[linux-2.6-block.git] / fs / xfs / libxfs / xfs_ag_resv.c

// SPDX-License-Identifier: GPL-2.0+
/*
 * Copyright (C) 2016 Oracle.  All Rights Reserved.
 * Author: Darrick J. Wong <darrick.wong@oracle.com>
 */
#include "xfs.h"
#include "xfs_fs.h"
#include "xfs_shared.h"
#include "xfs_format.h"
#include "xfs_log_format.h"
#include "xfs_trans_resv.h"
#include "xfs_mount.h"
#include "xfs_alloc.h"
#include "xfs_errortag.h"
#include "xfs_error.h"
#include "xfs_trace.h"
#include "xfs_trans.h"
#include "xfs_rmap_btree.h"
#include "xfs_btree.h"
#include "xfs_refcount_btree.h"
#include "xfs_ialloc_btree.h"
#include "xfs_ag.h"
#include "xfs_ag_resv.h"

/*
 * Per-AG Block Reservations
 *
 * For some kinds of allocation group metadata structures, it is advantageous
 * to reserve a small number of blocks in each AG so that future expansions of
 * that data structure do not encounter ENOSPC because errors during a btree
 * split cause the filesystem to go offline.
 *
 * Prior to the introduction of reflink, this wasn't an issue because the free
 * space btrees maintain a reserve of space (the AGFL) to handle any expansion
 * that may be necessary; and allocations of other metadata (inodes, BMBT,
 * dir/attr) aren't restricted to a single AG.  However, with reflink it is
 * possible to allocate all the space in an AG, have subsequent reflink/CoW
 * activity expand the refcount btree, and discover that there's no space left
 * to handle that expansion.  Since we can calculate the maximum size of the
 * refcount btree, we can reserve space for it and avoid ENOSPC.
 *
 * Handling per-AG reservations consists of three changes to the allocator's
 * behavior:  First, because these reservations are always needed, we decrease
 * the ag_max_usable counter to reflect the size of the AG after the reserved
 * blocks are taken.  Second, the reservations must be reflected in the
 * fdblocks count to maintain proper accounting.  Third, each AG must maintain
 * its own reserved block counter so that we can calculate the amount of space
 * that must remain free to maintain the reservations.  Fourth, the "remaining
 * reserved blocks" count must be used when calculating the length of the
 * longest free extent in an AG and to clamp maxlen in the per-AG allocation
 * functions.  In other words, we maintain a virtual allocation via in-core
 * accounting tricks so that we don't have to clean up after a crash. :)
 *
 * Reserved blocks can be managed by passing one of the enum xfs_ag_resv_type
 * values via struct xfs_alloc_arg or directly to the xfs_free_extent
 * function.  It might seem a little funny to maintain a reservoir of blocks
 * to feed another reservoir, but the AGFL only holds enough blocks to get
 * through the next transaction.  The per-AG reservation is to ensure (we
 * hope) that each AG never runs out of blocks.  Each data structure wanting
 * to use the reservation system should update ask/used in xfs_ag_resv_init.
 */

/*
 * Are we critically low on blocks?  For now we'll define that as the number
 * of blocks we can get our hands on being less than 10% of what we reserved
 * or less than some arbitrary number (maximum btree height).
 */
bool
xfs_ag_resv_critical(
	struct xfs_perag		*pag,
	enum xfs_ag_resv_type		type)
{
	xfs_extlen_t			avail;
	xfs_extlen_t			orig;

	switch (type) {
	case XFS_AG_RESV_METADATA:
		avail = pag->pagf_freeblks - pag->pag_rmapbt_resv.ar_reserved;
		orig = pag->pag_meta_resv.ar_asked;
		break;
	case XFS_AG_RESV_RMAPBT:
		avail = pag->pagf_freeblks + pag->pagf_flcount -
			pag->pag_meta_resv.ar_reserved;
		orig = pag->pag_rmapbt_resv.ar_asked;
		break;
	default:
		ASSERT(0);
		return false;
	}

	trace_xfs_ag_resv_critical(pag, type, avail);

