Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Apr 14-17, 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}
v6.8
   1// SPDX-License-Identifier: GPL-2.0
   2#include <linux/ceph/ceph_debug.h>
   3
   4#include <linux/backing-dev.h>
   5#include <linux/fs.h>
   6#include <linux/mm.h>
   7#include <linux/swap.h>
   8#include <linux/pagemap.h>
 
   9#include <linux/slab.h>
  10#include <linux/pagevec.h>
  11#include <linux/task_io_accounting_ops.h>
  12#include <linux/signal.h>
  13#include <linux/iversion.h>
  14#include <linux/ktime.h>
  15#include <linux/netfs.h>
  16
  17#include "super.h"
  18#include "mds_client.h"
  19#include "cache.h"
  20#include "metric.h"
  21#include "crypto.h"
  22#include <linux/ceph/osd_client.h>
  23#include <linux/ceph/striper.h>
  24
  25/*
  26 * Ceph address space ops.
  27 *
  28 * There are a few funny things going on here.
  29 *
  30 * The page->private field is used to reference a struct
  31 * ceph_snap_context for _every_ dirty page.  This indicates which
  32 * snapshot the page was logically dirtied in, and thus which snap
  33 * context needs to be associated with the osd write during writeback.
  34 *
  35 * Similarly, struct ceph_inode_info maintains a set of counters to
  36 * count dirty pages on the inode.  In the absence of snapshots,
  37 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  38 *
  39 * When a snapshot is taken (that is, when the client receives
  40 * notification that a snapshot was taken), each inode with caps and
  41 * with dirty pages (dirty pages implies there is a cap) gets a new
  42 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  43 * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  44 * moved to capsnap->dirty. (Unless a sync write is currently in
  45 * progress.  In that case, the capsnap is said to be "pending", new
  46 * writes cannot start, and the capsnap isn't "finalized" until the
  47 * write completes (or fails) and a final size/mtime for the inode for
  48 * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  49 *
  50 * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  51 * we look for the first capsnap in i_cap_snaps and write out pages in
  52 * that snap context _only_.  Then we move on to the next capsnap,
  53 * eventually reaching the "live" or "head" context (i.e., pages that
  54 * are not yet snapped) and are writing the most recently dirtied
  55 * pages.
  56 *
  57 * Invalidate and so forth must take care to ensure the dirty page
  58 * accounting is preserved.
  59 */
  60
  61#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  62#define CONGESTION_OFF_THRESH(congestion_kb)				\
  63	(CONGESTION_ON_THRESH(congestion_kb) -				\
  64	 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  65
  66static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
  67					struct folio **foliop, void **_fsdata);
  68
  69static inline struct ceph_snap_context *page_snap_context(struct page *page)
  70{
  71	if (PagePrivate(page))
  72		return (void *)page->private;
  73	return NULL;
  74}
  75
  76/*
  77 * Dirty a page.  Optimistically adjust accounting, on the assumption
  78 * that we won't race with invalidate.  If we do, readjust.
  79 */
  80static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
  81{
  82	struct inode *inode = mapping->host;
  83	struct ceph_client *cl = ceph_inode_to_client(inode);
  84	struct ceph_inode_info *ci;
 
  85	struct ceph_snap_context *snapc;
  86
  87	if (folio_test_dirty(folio)) {
  88		doutc(cl, "%llx.%llx %p idx %lu -- already dirty\n",
  89		      ceph_vinop(inode), folio, folio->index);
  90		VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
  91		return false;
 
 
  92	}
  93
 
  94	ci = ceph_inode(inode);
  95
 
 
 
 
 
 
  96	/* dirty the head */
  97	spin_lock(&ci->i_ceph_lock);
  98	BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  99	if (__ceph_have_pending_cap_snap(ci)) {
 100		struct ceph_cap_snap *capsnap =
 101				list_last_entry(&ci->i_cap_snaps,
 102						struct ceph_cap_snap,
 103						ci_item);
 104		snapc = ceph_get_snap_context(capsnap->context);
 105		capsnap->dirty_pages++;
 106	} else {
 107		BUG_ON(!ci->i_head_snapc);
 108		snapc = ceph_get_snap_context(ci->i_head_snapc);
 109		++ci->i_wrbuffer_ref_head;
 110	}
 111	if (ci->i_wrbuffer_ref == 0)
 112		ihold(inode);
 113	++ci->i_wrbuffer_ref;
 114	doutc(cl, "%llx.%llx %p idx %lu head %d/%d -> %d/%d "
 115	      "snapc %p seq %lld (%d snaps)\n",
 116	      ceph_vinop(inode), folio, folio->index,
 117	      ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 118	      ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 119	      snapc, snapc->seq, snapc->num_snaps);
 120	spin_unlock(&ci->i_ceph_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 121
 122	/*
 123	 * Reference snap context in folio->private.  Also set
 124	 * PagePrivate so that we get invalidate_folio callback.
 125	 */
 126	VM_WARN_ON_FOLIO(folio->private, folio);
 127	folio_attach_private(folio, snapc);
 128
 129	return ceph_fscache_dirty_folio(mapping, folio);
 
 130}
 131
 132/*
 133 * If we are truncating the full folio (i.e. offset == 0), adjust the
 134 * dirty folio counters appropriately.  Only called if there is private
 135 * data on the folio.
 136 */
 137static void ceph_invalidate_folio(struct folio *folio, size_t offset,
 138				size_t length)
 139{
 140	struct inode *inode = folio->mapping->host;
 141	struct ceph_client *cl = ceph_inode_to_client(inode);
 142	struct ceph_inode_info *ci = ceph_inode(inode);
 143	struct ceph_snap_context *snapc;
 
 
 
 
 144
 
 145
 146	if (offset != 0 || length != folio_size(folio)) {
 147		doutc(cl, "%llx.%llx idx %lu partial dirty page %zu~%zu\n",
 148		      ceph_vinop(inode), folio->index, offset, length);
 149		return;
 150	}
 
 
 151
 152	WARN_ON(!folio_test_locked(folio));
 153	if (folio_test_private(folio)) {
 154		doutc(cl, "%llx.%llx idx %lu full dirty page\n",
 155		      ceph_vinop(inode), folio->index);
 156
 157		snapc = folio_detach_private(folio);
 
 
 
 158		ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 159		ceph_put_snap_context(snapc);
 
 
 
 
 
 160	}
 161
 162	netfs_invalidate_folio(folio, offset, length);
 163}
 164
 165static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
 
 166{
 167	struct inode *inode = rreq->inode;
 168	struct ceph_inode_info *ci = ceph_inode(inode);
 169	struct ceph_file_layout *lo = &ci->i_layout;
 170	unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
 171	loff_t end = rreq->start + rreq->len, new_end;
 172	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
 173	unsigned long max_len;
 174	u32 blockoff;
 175
 176	if (priv) {
 177		/* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
 178		if (priv->file_ra_disabled)
 179			max_pages = 0;
 180		else
 181			max_pages = priv->file_ra_pages;
 182
 183	}
 184
 185	/* Readahead is disabled */
 186	if (!max_pages)
 187		return;
 188
 189	max_len = max_pages << PAGE_SHIFT;
 190
 191	/*
 192	 * Try to expand the length forward by rounding up it to the next
 193	 * block, but do not exceed the file size, unless the original
 194	 * request already exceeds it.
 195	 */
 196	new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
 197	if (new_end > end && new_end <= rreq->start + max_len)
 198		rreq->len = new_end - rreq->start;
 199
 200	/* Try to expand the start downward */
 201	div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
 202	if (rreq->len + blockoff <= max_len) {
 203		rreq->start -= blockoff;
 204		rreq->len += blockoff;
 205	}
 206}
 207
 208static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
 
 
 
 209{
 210	struct inode *inode = subreq->rreq->inode;
 211	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 212	struct ceph_inode_info *ci = ceph_inode(inode);
 213	u64 objno, objoff;
 214	u32 xlen;
 215
 216	/* Truncate the extent at the end of the current block */
 217	ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
 218				      &objno, &objoff, &xlen);
 219	subreq->len = min(xlen, fsc->mount_options->rsize);
 220	return true;
 221}
 222
 223static void finish_netfs_read(struct ceph_osd_request *req)
 224{
 225	struct inode *inode = req->r_inode;
 226	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 227	struct ceph_client *cl = fsc->client;
 228	struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 229	struct netfs_io_subrequest *subreq = req->r_priv;
 230	struct ceph_osd_req_op *op = &req->r_ops[0];
 231	int err = req->r_result;
 232	bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
 233
 234	ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
 235				 req->r_end_latency, osd_data->length, err);
 236
 237	doutc(cl, "result %d subreq->len=%zu i_size=%lld\n", req->r_result,
 238	      subreq->len, i_size_read(req->r_inode));
 239
 240	/* no object means success but no data */
 241	if (err == -ENOENT)
 242		err = 0;
 243	else if (err == -EBLOCKLISTED)
 244		fsc->blocklisted = true;
 245
 246	if (err >= 0) {
 247		if (sparse && err > 0)
 248			err = ceph_sparse_ext_map_end(op);
 249		if (err < subreq->len)
 250			__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 251		if (IS_ENCRYPTED(inode) && err > 0) {
 252			err = ceph_fscrypt_decrypt_extents(inode,
 253					osd_data->pages, subreq->start,
 254					op->extent.sparse_ext,
 255					op->extent.sparse_ext_cnt);
 256			if (err > subreq->len)
 257				err = subreq->len;
 258		}
 259	}
 
