_return(&fsc->writeback_count) <
901 			     CONGESTION_OFF_THRESH(
902 					fsc->mount_options->congestion_kb))
903 				fsc->write_congested = false;
904 
905 			ceph_put_snap_context(detach_page_private(page));
906 			end_page_writeback(page);
907 			doutc(cl, "unlocking %p\n", page);
908 
909 			if (remove_page)
910 				generic_error_remove_page(inode->i_mapping,
911 							  page);
912 
913 			unlock_page(page);
914 		}
915 		doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
916 		      ceph_vinop(inode), osd_data->length,
917 		      rc >= 0 ? num_pages : 0);
918 
919 		release_pages(osd_data->pages, num_pages);
920 	}
921 
922 	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
923 				  req->r_end_latency, len, rc);
924 
925 	ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
926 
927 	osd_data = osd_req_op_extent_osd_data(req, 0);
928 	if (osd_data->pages_from_pool)
929 		mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
930 	else
931 		kfree(osd_data->pages);
932 	ceph_osdc_put_request(req);
933 	ceph_dec_osd_stopping_blocker(fsc->mdsc);
934 }
935 
936 /*
937  * initiate async writeback
938  */
939 static int ceph_writepages_start(struct address_space *mapping,
940 				 struct writeback_control *wbc)
941 {
942 	struct inode *inode = mapping->host;
943 	struct ceph_inode_info *ci = ceph_inode(inode);
944 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
945 	struct ceph_client *cl = fsc->client;
946 	struct ceph_vino vino = ceph_vino(inode);
947 	pgoff_t index, start_index, end = -1;
948 	struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
949 	struct folio_batch fbatch;
950 	int rc = 0;
951 	unsigned int wsize = i_blocksize(inode);
952 	struct ceph_osd_request *req = NULL;
953 	struct ceph_writeback_ctl ceph_wbc;
954 	bool should_loop, range_whole = false;
955 	bool done = false;
956 	bool caching = ceph_is_cache_enabled(inode);
957 	xa_mark_t tag;
958 
959 	if (wbc->sync_mode == WB_SYNC_NONE &&
960 	    fsc->write_congested)
961 		return 0;
962 
963 	doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
964 	      wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
965 	      (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
966 
967 	if (ceph_inode_is_shutdown(inode)) {
968 		if (ci->i_wrbuffer_ref > 0) {
969 			pr_warn_ratelimited_client(cl,
970 				"%llx.%llx %lld forced umount\n",
971 				ceph_vinop(inode), ceph_ino(inode));
972 		}
973 		mapping_set_error(mapping, -EIO);
974 		return -EIO; /* we're in a forced umount, don't write! */
975 	}
976 	if (fsc->mount_options->wsize < wsize)
977 		wsize = fsc->mount_options->wsize;
978 
979 	folio_batch_init(&fbatch);
980 
981 	start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
982 	index = start_index;
983 
984 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
985 		tag = PAGECACHE_TAG_TOWRITE;
986 	} else {
987 		tag = PAGECACHE_TAG_DIRTY;
988 	}
989 retry:
990 	/* find oldest snap context with dirty data */
991 	snapc = get_oldest_context(inode, &ceph_wbc, NULL);
992 	if (!snapc) {
993 		/* hmm, why does writepages get called when there
994 		   is no dirty data? */
995 		doutc(cl, " no snap context with dirty data?\n");
996 		goto out;
997 	}
998 	doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
999 	      snapc->seq, snapc->num_snaps);
1000 
1001 	should_loop = false;
1002 	if (ceph_wbc.head_snapc && snapc != last_snapc) {
1003 		/* where to start/end? */
1004 		if (wbc->range_cyclic) {
1005 			index = start_index;
1006 			end = -1;
1007 			if (index > 0)
1008 				should_loop = true;
1009 			doutc(cl, " cyclic, start at %lu\n", index);
1010 		} else {
1011 			index = wbc->range_start >> PAGE_SHIFT;
1012 			end = wbc->range_end >> PAGE_SHIFT;
1013 			if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1014 				range_whole = true;
1015 			doutc(cl, " not cyclic, %lu to %lu\n", index, end);
1016 		}
1017 	} else if (!ceph_wbc.head_snapc) {
1018 		/* Do not respect wbc->range_{start,end}. Dirty pages
1019 		 * in that range can be associated with newer snapc.