	/* Critically low if less than 10% or max btree height remains. */
	return XFS_TEST_ERROR(avail < orig / 10 ||
			      avail < pag->pag_mount->m_agbtree_maxlevels,
			pag->pag_mount, XFS_ERRTAG_AG_RESV_CRITICAL);
}

/*
 * How many blocks are reserved but not used, and therefore must not be
 * allocated away?
 */
xfs_extlen_t
xfs_ag_resv_needed(
	struct xfs_perag		*pag,
	enum xfs_ag_resv_type		type)
{
	xfs_extlen_t			len;

	len = pag->pag_meta_resv.ar_reserved + pag->pag_rmapbt_resv.ar_reserved;
	switch (type) {
	case XFS_AG_RESV_METADATA:
	case XFS_AG_RESV_RMAPBT:
		len -= xfs_perag_resv(pag, type)->ar_reserved;
		break;
	case XFS_AG_RESV_NONE:
		/* empty */
		break;
	default:
		ASSERT(0);
	}

	trace_xfs_ag_resv_needed(pag, type, len);

	return len;
}

/* Clean out a reservation */
static int
__xfs_ag_resv_free(
	struct xfs_perag		*pag,
	enum xfs_ag_resv_type		type)
{
	struct xfs_ag_resv		*resv;
	xfs_extlen_t			oldresv;
	int				error;

	trace_xfs_ag_resv_free(pag, type, 0);

	resv = xfs_perag_resv(pag, type);
	if (pag->pag_agno == 0)
		pag->pag_mount->m_ag_max_usable += resv->ar_asked;
	/*
	 * RMAPBT blocks come from the AGFL and AGFL blocks are always
	 * considered "free", so whatever was reserved at mount time must be
	 * given back at umount.
	 */
	if (type == XFS_AG_RESV_RMAPBT)
		oldresv = resv->ar_orig_reserved;
	else
		oldresv = resv->ar_reserved;
	error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
	resv->ar_reserved = 0;
	resv->ar_asked = 0;
	resv->ar_orig_reserved = 0;

	if (error)
		trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
				error, _RET_IP_);
	return error;
}

/* Free a per-AG reservation. */
int
xfs_ag_resv_free(
	struct xfs_perag		*pag)
{
	int				error;
	int				err2;

	error = __xfs_ag_resv_free(pag, XFS_AG_RESV_RMAPBT);
	err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
	if (err2 && !error)
		error = err2;
	return error;
}

static int
__xfs_ag_resv_init(
	struct xfs_perag		*pag,
	enum xfs_ag_resv_type		type,
	xfs_extlen_t			ask,
	xfs_extlen_t			used)
{
	struct xfs_mount		*mp = pag->pag_mount;
	struct xfs_ag_resv		*resv;
	int				error;
	xfs_extlen_t			hidden_space;

	if (used > ask)
		ask = used;

	switch (type) {
	case XFS_AG_RESV_RMAPBT:
		/*
		 * Space taken by the rmapbt is not subtracted from fdblocks
		 * because the rmapbt lives in the free space.  Here we must
		 * subtract the entire reservation from fdblocks so that we
		 * always have blocks available for rmapbt expansion.
		 */
		hidden_space = ask;
		break;
	case XFS_AG_RESV_METADATA:
		/*
		 * Space taken by all other metadata btrees are accounted
		 * on-disk as used space.  We therefore only hide the space
		 * that is reserved but not used by the trees.
		 */
		hidden_space = ask - used;
		break;
	default:
		ASSERT(0);
		return -EINVAL;
	}

	if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_AG_RESV_FAIL))
		error = -ENOSPC;
	else
		error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space, true);
	if (error) {
		trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
				error, _RET_IP_);
		xfs_warn(mp,
"Per-AG reservation for AG %u failed.  Filesystem may run out of space.",
				pag->pag_agno);
		return error;
	}