 260
 261	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
 262		ceph_put_page_vector(osd_data->pages,
 263				     calc_pages_for(osd_data->alignment,
 264					osd_data->length), false);
 265	}
 266	netfs_subreq_terminated(subreq, err, false);
 267	iput(req->r_inode);
 268	ceph_dec_osd_stopping_blocker(fsc->mdsc);
 269}
 270
 271static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
 272{
 273	struct netfs_io_request *rreq = subreq->rreq;
 274	struct inode *inode = rreq->inode;
 275	struct ceph_mds_reply_info_parsed *rinfo;
 276	struct ceph_mds_reply_info_in *iinfo;
 277	struct ceph_mds_request *req;
 278	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 279	struct ceph_inode_info *ci = ceph_inode(inode);
 280	struct iov_iter iter;
 281	ssize_t err = 0;
 282	size_t len;
 283	int mode;
 284
 285	__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 286	__clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
 287
 288	if (subreq->start >= inode->i_size)
 289		goto out;
 290
 291	/* We need to fetch the inline data. */
 292	mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
 293	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
 294	if (IS_ERR(req)) {
 295		err = PTR_ERR(req);
 296		goto out;
 297	}
 298	req->r_ino1 = ci->i_vino;
 299	req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
 300	req->r_num_caps = 2;
 301
 302	err = ceph_mdsc_do_request(mdsc, NULL, req);
 303	if (err < 0)
 304		goto out;
 305
 306	rinfo = &req->r_reply_info;
 307	iinfo = &rinfo->targeti;
 308	if (iinfo->inline_version == CEPH_INLINE_NONE) {
 309		/* The data got uninlined */
 310		ceph_mdsc_put_request(req);
 311		return false;
 312	}
 313
 314	len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
 315	iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
 316	err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
 317	if (err == 0)
 318		err = -EFAULT;
 319
 320	ceph_mdsc_put_request(req);
 321out:
 322	netfs_subreq_terminated(subreq, err, false);
 323	return true;
 324}
 325
 326static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 
 
 
 
 327{
 328	struct netfs_io_request *rreq = subreq->rreq;
 329	struct inode *inode = rreq->inode;
 330	struct ceph_inode_info *ci = ceph_inode(inode);
 331	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 332	struct ceph_client *cl = fsc->client;
 333	struct ceph_osd_request *req = NULL;
 334	struct ceph_vino vino = ceph_vino(inode);
 335	struct iov_iter iter;
 336	int err = 0;
 337	u64 len = subreq->len;
 338	bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
 339	u64 off = subreq->start;
 340	int extent_cnt;
 341
 342	if (ceph_inode_is_shutdown(inode)) {
 343		err = -EIO;
 344		goto out;
 345	}
 346
 347	if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
 348		return;
 349
 350	ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
 351
 352	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
 353			off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
 354			CEPH_OSD_FLAG_READ, NULL, ci->i_truncate_seq,
 355			ci->i_truncate_size, false);
 356	if (IS_ERR(req)) {
 357		err = PTR_ERR(req);
 358		req = NULL;
 359		goto out;
 360	}
 361
 362	if (sparse) {
 363		extent_cnt = __ceph_sparse_read_ext_count(inode, len);
 364		err = ceph_alloc_sparse_ext_map(&req->r_ops[0], extent_cnt);
 365		if (err)
 366			goto out;
 367	}
 368
 369	doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
 370	      ceph_vinop(inode), subreq->start, subreq->len, len);
 371
 372	iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
 373
 374	/*
 375	 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
 376	 * encrypted inodes. We'd need infrastructure that handles an iov_iter
 377	 * instead of page arrays, and we don't have that as of yet. Once the
 378	 * dust settles on the write helpers and encrypt/decrypt routines for
 379	 * netfs, we should be able to rework this.
 380	 */
 381	if (IS_ENCRYPTED(inode)) {
 382		struct page **pages;
 383		size_t page_off;
 384
 385		err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
 386		if (err < 0) {
 387			doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
 388			      ceph_vinop(inode), err);
 389			goto out;
 390		}
 391
 392		/* should always give us a page-aligned read */
 393		WARN_ON_ONCE(page_off);
 394		len = err;
 395		err = 0;
 396
 397		osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
 398						 false);
 399	} else {
 400		osd_req_op_extent_osd_iter(req, 0, &iter);
 401	}
 402	if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
 403		err = -EIO;
 404		goto out;
 405	}
 406	req->r_callback = finish_netfs_read;
 407	req->r_priv = subreq;
 408	req->r_inode = inode;
 409	ihold(inode);
 410
 411	ceph_osdc_start_request(req->r_osdc, req);
 412out:
 413	ceph_osdc_put_request(req);
 414	if (err)
 415		netfs_subreq_terminated(subreq, err, false);
 416	doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
 417}
 418
 419static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 
 
 
 
 
 420{
 421	struct inode *inode = rreq->inode;
 422	struct ceph_client *cl = ceph_inode_to_client(inode);
 423	int got = 0, want = CEPH_CAP_FILE_CACHE;
 424	struct ceph_netfs_request_data *priv;
 425	int ret = 0;
 
 
 
 426
 427	if (rreq->origin != NETFS_READAHEAD)
 428		return 0;
 429
 430	priv = kzalloc(sizeof(*priv), GFP_NOFS);
 431	if (!priv)
 432		return -ENOMEM;
 433
 434	if (file) {
 435		struct ceph_rw_context *rw_ctx;
 436		struct ceph_file_info *fi = file->private_data;
 437
 438		priv->file_ra_pages = file->f_ra.ra_pages;
 439		priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
 440
 441		rw_ctx = ceph_find_rw_context(fi);
 442		if (rw_ctx) {
 443			rreq->netfs_priv = priv;
 444			return 0;
 445		}
 446	}
 447
 448	/*
 449	 * readahead callers do not necessarily hold Fcb caps
 450	 * (e.g. fadvise, madvise).
 451	 */
 452	ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 453	if (ret < 0) {
 454		doutc(cl, "%llx.%llx, error getting cap\n", ceph_vinop(inode));
 
 
 
 
 
 
 
 455		goto out;
 456	}
 457
 458	if (!(got & want)) {
 459		doutc(cl, "%llx.%llx, no cache cap\n", ceph_vinop(inode));
 460		ret = -EACCES;
 461		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 462	}
 463	if (ret == 0) {
 464		ret = -EACCES;
 465		goto out;
 466	}
 467
 468	priv->caps = got;
 469	rreq->netfs_priv = priv;
 470
 471out:
 472	if (ret < 0)
 473		kfree(priv);
 474
 475	return ret;
 476}
 477
 478static void ceph_netfs_free_request(struct netfs_io_request *rreq)
 479{
 480	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
 481
 482	if (!priv)
 483		return;
 484
 485	if (priv->caps)
 486		ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
 487	kfree(priv);
 488	rreq->netfs_priv = NULL;
 489}
 490
 491const struct netfs_request_ops ceph_netfs_ops = {
 492	.init_request		= ceph_init_request,
 493	.free_request		= ceph_netfs_free_request,
 494	.issue_read		= ceph_netfs_issue_read,
 495	.expand_readahead	= ceph_netfs_expand_readahead,
 496	.clamp_length		= ceph_netfs_clamp_length,
 497	.check_write_begin	= ceph_netfs_check_write_begin,
 498};
 499
 500#ifdef CONFIG_CEPH_FSCACHE
 501static void ceph_set_page_fscache(struct page *page)
 502{
 503	set_page_fscache(page);
 504}
 505
 506static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
 507{
 508	struct inode *inode = priv;
 509
 510	if (IS_ERR_VALUE(error) && error != -ENOBUFS)
 511		ceph_fscache_invalidate(inode, false);
 512}
 513
 514static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 515{
 516	struct ceph_inode_info *ci = ceph_inode(inode);
 517	struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
 518
 519	fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
 520			       ceph_fscache_write_terminated, inode, caching);
 521}
 522#else
 523static inline void ceph_set_page_fscache(struct page *page)
 524{
 525}
 526
 527static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 528{
 529}
 530#endif /* CONFIG_CEPH_FSCACHE */
 531
 532struct ceph_writeback_ctl
 533{
 534	loff_t i_size;
 535	u64 truncate_size;
 536	u32 truncate_seq;
 537	bool size_stable;
 538	bool head_snapc;
 539};
 540
 541/*
 542 * Get ref for the oldest snapc for an inode with dirty data... that is, the
 543 * only snap context we are allowed to write back.
 544 */
 545static struct ceph_snap_context *
 546get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 547		   struct ceph_snap_context *page_snapc)
 548{
 549	struct ceph_inode_info *ci = ceph_inode(inode);
 550	struct ceph_client *cl = ceph_inode_to_client(inode);
 551	struct ceph_snap_context *snapc = NULL;
 552	struct ceph_cap_snap *capsnap = NULL;
 553
 554	spin_lock(&ci->i_ceph_lock);
 555	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 556		doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
 557		      capsnap, capsnap->context, capsnap->dirty_pages);
 558		if (!capsnap->dirty_pages)
 559			continue;
 560
 561		/* get i_size, truncate_{seq,size} for page_snapc? */
 562		if (snapc && capsnap->context != page_snapc)
 563			continue;
 564
 565		if (ctl) {
 566			if (capsnap->writing) {
 567				ctl->i_size = i_size_read(inode);
 568				ctl->size_stable = false;
 569			} else {
 570				ctl->i_size = capsnap->size;
 571				ctl->size_stable = true;
 572			}
 573			ctl->truncate_size = capsnap->truncate_size;
 574			ctl->truncate_seq = capsnap->truncate_seq;
 575			ctl->head_snapc = false;
 576		}
 577
 578		if (snapc)
 579			break;
 580
 581		snapc = ceph_get_snap_context(capsnap->context);
 582		if (!page_snapc ||
 583		    page_snapc == snapc ||
 584		    page_snapc->seq > snapc->seq)
 585			break;
 586	}
 587	if (!snapc && ci->i_wrbuffer_ref_head) {
 588		snapc = ceph_get_snap_context(ci->i_head_snapc);
 589		doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
 590		      ci->i_wrbuffer_ref_head);
 591		if (ctl) {
 592			ctl->i_size = i_size_read(inode);
 593			ctl->truncate_size = ci->i_truncate_size;
 594			ctl->truncate_seq = ci->i_truncate_seq;
 595			ctl->size_stable = false;
 596			ctl->head_snapc = true;
 597		}
 598	}
 599	spin_unlock(&ci->i_ceph_lock);
 600	return snapc;
 601}
 602
 603static u64 get_writepages_data_length(struct inode *inode,
 604				      struct page *page, u64 start)
 605{
 606	struct ceph_inode_info *ci = ceph_inode(inode);
 607	struct ceph_snap_context *snapc;
 608	struct ceph_cap_snap *capsnap = NULL;
 609	u64 end = i_size_read(inode);
 610	u64 ret;
 611
 612	snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
 613	if (snapc != ci->i_head_snapc) {
 614		bool found = false;
 615		spin_lock(&ci->i_ceph_lock);
 616		list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 617			if (capsnap->context == snapc) {
 618				if (!capsnap->writing)
 619					end = capsnap->size;
 620				found = true;
 621				break;
 622			}
 623		}
 624		spin_unlock(&ci->i_ceph_lock);
 625		WARN_ON(!found);
 626	}
 627	if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
 628		end = ceph_fscrypt_page_offset(page) + thp_size(page);
 629	ret = end > start ? end - start : 0;
 630	if (ret && fscrypt_is_bounce_page(page))
 631		ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
 632	return ret;
 633}
 634
 635/*
 636 * Write a single page, but leave the page locked.
 637 *
 638 * If we get a write error, mark the mapping for error, but still adjust the
 639 * dirty page accounting (i.e., page is no longer dirty).
 640 */
 641static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 642{
 643	struct folio *folio = page_folio(page);
 644	struct inode *inode = page->mapping->host;
 645	struct ceph_inode_info *ci = ceph_inode(inode);
 646	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 647	struct ceph_client *cl = fsc->client;
 