1020 		 * They are not writeable until we write all dirty pages
1021 		 * associated with 'snapc' get written */
1022 		if (index > 0)
1023 			should_loop = true;
1024 		doutc(cl, " non-head snapc, range whole\n");
1025 	}
1026 
1027 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1028 		tag_pages_for_writeback(mapping, index, end);
1029 
1030 	ceph_put_snap_context(last_snapc);
1031 	last_snapc = snapc;
1032 
1033 	while (!done && index <= end) {
1034 		int num_ops = 0, op_idx;
1035 		unsigned i, nr_folios, max_pages, locked_pages = 0;
1036 		struct page **pages = NULL, **data_pages;
1037 		struct page *page;
1038 		pgoff_t strip_unit_end = 0;
1039 		u64 offset = 0, len = 0;
1040 		bool from_pool = false;
1041 
1042 		max_pages = wsize >> PAGE_SHIFT;
1043 
1044 get_more_pages:
1045 		nr_folios = filemap_get_folios_tag(mapping, &index,
1046 						   end, tag, &fbatch);
1047 		doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1048 		if (!nr_folios && !locked_pages)
1049 			break;
1050 		for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1051 			page = &fbatch.folios[i]->page;
1052 			doutc(cl, "? %p idx %lu\n", page, page->index);
1053 			if (locked_pages == 0)
1054 				lock_page(page);  /* first page */
1055 			else if (!trylock_page(page))
1056 				break;
1057 
1058 			/* only dirty pages, or our accounting breaks */
1059 			if (unlikely(!PageDirty(page)) ||
1060 			    unlikely(page->mapping != mapping)) {
1061 				doutc(cl, "!dirty or !mapping %p\n", page);
1062 				unlock_page(page);
1063 				continue;
1064 			}
1065 			/* only if matching snap context */
1066 			pgsnapc = page_snap_context(page);
1067 			if (pgsnapc != snapc) {
1068 				doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1069 				      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1070 				if (!should_loop &&
1071 				    !ceph_wbc.head_snapc &&
1072 				    wbc->sync_mode != WB_SYNC_NONE)
1073 					should_loop = true;
1074 				unlock_page(page);
1075 				continue;
1076 			}
1077 			if (page_offset(page) >= ceph_wbc.i_size) {
1078 				struct folio *folio = page_folio(page);
1079 
1080 				doutc(cl, "folio at %lu beyond eof %llu\n",
1081 				      folio->index, ceph_wbc.i_size);
1082 				if ((ceph_wbc.size_stable ||
1083 				    folio_pos(folio) >= i_size_read(inode)) &&
1084 				    folio_clear_dirty_for_io(folio))
1085 					folio_invalidate(folio, 0,
1086 							folio_size(folio));
1087 				folio_unlock(folio);
1088 				continue;
1089 			}
1090 			if (strip_unit_end && (page->index > strip_unit_end)) {
1091 				doutc(cl, "end of strip unit %p\n", page);
1092 				unlock_page(page);
1093 				break;
1094 			}
1095 			if (PageWriteback(page) || PageFsCache(page)) {
1096 				if (wbc->sync_mode == WB_SYNC_NONE) {
1097 					doutc(cl, "%p under writeback\n", page);
1098 					unlock_page(page);
1099 					continue;
1100 				}
1101 				doutc(cl, "waiting on writeback %p\n", page);
1102 				wait_on_page_writeback(page);
1103 				wait_on_page_fscache(page);
1104 			}
1105 
1106 			if (!clear_page_dirty_for_io(page)) {
1107 				doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1108 				unlock_page(page);
1109 				continue;
1110 			}
1111 
1112 			/*
1113 			 * We have something to write.  If this is
1114 			 * the first locked page this time through,
1115 			 * calculate max possinle write size and
1116 			 * allocate a page array
1117 			 */
1118 			if (locked_pages == 0) {
1119 				u64 objnum;
1120 				u64 objoff;
1121 				u32 xlen;
1122 
1123 				/* prepare async write request */
1124 				offset = (u64)page_offset(page);
1125 				ceph_calc_file_object_mapping(&ci->i_layout,
1126 							      offset, wsize,
1127 							      &objnum, &objoff,
1128 							      &xlen);
1129 				len = xlen;
1130 
1131 				num_ops = 1;
1132 				strip_unit_end = page->index +
1133 					((len - 1) >> PAGE_SHIFT);
1134 
1135 				BUG_ON(pages);
1136 				max_pages = calc_pages_for(0, (u64)len);
1137 				pages = kmalloc_array(max_pages,
1138 						      sizeof(*pages),
1139 						      GFP_NOFS);
1140 				if (!pages) {
1141 					from_pool = true;
1142 					pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1143 					BUG_ON(!pages);
1144 				}
1145 
1146 				len = 0;
1147 			} else if (page->index !=
1148 				   (offset + len) >> PAGE_SHIFT) {
1149 				if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1150 							     CEPH_OSD_MAX_OPS)) {
1151 					redirty_page_for_writepage(wbc, page);
1152 					unlock_page(page);
1153 					break;
1154 				}
1155 
1156 				num_ops++;
1157 				offset = (u64)page_offset(page);
1158 				len = 0;
1159 			}
1160 
1161 			/* note position of first page in fbatch */
1162 			doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1163 			      ceph_vinop(inode), page, page->index);
1164 
1165 			if (atomic_long_inc_return(&fsc->writeback_count) >
1166 			    CONGESTION_ON_THRESH(
1167 				    fsc->mount_options->congestion_kb))
1168 				fsc->write_congested = true;
1169 
1170 			if (IS_ENCRYPTED(inode)) {
1171 				pages[locked_pages] =
1172 					fscrypt_encrypt_pagecache_blocks(page,
1173 						PAGE_SIZE, 0,
1174 						locked_pages ? GFP_NOWAIT : GFP_NOFS);
1175 				if (IS_ERR(pages[locked_pages])) {
1176 					if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1177 						pr_err_client(cl,
1178 							"inode->i_blkbits=%hhu\n",
1179 							inode->i_blkbits);
1180 					/* better not fail on first page! */
1181 					BUG_ON(locked_pages == 0);
1182 					pages[locked_pages] = NULL;
1183 					redirty_page_for_writepage(wbc, page);
1184 					unlock_page(page);
1185 					break;
1186 				}
1187 				++locked_pages;
1188 			} else {
1189 				pages[locked_pages++] = page;
1190 			}
1191 
1192 			fbatch.folios[i] = NULL;
1193 			len += thp_size(page);
1194 		}
1195 
1196 		/* did we get anything? */
1197 		if (!locked_pages)
1198 			goto release_folios;
1199 		if (i) {
1200 			unsigned j, n = 0;
1201 			/* shift unused page to beginning of fbatch */
1202 			for (j = 0; j < nr_folios; j++) {
1203 				if (!fbatch.folios[j])
1204 					continue;
1205 				if (n < j)
1206 					fbatch.folios[n] = fbatch.folios[j];
1207 				n++;
1208 			}
1209 			fbatch.nr = n;
1210 
1211 			if (nr_folios && i == nr_folios &&
1212 			    locked_pages < max_pages) {
1213 				doutc(cl, "reached end fbatch, trying for more\n");
1214 				folio_batch_release(&fbatch);
1215 				goto get_more_pages;
1216 			}
1217 		}
1218 
1219 new_request:
1220 		offset = ceph_fscrypt_page_offset(pages[0]);
1221 		len = wsize;
1222 
1223 		req = ceph_osdc_new_request(&fsc->client->osdc,
1224 					&ci->i_layout, vino,
1225 					offset, &len, 0, num_ops,
1226 					CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1227 					snapc, ceph_wbc.truncate_seq,
1228 					ceph_wbc.truncate_size, false);
1229 		if (IS_ERR(req)) {
1230 			req = ceph_osdc_new_request(&fsc->client->osdc,
1231 						&ci->i_layout, vino,
1232 						offset, &len, 0,
1233 						min(num_ops,
1234 						    CEPH_OSD_SLAB_OPS),
1235 						CEPH_OSD_OP_WRITE,
1236 						CEPH_OSD_FLAG_WRITE,
1237 						snapc, ceph_wbc.truncate_seq,
1238 						ceph_wbc.truncate_size, true);
1239 			BUG_ON(IS_ERR(req));
1240 		}
1241 		BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1242 			     thp_size(pages[locked_pages - 1]) - offset);
1243 
1244 		if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1245 			rc = -EIO;
1246 			goto release_folios;
1247 		}
1248 		req->r_callback = writepages_finish;
1249 		req->r_inode = inode;
1250 
1251 		/* Format the osd request message and submit the write */
1252 		len = 0;
1253 		data_pages = pages;
1254 		op_idx = 0;
1255 		for (i = 0; i < locked_pages; i++) {
1256 			struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1257 
1258 			u64 cur_offset = page_offset(page);
1259 			/*
1260 			 * Discontinuity in page range? Ceph can handle that by just passing
1261 			 * multiple extents in the write op.