	/*
	 * Reduce the maximum per-AG allocation length by however much we're
	 * trying to reserve for an AG.  Since this is a filesystem-wide
	 * counter, we only make the adjustment for AG 0.  This assumes that
	 * there aren't any AGs hungrier for per-AG reservation than AG 0.
	 */
	if (pag->pag_agno == 0)
		mp->m_ag_max_usable -= ask;

	resv = xfs_perag_resv(pag, type);
	resv->ar_asked = ask;
	resv->ar_orig_reserved = hidden_space;
	resv->ar_reserved = ask - used;

	trace_xfs_ag_resv_init(pag, type, ask);
	return 0;
}

/* Create a per-AG block reservation. */
int
xfs_ag_resv_init(
	struct xfs_perag		*pag,
	struct xfs_trans		*tp)
{
	struct xfs_mount		*mp = pag->pag_mount;
	xfs_extlen_t			ask;
	xfs_extlen_t			used;
	int				error = 0, error2;
	bool				has_resv = false;

	/* Create the metadata reservation. */
	if (pag->pag_meta_resv.ar_asked == 0) {
		ask = used = 0;

		error = xfs_refcountbt_calc_reserves(mp, tp, pag, &ask, &used);
		if (error)
			goto out;

		error = xfs_finobt_calc_reserves(mp, tp, pag, &ask, &used);
		if (error)
			goto out;

		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
				ask, used);
		if (error) {
			/*
			 * Because we didn't have per-AG reservations when the
			 * finobt feature was added we might not be able to
			 * reserve all needed blocks.  Warn and fall back to the
			 * old and potentially buggy code in that case, but
			 * ensure we do have the reservation for the refcountbt.
			 */
			ask = used = 0;

			mp->m_finobt_nores = true;

			error = xfs_refcountbt_calc_reserves(mp, tp, pag, &ask,
					&used);
			if (error)
				goto out;

			error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
					ask, used);
			if (error)
				goto out;
		}
		if (ask)
			has_resv = true;
	}

	/* Create the RMAPBT metadata reservation */
	if (pag->pag_rmapbt_resv.ar_asked == 0) {
		ask = used = 0;

		error = xfs_rmapbt_calc_reserves(mp, tp, pag, &ask, &used);
		if (error)
			goto out;

		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_RMAPBT, ask, used);
		if (error)
			goto out;
		if (ask)
			has_resv = true;
	}

out:
	/*
	 * Initialize the pagf if we have at least one active reservation on the
	 * AG. This may have occurred already via reservation calculation, but
	 * fall back to an explicit init to ensure the in-core allocbt usage
	 * counters are initialized as soon as possible. This is important
	 * because filesystems with large perag reservations are susceptible to
	 * free space reservation problems that the allocbt counter is used to
	 * address.
	 */
	if (has_resv) {
		error2 = xfs_alloc_read_agf(pag, tp, 0, NULL);
		if (error2)
			return error2;

		/*
		 * If there isn't enough space in the AG to satisfy the
		 * reservation, let the caller know that there wasn't enough
		 * space.  Callers are responsible for deciding what to do
		 * next, since (in theory) we can stumble along with
		 * insufficient reservation if data blocks are being freed to
		 * replenish the AG's free space.
		 */
		if (!error &&
		    xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
		    xfs_perag_resv(pag, XFS_AG_RESV_RMAPBT)->ar_reserved >
		    pag->pagf_freeblks + pag->pagf_flcount)
			error = -ENOSPC;
	}

	return error;
}

/* Allocate a block from the reservation. */
void
xfs_ag_resv_alloc_extent(
	struct xfs_perag		*pag,
	enum xfs_ag_resv_type		type,
	struct xfs_alloc_arg		*args)
{
	struct xfs_ag_resv		*resv;
	xfs_extlen_t			len;
	uint				field;

	trace_xfs_ag_resv_alloc_extent(pag, type, args->len);

