1#include <linux/ceph/ceph_debug.h>
  2
  3#include <linux/module.h>
  4#include <linux/sched.h>
  5#include <linux/slab.h>
  6#include <linux/file.h>
  7#include <linux/namei.h>
  8#include <linux/writeback.h>
  9
 10#include "super.h"
 11#include "mds_client.h"
 12
 13/*
 14 * Ceph file operations
 15 *
 16 * Implement basic open/close functionality, and implement
 17 * read/write.
 18 *
 19 * We implement three modes of file I/O:
 20 *  - buffered uses the generic_file_aio_{read,write} helpers
 21 *
 22 *  - synchronous is used when there is multi-client read/write
 23 *    sharing, avoids the page cache, and synchronously waits for an
 24 *    ack from the OSD.
 25 *
 26 *  - direct io takes the variant of the sync path that references
 27 *    user pages directly.
 28 *
 29 * fsync() flushes and waits on dirty pages, but just queues metadata
 30 * for writeback: since the MDS can recover size and mtime there is no
 31 * need to wait for MDS acknowledgement.
 32 */
 33
 34
 35/*
 36 * Prepare an open request.  Preallocate ceph_cap to avoid an
 37 * inopportune ENOMEM later.
 38 */
 39static struct ceph_mds_request *
 40prepare_open_request(struct super_block *sb, int flags, int create_mode)
 41{
 42	struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
 43	struct ceph_mds_client *mdsc = fsc->mdsc;
 44	struct ceph_mds_request *req;
 45	int want_auth = USE_ANY_MDS;
 46	int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 47
 48	if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 49		want_auth = USE_AUTH_MDS;
 50
 51	req = ceph_mdsc_create_request(mdsc, op, want_auth);
 52	if (IS_ERR(req))
 53		goto out;
 54	req->r_fmode = ceph_flags_to_mode(flags);
 55	req->r_args.open.flags = cpu_to_le32(flags);
 56	req->r_args.open.mode = cpu_to_le32(create_mode);
 57	req->r_args.open.preferred = cpu_to_le32(-1);
 58out:
 59	return req;
 60}
 61
 62/*
 63 * initialize private struct file data.
 64 * if we fail, clean up by dropping fmode reference on the ceph_inode
 65 */
 66static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 67{
 68	struct ceph_file_info *cf;
 69	int ret = 0;
 70
 71	switch (inode->i_mode & S_IFMT) {
 72	case S_IFREG:
 73	case S_IFDIR:
 74		dout("init_file %p %p 0%o (regular)\n", inode, file,
 75		     inode->i_mode);
 76		cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO);
 77		if (cf == NULL) {
 78			ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 79			return -ENOMEM;
 80		}
 81		cf->fmode = fmode;
 82		cf->next_offset = 2;
 83		file->private_data = cf;
 84		BUG_ON(inode->i_fop->release != ceph_release);
 85		break;
 86
 87	case S_IFLNK:
 88		dout("init_file %p %p 0%o (symlink)\n", inode, file,
 89		     inode->i_mode);
 90		ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 91		break;
 92
 93	default:
 94		dout("init_file %p %p 0%o (special)\n", inode, file,
 95		     inode->i_mode);
 96		/*
 97		 * we need to drop the open ref now, since we don't
 98		 * have .release set to ceph_release.
 99		 */
100		ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
101		BUG_ON(inode->i_fop->release == ceph_release);
102
103		/* call the proper open fop */
104		ret = inode->i_fop->open(inode, file);
105	}
106	return ret;
107}
108
109/*
110 * If the filp already has private_data, that means the file was
111 * already opened by intent during lookup, and we do nothing.
112 *
113 * If we already have the requisite capabilities, we can satisfy
114 * the open request locally (no need to request new caps from the
115 * MDS).  We do, however, need to inform the MDS (asynchronously)
116 * if our wanted caps set expands.