 
 
 648	struct ceph_snap_context *snapc, *oldest;
 649	loff_t page_off = page_offset(page);
 650	int err;
 651	loff_t len = thp_size(page);
 652	loff_t wlen;
 653	struct ceph_writeback_ctl ceph_wbc;
 654	struct ceph_osd_client *osdc = &fsc->client->osdc;
 655	struct ceph_osd_request *req;
 656	bool caching = ceph_is_cache_enabled(inode);
 657	struct page *bounce_page = NULL;
 658
 659	doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
 660	      page->index);
 661
 662	if (ceph_inode_is_shutdown(inode))
 663		return -EIO;
 
 
 
 
 
 
 664
 665	/* verify this is a writeable snap context */
 666	snapc = page_snap_context(page);
 667	if (!snapc) {
 668		doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
 669		      page);
 670		return 0;
 671	}
 672	oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 673	if (snapc->seq > oldest->seq) {
 674		doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
 675		      ceph_vinop(inode), page, snapc);
 676		/* we should only noop if called by kswapd */
 677		WARN_ON(!(current->flags & PF_MEMALLOC));
 678		ceph_put_snap_context(oldest);
 679		redirty_page_for_writepage(wbc, page);
 680		return 0;
 681	}
 682	ceph_put_snap_context(oldest);
 683
 684	/* is this a partial page at end of file? */
 685	if (page_off >= ceph_wbc.i_size) {
 686		doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
 687		      ceph_vinop(inode), folio->index, ceph_wbc.i_size);
 688		folio_invalidate(folio, 0, folio_size(folio));
 689		return 0;
 690	}
 691
 692	if (ceph_wbc.i_size < page_off + len)
 693		len = ceph_wbc.i_size - page_off;
 694
 695	wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
 696	doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 697	      ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
 698	      snapc->seq);
 699
 700	if (atomic_long_inc_return(&fsc->writeback_count) >
 701	    CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 702		fsc->write_congested = true;
 703
 704	req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
 705				    page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
 706				    CEPH_OSD_FLAG_WRITE, snapc,
 707				    ceph_wbc.truncate_seq,
 708				    ceph_wbc.truncate_size, true);
 709	if (IS_ERR(req)) {
 710		redirty_page_for_writepage(wbc, page);
 711		return PTR_ERR(req);
 712	}
 713
 714	if (wlen < len)
 715		len = wlen;
 716
 717	set_page_writeback(page);
 718	if (caching)
 719		ceph_set_page_fscache(page);
 720	ceph_fscache_write_to_cache(inode, page_off, len, caching);
 721
 722	if (IS_ENCRYPTED(inode)) {
 723		bounce_page = fscrypt_encrypt_pagecache_blocks(page,
 724						    CEPH_FSCRYPT_BLOCK_SIZE, 0,
 725						    GFP_NOFS);
 726		if (IS_ERR(bounce_page)) {
 727			redirty_page_for_writepage(wbc, page);
 728			end_page_writeback(page);
 729			ceph_osdc_put_request(req);
 730			return PTR_ERR(bounce_page);
 731		}
 732	}
 733
 734	/* it may be a short write due to an object boundary */
 735	WARN_ON_ONCE(len > thp_size(page));
 736	osd_req_op_extent_osd_data_pages(req, 0,
 737			bounce_page ? &bounce_page : &page, wlen, 0,
 738			false, false);
 739	doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
 740	      ceph_vinop(inode), page_off, len, wlen,
 741	      IS_ENCRYPTED(inode) ? "" : "not ");
 742
 743	req->r_mtime = inode_get_mtime(inode);
 744	ceph_osdc_start_request(osdc, req);
 745	err = ceph_osdc_wait_request(osdc, req);
 746
 747	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 748				  req->r_end_latency, len, err);
 749	fscrypt_free_bounce_page(bounce_page);
 750	ceph_osdc_put_request(req);
 751	if (err == 0)
 752		err = len;
 753
 754	if (err < 0) {
 755		struct writeback_control tmp_wbc;
 756		if (!wbc)
 757			wbc = &tmp_wbc;
 758		if (err == -ERESTARTSYS) {
 759			/* killed by SIGKILL */
 760			doutc(cl, "%llx.%llx interrupted page %p\n",
 761			      ceph_vinop(inode), page);
 762			redirty_page_for_writepage(wbc, page);
 763			end_page_writeback(page);
 764			return err;
 765		}
 766		if (err == -EBLOCKLISTED)
 767			fsc->blocklisted = true;
 768		doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
 769		      ceph_vinop(inode), err, page);
 770		mapping_set_error(&inode->i_data, err);
 771		wbc->pages_skipped++;
 
 772	} else {
 773		doutc(cl, "%llx.%llx cleaned page %p\n",
 774		      ceph_vinop(inode), page);
 775		err = 0;  /* vfs expects us to return 0 */
 776	}
 777	oldest = detach_page_private(page);
 778	WARN_ON_ONCE(oldest != snapc);
 779	end_page_writeback(page);
 780	ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 781	ceph_put_snap_context(snapc);  /* page's reference */
 782
 783	if (atomic_long_dec_return(&fsc->writeback_count) <
 784	    CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 785		fsc->write_congested = false;
 786
 787	return err;
 788}
 789
 790static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 791{
 792	int err;
 793	struct inode *inode = page->mapping->host;
 794	BUG_ON(!inode);
 795	ihold(inode);
 
 
 
 
 
 796
 797	if (wbc->sync_mode == WB_SYNC_NONE &&
 798	    ceph_inode_to_fs_client(inode)->write_congested)
 799		return AOP_WRITEPAGE_ACTIVATE;
 800
 801	wait_on_page_fscache(page);
 
 
 
 
 
 
 
 802
 803	err = writepage_nounlock(page, wbc);
 804	if (err == -ERESTARTSYS) {
 805		/* direct memory reclaimer was killed by SIGKILL. return 0
 806		 * to prevent caller from setting mapping/page error */
 807		err = 0;
 808	}
 809	unlock_page(page);
 810	iput(inode);
 811	return err;
 812}
 813
 
 814/*
 815 * async writeback completion handler.
 816 *
 817 * If we get an error, set the mapping error bit, but not the individual
 818 * page error bits.
 819 */
 820static void writepages_finish(struct ceph_osd_request *req)
 
 821{
 822	struct inode *inode = req->r_inode;
 
 
 823	struct ceph_inode_info *ci = ceph_inode(inode);
 824	struct ceph_client *cl = ceph_inode_to_client(inode);
 825	struct ceph_osd_data *osd_data;
 826	struct page *page;
 827	int num_pages, total_pages = 0;
 828	int i, j;
 829	int rc = req->r_result;
 830	struct ceph_snap_context *snapc = req->r_snapc;
 831	struct address_space *mapping = inode->i_mapping;
 832	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 833	unsigned int len = 0;
 834	bool remove_page;
 
