Linux Audio

Check our new training course

In-person Linux kernel drivers training

Jun 16-20, 2025
Register
Loading...
v3.1
 
   1#include <linux/ceph/ceph_debug.h>
   2
   3#include <linux/backing-dev.h>
   4#include <linux/fs.h>
   5#include <linux/mm.h>
   6#include <linux/pagemap.h>
   7#include <linux/writeback.h>	/* generic_writepages */
   8#include <linux/slab.h>
   9#include <linux/pagevec.h>
  10#include <linux/task_io_accounting_ops.h>
 
 
 
 
  11
  12#include "super.h"
  13#include "mds_client.h"
 
 
  14#include <linux/ceph/osd_client.h>
 
  15
  16/*
  17 * Ceph address space ops.
  18 *
  19 * There are a few funny things going on here.
  20 *
  21 * The page->private field is used to reference a struct
  22 * ceph_snap_context for _every_ dirty page.  This indicates which
  23 * snapshot the page was logically dirtied in, and thus which snap
  24 * context needs to be associated with the osd write during writeback.
  25 *
  26 * Similarly, struct ceph_inode_info maintains a set of counters to
  27 * count dirty pages on the inode.  In the absence of snapshots,
  28 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  29 *
  30 * When a snapshot is taken (that is, when the client receives
  31 * notification that a snapshot was taken), each inode with caps and
  32 * with dirty pages (dirty pages implies there is a cap) gets a new
  33 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  34 * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  35 * moved to capsnap->dirty. (Unless a sync write is currently in
  36 * progress.  In that case, the capsnap is said to be "pending", new
  37 * writes cannot start, and the capsnap isn't "finalized" until the
  38 * write completes (or fails) and a final size/mtime for the inode for
  39 * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  40 *
  41 * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  42 * we look for the first capsnap in i_cap_snaps and write out pages in
  43 * that snap context _only_.  Then we move on to the next capsnap,
  44 * eventually reaching the "live" or "head" context (i.e., pages that
  45 * are not yet snapped) and are writing the most recently dirtied
  46 * pages.
  47 *
  48 * Invalidate and so forth must take care to ensure the dirty page
  49 * accounting is preserved.
  50 */
  51
  52#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  53#define CONGESTION_OFF_THRESH(congestion_kb)				\
  54	(CONGESTION_ON_THRESH(congestion_kb) -				\
  55	 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  56
 
 
  57
 
 
 
 
 
 
  58
  59/*
  60 * Dirty a page.  Optimistically adjust accounting, on the assumption
  61 * that we won't race with invalidate.  If we do, readjust.
  62 */
  63static int ceph_set_page_dirty(struct page *page)
  64{
  65	struct address_space *mapping = page->mapping;
  66	struct inode *inode;
  67	struct ceph_inode_info *ci;
  68	int undo = 0;
  69	struct ceph_snap_context *snapc;
  70
  71	if (unlikely(!mapping))
  72		return !TestSetPageDirty(page);
  73
  74	if (TestSetPageDirty(page)) {
  75		dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  76		     mapping->host, page, page->index);
 
  77		return 0;
  78	}
  79
  80	inode = mapping->host;
  81	ci = ceph_inode(inode);
  82
  83	/*
  84	 * Note that we're grabbing a snapc ref here without holding
  85	 * any locks!
  86	 */
  87	snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  88
  89	/* dirty the head */
  90	spin_lock(&inode->i_lock);
  91	if (ci->i_head_snapc == NULL)
  92		ci->i_head_snapc = ceph_get_snap_context(snapc);
  93	++ci->i_wrbuffer_ref_head;
 
 
 
 
 
 
 
 
 
 
  94	if (ci->i_wrbuffer_ref == 0)
  95		ihold(inode);
  96	++ci->i_wrbuffer_ref;
  97	dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
  98	     "snapc %p seq %lld (%d snaps)\n",
  99	     mapping->host, page, page->index,
 100	     ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 101	     ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 102	     snapc, snapc->seq, snapc->num_snaps);
 103	spin_unlock(&inode->i_lock);
 104
 105	/* now adjust page */
 106	spin_lock_irq(&mapping->tree_lock);
 107	if (page->mapping) {	/* Race with truncate? */
 108		WARN_ON_ONCE(!PageUptodate(page));
 109		account_page_dirtied(page, page->mapping);
 110		radix_tree_tag_set(&mapping->page_tree,
 111				page_index(page), PAGECACHE_TAG_DIRTY);
 112
 113		/*
 114		 * Reference snap context in page->private.  Also set
 115		 * PagePrivate so that we get invalidatepage callback.
 116		 */
 117		page->private = (unsigned long)snapc;
 118		SetPagePrivate(page);
 119	} else {
 120		dout("ANON set_page_dirty %p (raced truncate?)\n", page);
 121		undo = 1;
 122	}
 123
 124	spin_unlock_irq(&mapping->tree_lock);
 125
 126	if (undo)
 127		/* whoops, we failed to dirty the page */
 128		ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 129
 130	__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 
 131
 132	BUG_ON(!PageDirty(page));
 133	return 1;
 134}
 135
 136/*
 137 * If we are truncating the full page (i.e. offset == 0), adjust the
 138 * dirty page counters appropriately.  Only called if there is private
 139 * data on the page.
 140 */
 141static void ceph_invalidatepage(struct page *page, unsigned long offset)
 
 142{
 143	struct inode *inode;
 144	struct ceph_inode_info *ci;
 145	struct ceph_snap_context *snapc = (void *)page->private;
 146
 147	BUG_ON(!PageLocked(page));
 148	BUG_ON(!page->private);
 149	BUG_ON(!PagePrivate(page));
 150	BUG_ON(!page->mapping);
 151
 152	inode = page->mapping->host;
 
 153
 154	/*
 155	 * We can get non-dirty pages here due to races between
 156	 * set_page_dirty and truncate_complete_page; just spit out a
 157	 * warning, in case we end up with accounting problems later.
 158	 */
 159	if (!PageDirty(page))
 160		pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 161
 162	if (offset == 0)
 163		ClearPageChecked(page);
 
 164
 165	ci = ceph_inode(inode);
 166	if (offset == 0) {
 167		dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
 168		     inode, page, page->index, offset);
 169		ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 170		ceph_put_snap_context(snapc);
 171		page->private = 0;
 172		ClearPagePrivate(page);
 173	} else {
 174		dout("%p invalidatepage %p idx %lu partial dirty page\n",
 175		     inode, page, page->index);
 
 
 
 
 
 
 176	}
 
 177}
 178
 179/* just a sanity check */
 180static int ceph_releasepage(struct page *page, gfp_t g)
 181{
 182	struct inode *inode = page->mapping ? page->mapping->host : NULL;
 183	dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 184	WARN_ON(PageDirty(page));
 185	WARN_ON(page->private);
 186	WARN_ON(PagePrivate(page));
 187	return 0;
 
 
 
 
 
 
 
 188}
 189
 190/*
 191 * read a single page, without unlocking it.
 192 */
 193static int readpage_nounlock(struct file *filp, struct page *page)
 194{
 195	struct inode *inode = filp->f_dentry->d_inode;
 
 196	struct ceph_inode_info *ci = ceph_inode(inode);
 197	struct ceph_osd_client *osdc = 
 198		&ceph_inode_to_client(inode)->client->osdc;
 199	int err = 0;
 200	u64 len = PAGE_CACHE_SIZE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 201
 202	dout("readpage inode %p file %p page %p index %lu\n",
 203	     inode, filp, page, page->index);
 204	err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 205				  page->index << PAGE_CACHE_SHIFT, &len,
 206				  ci->i_truncate_seq, ci->i_truncate_size,
 207				  &page, 1, 0);
 208	if (err == -ENOENT)
 209		err = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 210	if (err < 0) {
 211		SetPageError(page);
 212		goto out;
 213	} else if (err < PAGE_CACHE_SIZE) {
 214		/* zero fill remainder of page */
 215		zero_user_segment(page, err, PAGE_CACHE_SIZE);
 216	}
 217	SetPageUptodate(page);
 218
 
 
 
 
 
 
 
 
 
 
 
 
 
 219out:
 220	return err < 0 ? err : 0;
 
 
 
 221}
 222
 223static int ceph_readpage(struct file *filp, struct page *page)
 224{
 225	int r = readpage_nounlock(filp, page);
 226	unlock_page(page);
 227	return r;
 228}
 229
 230/*
 231 * Build a vector of contiguous pages from the provided page list.
 232 */
 233static struct page **page_vector_from_list(struct list_head *page_list,
 234					   unsigned *nr_pages)
 235{
 236	struct page **pages;
 237	struct page *page;
 238	int next_index, contig_pages = 0;
 239
 240	/* build page vector */
 241	pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
 242	if (!pages)
 243		return ERR_PTR(-ENOMEM);
 244
 245	BUG_ON(list_empty(page_list));
 246	next_index = list_entry(page_list->prev, struct page, lru)->index;
 247	list_for_each_entry_reverse(page, page_list, lru) {
 248		if (page->index == next_index) {
 249			dout("readpages page %d %p\n", contig_pages, page);
 250			pages[contig_pages] = page;
 251			contig_pages++;
 252			next_index++;
 253		} else {
 254			break;
 255		}
 256	}
 257	*nr_pages = contig_pages;
 258	return pages;
 259}
 260
 261/*
 262 * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 263 * the caller (VM) cleans them up.
 264 */
 265static int ceph_readpages(struct file *file, struct address_space *mapping,
 266			  struct list_head *page_list, unsigned nr_pages)
 
 
 
 
 
 
 
 267{
 268	struct inode *inode = file->f_dentry->d_inode;
 269	struct ceph_inode_info *ci = ceph_inode(inode);
 270	struct ceph_osd_client *osdc =
 271		&ceph_inode_to_client(inode)->client->osdc;
 272	int rc = 0;
 273	struct page **pages;
 274	loff_t offset;
 275	u64 len;
 276
 277	dout("readpages %p file %p nr_pages %d\n",
 278	     inode, file, nr_pages);
 279
 280	pages = page_vector_from_list(page_list, &nr_pages);
 281	if (IS_ERR(pages))
 282		return PTR_ERR(pages);
 283
 284	/* guess read extent */
 285	offset = pages[0]->index << PAGE_CACHE_SHIFT;
 286	len = nr_pages << PAGE_CACHE_SHIFT;
 287	rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 288				 offset, &len,
 289				 ci->i_truncate_seq, ci->i_truncate_size,
 290				 pages, nr_pages, 0);
 291	if (rc == -ENOENT)
 292		rc = 0;
 293	if (rc < 0)
 294		goto out;
 295
 296	for (; !list_empty(page_list) && len > 0;
 297	     rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
 298		struct page *page =
 299			list_entry(page_list->prev, struct page, lru);
 300
 301		list_del(&page->lru);
 302
 303		if (rc < (int)PAGE_CACHE_SIZE) {
 304			/* zero (remainder of) page */
 305			int s = rc < 0 ? 0 : rc;
 306			zero_user_segment(page, s, PAGE_CACHE_SIZE);
 307		}
 308
 309		if (add_to_page_cache_lru(page, mapping, page->index,
 310					  GFP_NOFS)) {
 311			page_cache_release(page);
 312			dout("readpages %p add_to_page_cache failed %p\n",
 313			     inode, page);
 314			continue;
 315		}
 316		dout("readpages %p adding %p idx %lu\n", inode, page,
 317		     page->index);
 318		flush_dcache_page(page);
 319		SetPageUptodate(page);
 320		unlock_page(page);
 321		page_cache_release(page);
 322	}
 323	rc = 0;
 324
 325out:
 326	kfree(pages);
 327	return rc;
 
 328}
 329
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 330/*
 331 * Get ref for the oldest snapc for an inode with dirty data... that is, the
 332 * only snap context we are allowed to write back.
 333 */
 334static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 335						    u64 *snap_size)
 
 336{
 337	struct ceph_inode_info *ci = ceph_inode(inode);
 338	struct ceph_snap_context *snapc = NULL;
 339	struct ceph_cap_snap *capsnap = NULL;
 340
 341	spin_lock(&inode->i_lock);
 342	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 343		dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 344		     capsnap->context, capsnap->dirty_pages);
 345		if (capsnap->dirty_pages) {
 346			snapc = ceph_get_snap_context(capsnap->context);
 347			if (snap_size)
 348				*snap_size = capsnap->size;
 349			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 350		}
 
 
 
 
 
 
 
 
 
 351	}
 352	if (!snapc && ci->i_wrbuffer_ref_head) {
 353		snapc = ceph_get_snap_context(ci->i_head_snapc);
 354		dout(" head snapc %p has %d dirty pages\n",
 355		     snapc, ci->i_wrbuffer_ref_head);
 
 
 
 
 
 
 
 356	}
 357	spin_unlock(&inode->i_lock);
 358	return snapc;
 359}
 360
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 361/*
 362 * Write a single page, but leave the page locked.
 363 *
 364 * If we get a write error, set the page error bit, but still adjust the
 365 * dirty page accounting (i.e., page is no longer dirty).
 366 */
 367static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 368{
 369	struct inode *inode;
 370	struct ceph_inode_info *ci;
 371	struct ceph_fs_client *fsc;
 372	struct ceph_osd_client *osdc;
 373	loff_t page_off = page->index << PAGE_CACHE_SHIFT;
 374	int len = PAGE_CACHE_SIZE;
 375	loff_t i_size;
 376	int err = 0;
 377	struct ceph_snap_context *snapc, *oldest;
 378	u64 snap_size = 0;
 379	long writeback_stat;
 