1262 			 */
1263 			if (offset + len != cur_offset) {
1264 				/* If it's full, stop here */
1265 				if (op_idx + 1 == req->r_num_ops)
1266 					break;
1267 
1268 				/* Kick off an fscache write with what we have so far. */
1269 				ceph_fscache_write_to_cache(inode, offset, len, caching);
1270 
1271 				/* Start a new extent */
1272 				osd_req_op_extent_dup_last(req, op_idx,
1273 							   cur_offset - offset);
1274 				doutc(cl, "got pages at %llu~%llu\n", offset,
1275 				      len);
1276 				osd_req_op_extent_osd_data_pages(req, op_idx,
1277 							data_pages, len, 0,
1278 							from_pool, false);
1279 				osd_req_op_extent_update(req, op_idx, len);
1280 
1281 				len = 0;
1282 				offset = cur_offset;
1283 				data_pages = pages + i;
1284 				op_idx++;
1285 			}
1286 
1287 			set_page_writeback(page);
1288 			if (caching)
1289 				ceph_set_page_fscache(page);
1290 			len += thp_size(page);
1291 		}
1292 		ceph_fscache_write_to_cache(inode, offset, len, caching);
1293 
1294 		if (ceph_wbc.size_stable) {
1295 			len = min(len, ceph_wbc.i_size - offset);
1296 		} else if (i == locked_pages) {
1297 			/* writepages_finish() clears writeback pages
1298 			 * according to the data length, so make sure
1299 			 * data length covers all locked pages */
1300 			u64 min_len = len + 1 - thp_size(page);
1301 			len = get_writepages_data_length(inode, pages[i - 1],
1302 							 offset);
1303 			len = max(len, min_len);
1304 		}
1305 		if (IS_ENCRYPTED(inode))
1306 			len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1307 
1308 		doutc(cl, "got pages at %llu~%llu\n", offset, len);
1309 
1310 		if (IS_ENCRYPTED(inode) &&
1311 		    ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1312 			pr_warn_client(cl,
1313 				"bad encrypted write offset=%lld len=%llu\n",
1314 				offset, len);
1315 
1316 		osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1317 						 0, from_pool, false);
1318 		osd_req_op_extent_update(req, op_idx, len);
1319 
1320 		BUG_ON(op_idx + 1 != req->r_num_ops);
1321 
1322 		from_pool = false;
1323 		if (i < locked_pages) {
1324 			BUG_ON(num_ops <= req->r_num_ops);
1325 			num_ops -= req->r_num_ops;
1326 			locked_pages -= i;
1327 
1328 			/* allocate new pages array for next request */
1329 			data_pages = pages;
1330 			pages = kmalloc_array(locked_pages, sizeof(*pages),
1331 					      GFP_NOFS);
1332 			if (!pages) {
1333 				from_pool = true;
1334 				pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1335 				BUG_ON(!pages);
1336 			}
1337 			memcpy(pages, data_pages + i,
1338 			       locked_pages * sizeof(*pages));
1339 			memset(data_pages + i, 0,
1340 			       locked_pages * sizeof(*pages));
1341 		} else {
1342 			BUG_ON(num_ops != req->r_num_ops);
1343 			index = pages[i - 1]->index + 1;
1344 			/* request message now owns the pages array */
1345 			pages = NULL;
1346 		}
1347 
1348 		req->r_mtime = inode_get_mtime(inode);
1349 		ceph_osdc_start_request(&fsc->client->osdc, req);
1350 		req = NULL;
1351 
1352 		wbc->nr_to_write -= i;
1353 		if (pages)
1354 			goto new_request;
1355 
1356 		/*
1357 		 * We stop writing back only if we are not doing
1358 		 * integrity sync. In case of integrity sync we have to
1359 		 * keep going until we have written all the pages
1360 		 * we tagged for writeback prior to entering this loop.
