xref: /linux/fs/ceph/addr.c (revision ca56a74a31e26d81a481304ed2f631e65883372b)
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 #include <trace/events/netfs.h>
17 
18 #include "super.h"
19 #include "mds_client.h"
20 #include "cache.h"
21 #include "metric.h"
22 #include "crypto.h"
23 #include <linux/ceph/osd_client.h>
24 #include <linux/ceph/striper.h>
25 
26 /*
27  * Ceph address space ops.
28  *
29  * There are a few funny things going on here.
30  *
31  * The page->private field is used to reference a struct
32  * ceph_snap_context for _every_ dirty page.  This indicates which
33  * snapshot the page was logically dirtied in, and thus which snap
34  * context needs to be associated with the osd write during writeback.
35  *
36  * Similarly, struct ceph_inode_info maintains a set of counters to
37  * count dirty pages on the inode.  In the absence of snapshots,
38  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
39  *
40  * When a snapshot is taken (that is, when the client receives
41  * notification that a snapshot was taken), each inode with caps and
42  * with dirty pages (dirty pages implies there is a cap) gets a new
43  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
44  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
45  * moved to capsnap->dirty. (Unless a sync write is currently in
46  * progress.  In that case, the capsnap is said to be "pending", new
47  * writes cannot start, and the capsnap isn't "finalized" until the
48  * write completes (or fails) and a final size/mtime for the inode for
49  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
50  *
51  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
52  * we look for the first capsnap in i_cap_snaps and write out pages in
53  * that snap context _only_.  Then we move on to the next capsnap,
54  * eventually reaching the "live" or "head" context (i.e., pages that
55  * are not yet snapped) and are writing the most recently dirtied
56  * pages.
57  *
58  * Invalidate and so forth must take care to ensure the dirty page
59  * accounting is preserved.
60  */
61 
62 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
63 #define CONGESTION_OFF_THRESH(congestion_kb)				\
64 	(CONGESTION_ON_THRESH(congestion_kb) -				\
65 	 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
66 
67 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
68 					struct folio **foliop, void **_fsdata);
69 
page_snap_context(struct page * page)70 static inline struct ceph_snap_context *page_snap_context(struct page *page)
71 {
72 	if (PagePrivate(page))
73 		return (void *)page->private;
74 	return NULL;
75 }
76 
77 /*
78  * Dirty a page.  Optimistically adjust accounting, on the assumption
79  * that we won't race with invalidate.  If we do, readjust.
80  */
ceph_dirty_folio(struct address_space * mapping,struct folio * folio)81 static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
82 {
83 	struct inode *inode = mapping->host;
84 	struct ceph_client *cl = ceph_inode_to_client(inode);
85 	struct ceph_inode_info *ci;
86 	struct ceph_snap_context *snapc;
87 
88 	if (folio_test_dirty(folio)) {
89 		doutc(cl, "%llx.%llx %p idx %lu -- already dirty\n",
90 		      ceph_vinop(inode), folio, folio->index);
91 		VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
92 		return false;
93 	}
94 
95 	ci = ceph_inode(inode);
96 
97 	/* dirty the head */
98 	spin_lock(&ci->i_ceph_lock);
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  */
ceph_invalidate_folio(struct folio * folio,size_t offset,size_t length)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 	netfs_invalidate_folio(folio, offset, length);
163 }
164 
ceph_netfs_expand_readahead(struct netfs_io_request * rreq)165 static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
166 {
167 	struct inode *inode = rreq->inode;
168 	struct ceph_inode_info *ci = ceph_inode(inode);
169 	struct ceph_file_layout *lo = &ci->i_layout;
170 	unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
171 	loff_t end = rreq->start + rreq->len, new_end;
172 	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
173 	unsigned long max_len;
174 	u32 blockoff;
175 
176 	if (priv) {
177 		/* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
178 		if (priv->file_ra_disabled)
179 			max_pages = 0;
180 		else
181 			max_pages = priv->file_ra_pages;
182 
183 	}
184 
185 	/* Readahead is disabled */
186 	if (!max_pages)
187 		return;
188 
189 	max_len = max_pages << PAGE_SHIFT;
190 
191 	/*
192 	 * Try to expand the length forward by rounding up it to the next
193 	 * block, but do not exceed the file size, unless the original
194 	 * request already exceeds it.
195 	 */
196 	new_end = umin(round_up(end, lo->stripe_unit), rreq->i_size);
197 	if (new_end > end && new_end <= rreq->start + max_len)
198 		rreq->len = new_end - rreq->start;
199 
200 	/* Try to expand the start downward */
201 	div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
202 	if (rreq->len + blockoff <= max_len) {
203 		rreq->start -= blockoff;
204 		rreq->len += blockoff;
205 	}
206 }
207 
finish_netfs_read(struct ceph_osd_request * req)208 static void finish_netfs_read(struct ceph_osd_request *req)
209 {
210 	struct inode *inode = req->r_inode;
211 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
212 	struct ceph_client *cl = fsc->client;
213 	struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
214 	struct netfs_io_subrequest *subreq = req->r_priv;
215 	struct ceph_osd_req_op *op = &req->r_ops[0];
216 	int err = req->r_result;
217 	bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
218 
219 	ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
220 				 req->r_end_latency, osd_data->length, err);
221 
222 	doutc(cl, "result %d subreq->len=%zu i_size=%lld\n", req->r_result,
223 	      subreq->len, i_size_read(req->r_inode));
224 
225 	/* no object means success but no data */
226 	if (err == -ENOENT) {
227 		__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
228 		__set_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
229 		err = 0;
230 	} else if (err == -EBLOCKLISTED) {
231 		fsc->blocklisted = true;
232 	}
233 
234 	if (err >= 0) {
235 		if (sparse && err > 0)
236 			err = ceph_sparse_ext_map_end(op);
237 		if (err < subreq->len &&
238 		    subreq->rreq->origin != NETFS_DIO_READ)
239 			__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
240 		if (IS_ENCRYPTED(inode) && err > 0) {
241 			err = ceph_fscrypt_decrypt_extents(inode,
242 					osd_data->pages, subreq->start,
243 					op->extent.sparse_ext,
244 					op->extent.sparse_ext_cnt);
245 			if (err > subreq->len)
246 				err = subreq->len;
247 		}
248 		if (err > 0)
249 			__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
250 	}
251 
252 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
253 		ceph_put_page_vector(osd_data->pages,
254 				     calc_pages_for(osd_data->alignment,
255 					osd_data->length), false);
256 	}
257 	if (err > 0) {
258 		subreq->transferred = err;
259 		err = 0;
260 	}
261 	subreq->error = err;
262 	trace_netfs_sreq(subreq, netfs_sreq_trace_io_progress);
263 	netfs_read_subreq_terminated(subreq);
264 	iput(req->r_inode);
265 	ceph_dec_osd_stopping_blocker(fsc->mdsc);
266 }
267 
ceph_netfs_issue_op_inline(struct netfs_io_subrequest * subreq)268 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
269 {
270 	struct netfs_io_request *rreq = subreq->rreq;
271 	struct inode *inode = rreq->inode;
272 	struct ceph_mds_reply_info_parsed *rinfo;
273 	struct ceph_mds_reply_info_in *iinfo;
274 	struct ceph_mds_request *req;
275 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
276 	struct ceph_inode_info *ci = ceph_inode(inode);
277 	ssize_t err = 0;
278 	size_t len;
279 	int mode;
280 
281 	if (rreq->origin != NETFS_DIO_READ)
282 		__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
283 	__clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
284 
285 	if (subreq->start >= inode->i_size)
286 		goto out;
287 
288 	/* We need to fetch the inline data. */
289 	mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
290 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
291 	if (IS_ERR(req)) {
292 		err = PTR_ERR(req);
293 		goto out;
294 	}
295 	req->r_ino1 = ci->i_vino;
296 	req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
297 	req->r_num_caps = 2;
298 
299 	trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
300 	err = ceph_mdsc_do_request(mdsc, NULL, req);
301 	if (err < 0)
302 		goto out;
303 
304 	rinfo = &req->r_reply_info;
305 	iinfo = &rinfo->targeti;
306 	if (iinfo->inline_version == CEPH_INLINE_NONE) {
307 		/* The data got uninlined */
308 		ceph_mdsc_put_request(req);
309 		return false;
310 	}
311 
312 	len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
313 	err = copy_to_iter(iinfo->inline_data + subreq->start, len, &subreq->io_iter);
314 	if (err == 0) {
315 		err = -EFAULT;
316 	} else {
317 		subreq->transferred += err;
318 		err = 0;
319 	}
320 
321 	ceph_mdsc_put_request(req);
322 out:
323 	subreq->error = err;
324 	trace_netfs_sreq(subreq, netfs_sreq_trace_io_progress);
325 	netfs_read_subreq_terminated(subreq);
326 	return true;
327 }
328 
ceph_netfs_prepare_read(struct netfs_io_subrequest * subreq)329 static int ceph_netfs_prepare_read(struct netfs_io_subrequest *subreq)
330 {
331 	struct netfs_io_request *rreq = subreq->rreq;
332 	struct inode *inode = rreq->inode;
333 	struct ceph_inode_info *ci = ceph_inode(inode);
334 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
335 	u64 objno, objoff;
336 	u32 xlen;
337 
338 	/* Truncate the extent at the end of the current block */
339 	ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
340 				      &objno, &objoff, &xlen);
341 	rreq->io_streams[0].sreq_max_len = umin(xlen, fsc->mount_options->rsize);
342 	return 0;
343 }
344 
ceph_netfs_issue_read(struct netfs_io_subrequest * subreq)345 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
346 {
347 	struct netfs_io_request *rreq = subreq->rreq;
348 	struct inode *inode = rreq->inode;
349 	struct ceph_inode_info *ci = ceph_inode(inode);
350 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
351 	struct ceph_client *cl = fsc->client;
352 	struct ceph_osd_request *req = NULL;
353 	struct ceph_vino vino = ceph_vino(inode);
354 	int err;
355 	u64 len;
356 	bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
357 	u64 off = subreq->start;
358 	int extent_cnt;
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 	// TODO: This rounding here is slightly dodgy.  It *should* work, for
369 	// now, as the cache only deals in blocks that are a multiple of
370 	// PAGE_SIZE and fscrypt blocks are at most PAGE_SIZE.  What needs to
371 	// happen is for the fscrypt driving to be moved into netfslib and the
372 	// data in the cache also to be stored encrypted.