 
 
 
 380
 381	dout("writepage %p idx %lu\n", page, page->index);
 382
 383	if (!page->mapping || !page->mapping->host) {
 384		dout("writepage %p - no mapping\n", page);
 385		return -EFAULT;
 386	}
 387	inode = page->mapping->host;
 388	ci = ceph_inode(inode);
 389	fsc = ceph_inode_to_client(inode);
 390	osdc = &fsc->client->osdc;
 391
 392	/* verify this is a writeable snap context */
 393	snapc = (void *)page->private;
 394	if (snapc == NULL) {
 395		dout("writepage %p page %p not dirty?\n", inode, page);
 396		goto out;
 397	}
 398	oldest = get_oldest_context(inode, &snap_size);
 399	if (snapc->seq > oldest->seq) {
 400		dout("writepage %p page %p snapc %p not writeable - noop\n",
 401		     inode, page, (void *)page->private);
 402		/* we should only noop if called by kswapd */
 403		WARN_ON((current->flags & PF_MEMALLOC) == 0);
 404		ceph_put_snap_context(oldest);
 405		goto out;
 
 406	}
 407	ceph_put_snap_context(oldest);
 408
 409	/* is this a partial page at end of file? */
 410	if (snap_size)
 411		i_size = snap_size;
 412	else
 413		i_size = i_size_read(inode);
 414	if (i_size < page_off + len)
 415		len = i_size - page_off;
 
 
 416
 417	dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 418	     inode, page, page->index, page_off, len, snapc);
 419
 420	writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 421	if (writeback_stat >
 422	    CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 423		set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 424
 425	set_page_writeback(page);
 426	err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 427				   &ci->i_layout, snapc,
 428				   page_off, len,
 429				   ci->i_truncate_seq, ci->i_truncate_size,
 430				   &inode->i_mtime,
 431				   &page, 1, 0, 0, true);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 432	if (err < 0) {
 433		dout("writepage setting page/mapping error %d %p\n", err, page);
 434		SetPageError(page);
 
 
 
 
 
 
 
 
 
 
 
 
 435		mapping_set_error(&inode->i_data, err);
 436		if (wbc)
 437			wbc->pages_skipped++;
 438	} else {
 439		dout("writepage cleaned page %p\n", page);
 440		err = 0;  /* vfs expects us to return 0 */
 441	}
 442	page->private = 0;
 443	ClearPagePrivate(page);
 444	end_page_writeback(page);
 445	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 446	ceph_put_snap_context(snapc);  /* page's reference */
 447out:
 
 
 
 
 448	return err;
 449}
 450
 451static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 452{
 453	int err;
 454	struct inode *inode = page->mapping->host;
 455	BUG_ON(!inode);
 456	ihold(inode);
 457	err = writepage_nounlock(page, wbc);
 
 
 
 
 
 458	unlock_page(page);
 459	iput(inode);
 460	return err;
 461}
 462
 463
 464/*
 465 * lame release_pages helper.  release_pages() isn't exported to
 466 * modules.
 467 */
 468static void ceph_release_pages(struct page **pages, int num)
 469{
 470	struct pagevec pvec;
 471	int i;
 472
 473	pagevec_init(&pvec, 0);
 474	for (i = 0; i < num; i++) {
 475		if (pagevec_add(&pvec, pages[i]) == 0)
 476			pagevec_release(&pvec);
 477	}
 478	pagevec_release(&pvec);
 479}
 480
 481
 482/*
 483 * async writeback completion handler.
 484 *
 485 * If we get an error, set the mapping error bit, but not the individual
 486 * page error bits.
 487 */
 488static void writepages_finish(struct ceph_osd_request *req,
 489			      struct ceph_msg *msg)
 490{
 491	struct inode *inode = req->r_inode;
 492	struct ceph_osd_reply_head *replyhead;
 493	struct ceph_osd_op *op;
 494	struct ceph_inode_info *ci = ceph_inode(inode);
 495	unsigned wrote;
 496	struct page *page;
 497	int i;
 
 
 498	struct ceph_snap_context *snapc = req->r_snapc;
 499	struct address_space *mapping = inode->i_mapping;
 500	__s32 rc = -EIO;
 501	u64 bytes = 0;
 502	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 503	long writeback_stat;
 504	unsigned issued = ceph_caps_issued(ci);
 505
 506	/* parse reply */
 507	replyhead = msg->front.iov_base;
 508	WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 509	op = (void *)(replyhead + 1);
 510	rc = le32_to_cpu(replyhead->result);
 511	bytes = le64_to_cpu(op->extent.length);
 512
 513	if (rc >= 0) {
 514		/*
 515		 * Assume we wrote the pages we originally sent.  The
 516		 * osd might reply with fewer pages if our writeback
 517		 * raced with a truncation and was adjusted at the osd,
 518		 * so don't believe the reply.
 519		 */
 520		wrote = req->r_num_pages;
 521	} else {
 522		wrote = 0;
 523		mapping_set_error(mapping, rc);
 
 
 
 
 
 524	}
 525	dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 526	     inode, rc, bytes, wrote);
 
 
 
 
 
 
 
 527
 528	/* clean all pages */
 529	for (i = 0; i < req->r_num_pages; i++) {
 530		page = req->r_pages[i];
 531		BUG_ON(!page);
 532		WARN_ON(!PageUptodate(page));
 533
 534		writeback_stat =
 535			atomic_long_dec_return(&fsc->writeback_count);
 536		if (writeback_stat <
 537		    CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 538			clear_bdi_congested(&fsc->backing_dev_info,
 539					    BLK_RW_ASYNC);
 540
 541		ceph_put_snap_context((void *)page->private);
 542		page->private = 0;
 543		ClearPagePrivate(page);
 544		dout("unlocking %d %p\n", i, page);
 545		end_page_writeback(page);
 
 
 
 
 
 
 
 
 
 
 
 
 546
 547		/*
 548		 * We lost the cache cap, need to truncate the page before
 549		 * it is unlocked, otherwise we'd truncate it later in the
 550		 * page truncation thread, possibly losing some data that
 551		 * raced its way in
 552		 */
 553		if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 554			generic_error_remove_page(inode->i_mapping, page);
 555
 556		unlock_page(page);
 557	}
 558	dout("%p wrote+cleaned %d pages\n", inode, wrote);
 559	ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
 560
 561	ceph_release_pages(req->r_pages, req->r_num_pages);
 562	if (req->r_pages_from_pool)
 563		mempool_free(req->r_pages,
 564			     ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 
 
 
 
 565	else
 566		kfree(req->r_pages);
 567	ceph_osdc_put_request(req);
 568}
 569
 570/*
 571 * allocate a page vec, either directly, or if necessary, via a the
 572 * mempool.  we avoid the mempool if we can because req->r_num_pages
 573 * may be less than the maximum write size.
 574 */
 575static void alloc_page_vec(struct ceph_fs_client *fsc,
 576			   struct ceph_osd_request *req)
 577{
 578	req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
 579			       GFP_NOFS);
 580	if (!req->r_pages) {
 581		req->r_pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS);
 582		req->r_pages_from_pool = 1;
 583		WARN_ON(!req->r_pages);
 584	}
 585}
 586
 587/*
 588 * initiate async writeback
 589 */
 590static int ceph_writepages_start(struct address_space *mapping,
 591				 struct writeback_control *wbc)
 592{
 593	struct inode *inode = mapping->host;
 594	struct ceph_inode_info *ci = ceph_inode(inode);
 595	struct ceph_fs_client *fsc;
 596	pgoff_t index, start, end;
 597	int range_whole = 0;
 598	int should_loop = 1;
 599	pgoff_t max_pages = 0, max_pages_ever = 0;
 600	struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 601	struct pagevec pvec;
 602	int done = 0;
 603	int rc = 0;
 604	unsigned wsize = 1 << inode->i_blkbits;
 605	struct ceph_osd_request *req = NULL;
 606	int do_sync;
 607	u64 snap_size = 0;
 
 608
 609	/*
 610	 * Include a 'sync' in the OSD request if this is a data
 611	 * integrity write (e.g., O_SYNC write or fsync()), or if our
 612	 * cap is being revoked.
 613	 */
 614	do_sync = wbc->sync_mode == WB_SYNC_ALL;
 615	if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 616		do_sync = 1;
 617	dout("writepages_start %p dosync=%d (mode=%s)\n",
 618	     inode, do_sync,
 619	     wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 620	     (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 621
 622	fsc = ceph_inode_to_client(inode);
 623	if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 624		pr_warning("writepage_start %p on forced umount\n", inode);
 
 
 
 
 625		return -EIO; /* we're in a forced umount, don't write! */
 626	}
 627	if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 628		wsize = fsc->mount_options->wsize;
 629	if (wsize < PAGE_CACHE_SIZE)
 630		wsize = PAGE_CACHE_SIZE;
 631	max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 632
 633	pagevec_init(&pvec, 0);
 634
 635	/* where to start/end? */
 636	if (wbc->range_cyclic) {
 637		start = mapping->writeback_index; /* Start from prev offset */
 638		end = -1;
 639		dout(" cyclic, start at %lu\n", start);
 640	} else {
 641		start = wbc->range_start >> PAGE_CACHE_SHIFT;
 642		end = wbc->range_end >> PAGE_CACHE_SHIFT;
 643		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 644			range_whole = 1;
 645		should_loop = 0;
 646		dout(" not cyclic, %lu to %lu\n", start, end);
 647	}
 648	index = start;
 649
 650retry:
 651	/* find oldest snap context with dirty data */
 652	ceph_put_snap_context(snapc);
 653	snapc = get_oldest_context(inode, &snap_size);
 654	if (!snapc) {
 655		/* hmm, why does writepages get called when there
 656		   is no dirty data? */
 657		dout(" no snap context with dirty data?\n");
 658		goto out;
 659	}
 660	dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 661	     snapc, snapc->seq, snapc->num_snaps);
 662	if (last_snapc && snapc != last_snapc) {
 663		/* if we switched to a newer snapc, restart our scan at the
 664		 * start of the original file range. */
 665		dout("  snapc differs from last pass, restarting at %lu\n",
 666		     index);
 667		index = start;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 668	}
 
 
 669	last_snapc = snapc;
 670
 671	while (!done && index <= end) {
 672		unsigned i;
 673		int first;
 674		pgoff_t next;
 675		int pvec_pages, locked_pages;
 676		struct page *page;
 677		int want;
 678		u64 offset, len;
 679		struct ceph_osd_request_head *reqhead;
 680		struct ceph_osd_op *op;
 681		long writeback_stat;
 682
 683		next = 0;
 684		locked_pages = 0;
 685		max_pages = max_pages_ever;
 686
 687get_more_pages:
 688		first = -1;
 689		want = min(end - index,
 690			   min((pgoff_t)PAGEVEC_SIZE,
 691			       max_pages - (pgoff_t)locked_pages) - 1)
 692			+ 1;
 693		pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 694						PAGECACHE_TAG_DIRTY,
 695						want);
 696		dout("pagevec_lookup_tag got %d\n", pvec_pages);
 697		if (!pvec_pages && !locked_pages)
 698			break;
 699		for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 700			page = pvec.pages[i];
 701			dout("? %p idx %lu\n", page, page->index);
 702			if (locked_pages == 0)
 703				lock_page(page);  /* first page */
 704			else if (!trylock_page(page))
 705				break;
 706
 707			/* only dirty pages, or our accounting breaks */
 708			if (unlikely(!PageDirty(page)) ||
 709			    unlikely(page->mapping != mapping)) {
 710				dout("!dirty or !mapping %p\n", page);
 711				unlock_page(page);
 712				break;
 713			}
 714			if (!wbc->range_cyclic && page->index > end) {
 715				dout("end of range %p\n", page);
 716				done = 1;
 
 
 
 
 
 
 717				unlock_page(page);
 718				break;
 719			}
 720			if (next && (page->index != next)) {
 721				dout("not consecutive %p\n", page);
 
 
 
 
 
 
 722				unlock_page(page);
 723				break;
 724			}
 725			if (wbc->sync_mode != WB_SYNC_NONE) {
 726				dout("waiting on writeback %p\n", page);
 727				wait_on_page_writeback(page);
 728			}
 729			if ((snap_size && page_offset(page) > snap_size) ||
 730			    (!snap_size &&
 731			     page_offset(page) > i_size_read(inode))) {
 732				dout("%p page eof %llu\n", page, snap_size ?
 733				     snap_size : i_size_read(inode));
 734				done = 1;
 735				unlock_page(page);
 736				break;
 737			}
 738			if (PageWriteback(page)) {
 739				dout("%p under writeback\n", page);
 740				unlock_page(page);
 741				break;
 742			}
 743
 744			/* only if matching snap context */
 745			pgsnapc = (void *)page->private;
 746			if (pgsnapc->seq > snapc->seq) {
 747				dout("page snapc %p %lld > oldest %p %lld\n",
 748				     pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 749				unlock_page(page);
 750				if (!locked_pages)
 751					continue; /* keep looking for snap */
 752				break;
 753			}
 754
 755			if (!clear_page_dirty_for_io(page)) {
 756				dout("%p !clear_page_dirty_for_io\n", page);
 757				unlock_page(page);
 758				break;
 759			}
 760
 761			/* ok */
 
 
 
 
 
 762			if (locked_pages == 0) {
 
 
 