1361 		 */
1362 		if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1363 			done = true;
1364 
1365 release_folios:
1366 		doutc(cl, "folio_batch release on %d folios (%p)\n",
1367 		      (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1368 		folio_batch_release(&fbatch);
1369 	}
1370 
1371 	if (should_loop && !done) {
1372 		/* more to do; loop back to beginning of file */
1373 		doutc(cl, "looping back to beginning of file\n");
1374 		end = start_index - 1; /* OK even when start_index == 0 */
1375 
1376 		/* to write dirty pages associated with next snapc,
1377 		 * we need to wait until current writes complete */
1378 		if (wbc->sync_mode != WB_SYNC_NONE &&
1379 		    start_index == 0 && /* all dirty pages were checked */
1380 		    !ceph_wbc.head_snapc) {
1381 			struct page *page;
1382 			unsigned i, nr;
1383 			index = 0;
1384 			while ((index <= end) &&
1385 			       (nr = filemap_get_folios_tag(mapping, &index,
1386 						(pgoff_t)-1,
1387 						PAGECACHE_TAG_WRITEBACK,
1388 						&fbatch))) {
1389 				for (i = 0; i < nr; i++) {
1390 					page = &fbatch.folios[i]->page;
1391 					if (page_snap_context(page) != snapc)
1392 						continue;
1393 					wait_on_page_writeback(page);
1394 				}
1395 				folio_batch_release(&fbatch);
1396 				cond_resched();
1397 			}
1398 		}
1399 
1400 		start_index = 0;
1401 		index = 0;
1402 		goto retry;
1403 	}
1404 
1405 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1406 		mapping->writeback_index = index;
1407 
1408 out:
1409 	ceph_osdc_put_request(req);
1410 	ceph_put_snap_context(last_snapc);
1411 	doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1412 	      rc);
1413 	return rc;
1414 }
1415 
1416 
1417 
1418 /*
1419  * See if a given @snapc is either writeable, or already written.
1420  */
1421 static int context_is_writeable_or_written(struct inode *inode,
1422 					   struct ceph_snap_context *snapc)
1423 {
1424 	struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1425 	int ret = !oldest || snapc->seq <= oldest->seq;
1426 
1427 	ceph_put_snap_context(oldest);
1428 	return ret;
1429 }
1430 
1431 /**
1432  * ceph_find_incompatible - find an incompatible context and return it
1433  * @page: page being dirtied
1434  *
1435  * We are only allowed to write into/dirty a page if the page is
1436  * clean, or already dirty within the same snap context. Returns a
1437  * conflicting context if there is one, NULL if there isn't, or a
1438  * negative error code on other errors.
1439  *
1440  * Must be called with page lock held.
1441  */
1442 static struct ceph_snap_context *
1443 ceph_find_incompatible(struct page *page)
1444 {
1445 	struct inode *inode = page->mapping->host;
1446 	struct ceph_client *cl = ceph_inode_to_client(inode);
1447 	struct ceph_inode_info *ci = ceph_inode(inode);
1448 
1449 	if (ceph_inode_is_shutdown(inode)) {
1450 		doutc(cl, " %llx.%llx page %p is shutdown\n",
1451 		      ceph_vinop(inode), page);
1452 		return ERR_PTR(-ESTALE);
1453 	}
1454 
1455 	for (;;) {
1456 		struct ceph_snap_context *snapc, *oldest;
1457 
1458 		wait_on_page_writeback(page);
1459 
1460 		snapc = page_snap_context(page);
1461 		if (!snapc || snapc == ci->i_head_snapc)
1462 			break;
1463 
1464 		/*
1465 		 * this page is already dirty in another (older) snap
1466 		 * context!  is it writeable now?
1467 		 */
1468 		oldest = get_oldest_context(inode, NULL, NULL);
1469 		if (snapc->seq > oldest->seq) {
1470 			/* not writeable -- return it for the caller to deal with */
1471 			ceph_put_snap_context(oldest);
1472 			doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1473 			      ceph_vinop(inode), page, snapc);
1474 			return ceph_get_snap_context(snapc);
1475 		}
1476 		ceph_put_snap_context(oldest);
1477 
1478 		/* yay, writeable, do it now (without dropping page lock) */
1479 		doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1480 		      ceph_vinop(inode), page, snapc);
1481 		if (clear_page_dirty_for_io(page)) {
1482 			int r = writepage_nounlock(page, NULL);
1483 			if (r < 0)
1484 				return ERR_PTR(r);
1485 		}
1486 	}
1487 	return NULL;
1488 }
1489 
1490 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1491 					struct folio **foliop, void **_fsdata)
1492 {
1493 	struct inode *inode = file_inode(file);
1494 	struct ceph_inode_info *ci = ceph_inode(inode);
1495 	struct ceph_snap_context *snapc;
1496 
1497 	snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1498 	if (snapc) {
1499 		int r;
1500 
1501 		folio_unlock(*foliop);
1502 		folio_put(*foliop);
1503 		*foliop = NULL;
1504 		if (IS_ERR(snapc))
1505 			return PTR_ERR(snapc);
1506 
1507 		ceph_queue_writeback(inode);
1508 		r = wait_event_killable(ci->i_cap_wq,
1509 					context_is_writeable_or_written(inode, snapc));
1510 		ceph_put_snap_context(snapc);
1511 		return r == 0 ? -EAGAIN : r;
1512 	}
1513 	return 0;
1514 }
1515 
1516 /*
1517  * We are only allowed to write into/dirty the page if the page is
1518  * clean, or already dirty within the same snap context.