373 	len = subreq->len;
374 	ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
375 
376 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
377 			off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
378 			CEPH_OSD_FLAG_READ, NULL, ci->i_truncate_seq,
379 			ci->i_truncate_size, false);
380 	if (IS_ERR(req)) {
381 		err = PTR_ERR(req);
382 		req = NULL;
383 		goto out;
384 	}
385 
386 	if (sparse) {
387 		extent_cnt = __ceph_sparse_read_ext_count(inode, len);
388 		err = ceph_alloc_sparse_ext_map(&req->r_ops[0], extent_cnt);
389 		if (err)
390 			goto out;
391 	}
392 
393 	doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
394 	      ceph_vinop(inode), subreq->start, subreq->len, len);
395 
396 	/*
397 	 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
398 	 * encrypted inodes. We'd need infrastructure that handles an iov_iter
399 	 * instead of page arrays, and we don't have that as of yet. Once the
400 	 * dust settles on the write helpers and encrypt/decrypt routines for
401 	 * netfs, we should be able to rework this.
402 	 */
403 	if (IS_ENCRYPTED(inode)) {
404 		struct page **pages;
405 		size_t page_off;
406 
407 		err = iov_iter_get_pages_alloc2(&subreq->io_iter, &pages, len, &page_off);
408 		if (err < 0) {
409 			doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
410 			      ceph_vinop(inode), err);
411 			goto out;
412 		}
413 
414 		/* should always give us a page-aligned read */
415 		WARN_ON_ONCE(page_off);
416 		len = err;
417 		err = 0;
418 
419 		osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
420 						 false);
421 	} else {
422 		osd_req_op_extent_osd_iter(req, 0, &subreq->io_iter);
423 	}
424 	if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
425 		err = -EIO;
426 		goto out;
427 	}
428 	req->r_callback = finish_netfs_read;
429 	req->r_priv = subreq;
430 	req->r_inode = inode;
431 	ihold(inode);
432 
433 	trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
434 	ceph_osdc_start_request(req->r_osdc, req);
435 out:
436 	ceph_osdc_put_request(req);
437 	if (err) {
438 		subreq->error = err;
439 		netfs_read_subreq_terminated(subreq);
440 	}
441 	doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
442 }
443 
ceph_init_request(struct netfs_io_request * rreq,struct file * file)444 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
445 {
446 	struct inode *inode = rreq->inode;
447 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
448 	struct ceph_client *cl = ceph_inode_to_client(inode);
449 	int got = 0, want = CEPH_CAP_FILE_CACHE;
450 	struct ceph_netfs_request_data *priv;
451 	int ret = 0;
452 
453 	/* [DEPRECATED] Use PG_private_2 to mark folio being written to the cache. */
454 	__set_bit(NETFS_RREQ_USE_PGPRIV2, &rreq->flags);
455 
456 	if (rreq->origin != NETFS_READAHEAD)
457 		return 0;
458 
459 	priv = kzalloc(sizeof(*priv), GFP_NOFS);
460 	if (!priv)
461 		return -ENOMEM;
462 
463 	if (file) {
464 		struct ceph_rw_context *rw_ctx;
465 		struct ceph_file_info *fi = file->private_data;
466 
467 		priv->file_ra_pages = file->f_ra.ra_pages;
468 		priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
469 
470 		rw_ctx = ceph_find_rw_context(fi);
471 		if (rw_ctx) {
472 			rreq->netfs_priv = priv;
473 			return 0;
474 		}
475 	}
476 
477 	/*
478 	 * readahead callers do not necessarily hold Fcb caps
479 	 * (e.g. fadvise, madvise).
480 	 */
481 	ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
482 	if (ret < 0) {
483 		doutc(cl, "%llx.%llx, error getting cap\n", ceph_vinop(inode));
484 		goto out;
485 	}
486 
487 	if (!(got & want)) {
488 		doutc(cl, "%llx.%llx, no cache cap\n", ceph_vinop(inode));
489 		ret = -EACCES;
490 		goto out;
491 	}
492 	if (ret == 0) {
493 		ret = -EACCES;
494 		goto out;
495 	}
496 
497 	priv->caps = got;
498 	rreq->netfs_priv = priv;
499 	rreq->io_streams[0].sreq_max_len = fsc->mount_options->rsize;
500 
501 out:
502 	if (ret < 0) {
503 		if (got)
504 			ceph_put_cap_refs(ceph_inode(inode), got);
505 		kfree(priv);
506 	}
507 
508 	return ret;
509 }
510 
ceph_netfs_free_request(struct netfs_io_request * rreq)511 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
512 {
513 	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
514 
515 	if (!priv)
516 		return;
517 
518 	if (priv->caps)
519 		ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
520 	kfree(priv);
521 	rreq->netfs_priv = NULL;
522 }
523 
524 const struct netfs_request_ops ceph_netfs_ops = {
525 	.init_request		= ceph_init_request,
526 	.free_request		= ceph_netfs_free_request,
527 	.prepare_read		= ceph_netfs_prepare_read,
528 	.issue_read		= ceph_netfs_issue_read,
529 	.expand_readahead	= ceph_netfs_expand_readahead,
530 	.check_write_begin	= ceph_netfs_check_write_begin,
531 };
532 
533 #ifdef CONFIG_CEPH_FSCACHE
ceph_set_page_fscache(struct page * page)534 static void ceph_set_page_fscache(struct page *page)
535 {
536 	folio_start_private_2(page_folio(page)); /* [DEPRECATED] */
537 }
538 
ceph_fscache_write_terminated(void * priv,ssize_t error,bool was_async)539 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
540 {
541 	struct inode *inode = priv;
542 
543 	if (IS_ERR_VALUE(error) && error != -ENOBUFS)
544 		ceph_fscache_invalidate(inode, false);
545 }
546 
ceph_fscache_write_to_cache(struct inode * inode,u64 off,u64 len,bool caching)547 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
548 {
549 	struct ceph_inode_info *ci = ceph_inode(inode);
550 	struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
551 
552 	fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
553 			       ceph_fscache_write_terminated, inode, true, caching);
554 }
555 #else
ceph_set_page_fscache(struct page * page)556 static inline void ceph_set_page_fscache(struct page *page)
557 {
558 }
559 
ceph_fscache_write_to_cache(struct inode * inode,u64 off,u64 len,bool caching)560 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
561 {
562 }
563 #endif /* CONFIG_CEPH_FSCACHE */
564 
565 struct ceph_writeback_ctl
566 {
567 	loff_t i_size;
568 	u64 truncate_size;
569 	u32 truncate_seq;
570 	bool size_stable;
571 	bool head_snapc;
572 };
573 
574 /*
575  * Get ref for the oldest snapc for an inode with dirty data... that is, the
576  * only snap context we are allowed to write back.
577  */
578 static struct ceph_snap_context *
get_oldest_context(struct inode * inode,struct ceph_writeback_ctl * ctl,struct ceph_snap_context * page_snapc)579 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
580 		   struct ceph_snap_context *page_snapc)
581 {
582 	struct ceph_inode_info *ci = ceph_inode(inode);
583 	struct ceph_client *cl = ceph_inode_to_client(inode);
584 	struct ceph_snap_context *snapc = NULL;
585 	struct ceph_cap_snap *capsnap = NULL;
586 
587 	spin_lock(&ci->i_ceph_lock);
588 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
589 		doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
590 		      capsnap, capsnap->context, capsnap->dirty_pages);
591 		if (!capsnap->dirty_pages)
592 			continue;
593 
594 		/* get i_size, truncate_{seq,size} for page_snapc? */
595 		if (snapc && capsnap->context != page_snapc)
596 			continue;
597 
598 		if (ctl) {
599 			if (capsnap->writing) {
600 				ctl->i_size = i_size_read(inode);
601 				ctl->size_stable = false;
602 			} else {
603 				ctl->i_size = capsnap->size;
604 				ctl->size_stable = true;
605 			}
606 			ctl->truncate_size = capsnap->truncate_size;
607 			ctl->truncate_seq = capsnap->truncate_seq;
608 			ctl->head_snapc = false;
609 		}
610 
611 		if (snapc)
612 			break;
613 
614 		snapc = ceph_get_snap_context(capsnap->context);
615 		if (!page_snapc ||
616 		    page_snapc == snapc ||
617 		    page_snapc->seq > snapc->seq)
618 			break;
619 	}
620 	if (!snapc && ci->i_wrbuffer_ref_head) {
621 		snapc = ceph_get_snap_context(ci->i_head_snapc);
622 		doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
623 		      ci->i_wrbuffer_ref_head);
624 		if (ctl) {
625 			ctl->i_size = i_size_read(inode);
626 			ctl->truncate_size = ci->i_truncate_size;
627 			ctl->truncate_seq = ci->i_truncate_seq;
628 			ctl->size_stable = false;
629 			ctl->head_snapc = true;
630 		}
631 	}
632 	spin_unlock(&ci->i_ceph_lock);
633 	return snapc;
634 }
635 
get_writepages_data_length(struct inode * inode,struct page * page,u64 start)636 static u64 get_writepages_data_length(struct inode *inode,
637 				      struct page *page, u64 start)
638 {
639 	struct ceph_inode_info *ci = ceph_inode(inode);
640 	struct ceph_snap_context *snapc;
641 	struct ceph_cap_snap *capsnap = NULL;
642 	u64 end = i_size_read(inode);
643 	u64 ret;
644 
645 	snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
646 	if (snapc != ci->i_head_snapc) {
647 		bool found = false;
648 		spin_lock(&ci->i_ceph_lock);
649 		list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
650 			if (capsnap->context == snapc) {
651 				if (!capsnap->writing)
652 					end = capsnap->size;
653 				found = true;
654 				break;
655 			}
656 		}
657 		spin_unlock(&ci->i_ceph_lock);
658 		WARN_ON(!found);
659 	}
660 	if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
661 		end = ceph_fscrypt_page_offset(page) + thp_size(page);
662 	ret = end > start ? end - start : 0;
663 	if (ret && fscrypt_is_bounce_page(page))
664 		ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
665 	return ret;
666 }
667 
668 /*
669  * Write a single page, but leave the page locked.