 
 763				/* prepare async write request */
 764				offset = (unsigned long long)page->index
 765					<< PAGE_CACHE_SHIFT;
 766				len = wsize;
 767				req = ceph_osdc_new_request(&fsc->client->osdc,
 768					    &ci->i_layout,
 769					    ceph_vino(inode),
 770					    offset, &len,
 771					    CEPH_OSD_OP_WRITE,
 772					    CEPH_OSD_FLAG_WRITE |
 773						    CEPH_OSD_FLAG_ONDISK,
 774					    snapc, do_sync,
 775					    ci->i_truncate_seq,
 776					    ci->i_truncate_size,
 777					    &inode->i_mtime, true, 1, 0);
 
 
 
 
 
 
 
 778
 779				if (!req) {
 780					rc = -ENOMEM;
 
 
 
 
 781					unlock_page(page);
 782					break;
 783				}
 784
 785				max_pages = req->r_num_pages;
 786
 787				alloc_page_vec(fsc, req);
 788				req->r_callback = writepages_finish;
 789				req->r_inode = inode;
 790			}
 791
 792			/* note position of first page in pvec */
 793			if (first < 0)
 794				first = i;
 795			dout("%p will write page %p idx %lu\n",
 796			     inode, page, page->index);
 797
 798			writeback_stat =
 799			       atomic_long_inc_return(&fsc->writeback_count);
 800			if (writeback_stat > CONGESTION_ON_THRESH(
 801				    fsc->mount_options->congestion_kb)) {
 802				set_bdi_congested(&fsc->backing_dev_info,
 803						  BLK_RW_ASYNC);
 804			}
 805
 806			set_page_writeback(page);
 807			req->r_pages[locked_pages] = page;
 808			locked_pages++;
 809			next = page->index + 1;
 
 810		}
 811
 812		/* did we get anything? */
 813		if (!locked_pages)
 814			goto release_pvec_pages;
 815		if (i) {
 816			int j;
 817			BUG_ON(!locked_pages || first < 0);
 
 
 
 
 
 
 
 
 818
 819			if (pvec_pages && i == pvec_pages &&
 820			    locked_pages < max_pages) {
 821				dout("reached end pvec, trying for more\n");
 822				pagevec_reinit(&pvec);
 823				goto get_more_pages;
 824			}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 825
 826			/* shift unused pages over in the pvec...  we
 827			 * will need to release them below. */
 828			for (j = i; j < pvec_pages; j++) {
 829				dout(" pvec leftover page %p\n",
 830				     pvec.pages[j]);
 831				pvec.pages[j-i+first] = pvec.pages[j];
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 832			}
 833			pvec.nr -= i-first;
 
 
 834		}
 835
 836		/* submit the write */
 837		offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
 838		len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 839			  (u64)locked_pages << PAGE_CACHE_SHIFT);
 840		dout("writepages got %d pages at %llu~%llu\n",
 841		     locked_pages, offset, len);
 842
 843		/* revise final length, page count */
 844		req->r_num_pages = locked_pages;
 845		reqhead = req->r_request->front.iov_base;
 846		op = (void *)(reqhead + 1);
 847		op->extent.length = cpu_to_le64(len);
 848		op->payload_len = cpu_to_le32(len);
 849		req->r_request->hdr.data_len = cpu_to_le32(len);
 850
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 851		rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 852		BUG_ON(rc);
 853		req = NULL;
 854
 855		/* continue? */
 856		index = next;
 857		wbc->nr_to_write -= locked_pages;
 858		if (wbc->nr_to_write <= 0)
 859			done = 1;
 
 
 
 
 
 
 
 860
 861release_pvec_pages:
 862		dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 863		     pvec.nr ? pvec.pages[0] : NULL);
 864		pagevec_release(&pvec);
 865
 866		if (locked_pages && !done)
 867			goto retry;
 868	}
 869
 870	if (should_loop && !done) {
 871		/* more to do; loop back to beginning of file */
 872		dout("writepages looping back to beginning of file\n");
 873		should_loop = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 874		index = 0;
 875		goto retry;
 876	}
 877
 878	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 879		mapping->writeback_index = index;
 880
 881out:
 882	if (req)
 883		ceph_osdc_put_request(req);
 884	ceph_put_snap_context(snapc);
 885	dout("writepages done, rc = %d\n", rc);
 886	return rc;
 887}
 888
 889
 890
 891/*
 892 * See if a given @snapc is either writeable, or already written.
 893 */
 894static int context_is_writeable_or_written(struct inode *inode,
 895					   struct ceph_snap_context *snapc)
 896{
 897	struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 898	int ret = !oldest || snapc->seq <= oldest->seq;
 899
 900	ceph_put_snap_context(oldest);
 901	return ret;
 902}
 903
 904/*
 905 * We are only allowed to write into/dirty the page if the page is
 906 * clean, or already dirty within the same snap context.
 907 *
 908 * called with page locked.
 909 * return success with page locked,
 910 * or any failure (incl -EAGAIN) with page unlocked.
 
 
 
 911 */
 912static int ceph_update_writeable_page(struct file *file,
 913			    loff_t pos, unsigned len,
 914			    struct page *page)
 915{
 916	struct inode *inode = file->f_dentry->d_inode;
 
 917	struct ceph_inode_info *ci = ceph_inode(inode);
 918	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 919	loff_t page_off = pos & PAGE_CACHE_MASK;
 920	int pos_in_page = pos & ~PAGE_CACHE_MASK;
 921	int end_in_page = pos_in_page + len;
 922	loff_t i_size;
 923	int r;
 924	struct ceph_snap_context *snapc, *oldest;
 925
 926retry_locked:
 927	/* writepages currently holds page lock, but if we change that later, */
 928	wait_on_page_writeback(page);
 929
 930	/* check snap context */
 931	BUG_ON(!ci->i_snap_realm);
 932	down_read(&mdsc->snap_rwsem);
 933	BUG_ON(!ci->i_snap_realm->cached_context);
 934	snapc = (void *)page->private;
 935	if (snapc && snapc != ci->i_head_snapc) {
 
 
 
 
 936		/*
 937		 * this page is already dirty in another (older) snap
 938		 * context!  is it writeable now?
 939		 */
 940		oldest = get_oldest_context(inode, NULL);
 941		up_read(&mdsc->snap_rwsem);
 942
 943		if (snapc->seq > oldest->seq) {
 
 944			ceph_put_snap_context(oldest);
 945			dout(" page %p snapc %p not current or oldest\n",
 946			     page, snapc);
 947			/*
 948			 * queue for writeback, and wait for snapc to
 949			 * be writeable or written
 950			 */
 951			snapc = ceph_get_snap_context(snapc);
 952			unlock_page(page);
 953			ceph_queue_writeback(inode);
 954			r = wait_event_interruptible(ci->i_cap_wq,
 955			       context_is_writeable_or_written(inode, snapc));
 956			ceph_put_snap_context(snapc);
 957			if (r == -ERESTARTSYS)
 958				return r;
 959			return -EAGAIN;
 960		}
 961		ceph_put_snap_context(oldest);
 962
 963		/* yay, writeable, do it now (without dropping page lock) */
 964		dout(" page %p snapc %p not current, but oldest\n",
 965		     page, snapc);
 966		if (!clear_page_dirty_for_io(page))
 967			goto retry_locked;
 968		r = writepage_nounlock(page, NULL);
 969		if (r < 0)
 970			goto fail_nosnap;
 971		goto retry_locked;
 972	}
 973
 974	if (PageUptodate(page)) {
 975		dout(" page %p already uptodate\n", page);
 976		return 0;
 977	}
 
 
 978
 979	/* full page? */
 980	if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
 981		return 0;
 
 
 
 982
 983	/* past end of file? */
 984	i_size = inode->i_size;   /* caller holds i_mutex */
 
 985
 986	if (i_size + len > inode->i_sb->s_maxbytes) {
 987		/* file is too big */
 988		r = -EINVAL;
 989		goto fail;
 990	}
 991
 992	if (page_off >= i_size ||
 993	    (pos_in_page == 0 && (pos+len) >= i_size &&
 994	     end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
 995		dout(" zeroing %p 0 - %d and %d - %d\n",
 996		     page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
 997		zero_user_segments(page,
 998				   0, pos_in_page,
 999				   end_in_page, PAGE_CACHE_SIZE);
1000		return 0;
1001	}
1002
1003	/* we need to read it. */
1004	up_read(&mdsc->snap_rwsem);
1005	r = readpage_nounlock(file, page);
1006	if (r < 0)
1007		goto fail_nosnap;
1008	goto retry_locked;
1009
1010fail:
1011	up_read(&mdsc->snap_rwsem);
1012fail_nosnap:
1013	unlock_page(page);
1014	return r;
1015}
1016
1017/*
1018 * We are only allowed to write into/dirty the page if the page is
1019 * clean, or already dirty within the same snap context.
1020 */
1021static int ceph_write_begin(struct file *file, struct address_space *mapping,
1022			    loff_t pos, unsigned len, unsigned flags,
1023			    struct page **pagep, void **fsdata)
1024{
1025	struct inode *inode = file->f_dentry->d_inode;
1026	struct page *page;
1027	pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 
1028	int r;
1029
1030	do {
1031		/* get a page */
1032		page = grab_cache_page_write_begin(mapping, index, 0);
 
 
 
1033		if (!page)
1034			return -ENOMEM;
1035		*pagep = page;
1036
1037		dout("write_begin file %p inode %p page %p %d~%d\n", file,
1038		     inode, page, (int)pos, (int)len);
1039
1040		r = ceph_update_writeable_page(file, pos, len, page);
1041	} while (r == -EAGAIN);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1042
 
 
 
 
 
 
 
 
 
 
 
 
1043	return r;
1044}
1045
1046/*
1047 * we don't do anything in here that simple_write_end doesn't do
1048 * except adjust dirty page accounting and drop read lock on
1049 * mdsc->snap_rwsem.
1050 */
1051static int ceph_write_end(struct file *file, struct address_space *mapping,
1052			  loff_t pos, unsigned len, unsigned copied,
1053			  struct page *page, void *fsdata)
1054{
1055	struct inode *inode = file->f_dentry->d_inode;
1056	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1057	struct ceph_mds_client *mdsc = fsc->mdsc;
1058	unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1059	int check_cap = 0;
1060
1061	dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1062	     inode, page, (int)pos, (int)copied, (int)len);
1063
1064	/* zero the stale part of the page if we did a short copy */
1065	if (copied < len)
1066		zero_user_segment(page, from+copied, len);
 
 
 
 
 
1067
1068	/* did file size increase? */
1069	/* (no need for i_size_read(); we caller holds i_mutex */
1070	if (pos+copied > inode->i_size)
1071		check_cap = ceph_inode_set_size(inode, pos+copied);
1072
1073	if (!PageUptodate(page))
1074		SetPageUptodate(page);
1075
1076	set_page_dirty(page);
1077
 
1078	unlock_page(page);
1079	up_read(&mdsc->snap_rwsem);
1080	page_cache_release(page);
1081
1082	if (check_cap)
1083		ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1084
1085	return copied;
1086}
1087
1088/*
1089 * we set .direct_IO to indicate direct io is supported, but since we
1090 * intercept O_DIRECT reads and writes early, this function should
1091 * never get called.
1092 */
1093static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1094			      const struct iovec *iov,
1095			      loff_t pos, unsigned long nr_segs)
1096{
1097	WARN_ON(1);
1098	return -EINVAL;
1099}
1100
1101const struct address_space_operations ceph_aops = {
1102	.readpage = ceph_readpage,
1103	.readpages = ceph_readpages,
1104	.writepage = ceph_writepage,
1105	.writepages = ceph_writepages_start,
1106	.write_begin = ceph_write_begin,
1107	.write_end = ceph_write_end,
1108	.set_page_dirty = ceph_set_page_dirty,
1109	.invalidatepage = ceph_invalidatepage,
1110	.releasepage = ceph_releasepage,
1111	.direct_IO = ceph_direct_io,
1112};
1113
 
 
 
 
 
 
 
 
 
 
 
1114
1115/*
1116 * vm ops
1117 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1118
1119/*
1120 * Reuse write_begin here for simplicity.
1121 */
1122static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1123{
1124	struct inode *inode = vma->vm_file->f_dentry->d_inode;
 
 
 
 
1125	struct page *page = vmf->page;
1126	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1127	loff_t off = page->index << PAGE_CACHE_SHIFT;
1128	loff_t size, len;
1129	int ret;
1130
1131	size = i_size_read(inode);
1132	if (off + PAGE_CACHE_SIZE <= size)
1133		len = PAGE_CACHE_SIZE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1134	else
1135		len = size & ~PAGE_CACHE_MASK;
 
 
 
 
 
 
 
 
 
 
 
 
1136
1137	dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1138	     off, len, page, page->index);
1139
1140	lock_page(page);
1141
1142	ret = VM_FAULT_NOPAGE;
1143	if ((off > size) ||
1144	    (page->mapping != inode->i_mapping))
1145		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1146
1147	ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1148	if (ret == 0) {
1149		/* success.  we'll keep the page locked. */
1150		set_page_dirty(page);
1151		up_read(&mdsc->snap_rwsem);
1152		ret = VM_FAULT_LOCKED;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1153	} else {
1154		if (ret == -ENOMEM)
1155			ret = VM_FAULT_OOM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1156		else
1157			ret = VM_FAULT_SIGBUS;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1158	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1159out:
1160	dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1161	if (ret != VM_FAULT_LOCKED)
1162		unlock_page(page);
1163	return ret;
 