1519  */
1520 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1521 			    loff_t pos, unsigned len,
1522 			    struct page **pagep, void **fsdata)
1523 {
1524 	struct inode *inode = file_inode(file);
1525 	struct ceph_inode_info *ci = ceph_inode(inode);
1526 	struct folio *folio = NULL;
1527 	int r;
1528 
1529 	r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1530 	if (r < 0)
1531 		return r;
1532 
1533 	folio_wait_fscache(folio);
1534 	WARN_ON_ONCE(!folio_test_locked(folio));
1535 	*pagep = &folio->page;
1536 	return 0;
1537 }
1538 
1539 /*
1540  * we don't do anything in here that simple_write_end doesn't do
1541  * except adjust dirty page accounting
1542  */
1543 static int ceph_write_end(struct file *file, struct address_space *mapping,
1544 			  loff_t pos, unsigned len, unsigned copied,
1545 			  struct page *subpage, void *fsdata)
1546 {
1547 	struct folio *folio = page_folio(subpage);
1548 	struct inode *inode = file_inode(file);
1549 	struct ceph_client *cl = ceph_inode_to_client(inode);
1550 	bool check_cap = false;
1551 
1552 	doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1553 	      file, folio, (int)pos, (int)copied, (int)len);
1554 
1555 	if (!folio_test_uptodate(folio)) {
1556 		/* just return that nothing was copied on a short copy */
1557 		if (copied < len) {
1558 			copied = 0;
1559 			goto out;
1560 		}
1561 		folio_mark_uptodate(folio);
1562 	}
1563 
1564 	/* did file size increase? */
1565 	if (pos+copied > i_size_read(inode))
1566 		check_cap = ceph_inode_set_size(inode, pos+copied);
1567 
1568 	folio_mark_dirty(folio);
1569 
1570 out:
1571 	folio_unlock(folio);
1572 	folio_put(folio);
1573 
1574 	if (check_cap)
1575 		ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1576 
1577 	return copied;
1578 }
1579 
1580 const struct address_space_operations ceph_aops = {
1581 	.read_folio = netfs_read_folio,
1582 	.readahead = netfs_readahead,
1583 	.writepage = ceph_writepage,
1584 	.writepages = ceph_writepages_start,
1585 	.write_begin = ceph_write_begin,
1586 	.write_end = ceph_write_end,
1587 	.dirty_folio = ceph_dirty_folio,
1588 	.invalidate_folio = ceph_invalidate_folio,
1589 	.release_folio = ceph_release_folio,
1590 	.direct_IO = noop_direct_IO,
1591 };
1592 
1593 static void ceph_block_sigs(sigset_t *oldset)
1594 {
1595 	sigset_t mask;
1596 	siginitsetinv(&mask, sigmask(SIGKILL));
1597 	sigprocmask(SIG_BLOCK, &mask, oldset);
1598 }
1599 
1600 static void ceph_restore_sigs(sigset_t *oldset)
1601 {
1602 	sigprocmask(SIG_SETMASK, oldset, NULL);
1603 }
1604 
1605 /*
1606  * vm ops
1607  */
1608 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1609 {
1610 	struct vm_area_struct *vma = vmf->vma;
1611 	struct inode *inode = file_inode(vma->vm_file);
1612 	struct ceph_inode_info *ci = ceph_inode(inode);
1613 	struct ceph_client *cl = ceph_inode_to_client(inode);
1614 	struct ceph_file_info *fi = vma->vm_file->private_data;
1615 	loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1616 	int want, got, err;
1617 	sigset_t oldset;
1618 	vm_fault_t ret = VM_FAULT_SIGBUS;
1619 
1620 	if (ceph_inode_is_shutdown(inode))
1621 		return ret;
1622 
1623 	ceph_block_sigs(&oldset);
1624 
1625 	doutc(cl, "%llx.%llx %llu trying to get caps\n",
1626 	      ceph_vinop(inode), off);
1627 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1628 		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1629 	else
1630 		want = CEPH_CAP_FILE_CACHE;
1631 
1632 	got = 0;
1633 	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1634 	if (err < 0)
1635 		goto out_restore;
1636 
1637 	doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1638 	      off, ceph_cap_string(got));
1639 
1640 	if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1641 	    !ceph_has_inline_data(ci)) {
1642 		CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1643 		ceph_add_rw_context(fi, &rw_ctx);
1644 		ret = filemap_fault(vmf);
1645 		ceph_del_rw_context(fi, &rw_ctx);
1646 		doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1647 		      ceph_vinop(inode), off, ceph_cap_string(got), ret);
1648 	} else
1649 		err = -EAGAIN;
1650 
1651 	ceph_put_cap_refs(ci, got);
1652 
1653 	if (err != -EAGAIN)
1654 		goto out_restore;
1655 
1656 	/* read inline data */
1657 	if (off >= PAGE_SIZE) {
1658 		/* does not support inline data > PAGE_SIZE */
1659 		ret = VM_FAULT_SIGBUS;
1660 	} else {
1661 		struct address_space *mapping = inode->i_mapping;
1662 		struct page *page;
1663 
1664 		filemap_invalidate_lock_shared(mapping);
1665 		page = find_or_create_page(mapping, 0,
1666 				mapping_gfp_constraint(mapping, ~__GFP_FS));
1667 		if (!page) {
1668 			ret = VM_FAULT_OOM;
1669 			goto out_inline;
1670 		}
1671 		err = __ceph_do_getattr(inode, page,
1672 					 CEPH_STAT_CAP_INLINE_DATA, true);
1673 		if (err < 0 || off >= i_size_read(inode)) {
1674 			unlock_page(page);
1675 			put_page(page);
1676 			ret = vmf_error(err);
1677 			goto out_inline;
1678 		}
1679 		if (err < PAGE_SIZE)
1680 			zero_user_segment(page, err, PAGE_SIZE);
1681 		else
1682 			flush_dcache_page(page);
1683 		SetPageUptodate(page);
1684 		vmf->page = page;
1685 		ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1686 out_inline:
1687 		filemap_invalidate_unlock_shared(mapping);
1688 		doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1689 		      ceph_vinop(inode), off, ret);
1690 	}
1691 out_restore:
1692 	ceph_restore_sigs(&oldset);
1693 	if (err < 0)
1694 		ret = vmf_error(err);