670  *
671  * If we get a write error, mark the mapping for error, but still adjust the
672  * dirty page accounting (i.e., page is no longer dirty).
673  */
writepage_nounlock(struct page * page,struct writeback_control * wbc)674 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
675 {
676 	struct folio *folio = page_folio(page);
677 	struct inode *inode = page->mapping->host;
678 	struct ceph_inode_info *ci = ceph_inode(inode);
679 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
680 	struct ceph_client *cl = fsc->client;
681 	struct ceph_snap_context *snapc, *oldest;
682 	loff_t page_off = page_offset(page);
683 	int err;
684 	loff_t len = thp_size(page);
685 	loff_t wlen;
686 	struct ceph_writeback_ctl ceph_wbc;
687 	struct ceph_osd_client *osdc = &fsc->client->osdc;
688 	struct ceph_osd_request *req;
689 	bool caching = ceph_is_cache_enabled(inode);
690 	struct page *bounce_page = NULL;
691 
692 	doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
693 	      page->index);
694 
695 	if (ceph_inode_is_shutdown(inode))
696 		return -EIO;
697 
698 	/* verify this is a writeable snap context */
699 	snapc = page_snap_context(page);
700 	if (!snapc) {
701 		doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
702 		      page);
703 		return 0;
704 	}
705 	oldest = get_oldest_context(inode, &ceph_wbc, snapc);
706 	if (snapc->seq > oldest->seq) {
707 		doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
708 		      ceph_vinop(inode), page, snapc);
709 		/* we should only noop if called by kswapd */
710 		WARN_ON(!(current->flags & PF_MEMALLOC));
711 		ceph_put_snap_context(oldest);
712 		redirty_page_for_writepage(wbc, page);
713 		return 0;
714 	}
715 	ceph_put_snap_context(oldest);
716 
717 	/* is this a partial page at end of file? */
718 	if (page_off >= ceph_wbc.i_size) {
719 		doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
720 		      ceph_vinop(inode), folio->index, ceph_wbc.i_size);
721 		folio_invalidate(folio, 0, folio_size(folio));
722 		return 0;
723 	}
724 
725 	if (ceph_wbc.i_size < page_off + len)
726 		len = ceph_wbc.i_size - page_off;
727 
728 	wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
729 	doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
730 	      ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
731 	      snapc->seq);
732 
733 	if (atomic_long_inc_return(&fsc->writeback_count) >
734 	    CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
735 		fsc->write_congested = true;
736 
737 	req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
738 				    page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
739 				    CEPH_OSD_FLAG_WRITE, snapc,
740 				    ceph_wbc.truncate_seq,
741 				    ceph_wbc.truncate_size, true);
742 	if (IS_ERR(req)) {
743 		redirty_page_for_writepage(wbc, page);
744 		return PTR_ERR(req);
745 	}
746 
747 	if (wlen < len)
748 		len = wlen;
749 
750 	set_page_writeback(page);
751 	if (caching)
752 		ceph_set_page_fscache(page);
753 	ceph_fscache_write_to_cache(inode, page_off, len, caching);
754 
755 	if (IS_ENCRYPTED(inode)) {
756 		bounce_page = fscrypt_encrypt_pagecache_blocks(page,
757 						    CEPH_FSCRYPT_BLOCK_SIZE, 0,
758 						    GFP_NOFS);
759 		if (IS_ERR(bounce_page)) {
760 			redirty_page_for_writepage(wbc, page);
761 			end_page_writeback(page);
762 			ceph_osdc_put_request(req);
763 			return PTR_ERR(bounce_page);
764 		}
765 	}
766 
767 	/* it may be a short write due to an object boundary */
768 	WARN_ON_ONCE(len > thp_size(page));
769 	osd_req_op_extent_osd_data_pages(req, 0,
770 			bounce_page ? &bounce_page : &page, wlen, 0,
771 			false, false);
772 	doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
773 	      ceph_vinop(inode), page_off, len, wlen,
774 	      IS_ENCRYPTED(inode) ? "" : "not ");
775 
776 	req->r_mtime = inode_get_mtime(inode);
777 	ceph_osdc_start_request(osdc, req);
778 	err = ceph_osdc_wait_request(osdc, req);
779 
780 	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
781 				  req->r_end_latency, len, err);
782 	fscrypt_free_bounce_page(bounce_page);
783 	ceph_osdc_put_request(req);
784 	if (err == 0)
785 		err = len;
786 
787 	if (err < 0) {
788 		struct writeback_control tmp_wbc;
789 		if (!wbc)
790 			wbc = &tmp_wbc;
791 		if (err == -ERESTARTSYS) {
792 			/* killed by SIGKILL */
793 			doutc(cl, "%llx.%llx interrupted page %p\n",
794 			      ceph_vinop(inode), page);
795 			redirty_page_for_writepage(wbc, page);
796 			end_page_writeback(page);
797 			return err;
798 		}
799 		if (err == -EBLOCKLISTED)
800 			fsc->blocklisted = true;
801 		doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
802 		      ceph_vinop(inode), err, page);
803 		mapping_set_error(&inode->i_data, err);
804 		wbc->pages_skipped++;
805 	} else {
806 		doutc(cl, "%llx.%llx cleaned page %p\n",
807 		      ceph_vinop(inode), page);
808 		err = 0;  /* vfs expects us to return 0 */
809 	}
810 	oldest = detach_page_private(page);
811 	WARN_ON_ONCE(oldest != snapc);
812 	end_page_writeback(page);
813 	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
814 	ceph_put_snap_context(snapc);  /* page's reference */
815 
816 	if (atomic_long_dec_return(&fsc->writeback_count) <
817 	    CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
818 		fsc->write_congested = false;
819 
820 	return err;
821 }
822 
ceph_writepage(struct page * page,struct writeback_control * wbc)823 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
824 {
825 	int err;
826 	struct inode *inode = page->mapping->host;
827 	BUG_ON(!inode);
828 	ihold(inode);
829 
830 	if (wbc->sync_mode == WB_SYNC_NONE &&
831 	    ceph_inode_to_fs_client(inode)->write_congested) {
832 		redirty_page_for_writepage(wbc, page);
833 		return AOP_WRITEPAGE_ACTIVATE;
834 	}
835 
836 	folio_wait_private_2(page_folio(page)); /* [DEPRECATED] */
837 
838 	err = writepage_nounlock(page, wbc);
839 	if (err == -ERESTARTSYS) {
840 		/* direct memory reclaimer was killed by SIGKILL. return 0
841 		 * to prevent caller from setting mapping/page error */
842 		err = 0;
843 	}
844 	unlock_page(page);
845 	iput(inode);
846 	return err;
847 }
848 
849 /*
850  * async writeback completion handler.
851  *
852  * If we get an error, set the mapping error bit, but not the individual
853  * page error bits.