 
 
 
 
 
 
1164}
1165
1166static struct vm_operations_struct ceph_vmops = {
1167	.fault		= filemap_fault,
1168	.page_mkwrite	= ceph_page_mkwrite,
1169};
1170
1171int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1172{
1173	struct address_space *mapping = file->f_mapping;
1174
1175	if (!mapping->a_ops->readpage)
1176		return -ENOEXEC;
1177	file_accessed(file);
1178	vma->vm_ops = &ceph_vmops;
1179	vma->vm_flags |= VM_CAN_NONLINEAR;
1180	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1181}
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0
   2#include <linux/ceph/ceph_debug.h>
   3
   4#include <linux/backing-dev.h>
   5#include <linux/fs.h>
   6#include <linux/mm.h>
   7#include <linux/pagemap.h>
   8#include <linux/writeback.h>	/* generic_writepages */
   9#include <linux/slab.h>
  10#include <linux/pagevec.h>
  11#include <linux/task_io_accounting_ops.h>
  12#include <linux/signal.h>
  13#include <linux/iversion.h>
  14#include <linux/ktime.h>
  15#include <linux/netfs.h>
  16
  17#include "super.h"
  18#include "mds_client.h"
  19#include "cache.h"
  20#include "metric.h"
  21#include <linux/ceph/osd_client.h>
  22#include <linux/ceph/striper.h>
  23
  24/*
  25 * Ceph address space ops.
  26 *
  27 * There are a few funny things going on here.
  28 *
  29 * The page->private field is used to reference a struct
  30 * ceph_snap_context for _every_ dirty page.  This indicates which
  31 * snapshot the page was logically dirtied in, and thus which snap
  32 * context needs to be associated with the osd write during writeback.
  33 *
  34 * Similarly, struct ceph_inode_info maintains a set of counters to
  35 * count dirty pages on the inode.  In the absence of snapshots,
  36 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  37 *
  38 * When a snapshot is taken (that is, when the client receives
  39 * notification that a snapshot was taken), each inode with caps and
  40 * with dirty pages (dirty pages implies there is a cap) gets a new
  41 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  42 * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  43 * moved to capsnap->dirty. (Unless a sync write is currently in
  44 * progress.  In that case, the capsnap is said to be "pending", new
  45 * writes cannot start, and the capsnap isn't "finalized" until the
  46 * write completes (or fails) and a final size/mtime for the inode for
  47 * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  48 *
  49 * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  50 * we look for the first capsnap in i_cap_snaps and write out pages in
  51 * that snap context _only_.  Then we move on to the next capsnap,
  52 * eventually reaching the "live" or "head" context (i.e., pages that
  53 * are not yet snapped) and are writing the most recently dirtied
  54 * pages.
  55 *
  56 * Invalidate and so forth must take care to ensure the dirty page
  57 * accounting is preserved.
  58 */
  59
  60#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  61#define CONGESTION_OFF_THRESH(congestion_kb)				\
  62	(CONGESTION_ON_THRESH(congestion_kb) -				\
  63	 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  64
  65static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
  66					struct page *page, void **_fsdata);
  67
  68static inline struct ceph_snap_context *page_snap_context(struct page *page)
  69{
  70	if (PagePrivate(page))
  71		return (void *)page->private;
  72	return NULL;
  73}
  74
  75/*
  76 * Dirty a page.  Optimistically adjust accounting, on the assumption
  77 * that we won't race with invalidate.  If we do, readjust.
  78 */
  79static int ceph_set_page_dirty(struct page *page)
  80{
  81	struct address_space *mapping = page->mapping;
  82	struct inode *inode;
  83	struct ceph_inode_info *ci;
 
  84	struct ceph_snap_context *snapc;
  85
  86	if (PageDirty(page)) {
 
 
 
  87		dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  88		     mapping->host, page, page->index);
  89		BUG_ON(!PagePrivate(page));
  90		return 0;
  91	}
  92
  93	inode = mapping->host;
  94	ci = ceph_inode(inode);
  95
 
 
 
 
 
 
  96	/* dirty the head */
  97	spin_lock(&ci->i_ceph_lock);
  98	BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  99	if (__ceph_have_pending_cap_snap(ci)) {
 100		struct ceph_cap_snap *capsnap =
 101				list_last_entry(&ci->i_cap_snaps,
 102						struct ceph_cap_snap,
 103						ci_item);
 104		snapc = ceph_get_snap_context(capsnap->context);
 105		capsnap->dirty_pages++;
 106	} else {
 107		BUG_ON(!ci->i_head_snapc);
 108		snapc = ceph_get_snap_context(ci->i_head_snapc);
 109		++ci->i_wrbuffer_ref_head;
 110	}
 111	if (ci->i_wrbuffer_ref == 0)
 112		ihold(inode);
 113	++ci->i_wrbuffer_ref;
 114	dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 115	     "snapc %p seq %lld (%d snaps)\n",
 116	     mapping->host, page, page->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 page->private.  Also set
 124	 * PagePrivate so that we get invalidatepage callback.
 125	 */
 126	BUG_ON(PagePrivate(page));
 127	attach_page_private(page, snapc);
 128
 129	return __set_page_dirty_nobuffers(page);
 
 130}
 131
 132/*
 133 * If we are truncating the full page (i.e. offset == 0), adjust the
 134 * dirty page counters appropriately.  Only called if there is private
 135 * data on the page.
 136 */
 137static void ceph_invalidatepage(struct page *page, unsigned int offset,
 138				unsigned int length)
 139{
 140	struct inode *inode;
 141	struct ceph_inode_info *ci;
 142	struct ceph_snap_context *snapc;
 143
 144	wait_on_page_fscache(page);
 
 
 
 145
 146	inode = page->mapping->host;
 147	ci = ceph_inode(inode);
 148
 149	if (offset != 0 || length != thp_size(page)) {
 150		dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 151		     inode, page, page->index, offset, length);
 152		return;
 153	}
 
 
 154
 155	WARN_ON(!PageLocked(page));
 156	if (!PagePrivate(page))
 157		return;
 158
 159	dout("%p invalidatepage %p idx %lu full dirty page\n",
 160	     inode, page, page->index);
 161
 162	snapc = detach_page_private(page);
 163	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 164	ceph_put_snap_context(snapc);
 165}
 166
 167static int ceph_releasepage(struct page *page, gfp_t gfp)
 168{
 169	dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 170	     page, page->index, PageDirty(page) ? "" : "not ");
 171
 172	if (PageFsCache(page)) {
 173		if (!(gfp & __GFP_DIRECT_RECLAIM) || !(gfp & __GFP_FS))
 174			return 0;
 175		wait_on_page_fscache(page);
 176	}
 177	return !PagePrivate(page);
 178}
 179
 180static void ceph_netfs_expand_readahead(struct netfs_read_request *rreq)
 
 181{
 182	struct inode *inode = rreq->mapping->host;
 183	struct ceph_inode_info *ci = ceph_inode(inode);
 184	struct ceph_file_layout *lo = &ci->i_layout;
 185	u32 blockoff;
 186	u64 blockno;
 187
 188	/* Expand the start downward */
 189	blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
 190	rreq->start = blockno * lo->stripe_unit;
 191	rreq->len += blockoff;
 192
 193	/* Now, round up the length to the next block */
 194	rreq->len = roundup(rreq->len, lo->stripe_unit);
 195}
 196
 197static bool ceph_netfs_clamp_length(struct netfs_read_subrequest *subreq)
 
 
 
 198{
 199	struct inode *inode = subreq->rreq->mapping->host;
 200	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 201	struct ceph_inode_info *ci = ceph_inode(inode);
 202	u64 objno, objoff;
 203	u32 xlen;
 204
 205	/* Truncate the extent at the end of the current block */
 206	ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
 207				      &objno, &objoff, &xlen);
 208	subreq->len = min(xlen, fsc->mount_options->rsize);
 209	return true;
 210}
 211
 212static void finish_netfs_read(struct ceph_osd_request *req)
 213{
 214	struct ceph_fs_client *fsc = ceph_inode_to_client(req->r_inode);
 215	struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 216	struct netfs_read_subrequest *subreq = req->r_priv;
 217	int num_pages;
 218	int err = req->r_result;
 219
 220	ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
 221				 req->r_end_latency, osd_data->length, err);
 222
 223	dout("%s: result %d subreq->len=%zu i_size=%lld\n", __func__, req->r_result,
 224	     subreq->len, i_size_read(req->r_inode));
 225
 226	/* no object means success but no data */
 
 
 227	if (err == -ENOENT)
 228		err = 0;
 229	else if (err == -EBLOCKLISTED)
 230		fsc->blocklisted = true;
 231
 232	if (err >= 0 && err < subreq->len)
 233		__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 234
 235	netfs_subreq_terminated(subreq, err, true);
 236
 237	num_pages = calc_pages_for(osd_data->alignment, osd_data->length);
 238	ceph_put_page_vector(osd_data->pages, num_pages, false);
 239	iput(req->r_inode);
 240}
 241
 242static void ceph_netfs_issue_op(struct netfs_read_subrequest *subreq)
 243{
 244	struct netfs_read_request *rreq = subreq->rreq;
 245	struct inode *inode = rreq->mapping->host;
 246	struct ceph_inode_info *ci = ceph_inode(inode);
 247	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 248	struct ceph_osd_request *req;
 249	struct ceph_vino vino = ceph_vino(inode);
 250	struct iov_iter iter;
 251	struct page **pages;
 252	size_t page_off;
 253	int err = 0;
 254	u64 len = subreq->len;
 255
 256	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino, subreq->start, &len,
 257			0, 1, CEPH_OSD_OP_READ,
 258			CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
 259			NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
 260	if (IS_ERR(req)) {
 261		err = PTR_ERR(req);
 262		req = NULL;
 263		goto out;
 264	}
 265
 266	dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
 267	iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 268	err = iov_iter_get_pages_alloc(&iter, &pages, len, &page_off);
 269	if (err < 0) {
 270		dout("%s: iov_ter_get_pages_alloc returned %d\n", __func__, err);
 271		goto out;
 
 
 
 272	}
 
 273
 274	/* should always give us a page-aligned read */
 275	WARN_ON_ONCE(page_off);
 276	len = err;
 277
 278	osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 279	req->r_callback = finish_netfs_read;
 280	req->r_priv = subreq;
 281	req->r_inode = inode;
 282	ihold(inode);
 283
 284	err = ceph_osdc_start_request(req->r_osdc, req, false);
 285	if (err)
 286		iput(inode);
 287out:
 288	ceph_osdc_put_request(req);
 289	if (err)
 290		netfs_subreq_terminated(subreq, err, false);
 291	dout("%s: result %d\n", __func__, err);
 292}
 293
 294static void ceph_init_rreq(struct netfs_read_request *rreq, struct file *file)
 295{
 
 
 
 296}
 297
 298static void ceph_readahead_cleanup(struct address_space *mapping, void *priv)
 
 
 
 
 299{
 300	struct inode *inode = mapping->host;
 301	struct ceph_inode_info *ci = ceph_inode(inode);
 302	int got = (uintptr_t)priv;
 303
 304	if (got)
 305		ceph_put_cap_refs(ci, got);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 306}
 307
 308static const struct netfs_read_request_ops ceph_netfs_read_ops = {
 309	.init_rreq		= ceph_init_rreq,
 310	.is_cache_enabled	= ceph_is_cache_enabled,
 311	.begin_cache_operation	= ceph_begin_cache_operation,
 312	.issue_op		= ceph_netfs_issue_op,
 313	.expand_readahead	= ceph_netfs_expand_readahead,
 314	.clamp_length		= ceph_netfs_clamp_length,
 315	.check_write_begin	= ceph_netfs_check_write_begin,
 316	.cleanup		= ceph_readahead_cleanup,
 317};
 318
 319/* read a single page, without unlocking it. */
 320static int ceph_readpage(struct file *file, struct page *page)
 321{
 322	struct inode *inode = file_inode(file);
 323	struct ceph_inode_info *ci = ceph_inode(inode);
 324	struct ceph_vino vino = ceph_vino(inode);
 325	u64 off = page_offset(page);
 326	u64 len = thp_size(page);
 
 
 
 
 
 
 327
 328	if (ci->i_inline_version != CEPH_INLINE_NONE) {
 329		/*
 330		 * Uptodate inline data should have been added
 331		 * into page cache while getting Fcr caps.
 332		 */
 333		if (off == 0) {
 334			unlock_page(page);
 335			return -EINVAL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 336		}
 337		zero_user_segment(page, 0, thp_size(page));
 