1695 
1696 	return ret;
1697 }
1698 
1699 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1700 {
1701 	struct vm_area_struct *vma = vmf->vma;
1702 	struct inode *inode = file_inode(vma->vm_file);
1703 	struct ceph_client *cl = ceph_inode_to_client(inode);
1704 	struct ceph_inode_info *ci = ceph_inode(inode);
1705 	struct ceph_file_info *fi = vma->vm_file->private_data;
1706 	struct ceph_cap_flush *prealloc_cf;
1707 	struct page *page = vmf->page;
1708 	loff_t off = page_offset(page);
1709 	loff_t size = i_size_read(inode);
1710 	size_t len;
1711 	int want, got, err;
1712 	sigset_t oldset;
1713 	vm_fault_t ret = VM_FAULT_SIGBUS;
1714 
1715 	if (ceph_inode_is_shutdown(inode))
1716 		return ret;
1717 
1718 	prealloc_cf = ceph_alloc_cap_flush();
1719 	if (!prealloc_cf)
1720 		return VM_FAULT_OOM;
1721 
1722 	sb_start_pagefault(inode->i_sb);
1723 	ceph_block_sigs(&oldset);
1724 
1725 	if (off + thp_size(page) <= size)
1726 		len = thp_size(page);
1727 	else
1728 		len = offset_in_thp(page, size);
1729 
1730 	doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1731 	      ceph_vinop(inode), off, len, size);
1732 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1733 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1734 	else
1735 		want = CEPH_CAP_FILE_BUFFER;
1736 
1737 	got = 0;
1738 	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1739 	if (err < 0)
1740 		goto out_free;
1741 
1742 	doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1743 	      off, len, ceph_cap_string(got));
1744 
1745 	/* Update time before taking page lock */
1746 	file_update_time(vma->vm_file);
1747 	inode_inc_iversion_raw(inode);
1748 
1749 	do {
1750 		struct ceph_snap_context *snapc;
1751 
1752 		lock_page(page);
1753 
1754 		if (page_mkwrite_check_truncate(page, inode) < 0) {
1755 			unlock_page(page);
1756 			ret = VM_FAULT_NOPAGE;
1757 			break;
1758 		}
1759 
1760 		snapc = ceph_find_incompatible(page);
1761 		if (!snapc) {
1762 			/* success.  we'll keep the page locked. */
1763 			set_page_dirty(page);
1764 			ret = VM_FAULT_LOCKED;
1765 			break;
1766 		}
1767 
1768 		unlock_page(page);
1769 
1770 		if (IS_ERR(snapc)) {
1771 			ret = VM_FAULT_SIGBUS;
1772 			break;
1773 		}
1774 
1775 		ceph_queue_writeback(inode);
1776 		err = wait_event_killable(ci->i_cap_wq,
1777 				context_is_writeable_or_written(inode, snapc));
1778 		ceph_put_snap_context(snapc);
1779 	} while (err == 0);
1780 
1781 	if (ret == VM_FAULT_LOCKED) {
1782 		int dirty;
1783 		spin_lock(&ci->i_ceph_lock);
1784 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1785 					       &prealloc_cf);
1786 		spin_unlock(&ci->i_ceph_lock);
1787 		if (dirty)
1788 			__mark_inode_dirty(inode, dirty);
1789 	}
1790 
1791 	doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1792 	      ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1793 	ceph_put_cap_refs_async(ci, got);
1794 out_free:
1795 	ceph_restore_sigs(&oldset);
1796 	sb_end_pagefault(inode->i_sb);
1797 	ceph_free_cap_flush(prealloc_cf);
1798 	if (err < 0)
1799 		ret = vmf_error(err);
1800 	return ret;
1801 }
1802 
1803 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1804 			   char	*data, size_t len)
1805 {
1806 	struct ceph_client *cl = ceph_inode_to_client(inode);
1807 	struct address_space *mapping = inode->i_mapping;
1808 	struct page *page;
1809 
1810 	if (locked_page) {
1811 		page = locked_page;
1812 	} else {
1813 		if (i_size_read(inode) == 0)
1814 			return;
1815 		page = find_or_create_page(mapping, 0,
1816 					   mapping_gfp_constraint(mapping,
1817 					   ~__GFP_FS));
1818 		if (!page)
1819 			return;
1820 		if (PageUptodate(page)) {
1821 			unlock_page(page);
1822 			put_page(page);
1823 			return;
1824 		}
1825 	}
1826 
1827 	doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1828 	      ceph_vinop(inode), len, locked_page);
1829 
1830 	if (len > 0) {
1831 		void *kaddr = kmap_atomic(page);
1832 		memcpy(kaddr, data, len);
1833 		kunmap_atomic(kaddr);
1834 	}
1835 
1836 	if (page != locked_page) {
1837 		if (len < PAGE_SIZE)
1838 			zero_user_segment(page, len, PAGE_SIZE);
1839 		else
1840 			flush_dcache_page(page);
1841 
1842 		SetPageUptodate(page);
1843 		unlock_page(page);
1844 		put_page(page);
1845 	}
1846 }
1847 
1848 int ceph_uninline_data(struct file *file)
1849 {
1850 	struct inode *inode = file_inode(file);
1851 	struct ceph_inode_info *ci = ceph_inode(inode);
1852 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1853 	struct ceph_client *cl = fsc->client;
1854 	struct ceph_osd_request *req = NULL;
1855 	struct ceph_cap_flush *prealloc_cf = NULL;
1856 	struct folio *folio = NULL;
1857 	u64 inline_version = CEPH_INLINE_NONE;
1858 	struct page *pages[1];
1859 	int err = 0;
1860 	u64 len;
1861 
1862 	spin_lock(&ci->i_ceph_lock);
1863 	inline_version = ci->i_inline_version;
1864 	spin_unlock(&ci->i_ceph_lock);
1865 
1866 	doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1867 	      inline_version);
1868 
1869 	if (ceph_inode_is_shutdown(inode)) {
1870 		err = -EIO;
1871 		goto out;
1872 	}
1873 
1874 	if (inline_version == CEPH_INLINE_NONE)
1875 		return 0;
1876 
1877 	prealloc_cf = ceph_alloc_cap_flush();
1878 	if (!prealloc_cf)
1879 		return -ENOMEM;
1880 
1881 	if (inline_version == 1) /* initial version, no data */
1882 		goto out_uninline;
1883 
1884 	folio = read_mapping_folio(inode->i_mapping, 0, file);
1885 	if (IS_ERR(folio)) {
1886 		err = PTR_ERR(folio);
1887 		goto out;
1888 	}