 
 
 
 
 
 
 
 
 835
 836	doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
 837	if (rc < 0) {
 
 
 
 
 
 
 
 
 838		mapping_set_error(mapping, rc);
 839		ceph_set_error_write(ci);
 840		if (rc == -EBLOCKLISTED)
 841			fsc->blocklisted = true;
 842	} else {
 843		ceph_clear_error_write(ci);
 844	}
 845
 846	/*
 847	 * We lost the cache cap, need to truncate the page before
 848	 * it is unlocked, otherwise we'd truncate it later in the
 849	 * page truncation thread, possibly losing some data that
 850	 * raced its way in
 851	 */
 852	remove_page = !(ceph_caps_issued(ci) &
 853			(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 854
 855	/* clean all pages */
 856	for (i = 0; i < req->r_num_ops; i++) {
 857		if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
 858			pr_warn_client(cl,
 859				"%llx.%llx incorrect op %d req %p index %d tid %llu\n",
 860				ceph_vinop(inode), req->r_ops[i].op, req, i,
 861				req->r_tid);
 862			break;
 863		}
 
 
 
 
 
 
 
 
 
 864
 865		osd_data = osd_req_op_extent_osd_data(req, i);
 866		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 867		len += osd_data->length;
 868		num_pages = calc_pages_for((u64)osd_data->alignment,
 869					   (u64)osd_data->length);
 870		total_pages += num_pages;
 871		for (j = 0; j < num_pages; j++) {
 872			page = osd_data->pages[j];
 873			if (fscrypt_is_bounce_page(page)) {
 874				page = fscrypt_pagecache_page(page);
 875				fscrypt_free_bounce_page(osd_data->pages[j]);
 876				osd_data->pages[j] = page;
 877			}
 878			BUG_ON(!page);
 879			WARN_ON(!PageUptodate(page));
 880
 881			if (atomic_long_dec_return(&fsc->writeback_count) <
 882			     CONGESTION_OFF_THRESH(
 883					fsc->mount_options->congestion_kb))
 884				fsc->write_congested = false;
 885
 886			ceph_put_snap_context(detach_page_private(page));
 887			end_page_writeback(page);
 888			doutc(cl, "unlocking %p\n", page);
 889
 890			if (remove_page)
 891				generic_error_remove_folio(inode->i_mapping,
 892							  page_folio(page));
 893
 894			unlock_page(page);
 895		}
 896		doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
 897		      ceph_vinop(inode), osd_data->length,
 898		      rc >= 0 ? num_pages : 0);
 899
 900		release_pages(osd_data->pages, num_pages);
 901	}
 
 
 902
 903	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 904				  req->r_end_latency, len, rc);
 905
 906	ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 907
 908	osd_data = osd_req_op_extent_osd_data(req, 0);
 909	if (osd_data->pages_from_pool)
 910		mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 911	else
 912		kfree(osd_data->pages);
 913	ceph_osdc_put_request(req);
 914	ceph_dec_osd_stopping_blocker(fsc->mdsc);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 915}
 916
 917/*
 918 * initiate async writeback
 919 */
 920static int ceph_writepages_start(struct address_space *mapping,
 921				 struct writeback_control *wbc)
 922{
 923	struct inode *inode = mapping->host;
 924	struct ceph_inode_info *ci = ceph_inode(inode);
 925	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 926	struct ceph_client *cl = fsc->client;
 927	struct ceph_vino vino = ceph_vino(inode);
 928	pgoff_t index, start_index, end = -1;
 
 929	struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 930	struct folio_batch fbatch;
 
 931	int rc = 0;
 932	unsigned int wsize = i_blocksize(inode);
 933	struct ceph_osd_request *req = NULL;
 934	struct ceph_writeback_ctl ceph_wbc;
 935	bool should_loop, range_whole = false;
 936	bool done = false;
 937	bool caching = ceph_is_cache_enabled(inode);
 938	xa_mark_t tag;
 939
 940	if (wbc->sync_mode == WB_SYNC_NONE &&
 941	    fsc->write_congested)
 942		return 0;
 943
 944	doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
 945	      wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 946	      (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 947
 948	if (ceph_inode_is_shutdown(inode)) {
 949		if (ci->i_wrbuffer_ref > 0) {
 950			pr_warn_ratelimited_client(cl,
 951				"%llx.%llx %lld forced umount\n",
 952				ceph_vinop(inode), ceph_ino(inode));
 953		}
 954		mapping_set_error(mapping, -EIO);
 
 955		return -EIO; /* we're in a forced umount, don't write! */
 956	}
 957	if (fsc->mount_options->wsize < wsize)
 958		wsize = fsc->mount_options->wsize;
 959
 960	folio_batch_init(&fbatch);
 961
 962	start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 963	index = start_index;
 964
 965	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
 966		tag = PAGECACHE_TAG_TOWRITE;
 
 
 
 967	} else {
 968		tag = PAGECACHE_TAG_DIRTY;
 
 
 
 
 
 969	}
 
 
 970retry:
 971	/* find oldest snap context with dirty data */
 972	snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 
 973	if (!snapc) {
 974		/* hmm, why does writepages get called when there
 975		   is no dirty data? */
 976		doutc(cl, " no snap context with dirty data?\n");
 977		goto out;
 978	}
 979	doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
 980	      snapc->seq, snapc->num_snaps);
 981
 982	should_loop = false;
 983	if (ceph_wbc.head_snapc && snapc != last_snapc) {
 984		/* where to start/end? */
 985		if (wbc->range_cyclic) {
 986			index = start_index;
 987			end = -1;
 988			if (index > 0)
 989				should_loop = true;
 990			doutc(cl, " cyclic, start at %lu\n", index);
 991		} else {
 992			index = wbc->range_start >> PAGE_SHIFT;
 993			end = wbc->range_end >> PAGE_SHIFT;
 994			if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 995				range_whole = true;
 996			doutc(cl, " not cyclic, %lu to %lu\n", index, end);
 997		}
 998	} else if (!ceph_wbc.head_snapc) {
 999		/* Do not respect wbc->range_{start,end}. Dirty pages
1000		 * in that range can be associated with newer snapc.
1001		 * They are not writeable until we write all dirty pages
1002		 * associated with 'snapc' get written */
1003		if (index > 0)
1004			should_loop = true;
1005		doutc(cl, " non-head snapc, range whole\n");
1006	}
1007
1008	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1009		tag_pages_for_writeback(mapping, index, end);
1010
1011	ceph_put_snap_context(last_snapc);
1012	last_snapc = snapc;
1013
1014	while (!done && index <= end) {
1015		int num_ops = 0, op_idx;
1016		unsigned i, nr_folios, max_pages, locked_pages = 0;
1017		struct page **pages = NULL, **data_pages;
 
1018		struct page *page;
1019		pgoff_t strip_unit_end = 0;
1020		u64 offset = 0, len = 0;
1021		bool from_pool = false;
1022
1023		max_pages = wsize >> PAGE_SHIFT;
 
 
 
 
1024
1025get_more_pages:
1026		nr_folios = filemap_get_folios_tag(mapping, &index,
1027						   end, tag, &fbatch);
1028		doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1029		if (!nr_folios && !locked_pages)
 
 
 
 
 
 
1030			break;
1031		for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1032			page = &fbatch.folios[i]->page;
1033			doutc(cl, "? %p idx %lu\n", page, page->index);
1034			if (locked_pages == 0)
1035				lock_page(page);  /* first page */
1036			else if (!trylock_page(page))
1037				break;
1038
1039			/* only dirty pages, or our accounting breaks */
1040			if (unlikely(!PageDirty(page)) ||
1041			    unlikely(page->mapping != mapping)) {
1042				doutc(cl, "!dirty or !mapping %p\n", page);
1043				unlock_page(page);
1044				continue;
 
 
 
 
 
 
1045			}
1046			/* only if matching snap context */
1047			pgsnapc = page_snap_context(page);
1048			if (pgsnapc != snapc) {
1049				doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1050				      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1051				if (!should_loop &&
1052				    !ceph_wbc.head_snapc &&
1053				    wbc->sync_mode != WB_SYNC_NONE)
1054					should_loop = true;
1055				unlock_page(page);
1056				continue;
 
 
 
 
1057			}
1058			if (page_offset(page) >= ceph_wbc.i_size) {
1059				struct folio *folio = page_folio(page);
1060
1061				doutc(cl, "folio at %lu beyond eof %llu\n",
1062				      folio->index, ceph_wbc.i_size);
1063				if ((ceph_wbc.size_stable ||
1064				    folio_pos(folio) >= i_size_read(inode)) &&
1065				    folio_clear_dirty_for_io(folio))
1066					folio_invalidate(folio, 0,
1067							folio_size(folio));
1068				folio_unlock(folio);
1069				continue;
1070			}
1071			if (strip_unit_end && (page->index > strip_unit_end)) {
1072				doutc(cl, "end of strip unit %p\n", page);
1073				unlock_page(page);
1074				break;
1075			}
1076			if (PageWriteback(page) || PageFsCache(page)) {
1077				if (wbc->sync_mode == WB_SYNC_NONE) {
1078					doutc(cl, "%p under writeback\n", page);
1079					unlock_page(page);
1080					continue;
1081				}
1082				doutc(cl, "waiting on writeback %p\n", page);
1083				wait_on_page_writeback(page);
1084				wait_on_page_fscache(page);
 