 
 338		SetPageUptodate(page);
 339		unlock_page(page);
 340		return 0;
 341	}
 
 342
 343	dout("readpage ino %llx.%llx file %p off %llu len %llu page %p index %lu\n",
 344	     vino.ino, vino.snap, file, off, len, page, page->index);
 345
 346	return netfs_readpage(file, page, &ceph_netfs_read_ops, NULL);
 347}
 348
 349static void ceph_readahead(struct readahead_control *ractl)
 350{
 351	struct inode *inode = file_inode(ractl->file);
 352	struct ceph_file_info *fi = ractl->file->private_data;
 353	struct ceph_rw_context *rw_ctx;
 354	int got = 0;
 355	int ret = 0;
 356
 357	if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 358		return;
 359
 360	rw_ctx = ceph_find_rw_context(fi);
 361	if (!rw_ctx) {
 362		/*
 363		 * readahead callers do not necessarily hold Fcb caps
 364		 * (e.g. fadvise, madvise).
 365		 */
 366		int want = CEPH_CAP_FILE_CACHE;
 367
 368		ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 369		if (ret < 0)
 370			dout("start_read %p, error getting cap\n", inode);
 371		else if (!(got & want))
 372			dout("start_read %p, no cache cap\n", inode);
 373
 374		if (ret <= 0)
 375			return;
 376	}
 377	netfs_readahead(ractl, &ceph_netfs_read_ops, (void *)(uintptr_t)got);
 378}
 379
 380struct ceph_writeback_ctl
 381{
 382	loff_t i_size;
 383	u64 truncate_size;
 384	u32 truncate_seq;
 385	bool size_stable;
 386	bool head_snapc;
 387};
 388
 389/*
 390 * Get ref for the oldest snapc for an inode with dirty data... that is, the
 391 * only snap context we are allowed to write back.
 392 */
 393static struct ceph_snap_context *
 394get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 395		   struct ceph_snap_context *page_snapc)
 396{
 397	struct ceph_inode_info *ci = ceph_inode(inode);
 398	struct ceph_snap_context *snapc = NULL;
 399	struct ceph_cap_snap *capsnap = NULL;
 400
 401	spin_lock(&ci->i_ceph_lock);
 402	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 403		dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 404		     capsnap->context, capsnap->dirty_pages);
 405		if (!capsnap->dirty_pages)
 406			continue;
 407
 408		/* get i_size, truncate_{seq,size} for page_snapc? */
 409		if (snapc && capsnap->context != page_snapc)
 410			continue;
 411
 412		if (ctl) {
 413			if (capsnap->writing) {
 414				ctl->i_size = i_size_read(inode);
 415				ctl->size_stable = false;
 416			} else {
 417				ctl->i_size = capsnap->size;
 418				ctl->size_stable = true;
 419			}
 420			ctl->truncate_size = capsnap->truncate_size;
 421			ctl->truncate_seq = capsnap->truncate_seq;
 422			ctl->head_snapc = false;
 423		}
 424
 425		if (snapc)
 426			break;
 427
 428		snapc = ceph_get_snap_context(capsnap->context);
 429		if (!page_snapc ||
 430		    page_snapc == snapc ||
 431		    page_snapc->seq > snapc->seq)
 432			break;
 433	}
 434	if (!snapc && ci->i_wrbuffer_ref_head) {
 435		snapc = ceph_get_snap_context(ci->i_head_snapc);
 436		dout(" head snapc %p has %d dirty pages\n",
 437		     snapc, ci->i_wrbuffer_ref_head);
 438		if (ctl) {
 439			ctl->i_size = i_size_read(inode);
 440			ctl->truncate_size = ci->i_truncate_size;
 441			ctl->truncate_seq = ci->i_truncate_seq;
 442			ctl->size_stable = false;
 443			ctl->head_snapc = true;
 444		}
 445	}
 446	spin_unlock(&ci->i_ceph_lock);
 447	return snapc;
 448}
 449
 450static u64 get_writepages_data_length(struct inode *inode,
 451				      struct page *page, u64 start)
 452{
 453	struct ceph_inode_info *ci = ceph_inode(inode);
 454	struct ceph_snap_context *snapc = page_snap_context(page);
 455	struct ceph_cap_snap *capsnap = NULL;
 456	u64 end = i_size_read(inode);
 457
 458	if (snapc != ci->i_head_snapc) {
 459		bool found = false;
 460		spin_lock(&ci->i_ceph_lock);
 461		list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 462			if (capsnap->context == snapc) {
 463				if (!capsnap->writing)
 464					end = capsnap->size;
 465				found = true;
 466				break;
 467			}
 468		}
 469		spin_unlock(&ci->i_ceph_lock);
 470		WARN_ON(!found);
 471	}
 472	if (end > page_offset(page) + thp_size(page))
 473		end = page_offset(page) + thp_size(page);
 474	return end > start ? end - start : 0;
 475}
 476
 477/*
 478 * Write a single page, but leave the page locked.
 479 *
 480 * If we get a write error, mark the mapping for error, but still adjust the
 481 * dirty page accounting (i.e., page is no longer dirty).
 482 */
 483static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 484{
 485	struct inode *inode = page->mapping->host;
 486	struct ceph_inode_info *ci = ceph_inode(inode);
 487	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 
 
 
 
 
 488	struct ceph_snap_context *snapc, *oldest;
 489	loff_t page_off = page_offset(page);
 490	int err;
 491	loff_t len = thp_size(page);
 492	struct ceph_writeback_ctl ceph_wbc;
 493	struct ceph_osd_client *osdc = &fsc->client->osdc;
 494	struct ceph_osd_request *req;
 495
 496	dout("writepage %p idx %lu\n", page, page->index);
 497
 
 
 
 
 
 
 
 
 
 498	/* verify this is a writeable snap context */
 499	snapc = page_snap_context(page);
 500	if (!snapc) {
 501		dout("writepage %p page %p not dirty?\n", inode, page);
 502		return 0;
 503	}
 504	oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 505	if (snapc->seq > oldest->seq) {
 506		dout("writepage %p page %p snapc %p not writeable - noop\n",
 507		     inode, page, snapc);
 508		/* we should only noop if called by kswapd */
 509		WARN_ON(!(current->flags & PF_MEMALLOC));
 510		ceph_put_snap_context(oldest);
 511		redirty_page_for_writepage(wbc, page);
 512		return 0;
 513	}
 514	ceph_put_snap_context(oldest);
 515
 516	/* is this a partial page at end of file? */
 517	if (page_off >= ceph_wbc.i_size) {
 518		dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 519		page->mapping->a_ops->invalidatepage(page, 0, thp_size(page));
 520		return 0;
 521	}
 522
 523	if (ceph_wbc.i_size < page_off + len)
 524		len = ceph_wbc.i_size - page_off;
 525
 526	dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 527	     inode, page, page->index, page_off, len, snapc, snapc->seq);
 528
 529	if (atomic_long_inc_return(&fsc->writeback_count) >
 
 530	    CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 531		set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 532
 533	set_page_writeback(page);
 534	req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 535				    CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 536				    ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 537				    true);
 538	if (IS_ERR(req)) {
 539		redirty_page_for_writepage(wbc, page);
 540		end_page_writeback(page);
 541		return PTR_ERR(req);
 542	}
 543
 544	/* it may be a short write due to an object boundary */
 545	WARN_ON_ONCE(len > thp_size(page));
 546	osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 547	dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 548
 549	req->r_mtime = inode->i_mtime;
 550	err = ceph_osdc_start_request(osdc, req, true);
 551	if (!err)
 552		err = ceph_osdc_wait_request(osdc, req);
 553
 554	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 555				  req->r_end_latency, len, err);
 556
 557	ceph_osdc_put_request(req);
 558	if (err == 0)
 559		err = len;
 560
 561	if (err < 0) {
 562		struct writeback_control tmp_wbc;
 563		if (!wbc)
 564			wbc = &tmp_wbc;
 565		if (err == -ERESTARTSYS) {
 566			/* killed by SIGKILL */
 567			dout("writepage interrupted page %p\n", page);
 568			redirty_page_for_writepage(wbc, page);
 569			end_page_writeback(page);
 570			return err;
 571		}
 572		if (err == -EBLOCKLISTED)
 573			fsc->blocklisted = true;
 574		dout("writepage setting page/mapping error %d %p\n",
 575		     err, page);
 576		mapping_set_error(&inode->i_data, err);
 577		wbc->pages_skipped++;
 
 578	} else {
 579		dout("writepage cleaned page %p\n", page);
 580		err = 0;  /* vfs expects us to return 0 */
 581	}
 582	oldest = detach_page_private(page);
 583	WARN_ON_ONCE(oldest != snapc);
 584	end_page_writeback(page);
 585	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 586	ceph_put_snap_context(snapc);  /* page's reference */
 587
 588	if (atomic_long_dec_return(&fsc->writeback_count) <
 589	    CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 590		clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 591
 592	return err;
 593}
 594
 595static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 596{
 597	int err;
 598	struct inode *inode = page->mapping->host;
 599	BUG_ON(!inode);
 600	ihold(inode);
 601	err = writepage_nounlock(page, wbc);
 602	if (err == -ERESTARTSYS) {
 603		/* direct memory reclaimer was killed by SIGKILL. return 0
 604		 * to prevent caller from setting mapping/page error */
 605		err = 0;
 606	}
 607	unlock_page(page);
 608	iput(inode);
 609	return err;
 610}
 611
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 612/*
 613 * async writeback completion handler.
 614 *
 615 * If we get an error, set the mapping error bit, but not the individual
 616 * page error bits.
 617 */
 618static void writepages_finish(struct ceph_osd_request *req)
 
 619{
 620	struct inode *inode = req->r_inode;
 
 
 621	struct ceph_inode_info *ci = ceph_inode(inode);
 622	struct ceph_osd_data *osd_data;
 623	struct page *page;
 624	int num_pages, total_pages = 0;
 625	int i, j;
 626	int rc = req->r_result;
 627	struct ceph_snap_context *snapc = req->r_snapc;
 628	struct address_space *mapping = inode->i_mapping;
 
 
 629	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 630	unsigned int len = 0;
 631	bool remove_page;
 632
 633	dout("writepages_finish %p rc %d\n", inode, rc);
 634	if (rc < 0) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 635		mapping_set_error(mapping, rc);
 636		ceph_set_error_write(ci);
 637		if (rc == -EBLOCKLISTED)
 638			fsc->blocklisted = true;
 639	} else {
 640		ceph_clear_error_write(ci);
 641	}
 642
 643	/*
 644	 * We lost the cache cap, need to truncate the page before
 645	 * it is unlocked, otherwise we'd truncate it later in the
 646	 * page truncation thread, possibly losing some data that
 647	 * raced its way in
 648	 */
 649	remove_page = !(ceph_caps_issued(ci) &
 650			(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 651
 652	/* clean all pages */
 653	for (i = 0; i < req->r_num_ops; i++) {
 654		if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 655			break;
 
 656
 657		osd_data = osd_req_op_extent_osd_data(req, i);
 658		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 659		len += osd_data->length;
 660		num_pages = calc_pages_for((u64)osd_data->alignment,
 661					   (u64)osd_data->length);
 662		total_pages += num_pages;
 663		for (j = 0; j < num_pages; j++) {
 664			page = osd_data->pages[j];
 665			BUG_ON(!page);
 666			WARN_ON(!PageUptodate(page));
 667
 668			if (atomic_long_dec_return(&fsc->writeback_count) <
 669			     CONGESTION_OFF_THRESH(
 670					fsc->mount_options->congestion_kb))
 671				clear_bdi_congested(inode_to_bdi(inode),
 672						    BLK_RW_ASYNC);
 673
 674			ceph_put_snap_context(detach_page_private(page));
 675			end_page_writeback(page);
 676			dout("unlocking %p\n", page);
 677
 678			if (remove_page)
 679				generic_error_remove_page(inode->i_mapping,
 680							  page);
 681
 682			unlock_page(page);
 683		}
 684		dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 685		     inode, osd_data->length, rc >= 0 ? num_pages : 0);
 
 
 
 
 686
 687		release_pages(osd_data->pages, num_pages);
 688	}
 
 
 689
 690	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 691				  req->r_end_latency, len, rc);
 692
 693	ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 694
 695	osd_data = osd_req_op_extent_osd_data(req, 0);
 696	if (osd_data->pages_from_pool)
 697		mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 698	else
 699		kfree(osd_data->pages);
 700	ceph_osdc_put_request(req);
 701}
 702
 703/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 704 * initiate async writeback
 705 */
 706static int ceph_writepages_start(struct address_space *mapping,
 707				 struct writeback_control *wbc)
 708{
 709	struct inode *inode = mapping->host;
 710	struct ceph_inode_info *ci = ceph_inode(inode);
 711	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 712	struct ceph_vino vino = ceph_vino(inode);
 713	pgoff_t index, start_index, end = -1;
 
 
 714	struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 715	struct pagevec pvec;
 
 716	int rc = 0;
 717	unsigned int wsize = i_blocksize(inode);
 718	struct ceph_osd_request *req = NULL;
 719	struct ceph_writeback_ctl ceph_wbc;
 720	bool should_loop, range_whole = false;
 721	bool done = false;
 722
 723	dout("writepages_start %p (mode=%s)\n", inode,
 
 
 
 
 
 
 
 
 