117 */
118int ceph_open(struct inode *inode, struct file *file)
119{
120	struct ceph_inode_info *ci = ceph_inode(inode);
121	struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
122	struct ceph_mds_client *mdsc = fsc->mdsc;
123	struct ceph_mds_request *req;
124	struct ceph_file_info *cf = file->private_data;
125	struct inode *parent_inode = NULL;
126	int err;
127	int flags, fmode, wanted;
128
129	if (cf) {
130		dout("open file %p is already opened\n", file);
131		return 0;
132	}
133
134	/* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
135	flags = file->f_flags & ~(O_CREAT|O_EXCL);
136	if (S_ISDIR(inode->i_mode))
137		flags = O_DIRECTORY;  /* mds likes to know */
138
139	dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
140	     ceph_vinop(inode), file, flags, file->f_flags);
141	fmode = ceph_flags_to_mode(flags);
142	wanted = ceph_caps_for_mode(fmode);
143
144	/* snapped files are read-only */
145	if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
146		return -EROFS;
147
148	/* trivially open snapdir */
149	if (ceph_snap(inode) == CEPH_SNAPDIR) {
150		spin_lock(&inode->i_lock);
151		__ceph_get_fmode(ci, fmode);
152		spin_unlock(&inode->i_lock);
153		return ceph_init_file(inode, file, fmode);
154	}
155
156	/*
157	 * No need to block if we have caps on the auth MDS (for
158	 * write) or any MDS (for read).  Update wanted set
159	 * asynchronously.
160	 */
161	spin_lock(&inode->i_lock);
162	if (__ceph_is_any_real_caps(ci) &&
163	    (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
164		int mds_wanted = __ceph_caps_mds_wanted(ci);
165		int issued = __ceph_caps_issued(ci, NULL);
166
167		dout("open %p fmode %d want %s issued %s using existing\n",
168		     inode, fmode, ceph_cap_string(wanted),
169		     ceph_cap_string(issued));
170		__ceph_get_fmode(ci, fmode);
171		spin_unlock(&inode->i_lock);
172
173		/* adjust wanted? */
174		if ((issued & wanted) != wanted &&
175		    (mds_wanted & wanted) != wanted &&
176		    ceph_snap(inode) != CEPH_SNAPDIR)
177			ceph_check_caps(ci, 0, NULL);
178
179		return ceph_init_file(inode, file, fmode);
180	} else if (ceph_snap(inode) != CEPH_NOSNAP &&
181		   (ci->i_snap_caps & wanted) == wanted) {
182		__ceph_get_fmode(ci, fmode);
183		spin_unlock(&inode->i_lock);
184		return ceph_init_file(inode, file, fmode);
185	}
186	spin_unlock(&inode->i_lock);
187
188	dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
189	req = prepare_open_request(inode->i_sb, flags, 0);
190	if (IS_ERR(req)) {
191		err = PTR_ERR(req);
192		goto out;
193	}
194	req->r_inode = inode;
195	ihold(inode);
196	req->r_num_caps = 1;
197	if (flags & (O_CREAT|O_TRUNC))
198		parent_inode = ceph_get_dentry_parent_inode(file->f_dentry);
199	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
200	iput(parent_inode);
201	if (!err)
202		err = ceph_init_file(inode, file, req->r_fmode);
203	ceph_mdsc_put_request(req);
204	dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
205out:
206	return err;
207}
208
209
210/*
211 * Do a lookup + open with a single request.
212 *
213 * If this succeeds, but some subsequent check in the vfs
214 * may_open() fails, the struct *file gets cleaned up (i.e.
215 * ceph_release gets called).  So fear not!