854  */
writepages_finish(struct ceph_osd_request * req)855 static void writepages_finish(struct ceph_osd_request *req)
856 {
857 	struct inode *inode = req->r_inode;
858 	struct ceph_inode_info *ci = ceph_inode(inode);
859 	struct ceph_client *cl = ceph_inode_to_client(inode);
860 	struct ceph_osd_data *osd_data;
861 	struct page *page;
862 	int num_pages, total_pages = 0;
863 	int i, j;
864 	int rc = req->r_result;
865 	struct ceph_snap_context *snapc = req->r_snapc;
866 	struct address_space *mapping = inode->i_mapping;
867 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
868 	unsigned int len = 0;
869 	bool remove_page;
870 
871 	doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
872 	if (rc < 0) {
873 		mapping_set_error(mapping, rc);
874 		ceph_set_error_write(ci);
875 		if (rc == -EBLOCKLISTED)
876 			fsc->blocklisted = true;
877 	} else {
878 		ceph_clear_error_write(ci);
879 	}
880 
881 	/*
882 	 * We lost the cache cap, need to truncate the page before
883 	 * it is unlocked, otherwise we'd truncate it later in the
884 	 * page truncation thread, possibly losing some data that
885 	 * raced its way in
886 	 */
887 	remove_page = !(ceph_caps_issued(ci) &
888 			(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
889 
890 	/* clean all pages */
891 	for (i = 0; i < req->r_num_ops; i++) {
892 		if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
893 			pr_warn_client(cl,
894 				"%llx.%llx incorrect op %d req %p index %d tid %llu\n",
895 				ceph_vinop(inode), req->r_ops[i].op, req, i,
896 				req->r_tid);
897 			break;
898 		}
899 
900 		osd_data = osd_req_op_extent_osd_data(req, i);
901 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
902 		len += osd_data->length;
903 		num_pages = calc_pages_for((u64)osd_data->alignment,
904 					   (u64)osd_data->length);
905 		total_pages += num_pages;
906 		for (j = 0; j < num_pages; j++) {
907 			page = osd_data->pages[j];
908 			if (fscrypt_is_bounce_page(page)) {
909 				page = fscrypt_pagecache_page(page);
910 				fscrypt_free_bounce_page(osd_data->pages[j]);
911 				osd_data->pages[j] = page;
912 			}
913 			BUG_ON(!page);
914 			WARN_ON(!PageUptodate(page));
915 
916 			if (atomic_long_dec_return(&fsc->writeback_count) <
917 			     CONGESTION_OFF_THRESH(
918 					fsc->mount_options->congestion_kb))
919 				fsc->write_congested = false;
920 
921 			ceph_put_snap_context(detach_page_private(page));
922 			end_page_writeback(page);
923 			doutc(cl, "unlocking %p\n", page);
924 
925 			if (remove_page)
926 				generic_error_remove_folio(inode->i_mapping,
927 							  page_folio(page));
928 
929 			unlock_page(page);
930 		}
931 		doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
932 		      ceph_vinop(inode), osd_data->length,
933 		      rc >= 0 ? num_pages : 0);
934 
935 		release_pages(osd_data->pages, num_pages);
936 	}
937 
938 	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
939 				  req->r_end_latency, len, rc);
940 
941 	ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
942 
943 	osd_data = osd_req_op_extent_osd_data(req, 0);
944 	if (osd_data->pages_from_pool)
945 		mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
946 	else
947 		kfree(osd_data->pages);
948 	ceph_osdc_put_request(req);
949 	ceph_dec_osd_stopping_blocker(fsc->mdsc);
950 }
951 
952 /*
953  * initiate async writeback
954  */
ceph_writepages_start(struct address_space * mapping,struct writeback_control * wbc)955 static int ceph_writepages_start(struct address_space *mapping,
956 				 struct writeback_control *wbc)
957 {
958 	struct inode *inode = mapping->host;
959 	struct ceph_inode_info *ci = ceph_inode(inode);
960 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
961 	struct ceph_client *cl = fsc->client;
962 	struct ceph_vino vino = ceph_vino(inode);
963 	pgoff_t index, start_index, end = -1;
964 	struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
965 	struct folio_batch fbatch;
966 	int rc = 0;
967 	unsigned int wsize = i_blocksize(inode);
968 	struct ceph_osd_request *req = NULL;
969 	struct ceph_writeback_ctl ceph_wbc;
970 	bool should_loop, range_whole = false;
971 	bool done = false;
972 	bool caching = ceph_is_cache_enabled(inode);
973 	xa_mark_t tag;
974 
975 	if (wbc->sync_mode == WB_SYNC_NONE &&
976 	    fsc->write_congested)
977 		return 0;
978 
979 	doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
980 	      wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
981 	      (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
982 
983 	if (ceph_inode_is_shutdown(inode)) {
984 		if (ci->i_wrbuffer_ref > 0) {
985 			pr_warn_ratelimited_client(cl,
986 				"%llx.%llx %lld forced umount\n",
987 				ceph_vinop(inode), ceph_ino(inode));
988 		}
989 		mapping_set_error(mapping, -EIO);
990 		return -EIO; /* we're in a forced umount, don't write! */
991 	}
992 	if (fsc->mount_options->wsize < wsize)
993 		wsize = fsc->mount_options->wsize;
994 
995 	folio_batch_init(&fbatch);
996 
997 	start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
998 	index = start_index;
999 
1000 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
1001 		tag = PAGECACHE_TAG_TOWRITE;
1002 	} else {
1003 		tag = PAGECACHE_TAG_DIRTY;
1004 	}
1005 retry:
1006 	/* find oldest snap context with dirty data */
1007 	snapc = get_oldest_context(inode, &ceph_wbc, NULL);
1008 	if (!snapc) {
1009 		/* hmm, why does writepages get called when there
1010 		   is no dirty data? */
1011 		doutc(cl, " no snap context with dirty data?\n");
1012 		goto out;
1013 	}
1014 	doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
1015 	      snapc->seq, snapc->num_snaps);
1016 
1017 	should_loop = false;
1018 	if (ceph_wbc.head_snapc && snapc != last_snapc) {
1019 		/* where to start/end? */
1020 		if (wbc->range_cyclic) {
1021 			index = start_index;
1022 			end = -1;
1023 			if (index > 0)
1024 				should_loop = true;
1025 			doutc(cl, " cyclic, start at %lu\n", index);
1026 		} else {
1027 			index = wbc->range_start >> PAGE_SHIFT;
1028 			end = wbc->range_end >> PAGE_SHIFT;
1029 			if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1030 				range_whole = true;
1031 			doutc(cl, " not cyclic, %lu to %lu\n", index, end);
1032 		}
1033 	} else if (!ceph_wbc.head_snapc) {
1034 		/* Do not respect wbc->range_{start,end}. Dirty pages
1035 		 * in that range can be associated with newer snapc.
1036 		 * They are not writeable until we write all dirty pages
1037 		 * associated with 'snapc' get written */
1038 		if (index > 0)
1039 			should_loop = true;
1040 		doutc(cl, " non-head snapc, range whole\n");
1041 	}
1042 
1043 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1044 		tag_pages_for_writeback(mapping, index, end);
1045 
1046 	ceph_put_snap_context(last_snapc);
1047 	last_snapc = snapc;
1048 
1049 	while (!done && index <= end) {
1050 		int num_ops = 0, op_idx;
1051 		unsigned i, nr_folios, max_pages, locked_pages = 0;
1052 		struct page **pages = NULL, **data_pages;
1053 		struct page *page;
1054 		pgoff_t strip_unit_end = 0;
1055 		u64 offset = 0, len = 0;
1056 		bool from_pool = false;
1057 
1058 		max_pages = wsize >> PAGE_SHIFT;
1059 
1060 get_more_pages:
1061 		nr_folios = filemap_get_folios_tag(mapping, &index,
1062 						   end, tag, &fbatch);
1063 		doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1064 		if (!nr_folios && !locked_pages)
1065 			break;
1066 		for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1067 			struct folio *folio = fbatch.folios[i];
1068 
1069 			page = &folio->page;
1070 			doutc(cl, "? %p idx %lu\n", page, page->index);
1071 			if (locked_pages == 0)
1072 				lock_page(page);  /* first page */
1073 			else if (!trylock_page(page))
1074 				break;
1075 
1076 			/* only dirty pages, or our accounting breaks */
1077 			if (unlikely(!PageDirty(page)) ||
1078 			    unlikely(page->mapping != mapping)) {
1079 				doutc(cl, "!dirty or !mapping %p\n", page);
1080 				unlock_page(page);
1081 				continue;
1082 			}
1083 			/* only if matching snap context */
1084 			pgsnapc = page_snap_context(page);
1085 			if (pgsnapc != snapc) {
1086 				doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1087 				      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1088 				if (!should_loop &&
1089 				    !ceph_wbc.head_snapc &&
1090 				    wbc->sync_mode != WB_SYNC_NONE)
1091 					should_loop = true;
1092 				unlock_page(page);
1093 				continue;
1094 			}
1095 			if (page_offset(page) >= ceph_wbc.i_size) {
1096 				doutc(cl, "folio at %lu beyond eof %llu\n",
1097 				      folio->index, ceph_wbc.i_size);
1098 				if ((ceph_wbc.size_stable ||
1099 				    folio_pos(folio) >= i_size_read(inode)) &&
1100 				    folio_clear_dirty_for_io(folio))
1101 					folio_invalidate(folio, 0,
1102 							folio_size(folio));
1103 				folio_unlock(folio);
1104 				continue;
1105 			}
1106 			if (strip_unit_end && (page->index > strip_unit_end)) {
1107 				doutc(cl, "end of strip unit %p\n", page);
1108 				unlock_page(page);
1109 				break;
1110 			}
1111 			if (folio_test_writeback(folio) ||
1112 			    folio_test_private_2(folio) /* [DEPRECATED] */) {
1113 				if (wbc->sync_mode == WB_SYNC_NONE) {
1114 					doutc(cl, "%p under writeback\n", folio);
1115 					folio_unlock(folio);
1116 					continue;
1117 				}
1118 				doutc(cl, "waiting on writeback %p\n", folio);
1119 				folio_wait_writeback(folio);
1120 				folio_wait_private_2(folio); /* [DEPRECATED] */
1121 			}
1122 
1123 			if (!clear_page_dirty_for_io(page)) {
1124 				doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1125 				unlock_page(page);
1126 				continue;
1127 			}
1128 
1129 			/*
1130 			 * We have something to write.  If this is
1131 			 * the first locked page this time through,