1085			}
1086
1087			if (!clear_page_dirty_for_io(page)) {
1088				doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1089				unlock_page(page);
1090				continue;
1091			}
1092
1093			/*
1094			 * We have something to write.  If this is
1095			 * the first locked page this time through,
1096			 * calculate max possinle write size and
1097			 * allocate a page array
1098			 */
1099			if (locked_pages == 0) {
1100				u64 objnum;
1101				u64 objoff;
1102				u32 xlen;
1103
1104				/* prepare async write request */
1105				offset = (u64)page_offset(page);
1106				ceph_calc_file_object_mapping(&ci->i_layout,
1107							      offset, wsize,
1108							      &objnum, &objoff,
1109							      &xlen);
1110				len = xlen;
1111
1112				num_ops = 1;
1113				strip_unit_end = page->index +
1114					((len - 1) >> PAGE_SHIFT);
1115
1116				BUG_ON(pages);
1117				max_pages = calc_pages_for(0, (u64)len);
1118				pages = kmalloc_array(max_pages,
1119						      sizeof(*pages),
1120						      GFP_NOFS);
1121				if (!pages) {
1122					from_pool = true;
1123					pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1124					BUG_ON(!pages);
1125				}
1126
1127				len = 0;
1128			} else if (page->index !=
1129				   (offset + len) >> PAGE_SHIFT) {
1130				if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1131							     CEPH_OSD_MAX_OPS)) {
1132					redirty_page_for_writepage(wbc, page);
1133					unlock_page(page);
1134					break;
1135				}
1136
1137				num_ops++;
1138				offset = (u64)page_offset(page);
1139				len = 0;
 
 
1140			}
1141
1142			/* note position of first page in fbatch */
1143			doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1144			      ceph_vinop(inode), page, page->index);
1145
1146			if (atomic_long_inc_return(&fsc->writeback_count) >
1147			    CONGESTION_ON_THRESH(
1148				    fsc->mount_options->congestion_kb))
1149				fsc->write_congested = true;
1150
1151			if (IS_ENCRYPTED(inode)) {
1152				pages[locked_pages] =
1153					fscrypt_encrypt_pagecache_blocks(page,
1154						PAGE_SIZE, 0,
1155						locked_pages ? GFP_NOWAIT : GFP_NOFS);
1156				if (IS_ERR(pages[locked_pages])) {
1157					if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1158						pr_err_client(cl,
1159							"inode->i_blkbits=%hhu\n",
1160							inode->i_blkbits);
1161					/* better not fail on first page! */
1162					BUG_ON(locked_pages == 0);
1163					pages[locked_pages] = NULL;
1164					redirty_page_for_writepage(wbc, page);
1165					unlock_page(page);
1166					break;
1167				}
1168				++locked_pages;
1169			} else {
1170				pages[locked_pages++] = page;
1171			}
1172
1173			fbatch.folios[i] = NULL;
1174			len += thp_size(page);
 
 
1175		}
1176
1177		/* did we get anything? */
1178		if (!locked_pages)
1179			goto release_folios;
1180		if (i) {
1181			unsigned j, n = 0;
1182			/* shift unused page to beginning of fbatch */
1183			for (j = 0; j < nr_folios; j++) {
1184				if (!fbatch.folios[j])
1185					continue;
1186				if (n < j)
1187					fbatch.folios[n] = fbatch.folios[j];
1188				n++;
1189			}
1190			fbatch.nr = n;
1191
1192			if (nr_folios && i == nr_folios &&
1193			    locked_pages < max_pages) {
1194				doutc(cl, "reached end fbatch, trying for more\n");
1195				folio_batch_release(&fbatch);
1196				goto get_more_pages;
1197			}
1198		}
1199
1200new_request:
1201		offset = ceph_fscrypt_page_offset(pages[0]);
1202		len = wsize;
1203
1204		req = ceph_osdc_new_request(&fsc->client->osdc,
1205					&ci->i_layout, vino,
1206					offset, &len, 0, num_ops,
1207					CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1208					snapc, ceph_wbc.truncate_seq,
1209					ceph_wbc.truncate_size, false);
1210		if (IS_ERR(req)) {
1211			req = ceph_osdc_new_request(&fsc->client->osdc,
1212						&ci->i_layout, vino,
1213						offset, &len, 0,
1214						min(num_ops,
1215						    CEPH_OSD_SLAB_OPS),
1216						CEPH_OSD_OP_WRITE,
1217						CEPH_OSD_FLAG_WRITE,
1218						snapc, ceph_wbc.truncate_seq,
1219						ceph_wbc.truncate_size, true);
1220			BUG_ON(IS_ERR(req));
1221		}
1222		BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1223			     thp_size(pages[locked_pages - 1]) - offset);
1224
1225		if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1226			rc = -EIO;
1227			goto release_folios;
1228		}
1229		req->r_callback = writepages_finish;
1230		req->r_inode = inode;
1231
1232		/* Format the osd request message and submit the write */
1233		len = 0;
1234		data_pages = pages;
1235		op_idx = 0;
1236		for (i = 0; i < locked_pages; i++) {
1237			struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1238
1239			u64 cur_offset = page_offset(page);
1240			/*
1241			 * Discontinuity in page range? Ceph can handle that by just passing
1242			 * multiple extents in the write op.
1243			 */
1244			if (offset + len != cur_offset) {
1245				/* If it's full, stop here */
1246				if (op_idx + 1 == req->r_num_ops)
1247					break;
1248
1249				/* Kick off an fscache write with what we have so far. */
1250				ceph_fscache_write_to_cache(inode, offset, len, caching);
1251
1252				/* Start a new extent */
1253				osd_req_op_extent_dup_last(req, op_idx,
1254							   cur_offset - offset);
1255				doutc(cl, "got pages at %llu~%llu\n", offset,
1256				      len);
1257				osd_req_op_extent_osd_data_pages(req, op_idx,
1258							data_pages, len, 0,
1259							from_pool, false);
1260				osd_req_op_extent_update(req, op_idx, len);
1261
1262				len = 0;
1263				offset = cur_offset;
1264				data_pages = pages + i;
1265				op_idx++;
1266			}
1267
1268			set_page_writeback(page);
1269			if (caching)
1270				ceph_set_page_fscache(page);
1271			len += thp_size(page);
1272		}
1273		ceph_fscache_write_to_cache(inode, offset, len, caching);
1274
1275		if (ceph_wbc.size_stable) {
1276			len = min(len, ceph_wbc.i_size - offset);
1277		} else if (i == locked_pages) {
1278			/* writepages_finish() clears writeback pages
1279			 * according to the data length, so make sure
1280			 * data length covers all locked pages */
1281			u64 min_len = len + 1 - thp_size(page);
1282			len = get_writepages_data_length(inode, pages[i - 1],
1283							 offset);
1284			len = max(len, min_len);
1285		}
1286		if (IS_ENCRYPTED(inode))
1287			len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1288
1289		doutc(cl, "got pages at %llu~%llu\n", offset, len);
1290
1291		if (IS_ENCRYPTED(inode) &&
1292		    ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1293			pr_warn_client(cl,
1294				"bad encrypted write offset=%lld len=%llu\n",
1295				offset, len);
1296
1297		osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1298						 0, from_pool, false);
1299		osd_req_op_extent_update(req, op_idx, len);
1300
1301		BUG_ON(op_idx + 1 != req->r_num_ops);
1302
1303		from_pool = false;
1304		if (i < locked_pages) {
1305			BUG_ON(num_ops <= req->r_num_ops);
1306			num_ops -= req->r_num_ops;
1307			locked_pages -= i;
1308
1309			/* allocate new pages array for next request */
1310			data_pages = pages;
1311			pages = kmalloc_array(locked_pages, sizeof(*pages),
1312					      GFP_NOFS);
1313			if (!pages) {
1314				from_pool = true;
1315				pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1316				BUG_ON(!pages);
1317			}
1318			memcpy(pages, data_pages + i,
1319			       locked_pages * sizeof(*pages));
1320			memset(data_pages + i, 0,
1321			       locked_pages * sizeof(*pages));
1322		} else {
1323			BUG_ON(num_ops != req->r_num_ops);
1324			index = pages[i - 1]->index + 1;
1325			/* request message now owns the pages array */
1326			pages = NULL;
1327		}
1328
1329		req->r_mtime = inode_get_mtime(inode);
1330		ceph_osdc_start_request(&fsc->client->osdc, req);
1331		req = NULL;
1332
1333		wbc->nr_to_write -= i;
1334		if (pages)
1335			goto new_request;
 
 
 
 
 
 
 