1889 
1890 	folio_lock(folio);
1891 
1892 	len = i_size_read(inode);
1893 	if (len > folio_size(folio))
1894 		len = folio_size(folio);
1895 
1896 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1897 				    ceph_vino(inode), 0, &len, 0, 1,
1898 				    CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1899 				    NULL, 0, 0, false);
1900 	if (IS_ERR(req)) {
1901 		err = PTR_ERR(req);
1902 		goto out_unlock;
1903 	}
1904 
1905 	req->r_mtime = inode_get_mtime(inode);
1906 	ceph_osdc_start_request(&fsc->client->osdc, req);
1907 	err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1908 	ceph_osdc_put_request(req);
1909 	if (err < 0)
1910 		goto out_unlock;
1911 
1912 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1913 				    ceph_vino(inode), 0, &len, 1, 3,
1914 				    CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1915 				    NULL, ci->i_truncate_seq,
1916 				    ci->i_truncate_size, false);
1917 	if (IS_ERR(req)) {
1918 		err = PTR_ERR(req);
1919 		goto out_unlock;
1920 	}
1921 
1922 	pages[0] = folio_page(folio, 0);
1923 	osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1924 
1925 	{
1926 		__le64 xattr_buf = cpu_to_le64(inline_version);
1927 		err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1928 					    "inline_version", &xattr_buf,
1929 					    sizeof(xattr_buf),
1930 					    CEPH_OSD_CMPXATTR_OP_GT,
1931 					    CEPH_OSD_CMPXATTR_MODE_U64);
1932 		if (err)
1933 			goto out_put_req;
1934 	}
1935 
1936 	{
1937 		char xattr_buf[32];
1938 		int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1939 					 "%llu", inline_version);
1940 		err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1941 					    "inline_version",
1942 					    xattr_buf, xattr_len, 0, 0);
1943 		if (err)
1944 			goto out_put_req;
1945 	}
1946 
1947 	req->r_mtime = inode_get_mtime(inode);
1948 	ceph_osdc_start_request(&fsc->client->osdc, req);
1949 	err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1950 
1951 	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1952 				  req->r_end_latency, len, err);
1953 
1954 out_uninline:
1955 	if (!err) {
1956 		int dirty;
1957 
1958 		/* Set to CAP_INLINE_NONE and dirty the caps */
1959 		down_read(&fsc->mdsc->snap_rwsem);
1960 		spin_lock(&ci->i_ceph_lock);
1961 		ci->i_inline_version = CEPH_INLINE_NONE;
1962 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1963 		spin_unlock(&ci->i_ceph_lock);
1964 		up_read(&fsc->mdsc->snap_rwsem);
1965 		if (dirty)
1966 			__mark_inode_dirty(inode, dirty);
1967 	}
1968 out_put_req:
1969 	ceph_osdc_put_request(req);
1970 	if (err == -ECANCELED)
1971 		err = 0;
1972 out_unlock:
1973 	if (folio) {
1974 		folio_unlock(folio);
1975 		folio_put(folio);
1976 	}
1977 out:
1978 	ceph_free_cap_flush(prealloc_cf);
1979 	doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1980 	      ceph_vinop(inode), inline_version, err);
1981 	return err;
1982 }
1983 
1984 static const struct vm_operations_struct ceph_vmops = {
1985 	.fault		= ceph_filemap_fault,
1986 	.page_mkwrite	= ceph_page_mkwrite,
1987 };
1988 
1989 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1990 {
1991 	struct address_space *mapping = file->f_mapping;
1992 
1993 	if (!mapping->a_ops->read_folio)
1994 		return -ENOEXEC;
1995 	vma->vm_ops = &ceph_vmops;
1996 	return 0;
1997 }
1998 
1999 enum {
2000 	POOL_READ	= 1,
2001 	POOL_WRITE	= 2,
2002 };
2003 
2004 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
2005 				s64 pool, struct ceph_string *pool_ns)
2006 {
2007 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
2008 	struct ceph_mds_client *mdsc = fsc->mdsc;
2009 	struct ceph_client *cl = fsc->client;
2010 	struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
2011 	struct rb_node **p, *parent;
2012 	struct ceph_pool_perm *perm;
2013 	struct page **pages;
2014 	size_t pool_ns_len;
2015 	int err = 0, err2 = 0, have = 0;
2016 
2017 	down_read(&mdsc->pool_perm_rwsem);
2018 	p = &mdsc->pool_perm_tree.rb_node;
2019 	while (*p) {
2020 		perm = rb_entry(*p, struct ceph_pool_perm, node);
2021 		if (pool < perm->pool)
2022 			p = &(*p)->rb_left;
2023 		else if (pool > perm->pool)
2024 			p = &(*p)->rb_right;
2025 		else {
2026 			int ret = ceph_compare_string(pool_ns,
2027 						perm->pool_ns,
2028 						perm->pool_ns_len);
2029 			if (ret < 0)
2030 				p = &(*p)->rb_left;
2031 			else if (ret > 0)
2032 				p = &(*p)->rb_right;
2033 			else {
2034 				have = perm->perm;
2035 				break;
2036 			}
2037 		}
2038 	}
2039 	up_read(&mdsc->pool_perm_rwsem);
2040 	if (*p)
2041 		goto out;
2042 
2043 	if (pool_ns)
2044 		doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2045 		      (int)pool_ns->len, pool_ns->str);
2046 	else
2047 		doutc(cl, "pool %lld no perm cached\n", pool);
2048 
2049 	down_write(&mdsc->pool_perm_rwsem);
2050 	p = &mdsc->pool_perm_tree.rb_node;
2051 	parent = NULL;
2052 	while (*p) {
2053 		parent = *p;
2054 		perm = rb_entry(parent, struct ceph_pool_perm, node);
2055 		if (pool < perm->pool)
2056 			p = &(*p)->rb_left;
2057 		else if (pool > perm->pool)
2058 			p = &(*p)->rb_right;
2059 		else {
2060 			int ret = ceph_compare_string(pool_ns,
2061 						perm->pool_ns,
2062 						perm->pool_ns_len);
2063 			if (ret < 0)
2064 				p = &(*p)->rb_left;
2065 			else if (ret > 0)
2066 				p = &(*p)->rb_right;
2067 			else {
2068 				have = perm->perm;
2069 				break;
2070 			}
2071 		}
2072 	}
2073 	if (*p) {
2074 		up_write(&mdsc->pool_perm_rwsem);
2075 		goto out;