1132 			 * calculate max possinle write size and
1133 			 * allocate a page array
1134 			 */
1135 			if (locked_pages == 0) {
1136 				u64 objnum;
1137 				u64 objoff;
1138 				u32 xlen;
1139 
1140 				/* prepare async write request */
1141 				offset = (u64)page_offset(page);
1142 				ceph_calc_file_object_mapping(&ci->i_layout,
1143 							      offset, wsize,
1144 							      &objnum, &objoff,
1145 							      &xlen);
1146 				len = xlen;
1147 
1148 				num_ops = 1;
1149 				strip_unit_end = page->index +
1150 					((len - 1) >> PAGE_SHIFT);
1151 
1152 				BUG_ON(pages);
1153 				max_pages = calc_pages_for(0, (u64)len);
1154 				pages = kmalloc_array(max_pages,
1155 						      sizeof(*pages),
1156 						      GFP_NOFS);
1157 				if (!pages) {
1158 					from_pool = true;
1159 					pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1160 					BUG_ON(!pages);
1161 				}
1162 
1163 				len = 0;
1164 			} else if (page->index !=
1165 				   (offset + len) >> PAGE_SHIFT) {
1166 				if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1167 							     CEPH_OSD_MAX_OPS)) {
1168 					redirty_page_for_writepage(wbc, page);
1169 					unlock_page(page);
1170 					break;
1171 				}
1172 
1173 				num_ops++;
1174 				offset = (u64)page_offset(page);
1175 				len = 0;
1176 			}
1177 
1178 			/* note position of first page in fbatch */
1179 			doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1180 			      ceph_vinop(inode), page, page->index);
1181 
1182 			if (atomic_long_inc_return(&fsc->writeback_count) >
1183 			    CONGESTION_ON_THRESH(
1184 				    fsc->mount_options->congestion_kb))
1185 				fsc->write_congested = true;
1186 
1187 			if (IS_ENCRYPTED(inode)) {
1188 				pages[locked_pages] =
1189 					fscrypt_encrypt_pagecache_blocks(page,
1190 						PAGE_SIZE, 0,
1191 						locked_pages ? GFP_NOWAIT : GFP_NOFS);
1192 				if (IS_ERR(pages[locked_pages])) {
1193 					if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1194 						pr_err_client(cl,
1195 							"inode->i_blkbits=%hhu\n",
1196 							inode->i_blkbits);
1197 					/* better not fail on first page! */
1198 					BUG_ON(locked_pages == 0);
1199 					pages[locked_pages] = NULL;
1200 					redirty_page_for_writepage(wbc, page);
1201 					unlock_page(page);
1202 					break;
1203 				}
1204 				++locked_pages;
1205 			} else {
1206 				pages[locked_pages++] = page;
1207 			}
1208 
1209 			fbatch.folios[i] = NULL;
1210 			len += thp_size(page);
1211 		}
1212 
1213 		/* did we get anything? */
1214 		if (!locked_pages)
1215 			goto release_folios;
1216 		if (i) {
1217 			unsigned j, n = 0;
1218 			/* shift unused page to beginning of fbatch */
1219 			for (j = 0; j < nr_folios; j++) {
1220 				if (!fbatch.folios[j])
1221 					continue;
1222 				if (n < j)
1223 					fbatch.folios[n] = fbatch.folios[j];
1224 				n++;
1225 			}
1226 			fbatch.nr = n;
1227 
1228 			if (nr_folios && i == nr_folios &&
1229 			    locked_pages < max_pages) {
1230 				doutc(cl, "reached end fbatch, trying for more\n");
1231 				folio_batch_release(&fbatch);
1232 				goto get_more_pages;
1233 			}
1234 		}
1235 
1236 new_request:
1237 		offset = ceph_fscrypt_page_offset(pages[0]);
1238 		len = wsize;
1239 
1240 		req = ceph_osdc_new_request(&fsc->client->osdc,
1241 					&ci->i_layout, vino,
1242 					offset, &len, 0, num_ops,
1243 					CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1244 					snapc, ceph_wbc.truncate_seq,
1245 					ceph_wbc.truncate_size, false);
1246 		if (IS_ERR(req)) {
1247 			req = ceph_osdc_new_request(&fsc->client->osdc,
1248 						&ci->i_layout, vino,
1249 						offset, &len, 0,
1250 						min(num_ops,
1251 						    CEPH_OSD_SLAB_OPS),
1252 						CEPH_OSD_OP_WRITE,
1253 						CEPH_OSD_FLAG_WRITE,
1254 						snapc, ceph_wbc.truncate_seq,
1255 						ceph_wbc.truncate_size, true);
1256 			BUG_ON(IS_ERR(req));
1257 		}
1258 		BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1259 			     thp_size(pages[locked_pages - 1]) - offset);
1260 
1261 		if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1262 			rc = -EIO;
1263 			goto release_folios;
1264 		}
1265 		req->r_callback = writepages_finish;
1266 		req->r_inode = inode;
1267 
1268 		/* Format the osd request message and submit the write */
1269 		len = 0;
1270 		data_pages = pages;
1271 		op_idx = 0;
1272 		for (i = 0; i < locked_pages; i++) {
1273 			struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1274 
1275 			u64 cur_offset = page_offset(page);
1276 			/*
1277 			 * Discontinuity in page range? Ceph can handle that by just passing
1278 			 * multiple extents in the write op.
1279 			 */
1280 			if (offset + len != cur_offset) {
1281 				/* If it's full, stop here */
1282 				if (op_idx + 1 == req->r_num_ops)
1283 					break;
1284 
1285 				/* Kick off an fscache write with what we have so far. */
1286 				ceph_fscache_write_to_cache(inode, offset, len, caching);
1287 
1288 				/* Start a new extent */
1289 				osd_req_op_extent_dup_last(req, op_idx,
1290 							   cur_offset - offset);
1291 				doutc(cl, "got pages at %llu~%llu\n", offset,
1292 				      len);
1293 				osd_req_op_extent_osd_data_pages(req, op_idx,
1294 							data_pages, len, 0,
1295 							from_pool, false);
1296 				osd_req_op_extent_update(req, op_idx, len);
1297 
1298 				len = 0;
1299 				offset = cur_offset;
1300 				data_pages = pages + i;
1301 				op_idx++;
1302 			}
1303 
1304 			set_page_writeback(page);
1305 			if (caching)
1306 				ceph_set_page_fscache(page);
1307 			len += thp_size(page);
1308 		}
1309 		ceph_fscache_write_to_cache(inode, offset, len, caching);
1310 
1311 		if (ceph_wbc.size_stable) {
1312 			len = min(len, ceph_wbc.i_size - offset);
1313 		} else if (i == locked_pages) {
1314 			/* writepages_finish() clears writeback pages
1315 			 * according to the data length, so make sure
1316 			 * data length covers all locked pages */
1317 			u64 min_len = len + 1 - thp_size(page);
1318 			len = get_writepages_data_length(inode, pages[i - 1],
1319 							 offset);
1320 			len = max(len, min_len);
1321 		}
1322 		if (IS_ENCRYPTED(inode))
1323 			len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1324 
1325 		doutc(cl, "got pages at %llu~%llu\n", offset, len);
1326 
1327 		if (IS_ENCRYPTED(inode) &&
1328 		    ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1329 			pr_warn_client(cl,
1330 				"bad encrypted write offset=%lld len=%llu\n",
1331 				offset, len);
1332 
1333 		osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1334 						 0, from_pool, false);
1335 		osd_req_op_extent_update(req, op_idx, len);
1336 
1337 		BUG_ON(op_idx + 1 != req->r_num_ops);
1338 
1339 		from_pool = false;
1340 		if (i < locked_pages) {
1341 			BUG_ON(num_ops <= req->r_num_ops);
1342 			num_ops -= req->r_num_ops;
1343 			locked_pages -= i;
1344 
1345 			/* allocate new pages array for next request */
1346 			data_pages = pages;
1347 			pages = kmalloc_array(locked_pages, sizeof(*pages),
1348 					      GFP_NOFS);
1349 			if (!pages) {
1350 				from_pool = true;
1351 				pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1352 				BUG_ON(!pages);
1353 			}
1354 			memcpy(pages, data_pages + i,
1355 			       locked_pages * sizeof(*pages));
1356 			memset(data_pages + i, 0,
1357 			       locked_pages * sizeof(*pages));
1358 		} else {
1359 			BUG_ON(num_ops != req->r_num_ops);
1360 			index = pages[i - 1]->index + 1;
1361 			/* request message now owns the pages array */
1362 			pages = NULL;
1363 		}
1364 
1365 		req->r_mtime = inode_get_mtime(inode);
1366 		ceph_osdc_start_request(&fsc->client->osdc, req);
1367 		req = NULL;
1368 
1369 		wbc->nr_to_write -= i;
1370 		if (pages)
1371 			goto new_request;
1372 
1373 		/*
1374 		 * We stop writing back only if we are not doing
1375 		 * integrity sync. In case of integrity sync we have to
1376 		 * keep going until we have written all the pages
1377 		 * we tagged for writeback prior to entering this loop.
1378 		 */
1379 		if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1380 			done = true;
1381 
1382 release_folios:
1383 		doutc(cl, "folio_batch release on %d folios (%p)\n",
1384 		      (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1385 		folio_batch_release(&fbatch);
1386 	}
1387 
1388 	if (should_loop && !done) {
1389 		/* more to do; loop back to beginning of file */
1390 		doutc(cl, "looping back to beginning of file\n");
1391 		end = start_index - 1; /* OK even when start_index == 0 */
1392 
1393 		/* to write dirty pages associated with next snapc,
1394 		 * we need to wait until current writes complete */
1395 		if (wbc->sync_mode != WB_SYNC_NONE &&
1396 		    start_index == 0 && /* all dirty pages were checked */
1397 		    !ceph_wbc.head_snapc) {
1398 			struct page *page;
1399 			unsigned i, nr;
1400 			index = 0;
1401 			while ((index <= end) &&
1402 			       (nr = filemap_get_folios_tag(mapping, &index,
1403 						(pgoff_t)-1,
1404 						PAGECACHE_TAG_WRITEBACK,
1405 						&fbatch))) {
1406 				for (i = 0; i < nr; i++) {
1407 					page = &fbatch.folios[i]->page;
1408 					if (page_snap_context(page) != snapc)
1409 						continue;
1410 					wait_on_page_writeback(page);
1411 				}
1412 				folio_batch_release(&fbatch);
1413 				cond_resched();
1414 			}
1415 		}
1416 
1417 		start_index = 0;
1418 		index = 0;
1419 		goto retry;
1420 	}
1421 
1422 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1423 		mapping->writeback_index = index;
1424 
1425 out:
1426 	ceph_osdc_put_request(req);
1427 	ceph_put_snap_context(last_snapc);
1428 	doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1429 	      rc);
1430 	return rc;
1431 }
1432 
1433 
1434 
1435 /*
1436  * See if a given @snapc is either writeable, or already written.