216 */
217/*
218 * flags
219 *  path_lookup_open   -> LOOKUP_OPEN
220 *  path_lookup_create -> LOOKUP_OPEN|LOOKUP_CREATE
221 */
222struct dentry *ceph_lookup_open(struct inode *dir, struct dentry *dentry,
223				struct nameidata *nd, int mode,
224				int locked_dir)
225{
226	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
227	struct ceph_mds_client *mdsc = fsc->mdsc;
228	struct file *file;
229	struct ceph_mds_request *req;
230	struct dentry *ret;
231	int err;
232	int flags = nd->intent.open.flags;
233
234	dout("ceph_lookup_open dentry %p '%.*s' flags %d mode 0%o\n",
235	     dentry, dentry->d_name.len, dentry->d_name.name, flags, mode);
236
237	/* do the open */
238	req = prepare_open_request(dir->i_sb, flags, mode);
239	if (IS_ERR(req))
240		return ERR_CAST(req);
241	req->r_dentry = dget(dentry);
242	req->r_num_caps = 2;
243	if (flags & O_CREAT) {
244		req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
245		req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
246	}
247	req->r_locked_dir = dir;           /* caller holds dir->i_mutex */
248	err = ceph_mdsc_do_request(mdsc,
249				   (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
250				   req);
251	err = ceph_handle_snapdir(req, dentry, err);
252	if (err)
253		goto out;
254	if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
255		err = ceph_handle_notrace_create(dir, dentry);
256	if (err)
257		goto out;
258	file = lookup_instantiate_filp(nd, req->r_dentry, ceph_open);
259	if (IS_ERR(file))
260		err = PTR_ERR(file);
261out:
262	ret = ceph_finish_lookup(req, dentry, err);
263	ceph_mdsc_put_request(req);
264	dout("ceph_lookup_open result=%p\n", ret);
265	return ret;
266}
267
268int ceph_release(struct inode *inode, struct file *file)
269{
270	struct ceph_inode_info *ci = ceph_inode(inode);
271	struct ceph_file_info *cf = file->private_data;
272
273	dout("release inode %p file %p\n", inode, file);
274	ceph_put_fmode(ci, cf->fmode);
275	if (cf->last_readdir)
276		ceph_mdsc_put_request(cf->last_readdir);
277	kfree(cf->last_name);
278	kfree(cf->dir_info);
279	dput(cf->dentry);
280	kmem_cache_free(ceph_file_cachep, cf);
281
282	/* wake up anyone waiting for caps on this inode */
283	wake_up_all(&ci->i_cap_wq);
284	return 0;
285}
286
287/*
288 * Read a range of bytes striped over one or more objects.  Iterate over
289 * objects we stripe over.  (That's not atomic, but good enough for now.)
290 *
291 * If we get a short result from the OSD, check against i_size; we need to
292 * only return a short read to the caller if we hit EOF.
293 */
294static int striped_read(struct inode *inode,
295			u64 off, u64 len,
296			struct page **pages, int num_pages,
297			int *checkeof, bool o_direct,
298			unsigned long buf_align)
299{
300	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
301	struct ceph_inode_info *ci = ceph_inode(inode);
302	u64 pos, this_len;
303	int io_align, page_align;
304	int left, pages_left;
305	int read;
306	struct page **page_pos;
307	int ret;
308	bool hit_stripe, was_short;
309
310	/*
311	 * we may need to do multiple reads.  not atomic, unfortunately.
312	 */
313	pos = off;
314	left = len;
315	page_pos = pages;
316	pages_left = num_pages;
317	read = 0;
318	io_align = off & ~PAGE_MASK;
319
320more:
321	if (o_direct)
322		page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
323	else
324		page_align = pos & ~PAGE_MASK;
325	this_len = left;
326	ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
327				  &ci->i_layout, pos, &this_len,
328				  ci->i_truncate_seq,
329				  ci->i_truncate_size,
330				  page_pos, pages_left, page_align);
331	if (ret == -ENOENT)
332		ret = 0;
333	hit_stripe = this_len < left;
334	was_short = ret >= 0 && ret < this_len;
335	dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read,
336	     ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
337
338	if (ret > 0) {
339		int didpages = (page_align + ret) >> PAGE_CACHE_SHIFT;
340
341		if (read < pos - off) {
342			dout(" zero gap %llu to %llu\n", off + read, pos);
343			ceph_zero_page_vector_range(page_align + read,
344						    pos - off - read, pages);
345		}
346		pos += ret;
347		read = pos - off;
348		left -= ret;
349		page_pos += didpages;
350		pages_left -= didpages;
351
352		/* hit stripe? */
353		if (left && hit_stripe)
354			goto more;
355	}
356
357	if (was_short) {
358		/* did we bounce off eof? */
359		if (pos + left > inode->i_size)
360			*checkeof = 1;
361
362		/* zero trailing bytes (inside i_size) */
363		if (left > 0 && pos < inode->i_size) {
364			if (pos + left > inode->i_size)
365				left = inode->i_size - pos;
366
367			dout("zero tail %d\n", left);
368			ceph_zero_page_vector_range(page_align + read, left,
369						    pages);
370			read += left;
371		}
372	}
373
374	if (ret >= 0)
375		ret = read;
376	dout("striped_read returns %d\n", ret);
377	return ret;
378}
379
380/*
381 * Completely synchronous read and write methods.  Direct from __user
382 * buffer to osd, or directly to user pages (if O_DIRECT).