	switch (type) {
	case XFS_AG_RESV_AGFL:
		return;
	case XFS_AG_RESV_METADATA:
	case XFS_AG_RESV_RMAPBT:
		resv = xfs_perag_resv(pag, type);
		break;
	default:
		ASSERT(0);
		fallthrough;
	case XFS_AG_RESV_NONE:
		field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
				       XFS_TRANS_SB_FDBLOCKS;
		xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
		return;
	}

	len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
	resv->ar_reserved -= len;
	if (type == XFS_AG_RESV_RMAPBT)
		return;
	/* Allocations of reserved blocks only need on-disk sb updates... */
	xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
	/* ...but non-reserved blocks need in-core and on-disk updates. */
	if (args->len > len)
		xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
				-((int64_t)args->len - len));
}

/* Free a block to the reservation. */
void
xfs_ag_resv_free_extent(
	struct xfs_perag		*pag,
	enum xfs_ag_resv_type		type,
	struct xfs_trans		*tp,
	xfs_extlen_t			len)
{
	xfs_extlen_t			leftover;
	struct xfs_ag_resv		*resv;

	trace_xfs_ag_resv_free_extent(pag, type, len);

	switch (type) {
	case XFS_AG_RESV_AGFL:
		return;
	case XFS_AG_RESV_METADATA:
	case XFS_AG_RESV_RMAPBT:
		resv = xfs_perag_resv(pag, type);
		break;
	default:
		ASSERT(0);
		fallthrough;
	case XFS_AG_RESV_NONE:
		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
		return;
	}

	leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
	resv->ar_reserved += leftover;
	if (type == XFS_AG_RESV_RMAPBT)
		return;
	/* Freeing into the reserved pool only requires on-disk update... */
	xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
	/* ...but freeing beyond that requires in-core and on-disk update. */
	if (len > leftover)
		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
}
Commit	Line	Data
0b61f8a4	1	// SPDX-License-Identifier: GPL-2.0+
3fd129b6 DW	2	/*
3fd129b6 DW	3	* Copyright (C) 2016 Oracle. All Rights Reserved.
3fd129b6	4	* Author: Darrick J. Wong <darrick.wong@oracle.com>
3fd129b6 DW	5	*/
	6	#include "xfs.h"
	7	#include "xfs_fs.h"
	8	#include "xfs_shared.h"
	9	#include "xfs_format.h"
	10	#include "xfs_log_format.h"
	11	#include "xfs_trans_resv.h"
3fd129b6	12	#include "xfs_mount.h"
3fd129b6	13	#include "xfs_alloc.h"
e9e899a2	14	#include "xfs_errortag.h"
3fd129b6 DW	15	#include "xfs_error.h"
3fd129b6 DW	16	#include "xfs_trace.h"
3fd129b6	17	#include "xfs_trans.h"
3fd129b6 DW	18	#include "xfs_rmap_btree.h"
3fd129b6 DW	19	#include "xfs_btree.h"
84d69619	20	#include "xfs_refcount_btree.h"
76d771b4	21	#include "xfs_ialloc_btree.h"
9bbafc71	22	#include "xfs_ag.h"
5f213ddb	23	#include "xfs_ag_resv.h"
3fd129b6 DW	24
	25	/*
	26	* Per-AG Block Reservations
	27	*
	28	* For some kinds of allocation group metadata structures, it is advantageous
	29	* to reserve a small number of blocks in each AG so that future expansions of
	30	* that data structure do not encounter ENOSPC because errors during a btree
	31	* split cause the filesystem to go offline.
	32	*
	33	* Prior to the introduction of reflink, this wasn't an issue because the free
	34	* space btrees maintain a reserve of space (the AGFL) to handle any expansion
	35	* that may be necessary; and allocations of other metadata (inodes, BMBT,
	36	* dir/attr) aren't restricted to a single AG. However, with reflink it is
	37	* possible to allocate all the space in an AG, have subsequent reflink/CoW
	38	* activity expand the refcount btree, and discover that there's no space left
	39	* to handle that expansion. Since we can calculate the maximum size of the
	40	* refcount btree, we can reserve space for it and avoid ENOSPC.
	41	*
	42	* Handling per-AG reservations consists of three changes to the allocator's
	43	* behavior: First, because these reservations are always needed, we decrease
	44	* the ag_max_usable counter to reflect the size of the AG after the reserved
	45	* blocks are taken. Second, the reservations must be reflected in the
	46	* fdblocks count to maintain proper accounting. Third, each AG must maintain
	47	* its own reserved block counter so that we can calculate the amount of space
	48	* that must remain free to maintain the reservations. Fourth, the "remaining
	49	* reserved blocks" count must be used when calculating the length of the
	50	* longest free extent in an AG and to clamp maxlen in the per-AG allocation
	51	* functions. In other words, we maintain a virtual allocation via in-core
	52	* accounting tricks so that we don't have to clean up after a crash. :)
	53	*
	54	* Reserved blocks can be managed by passing one of the enum xfs_ag_resv_type
	55	* values via struct xfs_alloc_arg or directly to the xfs_free_extent
	56	* function. It might seem a little funny to maintain a reservoir of blocks
	57	* to feed another reservoir, but the AGFL only holds enough blocks to get
	58	* through the next transaction. The per-AG reservation is to ensure (we
	59	* hope) that each AG never runs out of blocks. Each data structure wanting
	60	* to use the reservation system should update ask/used in xfs_ag_resv_init.
	61	*/
	62
	63	/*
	64	* Are we critically low on blocks? For now we'll define that as the number