1437  */
context_is_writeable_or_written(struct inode * inode,struct ceph_snap_context * snapc)1438 static int context_is_writeable_or_written(struct inode *inode,
1439 					   struct ceph_snap_context *snapc)
1440 {
1441 	struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1442 	int ret = !oldest || snapc->seq <= oldest->seq;
1443 
1444 	ceph_put_snap_context(oldest);
1445 	return ret;
1446 }
1447 
1448 /**
1449  * ceph_find_incompatible - find an incompatible context and return it
1450  * @page: page being dirtied
1451  *
1452  * We are only allowed to write into/dirty a page if the page is
1453  * clean, or already dirty within the same snap context. Returns a
1454  * conflicting context if there is one, NULL if there isn't, or a
1455  * negative error code on other errors.
1456  *
1457  * Must be called with page lock held.
1458  */
1459 static struct ceph_snap_context *
ceph_find_incompatible(struct page * page)1460 ceph_find_incompatible(struct page *page)
1461 {
1462 	struct inode *inode = page->mapping->host;
1463 	struct ceph_client *cl = ceph_inode_to_client(inode);
1464 	struct ceph_inode_info *ci = ceph_inode(inode);
1465 
1466 	if (ceph_inode_is_shutdown(inode)) {
1467 		doutc(cl, " %llx.%llx page %p is shutdown\n",
1468 		      ceph_vinop(inode), page);
1469 		return ERR_PTR(-ESTALE);
1470 	}
1471 
1472 	for (;;) {
1473 		struct ceph_snap_context *snapc, *oldest;
1474 
1475 		wait_on_page_writeback(page);
1476 
1477 		snapc = page_snap_context(page);
1478 		if (!snapc || snapc == ci->i_head_snapc)
1479 			break;
1480 
1481 		/*
1482 		 * this page is already dirty in another (older) snap
1483 		 * context!  is it writeable now?
1484 		 */
1485 		oldest = get_oldest_context(inode, NULL, NULL);
1486 		if (snapc->seq > oldest->seq) {
1487 			/* not writeable -- return it for the caller to deal with */
1488 			ceph_put_snap_context(oldest);
1489 			doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1490 			      ceph_vinop(inode), page, snapc);
1491 			return ceph_get_snap_context(snapc);
1492 		}
1493 		ceph_put_snap_context(oldest);
1494 
1495 		/* yay, writeable, do it now (without dropping page lock) */
1496 		doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1497 		      ceph_vinop(inode), page, snapc);
1498 		if (clear_page_dirty_for_io(page)) {
1499 			int r = writepage_nounlock(page, NULL);
1500 			if (r < 0)
1501 				return ERR_PTR(r);
1502 		}
1503 	}
1504 	return NULL;
1505 }
1506 
ceph_netfs_check_write_begin(struct file * file,loff_t pos,unsigned int len,struct folio ** foliop,void ** _fsdata)1507 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1508 					struct folio **foliop, void **_fsdata)
1509 {
1510 	struct inode *inode = file_inode(file);
1511 	struct ceph_inode_info *ci = ceph_inode(inode);
1512 	struct ceph_snap_context *snapc;
1513 
1514 	snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1515 	if (snapc) {
1516 		int r;
1517 
1518 		folio_unlock(*foliop);
1519 		folio_put(*foliop);
1520 		*foliop = NULL;
1521 		if (IS_ERR(snapc))
1522 			return PTR_ERR(snapc);
1523 
1524 		ceph_queue_writeback(inode);
1525 		r = wait_event_killable(ci->i_cap_wq,
1526 					context_is_writeable_or_written(inode, snapc));
1527 		ceph_put_snap_context(snapc);
1528 		return r == 0 ? -EAGAIN : r;
1529 	}
1530 	return 0;
1531 }
1532 
1533 /*
1534  * We are only allowed to write into/dirty the page if the page is
1535  * clean, or already dirty within the same snap context.
1536  */
ceph_write_begin(struct file * file,struct address_space * mapping,loff_t pos,unsigned len,struct folio ** foliop,void ** fsdata)1537 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1538 			    loff_t pos, unsigned len,
1539 			    struct folio **foliop, void **fsdata)
1540 {
1541 	struct inode *inode = file_inode(file);
1542 	struct ceph_inode_info *ci = ceph_inode(inode);
1543 	int r;
1544 
1545 	r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, foliop, NULL);
1546 	if (r < 0)
1547 		return r;
1548 
1549 	folio_wait_private_2(*foliop); /* [DEPRECATED] */
1550 	WARN_ON_ONCE(!folio_test_locked(*foliop));
1551 	return 0;
1552 }
1553 
1554 /*
1555  * we don't do anything in here that simple_write_end doesn't do
1556  * except adjust dirty page accounting
1557  */
ceph_write_end(struct file * file,struct address_space * mapping,loff_t pos,unsigned len,unsigned copied,struct folio * folio,void * fsdata)1558 static int ceph_write_end(struct file *file, struct address_space *mapping,
1559 			  loff_t pos, unsigned len, unsigned copied,
1560 			  struct folio *folio, void *fsdata)
1561 {
1562 	struct inode *inode = file_inode(file);
1563 	struct ceph_client *cl = ceph_inode_to_client(inode);
1564 	bool check_cap = false;
1565 
1566 	doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1567 	      file, folio, (int)pos, (int)copied, (int)len);
1568 
1569 	if (!folio_test_uptodate(folio)) {
1570 		/* just return that nothing was copied on a short copy */
1571 		if (copied < len) {
1572 			copied = 0;
1573 			goto out;
1574 		}
1575 		folio_mark_uptodate(folio);
1576 	}
1577 
1578 	/* did file size increase? */
1579 	if (pos+copied > i_size_read(inode))
1580 		check_cap = ceph_inode_set_size(inode, pos+copied);
1581 
1582 	folio_mark_dirty(folio);
1583 
1584 out:
1585 	folio_unlock(folio);
1586 	folio_put(folio);
1587 
1588 	if (check_cap)
1589 		ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1590 
1591 	return copied;
1592 }
1593 
1594 const struct address_space_operations ceph_aops = {
1595 	.read_folio = netfs_read_folio,
1596 	.readahead = netfs_readahead,
1597 	.writepage = ceph_writepage,
1598 	.writepages = ceph_writepages_start,
1599 	.write_begin = ceph_write_begin,
1600 	.write_end = ceph_write_end,
1601 	.dirty_folio = ceph_dirty_folio,
1602 	.invalidate_folio = ceph_invalidate_folio,
1603 	.release_folio = netfs_release_folio,
1604 	.direct_IO = noop_direct_IO,
1605 };
1606 
ceph_block_sigs(sigset_t * oldset)1607 static void ceph_block_sigs(sigset_t *oldset)
1608 {
1609 	sigset_t mask;
1610 	siginitsetinv(&mask, sigmask(SIGKILL));
1611 	sigprocmask(SIG_BLOCK, &mask, oldset);
1612 }
1613 
ceph_restore_sigs(sigset_t * oldset)1614 static void ceph_restore_sigs(sigset_t *oldset)
1615 {
1616 	sigprocmask(SIG_SETMASK, oldset, NULL);
1617 }
1618 
1619 /*
1620  * vm ops
1621  */
ceph_filemap_fault(struct vm_fault * vmf)1622 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1623 {
1624 	struct vm_area_struct *vma = vmf->vma;
1625 	struct inode *inode = file_inode(vma->vm_file);
1626 	struct ceph_inode_info *ci = ceph_inode(inode);
1627 	struct ceph_client *cl = ceph_inode_to_client(inode);
1628 	struct ceph_file_info *fi = vma->vm_file->private_data;
1629 	loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1630 	int want, got, err;
1631 	sigset_t oldset;
1632 	vm_fault_t ret = VM_FAULT_SIGBUS;
1633 
1634 	if (ceph_inode_is_shutdown(inode))
1635 		return ret;
1636 
1637 	ceph_block_sigs(&oldset);
1638 
1639 	doutc(cl, "%llx.%llx %llu trying to get caps\n",
1640 	      ceph_vinop(inode), off);
1641 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1642 		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1643 	else
1644 		want = CEPH_CAP_FILE_CACHE;
1645 
1646 	got = 0;
1647 	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1648 	if (err < 0)
1649 		goto out_restore;
1650 
1651 	doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1652 	      off, ceph_cap_string(got));
1653 
1654 	if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1655 	    !ceph_has_inline_data(ci)) {
1656 		CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1657 		ceph_add_rw_context(fi, &rw_ctx);
1658 		ret = filemap_fault(vmf);
1659 		ceph_del_rw_context(fi, &rw_ctx);
1660 		doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1661 		      ceph_vinop(inode), off, ceph_cap_string(got), ret);
1662 	} else
1663 		err = -EAGAIN;
1664 
1665 	ceph_put_cap_refs(ci, got);
1666 
1667 	if (err != -EAGAIN)
1668 		goto out_restore;
1669 
1670 	/* read inline data */
1671 	if (off >= PAGE_SIZE) {
1672 		/* does not support inline data > PAGE_SIZE */
1673 		ret = VM_FAULT_SIGBUS;
1674 	} else {
1675 		struct address_space *mapping = inode->i_mapping;
1676 		struct page *page;
1677 
1678 		filemap_invalidate_lock_shared(mapping);
1679 		page = find_or_create_page(mapping, 0,
1680 				mapping_gfp_constraint(mapping, ~__GFP_FS));
1681 		if (!page) {
1682 			ret = VM_FAULT_OOM;
1683 			goto out_inline;
1684 		}
1685 		err = __ceph_do_getattr(inode, page,
1686 					 CEPH_STAT_CAP_INLINE_DATA, true);
1687 		if (err < 0 || off >= i_size_read(inode)) {
1688 			unlock_page(page);
1689 			put_page(page);
1690 			ret = vmf_error(err);