2076 	}
2077 
2078 	rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2079 					 1, false, GFP_NOFS);
2080 	if (!rd_req) {
2081 		err = -ENOMEM;
2082 		goto out_unlock;
2083 	}
2084 
2085 	rd_req->r_flags = CEPH_OSD_FLAG_READ;
2086 	osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2087 	rd_req->r_base_oloc.pool = pool;
2088 	if (pool_ns)
2089 		rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2090 	ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2091 
2092 	err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2093 	if (err)
2094 		goto out_unlock;
2095 
2096 	wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2097 					 1, false, GFP_NOFS);
2098 	if (!wr_req) {
2099 		err = -ENOMEM;
2100 		goto out_unlock;
2101 	}
2102 
2103 	wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2104 	osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2105 	ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2106 	ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2107 
2108 	err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2109 	if (err)
2110 		goto out_unlock;
2111 
2112 	/* one page should be large enough for STAT data */
2113 	pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2114 	if (IS_ERR(pages)) {
2115 		err = PTR_ERR(pages);
2116 		goto out_unlock;
2117 	}
2118 
2119 	osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2120 				     0, false, true);
2121 	ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2122 
2123 	wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2124 	ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2125 
2126 	err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2127 	err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2128 
2129 	if (err >= 0 || err == -ENOENT)
2130 		have |= POOL_READ;
2131 	else if (err != -EPERM) {
2132 		if (err == -EBLOCKLISTED)
2133 			fsc->blocklisted = true;
2134 		goto out_unlock;
2135 	}
2136 
2137 	if (err2 == 0 || err2 == -EEXIST)
2138 		have |= POOL_WRITE;
2139 	else if (err2 != -EPERM) {
2140 		if (err2 == -EBLOCKLISTED)
2141 			fsc->blocklisted = true;
2142 		err = err2;
2143 		goto out_unlock;
2144 	}
2145 
2146 	pool_ns_len = pool_ns ? pool_ns->len : 0;
2147 	perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2148 	if (!perm) {
2149 		err = -ENOMEM;
2150 		goto out_unlock;
2151 	}
2152 
2153 	perm->pool = pool;
2154 	perm->perm = have;
2155 	perm->pool_ns_len = pool_ns_len;
2156 	if (pool_ns_len > 0)
2157 		memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2158 	perm->pool_ns[pool_ns_len] = 0;
2159 
2160 	rb_link_node(&perm->node, parent, p);
2161 	rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2162 	err = 0;
2163 out_unlock:
2164 	up_write(&mdsc->pool_perm_rwsem);
2165 
2166 	ceph_osdc_put_request(rd_req);
2167 	ceph_osdc_put_request(wr_req);
2168 out:
2169 	if (!err)
2170 		err = have;
2171 	if (pool_ns)
2172 		doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2173 		      (int)pool_ns->len, pool_ns->str, err);
2174 	else
2175 		doutc(cl, "pool %lld result = %d\n", pool, err);
2176 	return err;
2177 }
2178 
2179 int ceph_pool_perm_check(struct inode *inode, int need)
2180 {
2181 	struct ceph_client *cl = ceph_inode_to_client(inode);
2182 	struct ceph_inode_info *ci = ceph_inode(inode);
2183 	struct ceph_string *pool_ns;
2184 	s64 pool;
2185 	int ret, flags;
2186 
2187 	/* Only need to do this for regular files */
2188 	if (!S_ISREG(inode->i_mode))
2189 		return 0;
2190 
2191 	if (ci->i_vino.snap != CEPH_NOSNAP) {
2192 		/*
2193 		 * Pool permission check needs to write to the first object.
2194 		 * But for snapshot, head of the first object may have alread
2195 		 * been deleted. Skip check to avoid creating orphan object.
2196 		 */
2197 		return 0;
2198 	}
2199 
2200 	if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2201 				NOPOOLPERM))
2202 		return 0;
2203 
2204 	spin_lock(&ci->i_ceph_lock);
2205 	flags = ci->i_ceph_flags;
2206 	pool = ci->i_layout.pool_id;
2207 	spin_unlock(&ci->i_ceph_lock);
2208 check:
2209 	if (flags & CEPH_I_POOL_PERM) {
2210 		if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2211 			doutc(cl, "pool %lld no read perm\n", pool);
2212 			return -EPERM;
2213 		}
2214 		if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2215 			doutc(cl, "pool %lld no write perm\n", pool);
2216 			return -EPERM;
2217 		}
2218 		return 0;
2219 	}
2220 
2221 	pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2222 	ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2223 	ceph_put_string(pool_ns);
2224 	if (ret < 0)
2225 		return ret;
2226 
2227 	flags = CEPH_I_POOL_PERM;
2228 	if (ret & POOL_READ)
2229 		flags |= CEPH_I_POOL_RD;
2230 	if (ret & POOL_WRITE)
2231 		flags |= CEPH_I_POOL_WR;
2232 
2233 	spin_lock(&ci->i_ceph_lock);
2234 	if (pool == ci->i_layout.pool_id &&
2235 	    pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2236 		ci->i_ceph_flags |= flags;
2237         } else {
2238 		pool = ci->i_layout.pool_id;
2239 		flags = ci->i_ceph_flags;
2240 	}
2241 	spin_unlock(&ci->i_ceph_lock);
2242 	goto check;
2243 }
2244 
2245 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2246 {
2247 	struct ceph_pool_perm *perm;
2248 	struct rb_node *n;
2249 
2250 	while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2251 		n = rb_first(&mdsc->pool_perm_tree);
2252 		perm = rb_entry(n, struct ceph_pool_perm, node);
2253 		rb_erase(n, &mdsc->pool_perm_tree);
2254 		kfree(perm);
2255 	}
2256 }
2257