	65	* of blocks we can get our hands on being less than 10% of what we reserved
	66	* or less than some arbitrary number (maximum btree height).
	67	*/
	68	bool
	69	xfs_ag_resv_critical(
	70	struct xfs_perag *pag,
	71	enum xfs_ag_resv_type type)
	72	{
	73	xfs_extlen_t avail;
	74	xfs_extlen_t orig;
	75
	76	switch (type) {
	77	case XFS_AG_RESV_METADATA:
21592863	78	avail = pag->pagf_freeblks - pag->pag_rmapbt_resv.ar_reserved;
3fd129b6 DW	79	orig = pag->pag_meta_resv.ar_asked;
3fd129b6 DW	80	break;
21592863	81	case XFS_AG_RESV_RMAPBT:
3fd129b6 DW	82	avail = pag->pagf_freeblks + pag->pagf_flcount -
3fd129b6 DW	83	pag->pag_meta_resv.ar_reserved;
21592863	84	orig = pag->pag_rmapbt_resv.ar_asked;
3fd129b6 DW	85	break;
	86	default:
	87	ASSERT(0);
	88	return false;
	89	}
	90
	91	trace_xfs_ag_resv_critical(pag, type, avail);
	92
	93	/* Critically low if less than 10% or max btree height remains. */
b74e15d7 DW	94	return XFS_TEST_ERROR(avail < orig / 10 \|\|
b74e15d7 DW	95	avail < pag->pag_mount->m_agbtree_maxlevels,
9e24cfd0	96	pag->pag_mount, XFS_ERRTAG_AG_RESV_CRITICAL);
3fd129b6 DW	97	}
	98
	99	/*
	100	* How many blocks are reserved but not used, and therefore must not be
	101	* allocated away?
	102	*/
	103	xfs_extlen_t
	104	xfs_ag_resv_needed(
	105	struct xfs_perag *pag,
	106	enum xfs_ag_resv_type type)
	107	{
	108	xfs_extlen_t len;
	109
21592863	110	len = pag->pag_meta_resv.ar_reserved + pag->pag_rmapbt_resv.ar_reserved;
3fd129b6 DW	111	switch (type) {
3fd129b6 DW	112	case XFS_AG_RESV_METADATA:
21592863	113	case XFS_AG_RESV_RMAPBT:
3fd129b6 DW	114	len -= xfs_perag_resv(pag, type)->ar_reserved;
	115	break;
	116	case XFS_AG_RESV_NONE:
	117	/* empty */
	118	break;
	119	default:
	120	ASSERT(0);
	121	}
	122
	123	trace_xfs_ag_resv_needed(pag, type, len);
	124
	125	return len;
	126	}
	127
	128	/* Clean out a reservation */
	129	static int
	130	__xfs_ag_resv_free(
	131	struct xfs_perag *pag,
	132	enum xfs_ag_resv_type type)
	133	{
	134	struct xfs_ag_resv *resv;
	135	xfs_extlen_t oldresv;
	136	int error;
	137
	138	trace_xfs_ag_resv_free(pag, type, 0);
	139
	140	resv = xfs_perag_resv(pag, type);
9789dd9e DW	141	if (pag->pag_agno == 0)
9789dd9e DW	142	pag->pag_mount->m_ag_max_usable += resv->ar_asked;
3fd129b6	143	/*
21592863 BF	144	* RMAPBT blocks come from the AGFL and AGFL blocks are always
	145	* considered "free", so whatever was reserved at mount time must be
	146	* given back at umount.
3fd129b6	147	*/
21592863	148	if (type == XFS_AG_RESV_RMAPBT)
3fd129b6 DW	149	oldresv = resv->ar_orig_reserved;
	150	else
	151	oldresv = resv->ar_reserved;
	152	error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
	153	resv->ar_reserved = 0;
	154	resv->ar_asked = 0;
d8cb5e42	155	resv->ar_orig_reserved = 0;
3fd129b6 DW	156
	157	if (error)
	158	trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
	159	error, _RET_IP_);
	160	return error;
	161	}
	162
	163	/* Free a per-AG reservation. */
	164	int
	165	xfs_ag_resv_free(
	166	struct xfs_perag *pag)
	167	{
	168	int error;
	169	int err2;
	170
21592863	171	error = __xfs_ag_resv_free(pag, XFS_AG_RESV_RMAPBT);
3fd129b6 DW	172	err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
	173	if (err2 && !error)
	174	error = err2;
	175	return error;
	176	}
	177
	178	static int
	179	__xfs_ag_resv_init(
	180	struct xfs_perag *pag,
	181	enum xfs_ag_resv_type type,
	182	xfs_extlen_t ask,
	183	xfs_extlen_t used)
	184	{
	185	struct xfs_mount *mp = pag->pag_mount;
	186	struct xfs_ag_resv *resv;
	187	int error;
d8cb5e42	188	xfs_extlen_t hidden_space;
3fd129b6	189
3fd129b6 DW	190	if (used > ask)
3fd129b6 DW	191	ask = used;
3fd129b6	192
d8cb5e42 DW	193	switch (type) {
	194	case XFS_AG_RESV_RMAPBT:
	195	/*
	196	* Space taken by the rmapbt is not subtracted from fdblocks
	197	* because the rmapbt lives in the free space. Here we must