1691 			goto out_inline;
1692 		}
1693 		if (err < PAGE_SIZE)
1694 			zero_user_segment(page, err, PAGE_SIZE);
1695 		else
1696 			flush_dcache_page(page);
1697 		SetPageUptodate(page);
1698 		vmf->page = page;
1699 		ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1700 out_inline:
1701 		filemap_invalidate_unlock_shared(mapping);
1702 		doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1703 		      ceph_vinop(inode), off, ret);
1704 	}
1705 out_restore:
1706 	ceph_restore_sigs(&oldset);
1707 	if (err < 0)
1708 		ret = vmf_error(err);
1709 
1710 	return ret;
1711 }
1712 
ceph_page_mkwrite(struct vm_fault * vmf)1713 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1714 {
1715 	struct vm_area_struct *vma = vmf->vma;
1716 	struct inode *inode = file_inode(vma->vm_file);
1717 	struct ceph_client *cl = ceph_inode_to_client(inode);
1718 	struct ceph_inode_info *ci = ceph_inode(inode);
1719 	struct ceph_file_info *fi = vma->vm_file->private_data;
1720 	struct ceph_cap_flush *prealloc_cf;
1721 	struct page *page = vmf->page;
1722 	loff_t off = page_offset(page);
1723 	loff_t size = i_size_read(inode);
1724 	size_t len;
1725 	int want, got, err;
1726 	sigset_t oldset;
1727 	vm_fault_t ret = VM_FAULT_SIGBUS;
1728 
1729 	if (ceph_inode_is_shutdown(inode))
1730 		return ret;
1731 
1732 	prealloc_cf = ceph_alloc_cap_flush();
1733 	if (!prealloc_cf)
1734 		return VM_FAULT_OOM;
1735 
1736 	sb_start_pagefault(inode->i_sb);
1737 	ceph_block_sigs(&oldset);
1738 
1739 	if (off + thp_size(page) <= size)
1740 		len = thp_size(page);
1741 	else
1742 		len = offset_in_thp(page, size);
1743 
1744 	doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1745 	      ceph_vinop(inode), off, len, size);
1746 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1747 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1748 	else
1749 		want = CEPH_CAP_FILE_BUFFER;
1750 
1751 	got = 0;
1752 	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1753 	if (err < 0)
1754 		goto out_free;
1755 
1756 	doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1757 	      off, len, ceph_cap_string(got));
1758 
1759 	/* Update time before taking page lock */
1760 	file_update_time(vma->vm_file);
1761 	inode_inc_iversion_raw(inode);
1762 
1763 	do {
1764 		struct ceph_snap_context *snapc;
1765 
1766 		lock_page(page);
1767 
1768 		if (page_mkwrite_check_truncate(page, inode) < 0) {
1769 			unlock_page(page);
1770 			ret = VM_FAULT_NOPAGE;
1771 			break;
1772 		}
1773 
1774 		snapc = ceph_find_incompatible(page);
1775 		if (!snapc) {
1776 			/* success.  we'll keep the page locked. */
1777 			set_page_dirty(page);
1778 			ret = VM_FAULT_LOCKED;
1779 			break;
1780 		}
1781 
1782 		unlock_page(page);
1783 
1784 		if (IS_ERR(snapc)) {
1785 			ret = VM_FAULT_SIGBUS;
1786 			break;
1787 		}
1788 
1789 		ceph_queue_writeback(inode);
1790 		err = wait_event_killable(ci->i_cap_wq,
1791 				context_is_writeable_or_written(inode, snapc));
1792 		ceph_put_snap_context(snapc);
1793 	} while (err == 0);
1794 
1795 	if (ret == VM_FAULT_LOCKED) {
1796 		int dirty;
1797 		spin_lock(&ci->i_ceph_lock);
1798 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1799 					       &prealloc_cf);
1800 		spin_unlock(&ci->i_ceph_lock);
1801 		if (dirty)
1802 			__mark_inode_dirty(inode, dirty);
1803 	}
1804 
1805 	doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1806 	      ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1807 	ceph_put_cap_refs_async(ci, got);
1808 out_free:
1809 	ceph_restore_sigs(&oldset);
1810 	sb_end_pagefault(inode->i_sb);
1811 	ceph_free_cap_flush(prealloc_cf);
1812 	if (err < 0)
1813 		ret = vmf_error(err);
1814 	return ret;
1815 }
1816 
ceph_fill_inline_data(struct inode * inode,struct page * locked_page,char * data,size_t len)1817 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1818 			   char	*data, size_t len)
1819 {
1820 	struct ceph_client *cl = ceph_inode_to_client(inode);
1821 	struct address_space *mapping = inode->i_mapping;
1822 	struct page *page;
1823 
1824 	if (locked_page) {
1825 		page = locked_page;
1826 	} else {
1827 		if (i_size_read(inode) == 0)
1828 			return;
1829 		page = find_or_create_page(mapping, 0,
1830 					   mapping_gfp_constraint(mapping,
1831 					   ~__GFP_FS));
1832 		if (!page)
1833 			return;
1834 		if (PageUptodate(page)) {
1835 			unlock_page(page);
1836 			put_page(page);
1837 			return;
1838 		}
1839 	}
1840 
1841 	doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1842 	      ceph_vinop(inode), len, locked_page);
1843 
1844 	if (len > 0) {
1845 		void *kaddr = kmap_atomic(page);
1846 		memcpy(kaddr, data, len);
1847 		kunmap_atomic(kaddr);
1848 	}
1849 
1850 	if (page != locked_page) {
1851 		if (len < PAGE_SIZE)
1852 			zero_user_segment(page, len, PAGE_SIZE);
1853 		else
1854 			flush_dcache_page(page);
1855 
1856 		SetPageUptodate(page);
1857 		unlock_page(page);
1858 		put_page(page);
1859 	}
1860 }
1861 
ceph_uninline_data(struct file * file)1862 int ceph_uninline_data(struct file *file)
1863 {
1864 	struct inode *inode = file_inode(file);
1865 	struct ceph_inode_info *ci = ceph_inode(inode);
1866 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1867 	struct ceph_client *cl = fsc->client;
1868 	struct ceph_osd_request *req = NULL;
1869 	struct ceph_cap_flush *prealloc_cf = NULL;
1870 	struct folio *folio = NULL;
1871 	u64 inline_version = CEPH_INLINE_NONE;
1872 	struct page *pages[1];
1873 	int err = 0;
1874 	u64 len;
1875 
1876 	spin_lock(&ci->i_ceph_lock);
1877 	inline_version = ci->i_inline_version;
1878 	spin_unlock(&ci->i_ceph_lock);
1879 
1880 	doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1881 	      inline_version);
1882 
1883 	if (ceph_inode_is_shutdown(inode)) {
1884 		err = -EIO;
1885 		goto out;
1886 	}
1887 
1888 	if (inline_version == CEPH_INLINE_NONE)
1889 		return 0;
1890 
1891 	prealloc_cf = ceph_alloc_cap_flush();
1892 	if (!prealloc_cf)
1893 		return -ENOMEM;
1894 
1895 	if (inline_version == 1) /* initial version, no data */
1896 		goto out_uninline;
1897 
1898 	folio = read_mapping_folio(inode->i_mapping, 0, file);
1899 	if (IS_ERR(folio)) {
1900 		err = PTR_ERR(folio);
1901 		goto out;
1902 	}
1903 
1904 	folio_lock(folio);
1905 
1906 	len = i_size_read(inode);
1907 	if (len > folio_size(folio))
1908 		len = folio_size(folio);
1909 
1910 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1911 				    ceph_vino(inode), 0, &len, 0, 1,
1912 				    CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1913 				    NULL, 0, 0, false);
1914 	if (IS_ERR(req)) {
1915 		err = PTR_ERR(req);
1916 		goto out_unlock;
1917 	}
1918 
1919 	req->r_mtime = inode_get_mtime(inode);
1920 	ceph_osdc_start_request(&fsc->client->osdc, req);
1921 	err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1922 	ceph_osdc_put_request(req);
1923 	if (err < 0)
1924 		goto out_unlock;
1925 
1926 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1927 				    ceph_vino(inode), 0, &len, 1, 3,
1928 				    CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1929 				    NULL, ci->i_truncate_seq,
1930 				    ci->i_truncate_size, false);
1931 	if (IS_ERR(req)) {
1932 		err = PTR_ERR(req);
1933 		goto out_unlock;
1934 	}
1935 
1936 	pages[0] = folio_page(folio, 0);
1937 	osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1938 
1939 	{
1940 		__le64 xattr_buf = cpu_to_le64(inline_version);
1941 		err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1942 					    "inline_version", &xattr_buf,
1943 					    sizeof(xattr_buf),
1944 					    CEPH_OSD_CMPXATTR_OP_GT,
1945 					    CEPH_OSD_CMPXATTR_MODE_U64);
1946 		if (err)
1947 			goto out_put_req;
1948 	}
1949 
1950 	{
1951 		char xattr_buf[32];
1952 		int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1953 					 "%llu", inline_version);
1954 		err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1955 					    "inline_version",
1956 					    xattr_buf, xattr_len, 0, 0);
1957 		if (err)
1958 			goto out_put_req;
1959 	}
1960 
1961 	req->r_mtime = inode_get_mtime(inode);
1962 	ceph_osdc_start_request(&fsc->client->osdc, req);
1963 	err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1964 
1965 	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1966 				  req->r_end_latency, len, err);
1967 
1968 out_uninline:
1969 	if (!err) {
1970 		int dirty;
1971 
1972 		/* Set to CAP_INLINE_NONE and dirty the caps */
1973 		down_read(&fsc->mdsc->snap_rwsem);
1974 		spin_lock(&ci->i_ceph_lock);
1975 		ci->i_inline_version = CEPH_INLINE_NONE;
1976 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1977 		spin_unlock(&ci->i_ceph_lock);
1978 		up_read(&fsc->mdsc->snap_rwsem);
1979 		if (dirty)
1980 			__mark_inode_dirty(inode, dirty);