383 *
384 * If the read spans object boundary, just do multiple reads.
385 */
386static ssize_t ceph_sync_read(struct file *file, char __user *data,
387			      unsigned len, loff_t *poff, int *checkeof)
388{
389	struct inode *inode = file->f_dentry->d_inode;
390	struct page **pages;
391	u64 off = *poff;
392	int num_pages, ret;
393
394	dout("sync_read on file %p %llu~%u %s\n", file, off, len,
395	     (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
396
397	if (file->f_flags & O_DIRECT) {
398		num_pages = calc_pages_for((unsigned long)data, len);
399		pages = ceph_get_direct_page_vector(data, num_pages, true);
400	} else {
401		num_pages = calc_pages_for(off, len);
402		pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
403	}
404	if (IS_ERR(pages))
405		return PTR_ERR(pages);
406
407	/*
408	 * flush any page cache pages in this range.  this
409	 * will make concurrent normal and sync io slow,
410	 * but it will at least behave sensibly when they are
411	 * in sequence.
412	 */
413	ret = filemap_write_and_wait(inode->i_mapping);
414	if (ret < 0)
415		goto done;
416
417	ret = striped_read(inode, off, len, pages, num_pages, checkeof,
418			   file->f_flags & O_DIRECT,
419			   (unsigned long)data & ~PAGE_MASK);
420
421	if (ret >= 0 && (file->f_flags & O_DIRECT) == 0)
422		ret = ceph_copy_page_vector_to_user(pages, data, off, ret);
423	if (ret >= 0)
424		*poff = off + ret;
425
426done:
427	if (file->f_flags & O_DIRECT)
428		ceph_put_page_vector(pages, num_pages, true);
429	else
430		ceph_release_page_vector(pages, num_pages);
431	dout("sync_read result %d\n", ret);
432	return ret;
433}
434
435/*
436 * Write commit callback, called if we requested both an ACK and
437 * ONDISK commit reply from the OSD.
438 */
439static void sync_write_commit(struct ceph_osd_request *req,
440			      struct ceph_msg *msg)
441{
442	struct ceph_inode_info *ci = ceph_inode(req->r_inode);
443
444	dout("sync_write_commit %p tid %llu\n", req, req->r_tid);
445	spin_lock(&ci->i_unsafe_lock);
446	list_del_init(&req->r_unsafe_item);
447	spin_unlock(&ci->i_unsafe_lock);
448	ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
449}
450
451/*
452 * Synchronous write, straight from __user pointer or user pages (if
453 * O_DIRECT).
454 *
455 * If write spans object boundary, just do multiple writes.  (For a
456 * correct atomic write, we should e.g. take write locks on all
457 * objects, rollback on failure, etc.)
