xfs_discard.c | Explore in Territory

// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (C) 2010, 2023 Red Hat, Inc.
 * All Rights Reserved.
 */
#include "xfs.h"
#include "xfs_shared.h"
#include "xfs_format.h"
#include "xfs_log_format.h"
#include "xfs_trans_resv.h"
#include "xfs_trans.h"
#include "xfs_mount.h"
#include "xfs_btree.h"
#include "xfs_alloc_btree.h"
#include "xfs_alloc.h"
#include "xfs_discard.h"
#include "xfs_error.h"
#include "xfs_extent_busy.h"
#include "xfs_trace.h"
#include "xfs_log.h"
#include "xfs_ag.h"
#include "xfs_health.h"
#include "xfs_rtbitmap.h"

/*
 * Notes on an efficient, low latency fstrim algorithm
 *
 * We need to walk the filesystem free space and issue discards on the free
 * space that meet the search criteria (size and location). We cannot issue
 * discards on extents that might be in use, or are so recently in use they are
 * still marked as busy. To serialise against extent state changes whilst we are
 * gathering extents to trim, we must hold the AGF lock to lock out other
 * allocations and extent free operations that might change extent state.
 *
 * However, we cannot just hold the AGF for the entire AG free space walk whilst
 * we issue discards on each free space that is found. Storage devices can have
 * extremely slow discard implementations (e.g. ceph RBD) and so walking a
 * couple of million free extents and issuing synchronous discards on each
 * extent can take a *long* time. Whilst we are doing this walk, nothing else
 * can access the AGF, and we can stall transactions and hence the log whilst
 * modifications wait for the AGF lock to be released. This can lead hung tasks
 * kicking the hung task timer and rebooting the system. This is bad.
 *
 * Hence we need to take a leaf from the bulkstat playbook. It takes the AGI
 * lock, gathers a range of inode cluster buffers that are allocated, drops the
 * AGI lock and then reads all the inode cluster buffers and processes them. It
 * loops doing this, using a cursor to keep track of where it is up to in the AG
 * for each iteration to restart the INOBT lookup from.
 *
 * We can't do this exactly with free space - once we drop the AGF lock, the
 * state of the free extent is out of our control and we cannot run a discard
 * safely on it in this situation. Unless, of course, we've marked the free
 * extent as busy and undergoing a discard operation whilst we held the AGF
 * locked.
 *
 * This is exactly how online discard works - free extents are marked busy when
 * they are freed, and once the extent free has been committed to the journal,
 * the busy extent record is marked as "undergoing discard" and the discard is
 * then issued on the free extent. Once the discard completes, the busy extent
 * record is removed and the extent is able to be allocated again.
 *
 * In the context of fstrim, if we find a free extent we need to discard, we
 * don't have to discard it immediately. All we need to do it record that free
 * extent as being busy and under discard, and all the allocation routines will
 * now avoid trying to allocate it. Hence if we mark the extent as busy under
 * the AGF lock, we can safely discard it without holding the AGF lock because
 * nothing will attempt to allocate that free space until the discard completes.
 *
 * This also allows us to issue discards asynchronously like we do with online
 * discard, and so for fast devices fstrim will run much faster as we can have
 * multiple discard operations in flight at once, as well as pipeline the free
 * extent search so that it overlaps in flight discard IO.
 */

struct workqueue_struct *xfs_discard_wq;

static void
xfs_discard_endio_work(
	struct work_struct	*work)
{ … }

/*
 * Queue up the actual completion to a thread to avoid IRQ-safe locking for
 * pagb_lock.
 */
static void
xfs_discard_endio(
	struct bio		*bio)
{ … }

/*
 * Walk the discard list and issue discards on all the busy extents in the
 * list. We plug and chain the bios so that we only need a single completion
 * call to clear all the busy extents once the discards are complete.
 */
int
xfs_discard_extents(
	struct xfs_mount	*mp,
	struct xfs_busy_extents	*extents)
{ … }

struct xfs_trim_cur { … };

static int
xfs_trim_gather_extents(
	struct xfs_perag	*pag,
	struct xfs_trim_cur	*tcur,
	struct xfs_busy_extents	*extents,
	uint64_t		*blocks_trimmed)
{ … }

static bool
xfs_trim_should_stop(void)
{ … }

/*
 * Iterate the free list gathering extents and discarding them. We need a cursor
 * for the repeated iteration of gather/discard loop, so use the longest extent
 * we found in the last batch as the key to start the next.
 */
static int
xfs_trim_perag_extents(
	struct xfs_perag	*pag,
	xfs_agblock_t		start,
	xfs_agblock_t		end,
	xfs_extlen_t		minlen,
	uint64_t		*blocks_trimmed)
{ … }

static int
xfs_trim_datadev_extents(
	struct xfs_mount	*mp,
	xfs_daddr_t		start,
	xfs_daddr_t		end,
	xfs_extlen_t		minlen,
	uint64_t		*blocks_trimmed)
{ … }

#ifdef CONFIG_XFS_RT
struct xfs_trim_rtdev { … };

struct xfs_rtx_busy { … };

static void
xfs_discard_free_rtdev_extents(
	struct xfs_trim_rtdev	*tr)
{ … }

/*
 * Walk the discard list and issue discards on all the busy extents in the
 * list. We plug and chain the bios so that we only need a single completion
 * call to clear all the busy extents once the discards are complete.
 */
static int
xfs_discard_rtdev_extents(
	struct xfs_mount	*mp,
	struct xfs_trim_rtdev	*tr)
{ … }

static int
xfs_trim_gather_rtextent(
	struct xfs_mount		*mp,
	struct xfs_trans		*tp,
	const struct xfs_rtalloc_rec	*rec,
	void				*priv)
{ … }

static int
xfs_trim_rtdev_extents(
	struct xfs_mount	*mp,
	xfs_daddr_t		start,
	xfs_daddr_t		end,
	xfs_daddr_t		minlen,
	uint64_t		*blocks_trimmed)
{ … }
#else
#define xfs_trim_rtdev_extents …
#endif /* CONFIG_XFS_RT */

/*
 * trim a range of the filesystem.
 *
 * Note: the parameters passed from userspace are byte ranges into the
 * filesystem which does not match to the format we use for filesystem block
 * addressing. FSB addressing is sparse (AGNO|AGBNO), while the incoming format
 * is a linear address range. Hence we need to use DADDR based conversions and
 * comparisons for determining the correct offset and regions to trim.
 *
 * The realtime device is mapped into the FITRIM "address space" immediately
 * after the data device.
 */
int
xfs_ioc_trim(
	struct xfs_mount		*mp,
	struct fstrim_range __user	*urange)
{ … }
linux/fs/xfs/xfs_discard.c