1336
1337		/*
1338		 * We stop writing back only if we are not doing
1339		 * integrity sync. In case of integrity sync we have to
1340		 * keep going until we have written all the pages
1341		 * we tagged for writeback prior to entering this loop.
1342		 */
1343		if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1344			done = true;
1345
1346release_folios:
1347		doutc(cl, "folio_batch release on %d folios (%p)\n",
1348		      (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1349		folio_batch_release(&fbatch);
1350	}
1351
1352	if (should_loop && !done) {
1353		/* more to do; loop back to beginning of file */
1354		doutc(cl, "looping back to beginning of file\n");
1355		end = start_index - 1; /* OK even when start_index == 0 */
1356
1357		/* to write dirty pages associated with next snapc,
1358		 * we need to wait until current writes complete */
1359		if (wbc->sync_mode != WB_SYNC_NONE &&
1360		    start_index == 0 && /* all dirty pages were checked */
1361		    !ceph_wbc.head_snapc) {
1362			struct page *page;
1363			unsigned i, nr;
1364			index = 0;
1365			while ((index <= end) &&
1366			       (nr = filemap_get_folios_tag(mapping, &index,
1367						(pgoff_t)-1,
1368						PAGECACHE_TAG_WRITEBACK,
1369						&fbatch))) {
1370				for (i = 0; i < nr; i++) {
1371					page = &fbatch.folios[i]->page;
1372					if (page_snap_context(page) != snapc)
1373						continue;
1374					wait_on_page_writeback(page);
1375				}
1376				folio_batch_release(&fbatch);
1377				cond_resched();
1378			}
1379		}
1380
1381		start_index = 0;
1382		index = 0;
1383		goto retry;
1384	}
1385
1386	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1387		mapping->writeback_index = index;
1388
1389out:
1390	ceph_osdc_put_request(req);
1391	ceph_put_snap_context(last_snapc);
1392	doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1393	      rc);
1394	return rc;
1395}
1396
1397
1398
1399/*
1400 * See if a given @snapc is either writeable, or already written.
1401 */
1402static int context_is_writeable_or_written(struct inode *inode,
1403					   struct ceph_snap_context *snapc)
1404{
1405	struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1406	int ret = !oldest || snapc->seq <= oldest->seq;
1407
1408	ceph_put_snap_context(oldest);
1409	return ret;
1410}
1411
1412/**
1413 * ceph_find_incompatible - find an incompatible context and return it
1414 * @page: page being dirtied
1415 *
1416 * We are only allowed to write into/dirty a page if the page is
1417 * clean, or already dirty within the same snap context. Returns a
1418 * conflicting context if there is one, NULL if there isn't, or a
1419 * negative error code on other errors.
1420 *
1421 * Must be called with page lock held.
 
 
1422 */
1423static struct ceph_snap_context *
1424ceph_find_incompatible(struct page *page)
 
1425{
1426	struct inode *inode = page->mapping->host;
1427	struct ceph_client *cl = ceph_inode_to_client(inode);
1428	struct ceph_inode_info *ci = ceph_inode(inode);
 
 
 
 
 
 
 
1429
1430	if (ceph_inode_is_shutdown(inode)) {
1431		doutc(cl, " %llx.%llx page %p is shutdown\n",
1432		      ceph_vinop(inode), page);
1433		return ERR_PTR(-ESTALE);
1434	}
1435
1436	for (;;) {
1437		struct ceph_snap_context *snapc, *oldest;
1438
1439		wait_on_page_writeback(page);
1440
1441		snapc = page_snap_context(page);
1442		if (!snapc || snapc == ci->i_head_snapc)
1443			break;
1444
1445		/*
1446		 * this page is already dirty in another (older) snap
1447		 * context!  is it writeable now?
1448		 */
1449		oldest = get_oldest_context(inode, NULL, NULL);
 
 
1450		if (snapc->seq > oldest->seq) {
1451			/* not writeable -- return it for the caller to deal with */
1452			ceph_put_snap_context(oldest);
1453			doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1454			      ceph_vinop(inode), page, snapc);
1455			return ceph_get_snap_context(snapc);
 
 
 
 
 
 
 
 
 
 
 
 
1456		}
1457		ceph_put_snap_context(oldest);
1458
1459		/* yay, writeable, do it now (without dropping page lock) */
1460		doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1461		      ceph_vinop(inode), page, snapc);
1462		if (clear_page_dirty_for_io(page)) {
1463			int r = writepage_nounlock(page, NULL);
1464			if (r < 0)
1465				return ERR_PTR(r);
1466		}
 
 
 
 
 
 
1467	}
1468	return NULL;
1469}
1470
1471static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1472					struct folio **foliop, void **_fsdata)
1473{
1474	struct inode *inode = file_inode(file);
1475	struct ceph_inode_info *ci = ceph_inode(inode);
1476	struct ceph_snap_context *snapc;
1477
1478	snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1479	if (snapc) {
1480		int r;
1481
1482		folio_unlock(*foliop);
1483		folio_put(*foliop);
1484		*foliop = NULL;
1485		if (IS_ERR(snapc))
1486			return PTR_ERR(snapc);
1487
1488		ceph_queue_writeback(inode);
1489		r = wait_event_killable(ci->i_cap_wq,
1490					context_is_writeable_or_written(inode, snapc));
1491		ceph_put_snap_context(snapc);
1492		return r == 0 ? -EAGAIN : r;
1493	}
1494	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
1495}
1496
1497/*
1498 * We are only allowed to write into/dirty the page if the page is
1499 * clean, or already dirty within the same snap context.
1500 */
1501static int ceph_write_begin(struct file *file, struct address_space *mapping,
1502			    loff_t pos, unsigned len,
1503			    struct page **pagep, void **fsdata)
1504{
1505	struct inode *inode = file_inode(file);
1506	struct ceph_inode_info *ci = ceph_inode(inode);
1507	struct folio *folio = NULL;
1508	int r;
1509
1510	r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1511	if (r < 0)
1512		return r;
 
 
 
 
 
 
 
 
 
1513
1514	folio_wait_fscache(folio);
1515	WARN_ON_ONCE(!folio_test_locked(folio));
1516	*pagep = &folio->page;
1517	return 0;
1518}
1519
1520/*
1521 * we don't do anything in here that simple_write_end doesn't do
1522 * except adjust dirty page accounting
 
1523 */
1524static int ceph_write_end(struct file *file, struct address_space *mapping,
1525			  loff_t pos, unsigned len, unsigned copied,
1526			  struct page *subpage, void *fsdata)
1527{
1528	struct folio *folio = page_folio(subpage);
1529	struct inode *inode = file_inode(file);
1530	struct ceph_client *cl = ceph_inode_to_client(inode);
1531	bool check_cap = false;
1532
1533	doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1534	      file, folio, (int)pos, (int)copied, (int)len);
1535
1536	if (!folio_test_uptodate(folio)) {
1537		/* just return that nothing was copied on a short copy */
1538		if (copied < len) {
1539			copied = 0;
1540			goto out;
1541		}
1542		folio_mark_uptodate(folio);
1543	}
1544
1545	/* did file size increase? */
1546	if (pos+copied > i_size_read(inode))
 
1547		check_cap = ceph_inode_set_size(inode, pos+copied);
1548
1549	folio_mark_dirty(folio);
 
 
 
1550
1551out:
1552	folio_unlock(folio);
1553	folio_put(folio);
1554
1555	if (check_cap)
1556		ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1557
1558	return copied;
1559}
1560
 
 
 
 
 
 
 
 
 
 
 
 
 