1981 	}
1982 out_put_req:
1983 	ceph_osdc_put_request(req);
1984 	if (err == -ECANCELED)
1985 		err = 0;
1986 out_unlock:
1987 	if (folio) {
1988 		folio_unlock(folio);
1989 		folio_put(folio);
1990 	}
1991 out:
1992 	ceph_free_cap_flush(prealloc_cf);
1993 	doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1994 	      ceph_vinop(inode), inline_version, err);
1995 	return err;
1996 }
1997 
1998 static const struct vm_operations_struct ceph_vmops = {
1999 	.fault		= ceph_filemap_fault,
2000 	.page_mkwrite	= ceph_page_mkwrite,
2001 };
2002 
ceph_mmap(struct file * file,struct vm_area_struct * vma)2003 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
2004 {
2005 	struct address_space *mapping = file->f_mapping;
2006 
2007 	if (!mapping->a_ops->read_folio)
2008 		return -ENOEXEC;
2009 	vma->vm_ops = &ceph_vmops;
2010 	return 0;
2011 }
2012 
2013 enum {
2014 	POOL_READ	= 1,
2015 	POOL_WRITE	= 2,
2016 };
2017 
__ceph_pool_perm_get(struct ceph_inode_info * ci,s64 pool,struct ceph_string * pool_ns)2018 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
2019 				s64 pool, struct ceph_string *pool_ns)
2020 {
2021 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
2022 	struct ceph_mds_client *mdsc = fsc->mdsc;
2023 	struct ceph_client *cl = fsc->client;
2024 	struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
2025 	struct rb_node **p, *parent;
2026 	struct ceph_pool_perm *perm;
2027 	struct page **pages;
2028 	size_t pool_ns_len;
2029 	int err = 0, err2 = 0, have = 0;
2030 
2031 	down_read(&mdsc->pool_perm_rwsem);
2032 	p = &mdsc->pool_perm_tree.rb_node;
2033 	while (*p) {
2034 		perm = rb_entry(*p, struct ceph_pool_perm, node);
2035 		if (pool < perm->pool)
2036 			p = &(*p)->rb_left;
2037 		else if (pool > perm->pool)
2038 			p = &(*p)->rb_right;
2039 		else {
2040 			int ret = ceph_compare_string(pool_ns,
2041 						perm->pool_ns,
2042 						perm->pool_ns_len);
2043 			if (ret < 0)
2044 				p = &(*p)->rb_left;
2045 			else if (ret > 0)
2046 				p = &(*p)->rb_right;
2047 			else {
2048 				have = perm->perm;
2049 				break;
2050 			}
2051 		}
2052 	}
2053 	up_read(&mdsc->pool_perm_rwsem);
2054 	if (*p)
2055 		goto out;
2056 
2057 	if (pool_ns)
2058 		doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2059 		      (int)pool_ns->len, pool_ns->str);
2060 	else
2061 		doutc(cl, "pool %lld no perm cached\n", pool);
2062 
2063 	down_write(&mdsc->pool_perm_rwsem);
2064 	p = &mdsc->pool_perm_tree.rb_node;
2065 	parent = NULL;
2066 	while (*p) {
2067 		parent = *p;
2068 		perm = rb_entry(parent, struct ceph_pool_perm, node);
2069 		if (pool < perm->pool)
2070 			p = &(*p)->rb_left;
2071 		else if (pool > perm->pool)
2072 			p = &(*p)->rb_right;
2073 		else {
2074 			int ret = ceph_compare_string(pool_ns,
2075 						perm->pool_ns,
2076 						perm->pool_ns_len);
2077 			if (ret < 0)
2078 				p = &(*p)->rb_left;
2079 			else if (ret > 0)
2080 				p = &(*p)->rb_right;
2081 			else {
2082 				have = perm->perm;
2083 				break;
2084 			}
2085 		}
2086 	}
2087 	if (*p) {
2088 		up_write(&mdsc->pool_perm_rwsem);
2089 		goto out;
2090 	}
2091 
2092 	rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2093 					 1, false, GFP_NOFS);
2094 	if (!rd_req) {
2095 		err = -ENOMEM;
2096 		goto out_unlock;
2097 	}
2098 
2099 	rd_req->r_flags = CEPH_OSD_FLAG_READ;
2100 	osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2101 	rd_req->r_base_oloc.pool = pool;
2102 	if (pool_ns)
2103 		rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2104 	ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2105 
2106 	err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2107 	if (err)
2108 		goto out_unlock;
2109 
2110 	wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2111 					 1, false, GFP_NOFS);
2112 	if (!wr_req) {
2113 		err = -ENOMEM;
2114 		goto out_unlock;
2115 	}
2116 
2117 	wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2118 	osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2119 	ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2120 	ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2121 
2122 	err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2123 	if (err)
2124 		goto out_unlock;
2125 
2126 	/* one page should be large enough for STAT data */
2127 	pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2128 	if (IS_ERR(pages)) {
2129 		err = PTR_ERR(pages);
2130 		goto out_unlock;
2131 	}
2132 
2133 	osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2134 				     0, false, true);
2135 	ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2136 
2137 	wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2138 	ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2139 
2140 	err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2141 	err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2142 
2143 	if (err >= 0 || err == -ENOENT)
2144 		have |= POOL_READ;
2145 	else if (err != -EPERM) {
2146 		if (err == -EBLOCKLISTED)
2147 			fsc->blocklisted = true;
2148 		goto out_unlock;
2149 	}
2150 
2151 	if (err2 == 0 || err2 == -EEXIST)
2152 		have |= POOL_WRITE;
2153 	else if (err2 != -EPERM) {
2154 		if (err2 == -EBLOCKLISTED)
2155 			fsc->blocklisted = true;
2156 		err = err2;
2157 		goto out_unlock;
2158 	}
2159 
2160 	pool_ns_len = pool_ns ? pool_ns->len : 0;
2161 	perm = kmalloc(struct_size(perm, pool_ns, pool_ns_len + 1), GFP_NOFS);
2162 	if (!perm) {
2163 		err = -ENOMEM;
2164 		goto out_unlock;
2165 	}
2166 
2167 	perm->pool = pool;
2168 	perm->perm = have;
2169 	perm->pool_ns_len = pool_ns_len;
2170 	if (pool_ns_len > 0)
2171 		memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2172 	perm->pool_ns[pool_ns_len] = 0;
2173 
2174 	rb_link_node(&perm->node, parent, p);
2175 	rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2176 	err = 0;
2177 out_unlock:
2178 	up_write(&mdsc->pool_perm_rwsem);
2179 
2180 	ceph_osdc_put_request(rd_req);
2181 	ceph_osdc_put_request(wr_req);
2182 out:
2183 	if (!err)
2184 		err = have;
2185 	if (pool_ns)
2186 		doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2187 		      (int)pool_ns->len, pool_ns->str, err);
2188 	else
2189 		doutc(cl, "pool %lld result = %d\n", pool, err);
2190 	return err;
2191 }
2192 
ceph_pool_perm_check(struct inode * inode,int need)2193 int ceph_pool_perm_check(struct inode *inode, int need)
2194 {
2195 	struct ceph_client *cl = ceph_inode_to_client(inode);
2196 	struct ceph_inode_info *ci = ceph_inode(inode);
2197 	struct ceph_string *pool_ns;
2198 	s64 pool;
2199 	int ret, flags;
2200 
2201 	/* Only need to do this for regular files */
2202 	if (!S_ISREG(inode->i_mode))
2203 		return 0;
2204 
2205 	if (ci->i_vino.snap != CEPH_NOSNAP) {
2206 		/*
2207 		 * Pool permission check needs to write to the first object.
2208 		 * But for snapshot, head of the first object may have already
2209 		 * been deleted. Skip check to avoid creating orphan object.
2210 		 */
2211 		return 0;
2212 	}
2213 
2214 	if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2215 				NOPOOLPERM))
2216 		return 0;
2217 
2218 	spin_lock(&ci->i_ceph_lock);
2219 	flags = ci->i_ceph_flags;
2220 	pool = ci->i_layout.pool_id;
2221 	spin_unlock(&ci->i_ceph_lock);
2222 check:
2223 	if (flags & CEPH_I_POOL_PERM) {
2224 		if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2225 			doutc(cl, "pool %lld no read perm\n", pool);
2226 			return -EPERM;
2227 		}
2228 		if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2229 			doutc(cl, "pool %lld no write perm\n", pool);
2230 			return -EPERM;
2231 		}
2232 		return 0;
2233 	}
2234 
2235 	pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2236 	ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2237 	ceph_put_string(pool_ns);
2238 	if (ret < 0)
2239 		return ret;
2240 
2241 	flags = CEPH_I_POOL_PERM;
2242 	if (ret & POOL_READ)
2243 		flags |= CEPH_I_POOL_RD;
2244 	if (ret & POOL_WRITE)
2245 		flags |= CEPH_I_POOL_WR;
2246 
2247 	spin_lock(&ci->i_ceph_lock);
2248 	if (pool == ci->i_layout.pool_id &&
2249 	    pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2250 		ci->i_ceph_flags |= flags;
2251         } else {
2252 		pool = ci->i_layout.pool_id;
2253 		flags = ci->i_ceph_flags;
2254 	}
2255 	spin_unlock(&ci->i_ceph_lock);
2256 	goto check;
2257 }
2258 
ceph_pool_perm_destroy(struct ceph_mds_client * mdsc)2259 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2260 {
2261 	struct ceph_pool_perm *perm;
2262 	struct rb_node *n;
2263 
2264 	while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2265 		n = rb_first(&mdsc->pool_perm_tree);
2266 		perm = rb_entry(n, struct ceph_pool_perm, node);
2267 		rb_erase(n, &mdsc->pool_perm_tree);
2268 		kfree(perm);
2269 	}
2270 }
2271