458 */
459static ssize_t ceph_sync_write(struct file *file, const char __user *data,
460			       size_t left, loff_t *offset)
461{
462	struct inode *inode = file->f_dentry->d_inode;
463	struct ceph_inode_info *ci = ceph_inode(inode);
464	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
465	struct ceph_osd_request *req;
466	struct page **pages;
467	int num_pages;
468	long long unsigned pos;
469	u64 len;
470	int written = 0;
471	int flags;
472	int do_sync = 0;
473	int check_caps = 0;
474	int page_align, io_align;
475	unsigned long buf_align;
476	int ret;
477	struct timespec mtime = CURRENT_TIME;
478
479	if (ceph_snap(file->f_dentry->d_inode) != CEPH_NOSNAP)
480		return -EROFS;
481
482	dout("sync_write on file %p %lld~%u %s\n", file, *offset,
483	     (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
484
485	if (file->f_flags & O_APPEND)
486		pos = i_size_read(inode);
487	else
488		pos = *offset;
489
490	ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left);
491	if (ret < 0)
492		return ret;
493
494	ret = invalidate_inode_pages2_range(inode->i_mapping,
495					    pos >> PAGE_CACHE_SHIFT,
496					    (pos + left) >> PAGE_CACHE_SHIFT);
497	if (ret < 0)
498		dout("invalidate_inode_pages2_range returned %d\n", ret);
499
500	flags = CEPH_OSD_FLAG_ORDERSNAP |
501		CEPH_OSD_FLAG_ONDISK |
502		CEPH_OSD_FLAG_WRITE;
503	if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0)
504		flags |= CEPH_OSD_FLAG_ACK;
505	else
506		do_sync = 1;
507
508	/*
509	 * we may need to do multiple writes here if we span an object
510	 * boundary.  this isn't atomic, unfortunately.  :(
511	 */
512more:
513	io_align = pos & ~PAGE_MASK;
514	buf_align = (unsigned long)data & ~PAGE_MASK;
515	len = left;
516	if (file->f_flags & O_DIRECT) {
517		/* write from beginning of first page, regardless of
518		   io alignment */
519		page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
520		num_pages = calc_pages_for((unsigned long)data, len);
521	} else {
522		page_align = pos & ~PAGE_MASK;
523		num_pages = calc_pages_for(pos, len);
524	}
525	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
526				    ceph_vino(inode), pos, &len,
527				    CEPH_OSD_OP_WRITE, flags,
528				    ci->i_snap_realm->cached_context,
529				    do_sync,
530				    ci->i_truncate_seq, ci->i_truncate_size,
531				    &mtime, false, 2, page_align);
532	if (!req)
533		return -ENOMEM;
534
535	if (file->f_flags & O_DIRECT) {
536		pages = ceph_get_direct_page_vector(data, num_pages, false);
537		if (IS_ERR(pages)) {
538			ret = PTR_ERR(pages);
539			goto out;
540		}
541
542		/*
543		 * throw out any page cache pages in this range. this
544		 * may block.
545		 */
546		truncate_inode_pages_range(inode->i_mapping, pos,
547					   (pos+len) | (PAGE_CACHE_SIZE-1));
548	} else {
549		pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
550		if (IS_ERR(pages)) {
551			ret = PTR_ERR(pages);
552			goto out;
553		}
554		ret = ceph_copy_user_to_page_vector(pages, data, pos, len);
555		if (ret < 0) {
556			ceph_release_page_vector(pages, num_pages);
557			goto out;
558		}
559
560		if ((file->f_flags & O_SYNC) == 0) {
561			/* get a second commit callback */
562			req->r_safe_callback = sync_write_commit;
563			req->r_own_pages = 1;
564		}
565	}
566	req->r_pages = pages;
567	req->r_num_pages = num_pages;
568	req->r_inode = inode;
569
570	ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
571	if (!ret) {
572		if (req->r_safe_callback) {
573			/*
574			 * Add to inode unsafe list only after we
575			 * start_request so that a tid has been assigned.
576			 */
577			spin_lock(&ci->i_unsafe_lock);
578			list_add_tail(&req->r_unsafe_item,
579				      &ci->i_unsafe_writes);
580			spin_unlock(&ci->i_unsafe_lock);
581			ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
582		}
583		
584		ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
585		if (ret < 0 && req->r_safe_callback) {
586			spin_lock(&ci->i_unsafe_lock);
587			list_del_init(&req->r_unsafe_item);
588			spin_unlock(&ci->i_unsafe_lock);
589			ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
590		}
591	}
592
593	if (file->f_flags & O_DIRECT)
594		ceph_put_page_vector(pages, num_pages, false);
595	else if (file->f_flags & O_SYNC)
596		ceph_release_page_vector(pages, num_pages);
597
598out:
599	ceph_osdc_put_request(req);
600	if (ret == 0) {
601		pos += len;
602		written += len;
603		left -= len;
604		data += written;
605		if (left)
606			goto more;
607
608		ret = written;
609		*offset = pos;
610		if (pos > i_size_read(inode))
611			check_caps = ceph_inode_set_size(inode, pos);
612		if (check_caps)
613			ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY,
614					NULL);
615	}
616	return ret;
617}
618
619/*
620 * Wrap generic_file_aio_read with checks for cap bits on the inode.