1561const struct address_space_operations ceph_aops = {
1562	.read_folio = netfs_read_folio,
1563	.readahead = netfs_readahead,
1564	.writepage = ceph_writepage,
1565	.writepages = ceph_writepages_start,
1566	.write_begin = ceph_write_begin,
1567	.write_end = ceph_write_end,
1568	.dirty_folio = ceph_dirty_folio,
1569	.invalidate_folio = ceph_invalidate_folio,
1570	.release_folio = netfs_release_folio,
1571	.direct_IO = noop_direct_IO,
1572};
1573
1574static void ceph_block_sigs(sigset_t *oldset)
1575{
1576	sigset_t mask;
1577	siginitsetinv(&mask, sigmask(SIGKILL));
1578	sigprocmask(SIG_BLOCK, &mask, oldset);
1579}
1580
1581static void ceph_restore_sigs(sigset_t *oldset)
1582{
1583	sigprocmask(SIG_SETMASK, oldset, NULL);
1584}
1585
1586/*
1587 * vm ops
1588 */
1589static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1590{
1591	struct vm_area_struct *vma = vmf->vma;
1592	struct inode *inode = file_inode(vma->vm_file);
1593	struct ceph_inode_info *ci = ceph_inode(inode);
1594	struct ceph_client *cl = ceph_inode_to_client(inode);
1595	struct ceph_file_info *fi = vma->vm_file->private_data;
1596	loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1597	int want, got, err;
1598	sigset_t oldset;
1599	vm_fault_t ret = VM_FAULT_SIGBUS;
1600
1601	if (ceph_inode_is_shutdown(inode))
1602		return ret;
1603
1604	ceph_block_sigs(&oldset);
1605
1606	doutc(cl, "%llx.%llx %llu trying to get caps\n",
1607	      ceph_vinop(inode), off);
1608	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1609		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1610	else
1611		want = CEPH_CAP_FILE_CACHE;
1612
1613	got = 0;
1614	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1615	if (err < 0)
1616		goto out_restore;
1617
1618	doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1619	      off, ceph_cap_string(got));
1620
1621	if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1622	    !ceph_has_inline_data(ci)) {
1623		CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1624		ceph_add_rw_context(fi, &rw_ctx);
1625		ret = filemap_fault(vmf);
1626		ceph_del_rw_context(fi, &rw_ctx);
1627		doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1628		      ceph_vinop(inode), off, ceph_cap_string(got), ret);
1629	} else
1630		err = -EAGAIN;
1631
1632	ceph_put_cap_refs(ci, got);
1633
1634	if (err != -EAGAIN)
1635		goto out_restore;
1636
1637	/* read inline data */
1638	if (off >= PAGE_SIZE) {
1639		/* does not support inline data > PAGE_SIZE */
1640		ret = VM_FAULT_SIGBUS;
1641	} else {
1642		struct address_space *mapping = inode->i_mapping;
1643		struct page *page;
1644
1645		filemap_invalidate_lock_shared(mapping);
1646		page = find_or_create_page(mapping, 0,
1647				mapping_gfp_constraint(mapping, ~__GFP_FS));
1648		if (!page) {
1649			ret = VM_FAULT_OOM;
1650			goto out_inline;
1651		}
1652		err = __ceph_do_getattr(inode, page,
1653					 CEPH_STAT_CAP_INLINE_DATA, true);
1654		if (err < 0 || off >= i_size_read(inode)) {
1655			unlock_page(page);
1656			put_page(page);
1657			ret = vmf_error(err);
1658			goto out_inline;
1659		}
1660		if (err < PAGE_SIZE)
1661			zero_user_segment(page, err, PAGE_SIZE);
1662		else
1663			flush_dcache_page(page);
1664		SetPageUptodate(page);
1665		vmf->page = page;
1666		ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1667out_inline:
1668		filemap_invalidate_unlock_shared(mapping);
1669		doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1670		      ceph_vinop(inode), off, ret);
1671	}
1672out_restore:
1673	ceph_restore_sigs(&oldset);
1674	if (err < 0)
1675		ret = vmf_error(err);
1676
1677	return ret;
1678}
1679
1680static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1681{
1682	struct vm_area_struct *vma = vmf->vma;
1683	struct inode *inode = file_inode(vma->vm_file);
1684	struct ceph_client *cl = ceph_inode_to_client(inode);
1685	struct ceph_inode_info *ci = ceph_inode(inode);
1686	struct ceph_file_info *fi = vma->vm_file->private_data;
1687	struct ceph_cap_flush *prealloc_cf;
1688	struct page *page = vmf->page;
1689	loff_t off = page_offset(page);
1690	loff_t size = i_size_read(inode);
1691	size_t len;
1692	int want, got, err;
1693	sigset_t oldset;
1694	vm_fault_t ret = VM_FAULT_SIGBUS;
1695
1696	if (ceph_inode_is_shutdown(inode))
1697		return ret;
1698
1699	prealloc_cf = ceph_alloc_cap_flush();
1700	if (!prealloc_cf)
1701		return VM_FAULT_OOM;
1702
1703	sb_start_pagefault(inode->i_sb);
1704	ceph_block_sigs(&oldset);
1705
1706	if (off + thp_size(page) <= size)
1707		len = thp_size(page);
1708	else
1709		len = offset_in_thp(page, size);
1710
1711	doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1712	      ceph_vinop(inode), off, len, size);
1713	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1714		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1715	else
1716		want = CEPH_CAP_FILE_BUFFER;
1717
1718	got = 0;
1719	err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1720	if (err < 0)
1721		goto out_free;
1722
1723	doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1724	      off, len, ceph_cap_string(got));
1725
1726	/* Update time before taking page lock */
1727	file_update_time(vma->vm_file);
1728	inode_inc_iversion_raw(inode);
1729
1730	do {
1731		struct ceph_snap_context *snapc;
 
 
1732
1733		lock_page(page);
1734
1735		if (page_mkwrite_check_truncate(page, inode) < 0) {
1736			unlock_page(page);
1737			ret = VM_FAULT_NOPAGE;
1738			break;
1739		}
1740
1741		snapc = ceph_find_incompatible(page);
1742		if (!snapc) {
1743			/* success.  we'll keep the page locked. */
1744			set_page_dirty(page);
1745			ret = VM_FAULT_LOCKED;
1746			break;
1747		}
1748
1749		unlock_page(page);
1750
1751		if (IS_ERR(snapc)) {
1752			ret = VM_FAULT_SIGBUS;
1753			break;
1754		}
1755
1756		ceph_queue_writeback(inode);
1757		err = wait_event_killable(ci->i_cap_wq,
1758				context_is_writeable_or_written(inode, snapc));
1759		ceph_put_snap_context(snapc);
1760	} while (err == 0);
1761
1762	if (ret == VM_FAULT_LOCKED) {
1763		int dirty;
1764		spin_lock(&ci->i_ceph_lock);
1765		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1766					       &prealloc_cf);
1767		spin_unlock(&ci->i_ceph_lock);
1768		if (dirty)
1769			__mark_inode_dirty(inode, dirty);
1770	}
1771
1772	doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1773	      ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1774	ceph_put_cap_refs_async(ci, got);
1775out_free:
1776	ceph_restore_sigs(&oldset);
1777	sb_end_pagefault(inode->i_sb);
1778	ceph_free_cap_flush(prealloc_cf);
1779	if (err < 0)
1780		ret = vmf_error(err);
1781	return ret;
1782}
1783
1784void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1785			   char	*data, size_t len)
1786{
1787	struct ceph_client *cl = ceph_inode_to_client(inode);
1788	struct address_space *mapping = inode->i_mapping;
1789	struct page *page;
1790
1791	if (locked_page) {
1792		page = locked_page;
1793	} else {
1794		if (i_size_read(inode) == 0)
1795			return;
1796		page = find_or_create_page(mapping, 0,
1797					   mapping_gfp_constraint(mapping,
1798					   ~__GFP_FS));
1799		if (!page)
1800			return;
1801		if (PageUptodate(page)) {
1802			unlock_page(page);
1803			put_page(page);
1804			return;
1805		}
1806	}
1807
1808	doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1809	      ceph_vinop(inode), len, locked_page);
1810
1811	if (len > 0) {
1812		void *kaddr = kmap_atomic(page);
1813		memcpy(kaddr, data, len);
1814		kunmap_atomic(kaddr);
1815	}
1816
1817	if (page != locked_page) {
1818		if (len < PAGE_SIZE)
1819			zero_user_segment(page, len, PAGE_SIZE);
1820		else
1821			flush_dcache_page(page);
1822
1823		SetPageUptodate(page);
1824		unlock_page(page);
1825		put_page(page);
1826	}
1827}
1828
1829int ceph_uninline_data(struct file *file)
1830{
1831	struct inode *inode = file_inode(file);
1832	struct ceph_inode_info *ci = ceph_inode(inode);
1833	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1834	struct ceph_client *cl = fsc->client;
1835	struct ceph_osd_request *req = NULL;
1836	struct ceph_cap_flush *prealloc_cf = NULL;
1837	struct folio *folio = NULL;
1838	u64 inline_version = CEPH_INLINE_NONE;
1839	struct page *pages[1];
1840	int err = 0;
1841	u64 len;
1842
1843	spin_lock(&ci->i_ceph_lock);
1844	inline_version = ci->i_inline_version;
1845	spin_unlock(&ci->i_ceph_lock);
1846
1847	doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1848	      inline_version);
1849
1850	if (ceph_inode_is_shutdown(inode)) {
1851		err = -EIO;
1852		goto out;
1853	}
1854
1855	if (inline_version == CEPH_INLINE_NONE)
1856		return 0;
1857
1858	prealloc_cf = ceph_alloc_cap_flush();
1859	if (!prealloc_cf)
1860		return -ENOMEM;
1861
1862	if (inline_version == 1) /* initial version, no data */
1863		goto out_uninline;
1864
1865	folio = read_mapping_folio(inode->i_mapping, 0, file);
1866	if (IS_ERR(folio)) {
1867		err = PTR_ERR(folio);
1868		goto out;
1869	}
1870
1871	folio_lock(folio);
1872
1873	len = i_size_read(inode);
1874	if (len > folio_size(folio))
1875		len = folio_size(folio);
1876
1877	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1878				    ceph_vino(inode), 0, &len, 0, 1,
1879				    CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1880				    NULL, 0, 0, false);
1881	if (IS_ERR(req)) {
1882		err = PTR_ERR(req);
1883		goto out_unlock;
1884	}
1885
1886	req->r_mtime = inode_get_mtime(inode);
1887	ceph_osdc_start_request(&fsc->client->osdc, req);
1888	err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1889	ceph_osdc_put_request(req);
1890	if (err < 0)
1891		goto out_unlock;
1892
1893	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1894				    ceph_vino(inode), 0, &len, 1, 3,
1895				    CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1896				    NULL, ci->i_truncate_seq,
1897				    ci->i_truncate_size, false);
1898	if (IS_ERR(req)) {
1899		err = PTR_ERR(req);
1900		goto out_unlock;
1901	}
1902
1903	pages[0] = folio_page(folio, 0);
1904	osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1905
1906	{
1907		__le64 xattr_buf = cpu_to_le64(inline_version);
1908		err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1909					    "inline_version", &xattr_buf,
1910					    sizeof(xattr_buf),
1911					    CEPH_OSD_CMPXATTR_OP_GT,
1912					    CEPH_OSD_CMPXATTR_MODE_U64);
1913		if (err)
1914			goto out_put_req;
1915	}
1916
1917	{
1918		char xattr_buf[32];
1919		int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1920					 "%llu", inline_version);
1921		err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1922					    "inline_version",
1923					    xattr_buf, xattr_len, 0, 0);
1924		if (err)
1925			goto out_put_req;
1926	}
1927
1928	req->r_mtime = inode_get_mtime(inode);
1929	ceph_osdc_start_request(&fsc->client->osdc, req);
1930	err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1931
1932	ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1933				  req->r_end_latency, len, err);
1934
1935out_uninline:
1936	if (!err) {
1937		int dirty;
1938
1939		/* Set to CAP_INLINE_NONE and dirty the caps */
1940		down_read(&fsc->mdsc->snap_rwsem);
1941		spin_lock(&ci->i_ceph_lock);
1942		ci->i_inline_version = CEPH_INLINE_NONE;
1943		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1944		spin_unlock(&ci->i_ceph_lock);
1945		up_read(&fsc->mdsc->snap_rwsem);
1946		if (dirty)
1947			__mark_inode_dirty(inode, dirty);
1948	}
1949out_put_req:
1950	ceph_osdc_put_request(req);
1951	if (err == -ECANCELED)
1952		err = 0;
1953out_unlock:
1954	if (folio) {
1955		folio_unlock(folio);
1956		folio_put(folio);
1957	}
1958out:
1959	ceph_free_cap_flush(prealloc_cf);
1960	doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1961	      ceph_vinop(inode), inline_version, err);
1962	return err;
1963}
1964
1965static const struct vm_operations_struct ceph_vmops = {
1966	.fault		= ceph_filemap_fault,
1967	.page_mkwrite	= ceph_page_mkwrite,
1968};
1969
1970int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1971{
1972	struct address_space *mapping = file->f_mapping;
1973
1974	if (!mapping->a_ops->read_folio)
1975		return -ENOEXEC;
 