 724	     wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 725	     (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 726
 727	if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
 728		if (ci->i_wrbuffer_ref > 0) {
 729			pr_warn_ratelimited(
 730				"writepage_start %p %lld forced umount\n",
 731				inode, ceph_ino(inode));
 732		}
 733		mapping_set_error(mapping, -EIO);
 734		return -EIO; /* we're in a forced umount, don't write! */
 735	}
 736	if (fsc->mount_options->wsize < wsize)
 737		wsize = fsc->mount_options->wsize;
 738
 739	pagevec_init(&pvec);
 740
 741	start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 742	index = start_index;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 743
 744retry:
 745	/* find oldest snap context with dirty data */
 746	snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 
 747	if (!snapc) {
 748		/* hmm, why does writepages get called when there
 749		   is no dirty data? */
 750		dout(" no snap context with dirty data?\n");
 751		goto out;
 752	}
 753	dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 754	     snapc, snapc->seq, snapc->num_snaps);
 755
 756	should_loop = false;
 757	if (ceph_wbc.head_snapc && snapc != last_snapc) {
 758		/* where to start/end? */
 759		if (wbc->range_cyclic) {
 760			index = start_index;
 761			end = -1;
 762			if (index > 0)
 763				should_loop = true;
 764			dout(" cyclic, start at %lu\n", index);
 765		} else {
 766			index = wbc->range_start >> PAGE_SHIFT;
 767			end = wbc->range_end >> PAGE_SHIFT;
 768			if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 769				range_whole = true;
 770			dout(" not cyclic, %lu to %lu\n", index, end);
 771		}
 772	} else if (!ceph_wbc.head_snapc) {
 773		/* Do not respect wbc->range_{start,end}. Dirty pages
 774		 * in that range can be associated with newer snapc.
 775		 * They are not writeable until we write all dirty pages
 776		 * associated with 'snapc' get written */
 777		if (index > 0)
 778			should_loop = true;
 779		dout(" non-head snapc, range whole\n");
 780	}
 781
 782	ceph_put_snap_context(last_snapc);
 783	last_snapc = snapc;
 784
 785	while (!done && index <= end) {
 786		int num_ops = 0, op_idx;
 787		unsigned i, pvec_pages, max_pages, locked_pages = 0;
 788		struct page **pages = NULL, **data_pages;
 
 789		struct page *page;
 790		pgoff_t strip_unit_end = 0;
 791		u64 offset = 0, len = 0;
 792		bool from_pool = false;
 793
 794		max_pages = wsize >> PAGE_SHIFT;
 
 
 
 
 795
 796get_more_pages:
 797		pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 798						end, PAGECACHE_TAG_DIRTY);
 799		dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 
 
 
 
 
 
 800		if (!pvec_pages && !locked_pages)
 801			break;
 802		for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 803			page = pvec.pages[i];
 804			dout("? %p idx %lu\n", page, page->index);
 805			if (locked_pages == 0)
 806				lock_page(page);  /* first page */
 807			else if (!trylock_page(page))
 808				break;
 809
 810			/* only dirty pages, or our accounting breaks */
 811			if (unlikely(!PageDirty(page)) ||
 812			    unlikely(page->mapping != mapping)) {
 813				dout("!dirty or !mapping %p\n", page);
 814				unlock_page(page);
 815				continue;
 816			}
 817			/* only if matching snap context */
 818			pgsnapc = page_snap_context(page);
 819			if (pgsnapc != snapc) {
 820				dout("page snapc %p %lld != oldest %p %lld\n",
 821				     pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 822				if (!should_loop &&
 823				    !ceph_wbc.head_snapc &&
 824				    wbc->sync_mode != WB_SYNC_NONE)
 825					should_loop = true;
 826				unlock_page(page);
 827				continue;
 828			}
 829			if (page_offset(page) >= ceph_wbc.i_size) {
 830				dout("%p page eof %llu\n",
 831				     page, ceph_wbc.i_size);
 832				if ((ceph_wbc.size_stable ||
 833				    page_offset(page) >= i_size_read(inode)) &&
 834				    clear_page_dirty_for_io(page))
 835					mapping->a_ops->invalidatepage(page,
 836								0, thp_size(page));
 837				unlock_page(page);
 838				continue;
 839			}
 840			if (strip_unit_end && (page->index > strip_unit_end)) {
 841				dout("end of strip unit %p\n", page);
 
 
 
 
 
 
 
 
 842				unlock_page(page);
 843				break;
 844			}
 845			if (PageWriteback(page)) {
 846				if (wbc->sync_mode == WB_SYNC_NONE) {
 847					dout("%p under writeback\n", page);
 848					unlock_page(page);
 849					continue;
 850				}
 851				dout("waiting on writeback %p\n", page);
 852				wait_on_page_writeback(page);
 
 
 
 
 
 
 
 853			}
 854
 855			if (!clear_page_dirty_for_io(page)) {
 856				dout("%p !clear_page_dirty_for_io\n", page);
 857				unlock_page(page);
 858				continue;
 859			}
 860
 861			/*
 862			 * We have something to write.  If this is
 863			 * the first locked page this time through,
 864			 * calculate max possinle write size and
 865			 * allocate a page array
 866			 */
 867			if (locked_pages == 0) {
 868				u64 objnum;
 869				u64 objoff;
 870				u32 xlen;
 871
 872				/* prepare async write request */
 873				offset = (u64)page_offset(page);
 874				ceph_calc_file_object_mapping(&ci->i_layout,
 875							      offset, wsize,
 876							      &objnum, &objoff,
 877							      &xlen);
 878				len = xlen;
 879
 880				num_ops = 1;
 881				strip_unit_end = page->index +
 882					((len - 1) >> PAGE_SHIFT);
 883
 884				BUG_ON(pages);
 885				max_pages = calc_pages_for(0, (u64)len);
 886				pages = kmalloc_array(max_pages,
 887						      sizeof(*pages),
 888						      GFP_NOFS);
 889				if (!pages) {
 890					from_pool = true;
 891					pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
 892					BUG_ON(!pages);
 893				}
 894
 895				len = 0;
 896			} else if (page->index !=
 897				   (offset + len) >> PAGE_SHIFT) {
 898				if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
 899							     CEPH_OSD_MAX_OPS)) {
 900					redirty_page_for_writepage(wbc, page);
 901					unlock_page(page);
 902					break;
 903				}
 904
 905				num_ops++;
 906				offset = (u64)page_offset(page);
 907				len = 0;
 
 
 908			}
 909
 910			/* note position of first page in pvec */
 
 
 911			dout("%p will write page %p idx %lu\n",
 912			     inode, page, page->index);
 913
 914			if (atomic_long_inc_return(&fsc->writeback_count) >
 915			    CONGESTION_ON_THRESH(
 
 916				    fsc->mount_options->congestion_kb)) {
 917				set_bdi_congested(inode_to_bdi(inode),
 918						  BLK_RW_ASYNC);
 919			}
 920
 921
 922			pages[locked_pages++] = page;
 923			pvec.pages[i] = NULL;
 924
 925			len += thp_size(page);
 926		}
 927
 928		/* did we get anything? */
 929		if (!locked_pages)
 930			goto release_pvec_pages;
 931		if (i) {
 932			unsigned j, n = 0;
 933			/* shift unused page to beginning of pvec */
 934			for (j = 0; j < pvec_pages; j++) {
 935				if (!pvec.pages[j])
 936					continue;
 937				if (n < j)
 938					pvec.pages[n] = pvec.pages[j];
 939				n++;
 940			}
 941			pvec.nr = n;
 942
 943			if (pvec_pages && i == pvec_pages &&
 944			    locked_pages < max_pages) {
 945				dout("reached end pvec, trying for more\n");
 946				pagevec_release(&pvec);
 947				goto get_more_pages;
 948			}
 949		}
 950
 951new_request:
 952		offset = page_offset(pages[0]);
 953		len = wsize;
 954
 955		req = ceph_osdc_new_request(&fsc->client->osdc,
 956					&ci->i_layout, vino,
 957					offset, &len, 0, num_ops,
 958					CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
 959					snapc, ceph_wbc.truncate_seq,
 960					ceph_wbc.truncate_size, false);
 961		if (IS_ERR(req)) {
 962			req = ceph_osdc_new_request(&fsc->client->osdc,
 963						&ci->i_layout, vino,
 964						offset, &len, 0,
 965						min(num_ops,
 966						    CEPH_OSD_SLAB_OPS),
 967						CEPH_OSD_OP_WRITE,
 968						CEPH_OSD_FLAG_WRITE,
 969						snapc, ceph_wbc.truncate_seq,
 970						ceph_wbc.truncate_size, true);
 971			BUG_ON(IS_ERR(req));
 972		}
 973		BUG_ON(len < page_offset(pages[locked_pages - 1]) +
 974			     thp_size(page) - offset);
 975
 976		req->r_callback = writepages_finish;
 977		req->r_inode = inode;
 978
 979		/* Format the osd request message and submit the write */
 980		len = 0;
 981		data_pages = pages;
 982		op_idx = 0;
 983		for (i = 0; i < locked_pages; i++) {
 984			u64 cur_offset = page_offset(pages[i]);
 985			if (offset + len != cur_offset) {
 986				if (op_idx + 1 == req->r_num_ops)
 987					break;
 988				osd_req_op_extent_dup_last(req, op_idx,
 989							   cur_offset - offset);
 990				dout("writepages got pages at %llu~%llu\n",
 991				     offset, len);
 992				osd_req_op_extent_osd_data_pages(req, op_idx,
 993							data_pages, len, 0,
 994							from_pool, false);
 995				osd_req_op_extent_update(req, op_idx, len);
 996
 997				len = 0;
 998				offset = cur_offset; 
 999				data_pages = pages + i;
1000				op_idx++;
1001			}
1002
1003			set_page_writeback(pages[i]);
1004			len += thp_size(page);
1005		}
1006
1007		if (ceph_wbc.size_stable) {
1008			len = min(len, ceph_wbc.i_size - offset);
1009		} else if (i == locked_pages) {
1010			/* writepages_finish() clears writeback pages
1011			 * according to the data length, so make sure
1012			 * data length covers all locked pages */
1013			u64 min_len = len + 1 - thp_size(page);
1014			len = get_writepages_data_length(inode, pages[i - 1],
1015							 offset);
1016			len = max(len, min_len);
1017		}
1018		dout("writepages got pages at %llu~%llu\n", offset, len);
 
 
1019
1020		osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1021						 0, from_pool, false);
1022		osd_req_op_extent_update(req, op_idx, len);
1023
1024		BUG_ON(op_idx + 1 != req->r_num_ops);
1025
1026		from_pool = false;
1027		if (i < locked_pages) {
1028			BUG_ON(num_ops <= req->r_num_ops);
1029			num_ops -= req->r_num_ops;
1030			locked_pages -= i;
1031
1032			/* allocate new pages array for next request */
1033			data_pages = pages;
1034			pages = kmalloc_array(locked_pages, sizeof(*pages),
1035					      GFP_NOFS);
1036			if (!pages) {
1037				from_pool = true;
1038				pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1039				BUG_ON(!pages);
1040			}
1041			memcpy(pages, data_pages + i,
1042			       locked_pages * sizeof(*pages));
1043			memset(data_pages + i, 0,
1044			       locked_pages * sizeof(*pages));
1045		} else {
1046			BUG_ON(num_ops != req->r_num_ops);
1047			index = pages[i - 1]->index + 1;
1048			/* request message now owns the pages array */
1049			pages = NULL;
1050		}
1051
1052		req->r_mtime = inode->i_mtime;
1053		rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1054		BUG_ON(rc);
1055		req = NULL;
1056
1057		wbc->nr_to_write -= i;
1058		if (pages)
1059			goto new_request;
1060
1061		/*
1062		 * We stop writing back only if we are not doing
1063		 * integrity sync. In case of integrity sync we have to
1064		 * keep going until we have written all the pages
1065		 * we tagged for writeback prior to entering this loop.
1066		 */
1067		if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1068			done = true;
1069
1070release_pvec_pages:
1071		dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1072		     pvec.nr ? pvec.pages[0] : NULL);
1073		pagevec_release(&pvec);
 
 
 
1074	}
1075
1076	if (should_loop && !done) {
1077		/* more to do; loop back to beginning of file */
1078		dout("writepages looping back to beginning of file\n");
1079		end = start_index - 1; /* OK even when start_index == 0 */
1080
1081		/* to write dirty pages associated with next snapc,
1082		 * we need to wait until current writes complete */
1083		if (wbc->sync_mode != WB_SYNC_NONE &&
1084		    start_index == 0 && /* all dirty pages were checked */
1085		    !ceph_wbc.head_snapc) {
1086			struct page *page;
1087			unsigned i, nr;
1088			index = 0;
1089			while ((index <= end) &&
1090			       (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1091						PAGECACHE_TAG_WRITEBACK))) {
1092				for (i = 0; i < nr; i++) {
1093					page = pvec.pages[i];
1094					if (page_snap_context(page) != snapc)
1095						continue;
1096					wait_on_page_writeback(page);
1097				}
1098				pagevec_release(&pvec);
1099				cond_resched();
1100			}
1101		}
1102
1103		start_index = 0;
1104		index = 0;
1105		goto retry;
1106	}
1107
1108	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1109		mapping->writeback_index = index;
1110
1111out:
1112	ceph_osdc_put_request(req);
1113	ceph_put_snap_context(last_snapc);
1114	dout("writepages dend - startone, rc = %d\n", rc);
 
1115	return rc;
1116}
1117
1118
1119
1120/*
1121 * See if a given @snapc is either writeable, or already written.
1122 */
1123static int context_is_writeable_or_written(struct inode *inode,
1124					   struct ceph_snap_context *snapc)
1125{
1126	struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1127	int ret = !oldest || snapc->seq <= oldest->seq;
1128
1129	ceph_put_snap_context(oldest);
1130	return ret;
1131}
1132
1133/**
1134 * ceph_find_incompatible - find an incompatible context and return it
1135 * @page: page being dirtied
1136 *
1137 * We are only allowed to write into/dirty a page if the page is
1138 * clean, or already dirty within the same snap context. Returns a
1139 * conflicting context if there is one, NULL if there isn't, or a
1140 * negative error code on other errors.
1141 *
1142 * Must be called with page lock held.
1143 */
1144static struct ceph_snap_context *
1145ceph_find_incompatible(struct page *page)
 