621 * Atomically grab references, so that those bits are not released
622 * back to the MDS mid-read.
623 *
624 * Hmm, the sync read case isn't actually async... should it be?
625 */
626static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov,
627			     unsigned long nr_segs, loff_t pos)
628{
629	struct file *filp = iocb->ki_filp;
630	struct ceph_file_info *fi = filp->private_data;
631	loff_t *ppos = &iocb->ki_pos;
632	size_t len = iov->iov_len;
633	struct inode *inode = filp->f_dentry->d_inode;
634	struct ceph_inode_info *ci = ceph_inode(inode);
635	void __user *base = iov->iov_base;
636	ssize_t ret;
637	int want, got = 0;
638	int checkeof = 0, read = 0;
639
640	dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
641	     inode, ceph_vinop(inode), pos, (unsigned)len, inode);
642again:
643	__ceph_do_pending_vmtruncate(inode);
644	if (fi->fmode & CEPH_FILE_MODE_LAZY)
645		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
646	else
647		want = CEPH_CAP_FILE_CACHE;
648	ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
649	if (ret < 0)
650		goto out;
651	dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
652	     inode, ceph_vinop(inode), pos, (unsigned)len,
653	     ceph_cap_string(got));
654
655	if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
656	    (iocb->ki_filp->f_flags & O_DIRECT) ||
657	    (inode->i_sb->s_flags & MS_SYNCHRONOUS) ||
658	    (fi->flags & CEPH_F_SYNC))
659		/* hmm, this isn't really async... */
660		ret = ceph_sync_read(filp, base, len, ppos, &checkeof);
661	else
662		ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
663
664out:
665	dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
666	     inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
667	ceph_put_cap_refs(ci, got);
668
669	if (checkeof && ret >= 0) {
670		int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
671
672		/* hit EOF or hole? */
673		if (statret == 0 && *ppos < inode->i_size) {
674			dout("aio_read sync_read hit hole, ppos %lld < size %lld, reading more\n", *ppos, inode->i_size);
675			read += ret;
676			base += ret;
677			len -= ret;
678			checkeof = 0;
679			goto again;
680		}
681	}
682	if (ret >= 0)
683		ret += read;
684
685	return ret;
686}
687
688/*
689 * Take cap references to avoid releasing caps to MDS mid-write.
690 *
691 * If we are synchronous, and write with an old snap context, the OSD
692 * may return EOLDSNAPC.  In that case, retry the write.. _after_
693 * dropping our cap refs and allowing the pending snap to logically
694 * complete _before_ this write occurs.
695 *
696 * If we are near ENOSPC, write synchronously.