1976	vma->vm_ops = &ceph_vmops;
 
1977	return 0;
1978}
1979
1980enum {
1981	POOL_READ	= 1,
1982	POOL_WRITE	= 2,
1983};
1984
1985static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1986				s64 pool, struct ceph_string *pool_ns)
1987{
1988	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
1989	struct ceph_mds_client *mdsc = fsc->mdsc;
1990	struct ceph_client *cl = fsc->client;
1991	struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1992	struct rb_node **p, *parent;
1993	struct ceph_pool_perm *perm;
1994	struct page **pages;
1995	size_t pool_ns_len;
1996	int err = 0, err2 = 0, have = 0;
1997
1998	down_read(&mdsc->pool_perm_rwsem);
1999	p = &mdsc->pool_perm_tree.rb_node;
2000	while (*p) {
2001		perm = rb_entry(*p, struct ceph_pool_perm, node);
2002		if (pool < perm->pool)
2003			p = &(*p)->rb_left;
2004		else if (pool > perm->pool)
2005			p = &(*p)->rb_right;
2006		else {
2007			int ret = ceph_compare_string(pool_ns,
2008						perm->pool_ns,
2009						perm->pool_ns_len);
2010			if (ret < 0)
2011				p = &(*p)->rb_left;
2012			else if (ret > 0)
2013				p = &(*p)->rb_right;
2014			else {
2015				have = perm->perm;
2016				break;
2017			}
2018		}
2019	}
2020	up_read(&mdsc->pool_perm_rwsem);
2021	if (*p)
2022		goto out;
2023
2024	if (pool_ns)
2025		doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2026		      (int)pool_ns->len, pool_ns->str);
2027	else
2028		doutc(cl, "pool %lld no perm cached\n", pool);
2029
2030	down_write(&mdsc->pool_perm_rwsem);
2031	p = &mdsc->pool_perm_tree.rb_node;
2032	parent = NULL;
2033	while (*p) {
2034		parent = *p;
2035		perm = rb_entry(parent, struct ceph_pool_perm, node);
2036		if (pool < perm->pool)
2037			p = &(*p)->rb_left;
2038		else if (pool > perm->pool)
2039			p = &(*p)->rb_right;
2040		else {
2041			int ret = ceph_compare_string(pool_ns,
2042						perm->pool_ns,
2043						perm->pool_ns_len);
2044			if (ret < 0)
2045				p = &(*p)->rb_left;
2046			else if (ret > 0)
2047				p = &(*p)->rb_right;
2048			else {
2049				have = perm->perm;
2050				break;
2051			}
2052		}
2053	}
2054	if (*p) {
2055		up_write(&mdsc->pool_perm_rwsem);
2056		goto out;
2057	}
2058
2059	rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2060					 1, false, GFP_NOFS);
2061	if (!rd_req) {
2062		err = -ENOMEM;
2063		goto out_unlock;
2064	}
2065
2066	rd_req->r_flags = CEPH_OSD_FLAG_READ;
2067	osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2068	rd_req->r_base_oloc.pool = pool;
2069	if (pool_ns)
2070		rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2071	ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2072
2073	err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2074	if (err)
2075		goto out_unlock;
2076
2077	wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2078					 1, false, GFP_NOFS);
2079	if (!wr_req) {
2080		err = -ENOMEM;
2081		goto out_unlock;
2082	}
2083
2084	wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2085	osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2086	ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2087	ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2088
2089	err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2090	if (err)
2091		goto out_unlock;
2092
2093	/* one page should be large enough for STAT data */
2094	pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2095	if (IS_ERR(pages)) {
2096		err = PTR_ERR(pages);
2097		goto out_unlock;
2098	}
2099
2100	osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2101				     0, false, true);
2102	ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2103
2104	wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2105	ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2106
2107	err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2108	err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2109
2110	if (err >= 0 || err == -ENOENT)
2111		have |= POOL_READ;
2112	else if (err != -EPERM) {
2113		if (err == -EBLOCKLISTED)
2114			fsc->blocklisted = true;
2115		goto out_unlock;
2116	}
2117
2118	if (err2 == 0 || err2 == -EEXIST)
2119		have |= POOL_WRITE;
2120	else if (err2 != -EPERM) {
2121		if (err2 == -EBLOCKLISTED)
2122			fsc->blocklisted = true;
2123		err = err2;
2124		goto out_unlock;
2125	}
2126
2127	pool_ns_len = pool_ns ? pool_ns->len : 0;
2128	perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2129	if (!perm) {
2130		err = -ENOMEM;
2131		goto out_unlock;
2132	}
2133
2134	perm->pool = pool;
2135	perm->perm = have;
2136	perm->pool_ns_len = pool_ns_len;
2137	if (pool_ns_len > 0)
2138		memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2139	perm->pool_ns[pool_ns_len] = 0;
2140
2141	rb_link_node(&perm->node, parent, p);
2142	rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2143	err = 0;
2144out_unlock:
2145	up_write(&mdsc->pool_perm_rwsem);
2146
2147	ceph_osdc_put_request(rd_req);
2148	ceph_osdc_put_request(wr_req);
2149out:
2150	if (!err)
2151		err = have;
2152	if (pool_ns)
2153		doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2154		      (int)pool_ns->len, pool_ns->str, err);
2155	else
2156		doutc(cl, "pool %lld result = %d\n", pool, err);
2157	return err;
2158}
2159
2160int ceph_pool_perm_check(struct inode *inode, int need)
2161{
2162	struct ceph_client *cl = ceph_inode_to_client(inode);
2163	struct ceph_inode_info *ci = ceph_inode(inode);
2164	struct ceph_string *pool_ns;
2165	s64 pool;
2166	int ret, flags;
2167
2168	/* Only need to do this for regular files */
2169	if (!S_ISREG(inode->i_mode))
2170		return 0;
2171
2172	if (ci->i_vino.snap != CEPH_NOSNAP) {
2173		/*
2174		 * Pool permission check needs to write to the first object.
2175		 * But for snapshot, head of the first object may have alread
2176		 * been deleted. Skip check to avoid creating orphan object.
2177		 */
2178		return 0;
2179	}
2180
2181	if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2182				NOPOOLPERM))
2183		return 0;
2184
2185	spin_lock(&ci->i_ceph_lock);
2186	flags = ci->i_ceph_flags;
2187	pool = ci->i_layout.pool_id;
2188	spin_unlock(&ci->i_ceph_lock);
2189check:
2190	if (flags & CEPH_I_POOL_PERM) {
2191		if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2192			doutc(cl, "pool %lld no read perm\n", pool);
2193			return -EPERM;
2194		}
2195		if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2196			doutc(cl, "pool %lld no write perm\n", pool);
2197			return -EPERM;
2198		}
2199		return 0;
2200	}
2201
2202	pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2203	ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2204	ceph_put_string(pool_ns);
2205	if (ret < 0)
2206		return ret;
2207
2208	flags = CEPH_I_POOL_PERM;
2209	if (ret & POOL_READ)
2210		flags |= CEPH_I_POOL_RD;
2211	if (ret & POOL_WRITE)
2212		flags |= CEPH_I_POOL_WR;
2213
2214	spin_lock(&ci->i_ceph_lock);
2215	if (pool == ci->i_layout.pool_id &&
2216	    pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2217		ci->i_ceph_flags |= flags;
2218        } else {
2219		pool = ci->i_layout.pool_id;
2220		flags = ci->i_ceph_flags;
2221	}
2222	spin_unlock(&ci->i_ceph_lock);
2223	goto check;
2224}
2225
2226void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2227{
2228	struct ceph_pool_perm *perm;
2229	struct rb_node *n;
2230
2231	while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2232		n = rb_first(&mdsc->pool_perm_tree);
2233		perm = rb_entry(n, struct ceph_pool_perm, node);
2234		rb_erase(n, &mdsc->pool_perm_tree);
2235		kfree(perm);
2236	}
2237}