	198	* subtract the entire reservation from fdblocks so that we
	199	* always have blocks available for rmapbt expansion.
	200	*/
	201	hidden_space = ask;
	202	break;
	203	case XFS_AG_RESV_METADATA:
	204	/*
	205	* Space taken by all other metadata btrees are accounted
	206	* on-disk as used space. We therefore only hide the space
	207	* that is reserved but not used by the trees.
	208	*/
	209	hidden_space = ask - used;
	210	break;
	211	default:
	212	ASSERT(0);
	213	return -EINVAL;
	214	}
2b92faed GX	215
	216	if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_AG_RESV_FAIL))
	217	error = -ENOSPC;
	218	else
	219	error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space, true);
4dfa2b84	220	if (error) {
3fd129b6 DW	221	trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
3fd129b6 DW	222	error, _RET_IP_);
76d771b4 CH	223	xfs_warn(mp,
	224	"Per-AG reservation for AG %u failed. Filesystem may run out of space.",
	225	pag->pag_agno);
4dfa2b84 CH	226	return error;
4dfa2b84 CH	227	}
3fd129b6	228
9789dd9e DW	229	/*
	230	* Reduce the maximum per-AG allocation length by however much we're
	231	* trying to reserve for an AG. Since this is a filesystem-wide
	232	* counter, we only make the adjustment for AG 0. This assumes that
	233	* there aren't any AGs hungrier for per-AG reservation than AG 0.
	234	*/
	235	if (pag->pag_agno == 0)
	236	mp->m_ag_max_usable -= ask;
4dfa2b84 CH	237
	238	resv = xfs_perag_resv(pag, type);
	239	resv->ar_asked = ask;
d8cb5e42 DW	240	resv->ar_orig_reserved = hidden_space;
d8cb5e42 DW	241	resv->ar_reserved = ask - used;
4dfa2b84 CH	242
	243	trace_xfs_ag_resv_init(pag, type, ask);
	244	return 0;
3fd129b6 DW	245	}
	246
	247	/* Create a per-AG block reservation. */
	248	int
	249	xfs_ag_resv_init(
ebcbef3a DW	250	struct xfs_perag *pag,
ebcbef3a DW	251	struct xfs_trans *tp)
3fd129b6	252	{
76d771b4	253	struct xfs_mount *mp = pag->pag_mount;
3fd129b6 DW	254	xfs_extlen_t ask;
3fd129b6 DW	255	xfs_extlen_t used;
2675ad38 BF	256	int error = 0, error2;
2675ad38 BF	257	bool has_resv = false;
3fd129b6 DW	258
	259	/* Create the metadata reservation. */
	260	if (pag->pag_meta_resv.ar_asked == 0) {
	261	ask = used = 0;
	262
30933120	263	error = xfs_refcountbt_calc_reserves(mp, tp, pag, &ask, &used);
84d69619 DW	264	if (error)
	265	goto out;
	266
30933120	267	error = xfs_finobt_calc_reserves(mp, tp, pag, &ask, &used);
3fd129b6 DW	268	if (error)
3fd129b6 DW	269	goto out;
76d771b4 CH	270
	271	error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
	272	ask, used);
	273	if (error) {
	274	/*
	275	* Because we didn't have per-AG reservations when the
	276	* finobt feature was added we might not be able to
	277	* reserve all needed blocks. Warn and fall back to the
	278	* old and potentially buggy code in that case, but
	279	* ensure we do have the reservation for the refcountbt.
	280	*/
	281	ask = used = 0;
	282
e1f6ca11	283	mp->m_finobt_nores = true;
76d771b4	284
30933120	285	error = xfs_refcountbt_calc_reserves(mp, tp, pag, &ask,
76d771b4 CH	286	&used);
	287	if (error)
	288	goto out;
	289
	290	error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
	291	ask, used);
	292	if (error)
	293	goto out;
	294	}
2675ad38 BF	295	if (ask)
2675ad38 BF	296	has_resv = true;
3fd129b6 DW	297	}
3fd129b6 DW	298
21592863 BF	299	/* Create the RMAPBT metadata reservation */
21592863 BF	300	if (pag->pag_rmapbt_resv.ar_asked == 0) {
3fd129b6 DW	301	ask = used = 0;
3fd129b6 DW	302
30933120	303	error = xfs_rmapbt_calc_reserves(mp, tp, pag, &ask, &used);
84d69619 DW	304	if (error)
	305	goto out;
	306
21592863	307	error = __xfs_ag_resv_init(pag, XFS_AG_RESV_RMAPBT, ask, used);
3fd129b6 DW	308	if (error)
3fd129b6 DW	309	goto out;
2675ad38 BF	310	if (ask)
2675ad38 BF	311	has_resv = true;
3fd129b6 DW	312	}
	313
	314	out:
2675ad38 BF	315	/*