697 */
698static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov,
699		       unsigned long nr_segs, loff_t pos)
700{
701	struct file *file = iocb->ki_filp;
702	struct ceph_file_info *fi = file->private_data;
703	struct inode *inode = file->f_dentry->d_inode;
704	struct ceph_inode_info *ci = ceph_inode(inode);
705	struct ceph_osd_client *osdc =
706		&ceph_sb_to_client(inode->i_sb)->client->osdc;
707	loff_t endoff = pos + iov->iov_len;
708	int want, got = 0;
709	int ret, err;
710
711	if (ceph_snap(inode) != CEPH_NOSNAP)
712		return -EROFS;
713
714retry_snap:
715	if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
716		return -ENOSPC;
717	__ceph_do_pending_vmtruncate(inode);
718	dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n",
719	     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
720	     inode->i_size);
721	if (fi->fmode & CEPH_FILE_MODE_LAZY)
722		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
723	else
724		want = CEPH_CAP_FILE_BUFFER;
725	ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, endoff);
726	if (ret < 0)
727		goto out_put;
728
729	dout("aio_write %p %llx.%llx %llu~%u  got cap refs on %s\n",
730	     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
731	     ceph_cap_string(got));
732
733	if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
734	    (iocb->ki_filp->f_flags & O_DIRECT) ||
735	    (inode->i_sb->s_flags & MS_SYNCHRONOUS) ||
736	    (fi->flags & CEPH_F_SYNC)) {
737		ret = ceph_sync_write(file, iov->iov_base, iov->iov_len,
738			&iocb->ki_pos);
739	} else {
740		/*
741		 * buffered write; drop Fw early to avoid slow
742		 * revocation if we get stuck on balance_dirty_pages
743		 */
744		int dirty;
745
746		spin_lock(&inode->i_lock);
747		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
748		spin_unlock(&inode->i_lock);
749		ceph_put_cap_refs(ci, got);
750
751		ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
752		if ((ret >= 0 || ret == -EIOCBQUEUED) &&
753		    ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host)
754		     || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
755			err = vfs_fsync_range(file, pos, pos + ret - 1, 1);
756			if (err < 0)
757				ret = err;
758		}
759
760		if (dirty)
761			__mark_inode_dirty(inode, dirty);
762		goto out;
763	}
764
765	if (ret >= 0) {
766		int dirty;
767		spin_lock(&inode->i_lock);
768		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
769		spin_unlock(&inode->i_lock);
770		if (dirty)
771			__mark_inode_dirty(inode, dirty);
772	}
773
774out_put:
775	dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
776	     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
777	     ceph_cap_string(got));
778	ceph_put_cap_refs(ci, got);
779
780out:
781	if (ret == -EOLDSNAPC) {
782		dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n",
783		     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len);
784		goto retry_snap;
785	}
786
787	return ret;
788}
789
790/*
791 * llseek.  be sure to verify file size on SEEK_END.
792 */
793static loff_t ceph_llseek(struct file *file, loff_t offset, int origin)
794{
795	struct inode *inode = file->f_mapping->host;
796	int ret;
797
798	mutex_lock(&inode->i_mutex);
799	__ceph_do_pending_vmtruncate(inode);
800	if (origin != SEEK_CUR || origin != SEEK_SET) {
801		ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
802		if (ret < 0) {
803			offset = ret;
804			goto out;
805		}
806	}
807
808	switch (origin) {
809	case SEEK_END:
810		offset += inode->i_size;
811		break;
812	case SEEK_CUR:
813		/*
814		 * Here we special-case the lseek(fd, 0, SEEK_CUR)
815		 * position-querying operation.  Avoid rewriting the "same"
816		 * f_pos value back to the file because a concurrent read(),
817		 * write() or lseek() might have altered it
818		 */
819		if (offset == 0) {
820			offset = file->f_pos;
821			goto out;
822		}
823		offset += file->f_pos;
824		break;
825	case SEEK_DATA:
826		if (offset >= inode->i_size) {
827			ret = -ENXIO;
828			goto out;
829		}
830		break;
831	case SEEK_HOLE:
832		if (offset >= inode->i_size) {
833			ret = -ENXIO;
834			goto out;
835		}
836		offset = inode->i_size;
837		break;
838	}
839
840	if (offset < 0 || offset > inode->i_sb->s_maxbytes) {
841		offset = -EINVAL;
842		goto out;
843	}
844
845	/* Special lock needed here? */
846	if (offset != file->f_pos) {
847		file->f_pos = offset;
848		file->f_version = 0;
849	}
850
851out:
852	mutex_unlock(&inode->i_mutex);
853	return offset;
854}
855
856const struct file_operations ceph_file_fops = {
857	.open = ceph_open,
858	.release = ceph_release,
859	.llseek = ceph_llseek,
860	.read = do_sync_read,
861	.write = do_sync_write,
862	.aio_read = ceph_aio_read,
863	.aio_write = ceph_aio_write,
864	.mmap = ceph_mmap,
865	.fsync = ceph_fsync,
866	.lock = ceph_lock,
867	.flock = ceph_flock,
868	.splice_read = generic_file_splice_read,
869	.splice_write = generic_file_splice_write,
870	.unlocked_ioctl = ceph_ioctl,
871	.compat_ioctl	= ceph_ioctl,
872};
873