1146{
1147	struct inode *inode = page->mapping->host;
1148	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1149	struct ceph_inode_info *ci = ceph_inode(inode);
 
 
 
 
 
 
 
1150
1151	if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
1152		dout(" page %p forced umount\n", page);
1153		return ERR_PTR(-EIO);
1154	}
1155
1156	for (;;) {
1157		struct ceph_snap_context *snapc, *oldest;
1158
1159		wait_on_page_writeback(page);
1160
1161		snapc = page_snap_context(page);
1162		if (!snapc || snapc == ci->i_head_snapc)
1163			break;
1164
1165		/*
1166		 * this page is already dirty in another (older) snap
1167		 * context!  is it writeable now?
1168		 */
1169		oldest = get_oldest_context(inode, NULL, NULL);
 
 
1170		if (snapc->seq > oldest->seq) {
1171			/* not writeable -- return it for the caller to deal with */
1172			ceph_put_snap_context(oldest);
1173			dout(" page %p snapc %p not current or oldest\n", page, snapc);
1174			return ceph_get_snap_context(snapc);
 
 
 
 
 
 
 
 
 
 
 
 
 
1175		}
1176		ceph_put_snap_context(oldest);
1177
1178		/* yay, writeable, do it now (without dropping page lock) */
1179		dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1180		if (clear_page_dirty_for_io(page)) {
1181			int r = writepage_nounlock(page, NULL);
1182			if (r < 0)
1183				return ERR_PTR(r);
1184		}
 
 
 
 
 
 
 
1185	}
1186	return NULL;
1187}
1188
1189static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1190					struct page *page, void **_fsdata)
1191{
1192	struct inode *inode = file_inode(file);
1193	struct ceph_inode_info *ci = ceph_inode(inode);
1194	struct ceph_snap_context *snapc;
1195
1196	snapc = ceph_find_incompatible(page);
1197	if (snapc) {
1198		int r;
1199
1200		unlock_page(page);
1201		put_page(page);
1202		if (IS_ERR(snapc))
1203			return PTR_ERR(snapc);
1204
1205		ceph_queue_writeback(inode);
1206		r = wait_event_killable(ci->i_cap_wq,
1207					context_is_writeable_or_written(inode, snapc));
1208		ceph_put_snap_context(snapc);
1209		return r == 0 ? -EAGAIN : r;
 
 
 
 
 
1210	}
1211	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
1212}
1213
1214/*
1215 * We are only allowed to write into/dirty the page if the page is
1216 * clean, or already dirty within the same snap context.
1217 */
1218static int ceph_write_begin(struct file *file, struct address_space *mapping,
1219			    loff_t pos, unsigned len, unsigned flags,
1220			    struct page **pagep, void **fsdata)
1221{
1222	struct inode *inode = file_inode(file);
1223	struct ceph_inode_info *ci = ceph_inode(inode);
1224	struct page *page = NULL;
1225	pgoff_t index = pos >> PAGE_SHIFT;
1226	int r;
1227
1228	/*
1229	 * Uninlining should have already been done and everything updated, EXCEPT
1230	 * for inline_version sent to the MDS.
1231	 */
1232	if (ci->i_inline_version != CEPH_INLINE_NONE) {
1233		page = grab_cache_page_write_begin(mapping, index, flags);
1234		if (!page)
1235			return -ENOMEM;
 
 
 
 
1236
1237		/*
1238		 * The inline_version on a new inode is set to 1. If that's the
1239		 * case, then the page is brand new and isn't yet Uptodate.
1240		 */
1241		r = 0;
1242		if (index == 0 && ci->i_inline_version != 1) {
1243			if (!PageUptodate(page)) {
1244				WARN_ONCE(1, "ceph: write_begin called on still-inlined inode (inline_version %llu)!\n",
1245					  ci->i_inline_version);
1246				r = -EINVAL;
1247			}
1248			goto out;
1249		}
1250		zero_user_segment(page, 0, thp_size(page));
1251		SetPageUptodate(page);
1252		goto out;
1253	}
1254
1255	r = netfs_write_begin(file, inode->i_mapping, pos, len, 0, &page, NULL,
1256			      &ceph_netfs_read_ops, NULL);
1257out:
1258	if (r == 0)
1259		wait_on_page_fscache(page);
1260	if (r < 0) {
1261		if (page)
1262			put_page(page);
1263	} else {
1264		WARN_ON_ONCE(!PageLocked(page));
1265		*pagep = page;
1266	}
1267	return r;
1268}
1269
1270/*
1271 * we don't do anything in here that simple_write_end doesn't do
1272 * except adjust dirty page accounting
 
1273 */
1274static int ceph_write_end(struct file *file, struct address_space *mapping,
1275			  loff_t pos, unsigned len, unsigned copied,
1276			  struct page *page, void *fsdata)
1277{
1278	struct inode *inode = file_inode(file);
1279	bool check_cap = false;
 
 
 
1280
1281	dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1282	     inode, page, (int)pos, (int)copied, (int)len);
1283
1284	/* zero the stale part of the page if we did a short copy */
1285	if (!PageUptodate(page)) {
1286		if (copied < len) {
1287			copied = 0;
1288			goto out;
1289		}
1290		SetPageUptodate(page);
1291	}
1292
1293	/* did file size increase? */
1294	if (pos+copied > i_size_read(inode))
 
1295		check_cap = ceph_inode_set_size(inode, pos+copied);
1296
 
 
 
1297	set_page_dirty(page);
1298
1299out:
1300	unlock_page(page);
1301	put_page(page);
 
1302
1303	if (check_cap)
1304		ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1305
1306	return copied;
1307}
1308
1309/*
1310 * we set .direct_IO to indicate direct io is supported, but since we
1311 * intercept O_DIRECT reads and writes early, this function should
1312 * never get called.
1313 */
1314static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
 
 
1315{
1316	WARN_ON(1);
1317	return -EINVAL;
1318}
1319
1320const struct address_space_operations ceph_aops = {
1321	.readpage = ceph_readpage,
1322	.readahead = ceph_readahead,
1323	.writepage = ceph_writepage,
1324	.writepages = ceph_writepages_start,
1325	.write_begin = ceph_write_begin,
1326	.write_end = ceph_write_end,
1327	.set_page_dirty = ceph_set_page_dirty,
1328	.invalidatepage = ceph_invalidatepage,
1329	.releasepage = ceph_releasepage,
1330	.direct_IO = ceph_direct_io,
1331};
1332
1333static void ceph_block_sigs(sigset_t *oldset)
1334{
1335	sigset_t mask;
1336	siginitsetinv(&mask, sigmask(SIGKILL));
1337	sigprocmask(SIG_BLOCK, &mask, oldset);
1338}
1339
1340static void ceph_restore_sigs(sigset_t *oldset)
1341{
1342	sigprocmask(SIG_SETMASK, oldset, NULL);
1343}
1344
1345/*
1346 * vm ops
1347 */
1348static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1349{
1350	struct vm_area_struct *vma = vmf->vma;
1351	struct inode *inode = file_inode(vma->vm_file);
1352	struct ceph_inode_info *ci = ceph_inode(inode);
1353	struct ceph_file_info *fi = vma->vm_file->private_data;
1354	loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1355	int want, got, err;
1356	sigset_t oldset;
1357	vm_fault_t ret = VM_FAULT_SIGBUS;
1358
1359	ceph_block_sigs(&oldset);
1360
1361	dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1362	     inode, ceph_vinop(inode), off);
1363	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1364		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1365	else
1366		want = CEPH_CAP_FILE_CACHE;
1367
1368	got = 0;
1369	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1370	if (err < 0)
1371		goto out_restore;
1372
1373	dout("filemap_fault %p %llu got cap refs on %s\n",
1374	     inode, off, ceph_cap_string(got));
1375
1376	if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1377	    ci->i_inline_version == CEPH_INLINE_NONE) {
1378		CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1379		ceph_add_rw_context(fi, &rw_ctx);
1380		ret = filemap_fault(vmf);
1381		ceph_del_rw_context(fi, &rw_ctx);
1382		dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1383		     inode, off, ceph_cap_string(got), ret);
1384	} else
1385		err = -EAGAIN;
1386
1387	ceph_put_cap_refs(ci, got);
1388
1389	if (err != -EAGAIN)
1390		goto out_restore;
1391
1392	/* read inline data */
1393	if (off >= PAGE_SIZE) {
1394		/* does not support inline data > PAGE_SIZE */
1395		ret = VM_FAULT_SIGBUS;
1396	} else {
1397		struct address_space *mapping = inode->i_mapping;
1398		struct page *page = find_or_create_page(mapping, 0,
1399						mapping_gfp_constraint(mapping,
1400						~__GFP_FS));
1401		if (!page) {
1402			ret = VM_FAULT_OOM;
1403			goto out_inline;
1404		}
1405		err = __ceph_do_getattr(inode, page,
1406					 CEPH_STAT_CAP_INLINE_DATA, true);
1407		if (err < 0 || off >= i_size_read(inode)) {
1408			unlock_page(page);
1409			put_page(page);
1410			ret = vmf_error(err);
1411			goto out_inline;
1412		}
1413		if (err < PAGE_SIZE)
1414			zero_user_segment(page, err, PAGE_SIZE);
1415		else
1416			flush_dcache_page(page);
1417		SetPageUptodate(page);
1418		vmf->page = page;
1419		ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1420out_inline:
1421		dout("filemap_fault %p %llu read inline data ret %x\n",
1422		     inode, off, ret);
1423	}
1424out_restore:
1425	ceph_restore_sigs(&oldset);
1426	if (err < 0)
1427		ret = vmf_error(err);
1428
1429	return ret;
1430}
1431
1432static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1433{
1434	struct vm_area_struct *vma = vmf->vma;
1435	struct inode *inode = file_inode(vma->vm_file);
1436	struct ceph_inode_info *ci = ceph_inode(inode);
1437	struct ceph_file_info *fi = vma->vm_file->private_data;
1438	struct ceph_cap_flush *prealloc_cf;
1439	struct page *page = vmf->page;
1440	loff_t off = page_offset(page);
1441	loff_t size = i_size_read(inode);
1442	size_t len;
1443	int want, got, err;
1444	sigset_t oldset;
1445	vm_fault_t ret = VM_FAULT_SIGBUS;
1446
1447	prealloc_cf = ceph_alloc_cap_flush();
1448	if (!prealloc_cf)
1449		return VM_FAULT_OOM;
1450
1451	sb_start_pagefault(inode->i_sb);
1452	ceph_block_sigs(&oldset);
1453
1454	if (ci->i_inline_version != CEPH_INLINE_NONE) {
1455		struct page *locked_page = NULL;
1456		if (off == 0) {
1457			lock_page(page);
1458			locked_page = page;
1459		}
1460		err = ceph_uninline_data(vma->vm_file, locked_page);
1461		if (locked_page)
1462			unlock_page(locked_page);
1463		if (err < 0)
1464			goto out_free;
1465	}
1466
1467	if (off + thp_size(page) <= size)
1468		len = thp_size(page);
1469	else
1470		len = offset_in_thp(page, size);
1471
1472	dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1473	     inode, ceph_vinop(inode), off, len, size);
1474	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1475		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1476	else
1477		want = CEPH_CAP_FILE_BUFFER;
1478
1479	got = 0;
1480	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1481	if (err < 0)
1482		goto out_free;
1483
1484	dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1485	     inode, off, len, ceph_cap_string(got));
1486
1487	/* Update time before taking page lock */
1488	file_update_time(vma->vm_file);
1489	inode_inc_iversion_raw(inode);
1490
1491	do {
1492		struct ceph_snap_context *snapc;
1493
1494		lock_page(page);
1495
1496		if (page_mkwrite_check_truncate(page, inode) < 0) {
1497			unlock_page(page);
1498			ret = VM_FAULT_NOPAGE;
1499			break;
1500		}
1501
1502		snapc = ceph_find_incompatible(page);
1503		if (!snapc) {
1504			/* success.  we'll keep the page locked. */
1505			set_page_dirty(page);
1506			ret = VM_FAULT_LOCKED;
1507			break;
1508		}
1509
1510		unlock_page(page);
1511
1512		if (IS_ERR(snapc)) {
1513			ret = VM_FAULT_SIGBUS;
1514			break;
1515		}
1516
1517		ceph_queue_writeback(inode);
1518		err = wait_event_killable(ci->i_cap_wq,
1519				context_is_writeable_or_written(inode, snapc));
1520		ceph_put_snap_context(snapc);
1521	} while (err == 0);
1522
1523	if (ret == VM_FAULT_LOCKED ||
1524	    ci->i_inline_version != CEPH_INLINE_NONE) {
1525		int dirty;
1526		spin_lock(&ci->i_ceph_lock);
1527		ci->i_inline_version = CEPH_INLINE_NONE;
1528		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1529					       &prealloc_cf);
1530		spin_unlock(&ci->i_ceph_lock);
1531		if (dirty)
1532			__mark_inode_dirty(inode, dirty);
1533	}
1534
1535	dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1536	     inode, off, len, ceph_cap_string(got), ret);
1537	ceph_put_cap_refs_async(ci, got);
1538out_free:
1539	ceph_restore_sigs(&oldset);
1540	sb_end_pagefault(inode->i_sb);
1541	ceph_free_cap_flush(prealloc_cf);
1542	if (err < 0)
1543		ret = vmf_error(err);
1544	return ret;
1545}
1546
1547void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1548			   char	*data, size_t len)
1549{
1550	struct address_space *mapping = inode->i_mapping;
1551	struct page *page;
1552
1553	if (locked_page) {
1554		page = locked_page;
1555	} else {
1556		if (i_size_read(inode) == 0)
1557			return;
1558		page = find_or_create_page(mapping, 0,
1559					   mapping_gfp_constraint(mapping,
1560					   ~__GFP_FS));
1561		if (!page)
1562			return;
1563		if (PageUptodate(page)) {
1564			unlock_page(page);
1565			put_page(page);
1566			return;
1567		}
1568	}
1569
1570	dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1571	     inode, ceph_vinop(inode), len, locked_page);
1572
1573	if (len > 0) {
1574		void *kaddr = kmap_atomic(page);
1575		memcpy(kaddr, data, len);
1576		kunmap_atomic(kaddr);
1577	}
1578
1579	if (page != locked_page) {
1580		if (len < PAGE_SIZE)
1581			zero_user_segment(page, len, PAGE_SIZE);
1582		else
1583			flush_dcache_page(page);
1584
1585		SetPageUptodate(page);
1586		unlock_page(page);
1587		put_page(page);
1588	}
1589}
1590
1591int ceph_uninline_data(struct file *filp, struct page *locked_page)
1592{
1593	struct inode *inode = file_inode(filp);
1594	struct ceph_inode_info *ci = ceph_inode(inode);
1595	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1596	struct ceph_osd_request *req;
1597	struct page *page = NULL;
1598	u64 len, inline_version;
1599	int err = 0;
1600	bool from_pagecache = false;
1601
1602	spin_lock(&ci->i_ceph_lock);
1603	inline_version = ci->i_inline_version;
1604	spin_unlock(&ci->i_ceph_lock);
1605
1606	dout("uninline_data %p %llx.%llx inline_version %llu\n",
1607	     inode, ceph_vinop(inode), inline_version);
1608
1609	if (inline_version == 1 || /* initial version, no data */
1610	    inline_version == CEPH_INLINE_NONE)
1611		goto out;
1612
1613	if (locked_page) {
1614		page = locked_page;
1615		WARN_ON(!PageUptodate(page));
1616	} else if (ceph_caps_issued(ci) &
1617		   (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1618		page = find_get_page(inode->i_mapping, 0);
1619		if (page) {
1620			if (PageUptodate(page)) {
1621				from_pagecache = true;
1622				lock_page(page);
1623			} else {
1624				put_page(page);
1625				page = NULL;
1626			}
1627		}
1628	}
1629
1630	if (page) {
1631		len = i_size_read(inode);
1632		if (len > PAGE_SIZE)
1633			len = PAGE_SIZE;
1634	} else {
1635		page = __page_cache_alloc(GFP_NOFS);
1636		if (!page) {
1637			err = -ENOMEM;
1638			goto out;
1639		}
1640		err = __ceph_do_getattr(inode, page,
1641					CEPH_STAT_CAP_INLINE_DATA, true);
1642		if (err < 0) {
1643			/* no inline data */
1644			if (err == -ENODATA)
1645				err = 0;
1646			goto out;
1647		}
1648		len = err;
1649	}
1650
1651	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1652				    ceph_vino(inode), 0, &len, 0, 1,
1653				    CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1654				    NULL, 0, 0, false);
1655	if (IS_ERR(req)) {
1656		err = PTR_ERR(req);
1657		goto out;
1658	}
1659
1660	req->r_mtime = inode->i_mtime;
1661	err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1662	if (!err)
1663		err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1664	ceph_osdc_put_request(req);
1665	if (err < 0)
1666		goto out;
1667
1668	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1669				    ceph_vino(inode), 0, &len, 1, 3,
1670				    CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1671				    NULL, ci->i_truncate_seq,
1672				    ci->i_truncate_size, false);
1673	if (IS_ERR(req)) {
1674		err = PTR_ERR(req);
1675		goto out;
1676	}
1677
1678	osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1679
1680	{
1681		__le64 xattr_buf = cpu_to_le64(inline_version);
1682		err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1683					    "inline_version", &xattr_buf,
1684					    sizeof(xattr_buf),
1685					    CEPH_OSD_CMPXATTR_OP_GT,
1686					    CEPH_OSD_CMPXATTR_MODE_U64);
1687		if (err)
1688			goto out_put;
1689	}
1690
1691	{
1692		char xattr_buf[32];
1693		int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1694					 "%llu", inline_version);
1695		err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1696					    "inline_version",
1697					    xattr_buf, xattr_len, 0, 0);
1698		if (err)
1699			goto out_put;
1700	}
1701
1702	req->r_mtime = inode->i_mtime;
1703	err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1704	if (!err)
1705		err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1706
1707	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1708				  req->r_end_latency, len, err);
1709
1710out_put:
1711	ceph_osdc_put_request(req);
1712	if (err == -ECANCELED)
1713		err = 0;
1714out:
1715	if (page && page != locked_page) {
1716		if (from_pagecache) {
1717			unlock_page(page);
1718			put_page(page);
1719		} else
1720			__free_pages(page, 0);
1721	}
1722
1723	dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1724	     inode, ceph_vinop(inode), inline_version, err);
1725	return err;
1726}
1727
1728static const struct vm_operations_struct ceph_vmops = {
1729	.fault		= ceph_filemap_fault,
1730	.page_mkwrite	= ceph_page_mkwrite,
1731};
1732
1733int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1734{
1735	struct address_space *mapping = file->f_mapping;
1736
1737	if (!mapping->a_ops->readpage)
1738		return -ENOEXEC;
1739	file_accessed(file);
1740	vma->vm_ops = &ceph_vmops;
 