	316	* Initialize the pagf if we have at least one active reservation on the
	317	* AG. This may have occurred already via reservation calculation, but
	318	* fall back to an explicit init to ensure the in-core allocbt usage
	319	* counters are initialized as soon as possible. This is important
	320	* because filesystems with large perag reservations are susceptible to
	321	* free space reservation problems that the allocbt counter is used to
	322	* address.
	323	*/
	324	if (has_resv) {
08d3e84f	325	error2 = xfs_alloc_read_agf(pag, tp, 0, NULL);
2675ad38 BF	326	if (error2)
2675ad38 BF	327	return error2;
0f934251 DW	328
	329	/*
	330	* If there isn't enough space in the AG to satisfy the
	331	* reservation, let the caller know that there wasn't enough
	332	* space. Callers are responsible for deciding what to do
	333	* next, since (in theory) we can stumble along with
	334	* insufficient reservation if data blocks are being freed to
	335	* replenish the AG's free space.
	336	*/
	337	if (!error &&
	338	xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
	339	xfs_perag_resv(pag, XFS_AG_RESV_RMAPBT)->ar_reserved >
	340	pag->pagf_freeblks + pag->pagf_flcount)
	341	error = -ENOSPC;
2675ad38	342	}
0f934251	343
3fd129b6 DW	344	return error;
	345	}
	346
	347	/* Allocate a block from the reservation. */
	348	void
	349	xfs_ag_resv_alloc_extent(
	350	struct xfs_perag *pag,
	351	enum xfs_ag_resv_type type,
	352	struct xfs_alloc_arg *args)
	353	{
	354	struct xfs_ag_resv *resv;
	355	xfs_extlen_t len;
	356	uint field;
	357
	358	trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
	359
	360	switch (type) {
0ab32086 BF	361	case XFS_AG_RESV_AGFL:
0ab32086 BF	362	return;
3fd129b6	363	case XFS_AG_RESV_METADATA:
21592863	364	case XFS_AG_RESV_RMAPBT:
3fd129b6 DW	365	resv = xfs_perag_resv(pag, type);
	366	break;
	367	default:
	368	ASSERT(0);
53004ee7	369	fallthrough;
3fd129b6 DW	370	case XFS_AG_RESV_NONE:
	371	field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
	372	XFS_TRANS_SB_FDBLOCKS;
	373	xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
	374	return;
	375	}
	376
	377	len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
	378	resv->ar_reserved -= len;
21592863	379	if (type == XFS_AG_RESV_RMAPBT)
3fd129b6 DW	380	return;
	381	/* Allocations of reserved blocks only need on-disk sb updates... */
	382	xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
	383	/* ...but non-reserved blocks need in-core and on-disk updates. */
	384	if (args->len > len)
	385	xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
	386	-((int64_t)args->len - len));
	387	}
	388
	389	/* Free a block to the reservation. */
	390	void
	391	xfs_ag_resv_free_extent(
	392	struct xfs_perag *pag,
	393	enum xfs_ag_resv_type type,
	394	struct xfs_trans *tp,
	395	xfs_extlen_t len)
	396	{
	397	xfs_extlen_t leftover;
	398	struct xfs_ag_resv *resv;
	399
	400	trace_xfs_ag_resv_free_extent(pag, type, len);
	401
	402	switch (type) {
0ab32086 BF	403	case XFS_AG_RESV_AGFL:
0ab32086 BF	404	return;
3fd129b6	405	case XFS_AG_RESV_METADATA:
21592863	406	case XFS_AG_RESV_RMAPBT:
3fd129b6 DW	407	resv = xfs_perag_resv(pag, type);
	408	break;
	409	default:
	410	ASSERT(0);
53004ee7	411	fallthrough;
3fd129b6 DW	412	case XFS_AG_RESV_NONE:
	413	xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
	414	return;
	415	}
	416
	417	leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
	418	resv->ar_reserved += leftover;
21592863	419	if (type == XFS_AG_RESV_RMAPBT)
3fd129b6 DW	420	return;
	421	/* Freeing into the reserved pool only requires on-disk update... */
	422	xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
	423	/* ...but freeing beyond that requires in-core and on-disk update. */
	424	if (len > leftover)
	425	xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
	426	}