1741	return 0;
1742}
1743
1744enum {
1745	POOL_READ	= 1,
1746	POOL_WRITE	= 2,
1747};
1748
1749static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1750				s64 pool, struct ceph_string *pool_ns)
1751{
1752	struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1753	struct ceph_mds_client *mdsc = fsc->mdsc;
1754	struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1755	struct rb_node **p, *parent;
1756	struct ceph_pool_perm *perm;
1757	struct page **pages;
1758	size_t pool_ns_len;
1759	int err = 0, err2 = 0, have = 0;
1760
1761	down_read(&mdsc->pool_perm_rwsem);
1762	p = &mdsc->pool_perm_tree.rb_node;
1763	while (*p) {
1764		perm = rb_entry(*p, struct ceph_pool_perm, node);
1765		if (pool < perm->pool)
1766			p = &(*p)->rb_left;
1767		else if (pool > perm->pool)
1768			p = &(*p)->rb_right;
1769		else {
1770			int ret = ceph_compare_string(pool_ns,
1771						perm->pool_ns,
1772						perm->pool_ns_len);
1773			if (ret < 0)
1774				p = &(*p)->rb_left;
1775			else if (ret > 0)
1776				p = &(*p)->rb_right;
1777			else {
1778				have = perm->perm;
1779				break;
1780			}
1781		}
1782	}
1783	up_read(&mdsc->pool_perm_rwsem);
1784	if (*p)
1785		goto out;
1786
1787	if (pool_ns)
1788		dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1789		     pool, (int)pool_ns->len, pool_ns->str);
1790	else
1791		dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1792
1793	down_write(&mdsc->pool_perm_rwsem);
1794	p = &mdsc->pool_perm_tree.rb_node;
1795	parent = NULL;
1796	while (*p) {
1797		parent = *p;
1798		perm = rb_entry(parent, struct ceph_pool_perm, node);
1799		if (pool < perm->pool)
1800			p = &(*p)->rb_left;
1801		else if (pool > perm->pool)
1802			p = &(*p)->rb_right;
1803		else {
1804			int ret = ceph_compare_string(pool_ns,
1805						perm->pool_ns,
1806						perm->pool_ns_len);
1807			if (ret < 0)
1808				p = &(*p)->rb_left;
1809			else if (ret > 0)
1810				p = &(*p)->rb_right;
1811			else {
1812				have = perm->perm;
1813				break;
1814			}
1815		}
1816	}
1817	if (*p) {
1818		up_write(&mdsc->pool_perm_rwsem);
1819		goto out;
1820	}
1821
1822	rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1823					 1, false, GFP_NOFS);
1824	if (!rd_req) {
1825		err = -ENOMEM;
1826		goto out_unlock;
1827	}
1828
1829	rd_req->r_flags = CEPH_OSD_FLAG_READ;
1830	osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1831	rd_req->r_base_oloc.pool = pool;
1832	if (pool_ns)
1833		rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1834	ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1835
1836	err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1837	if (err)
1838		goto out_unlock;
1839
1840	wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1841					 1, false, GFP_NOFS);
1842	if (!wr_req) {
1843		err = -ENOMEM;
1844		goto out_unlock;
1845	}
1846
1847	wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1848	osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1849	ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1850	ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1851
1852	err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1853	if (err)
1854		goto out_unlock;
1855
1856	/* one page should be large enough for STAT data */
1857	pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1858	if (IS_ERR(pages)) {
1859		err = PTR_ERR(pages);
1860		goto out_unlock;
1861	}
1862
1863	osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1864				     0, false, true);
1865	err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1866
1867	wr_req->r_mtime = ci->vfs_inode.i_mtime;
1868	err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1869
1870	if (!err)
1871		err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1872	if (!err2)
1873		err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1874
1875	if (err >= 0 || err == -ENOENT)
1876		have |= POOL_READ;
1877	else if (err != -EPERM) {
1878		if (err == -EBLOCKLISTED)
1879			fsc->blocklisted = true;
1880		goto out_unlock;
1881	}
1882
1883	if (err2 == 0 || err2 == -EEXIST)
1884		have |= POOL_WRITE;
1885	else if (err2 != -EPERM) {
1886		if (err2 == -EBLOCKLISTED)
1887			fsc->blocklisted = true;
1888		err = err2;
1889		goto out_unlock;
1890	}
1891
1892	pool_ns_len = pool_ns ? pool_ns->len : 0;
1893	perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1894	if (!perm) {
1895		err = -ENOMEM;
1896		goto out_unlock;
1897	}
1898
1899	perm->pool = pool;
1900	perm->perm = have;
1901	perm->pool_ns_len = pool_ns_len;
1902	if (pool_ns_len > 0)
1903		memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1904	perm->pool_ns[pool_ns_len] = 0;
1905
1906	rb_link_node(&perm->node, parent, p);
1907	rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1908	err = 0;
1909out_unlock:
1910	up_write(&mdsc->pool_perm_rwsem);
1911
1912	ceph_osdc_put_request(rd_req);
1913	ceph_osdc_put_request(wr_req);
1914out:
1915	if (!err)
1916		err = have;
1917	if (pool_ns)
1918		dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1919		     pool, (int)pool_ns->len, pool_ns->str, err);
1920	else
1921		dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1922	return err;
1923}
1924
1925int ceph_pool_perm_check(struct inode *inode, int need)
1926{
1927	struct ceph_inode_info *ci = ceph_inode(inode);
1928	struct ceph_string *pool_ns;
1929	s64 pool;
1930	int ret, flags;
1931
1932	/* Only need to do this for regular files */
1933	if (!S_ISREG(inode->i_mode))
1934		return 0;
1935
1936	if (ci->i_vino.snap != CEPH_NOSNAP) {
1937		/*
1938		 * Pool permission check needs to write to the first object.
1939		 * But for snapshot, head of the first object may have alread
1940		 * been deleted. Skip check to avoid creating orphan object.
1941		 */
1942		return 0;
1943	}
1944
1945	if (ceph_test_mount_opt(ceph_inode_to_client(inode),
1946				NOPOOLPERM))
1947		return 0;
1948
1949	spin_lock(&ci->i_ceph_lock);
1950	flags = ci->i_ceph_flags;
1951	pool = ci->i_layout.pool_id;
1952	spin_unlock(&ci->i_ceph_lock);
1953check:
1954	if (flags & CEPH_I_POOL_PERM) {
1955		if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1956			dout("ceph_pool_perm_check pool %lld no read perm\n",
1957			     pool);
1958			return -EPERM;
1959		}
1960		if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1961			dout("ceph_pool_perm_check pool %lld no write perm\n",
1962			     pool);
1963			return -EPERM;
1964		}
1965		return 0;
1966	}
1967
1968	pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1969	ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1970	ceph_put_string(pool_ns);
1971	if (ret < 0)
1972		return ret;
1973
1974	flags = CEPH_I_POOL_PERM;
1975	if (ret & POOL_READ)
1976		flags |= CEPH_I_POOL_RD;
1977	if (ret & POOL_WRITE)
1978		flags |= CEPH_I_POOL_WR;
1979
1980	spin_lock(&ci->i_ceph_lock);
1981	if (pool == ci->i_layout.pool_id &&
1982	    pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
1983		ci->i_ceph_flags |= flags;
1984        } else {
1985		pool = ci->i_layout.pool_id;
1986		flags = ci->i_ceph_flags;
1987	}
1988	spin_unlock(&ci->i_ceph_lock);
1989	goto check;
1990}
1991
1992void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1993{
1994	struct ceph_pool_perm *perm;
1995	struct rb_node *n;
1996
1997	while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1998		n = rb_first(&mdsc->pool_perm_tree);
1999		perm = rb_entry(n, struct ceph_pool_perm, node);
2000		rb_erase(n, &mdsc->pool_perm_tree);
2001		kfree(perm);
2002	}
2003}