Skip to contentSkip to conversation

xfs: fix inode allocation block res calculation precedence

PATCH In review Brian Foster 5 participants 11 messages Jul 15, 2020 – Jul 21, 2020 XFS filesystem

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xfs: fix inode allocation block res calculation precedence

The block reservation calculation for inode allocation is supposed to consist of the blocks required for the inode chunk plus (maxlevels-1) of the inode btree multiplied by the number of inode btrees in the fs (2 when finobt is enabled, 1 otherwise).

Instead, the macro returns (ialloc_blocks + 2) due to a precedence error in the calculation logic. This leads to block reservation overruns via generic/531 on small block filesystems with finobt enabled. Add braces to fix the calculation and reserve the appropriate number of blocks.

Fixes 9d43b180af67 ("xfs: update inode allocation/free transaction reservations for finobt")
---
 fs/xfs/libxfs/xfs_trans_space.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/xfs/libxfs/xfs_trans_space.h +1 -1
index 88221c7a04cc..c6df01a2a158 100644
@@ -57,7 +57,7 @@
 	XFS_DAREMOVE_SPACE_RES(mp, XFS_DATA_FORK)
 #define	XFS_IALLOC_SPACE_RES(mp)	\
 	(M_IGEO(mp)->ialloc_blks + \
-	 (xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1 * \
+	 ((xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1) * \
 	  (M_IGEO(mp)->inobt_maxlevels - 1)))
 
 /*
Signature · 1 line
2.21.3
#1 lore raw
Quoted 22 lines · Brian Foster

The block reservation calculation for inode allocation is supposed to consist of the blocks required for the inode chunk plus (maxlevels-1) of the inode btree multiplied by the number of inode btrees in the fs (2 when finobt is enabled, 1 otherwise).

Instead, the macro returns (ialloc_blocks + 2) due to a precedence error in the calculation logic. This leads to block reservation overruns via generic/531 on small block filesystems with finobt enabled. Add braces to fix the calculation and reserve the appropriate number of blocks.

Fixes 9d43b180af67 ("xfs: update inode allocation/free transaction reservations for finobt")
---
 fs/xfs/libxfs/xfs_trans_space.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/xfs/libxfs/xfs_trans_space.h +0 -0
index 88221c7a04cc..c6df01a2a158 100644

fs/xfs/libxfs/xfs_trans_space.h line 61

Quoted 6 lines
fs/xfs/libxfs/xfs_trans_space.h +1 -1
@@ -57,7 +57,7 @@
 	XFS_DAREMOVE_SPACE_RES(mp, XFS_DATA_FORK)
 #define	XFS_IALLOC_SPACE_RES(mp)	\
 	(M_IGEO(mp)->ialloc_blks + \
-	 (xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1 * \
+	 ((xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1) * \
 	  (M_IGEO(mp)->inobt_maxlevels - 1)))

Ugh. THese macros really need rewriting as static inline functions. This would not have happened if it were written as:

static inline int
xfs_ialloc_space_res(struct xfs_mount *mp)

{
	int	res = M_IGEO(mp)->ialloc_blks;

	res += M_IGEO(mp)->inobt_maxlevels - 1;
	if (xfs_sb_version_hasfinobt(&mp->m_sb))
		res += M_IGEO(mp)->inobt_maxlevels - 1;
	return res;
}

Next question: why is this even a macro that is calculated on demand instead of a read-only constant held in inode geometry calculated at mount time? Then it doesn't even need to be an inline function and can just be rolled into xfs_ialloc_setup_geometry()....

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
Reviewed-by
#2 lore raw
Quoted 44 lines · Dave Chinner
Quoted 22 lines · Brian Foster

The block reservation calculation for inode allocation is supposed to consist of the blocks required for the inode chunk plus (maxlevels-1) of the inode btree multiplied by the number of inode btrees in the fs (2 when finobt is enabled, 1 otherwise).

Instead, the macro returns (ialloc_blocks + 2) due to a precedence error in the calculation logic. This leads to block reservation overruns via generic/531 on small block filesystems with finobt enabled. Add braces to fix the calculation and reserve the appropriate number of blocks.

Fixes 9d43b180af67 ("xfs: update inode allocation/free transaction reservations for finobt")
---
 fs/xfs/libxfs/xfs_trans_space.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/xfs/libxfs/xfs_trans_space.h +0 -0
index 88221c7a04cc..c6df01a2a158 100644

fs/xfs/libxfs/xfs_trans_space.h line 61

Quoted 6 lines
fs/xfs/libxfs/xfs_trans_space.h +1 -1
@@ -57,7 +57,7 @@
 	XFS_DAREMOVE_SPACE_RES(mp, XFS_DATA_FORK)
 #define	XFS_IALLOC_SPACE_RES(mp)	\
 	(M_IGEO(mp)->ialloc_blks + \
-	 (xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1 * \
+	 ((xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1) * \
 	  (M_IGEO(mp)->inobt_maxlevels - 1)))

Ugh. THese macros really need rewriting as static inline functions. This would not have happened if it were written as:

static inline int
xfs_ialloc_space_res(struct xfs_mount *mp)

{
	int	res = M_IGEO(mp)->ialloc_blks;

	res += M_IGEO(mp)->inobt_maxlevels - 1;
	if (xfs_sb_version_hasfinobt(&mp->m_sb))
		res += M_IGEO(mp)->inobt_maxlevels - 1;
	return res;
}
Quoted 4 lines

Next question: why is this even a macro that is calculated on demand instead of a read-only constant held in inode geometry calculated at mount time? Then it doesn't even need to be an inline function and can just be rolled into xfs_ialloc_setup_geometry()....

Yeah, I hate those macros too. Fixing all that sounds like a <cough> cleanup series for someone, but in the meantime this is easy enough to backport to stable kernels.

--D

Quoted 6 lines

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#3 lore raw
Quoted 45 lines · Darrick J. Wong
Quoted 44 lines · Dave Chinner
Quoted 22 lines · Brian Foster

The block reservation calculation for inode allocation is supposed to consist of the blocks required for the inode chunk plus (maxlevels-1) of the inode btree multiplied by the number of inode btrees in the fs (2 when finobt is enabled, 1 otherwise).

Instead, the macro returns (ialloc_blocks + 2) due to a precedence error in the calculation logic. This leads to block reservation overruns via generic/531 on small block filesystems with finobt enabled. Add braces to fix the calculation and reserve the appropriate number of blocks.

Fixes 9d43b180af67 ("xfs: update inode allocation/free transaction reservations for finobt")
---
 fs/xfs/libxfs/xfs_trans_space.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/xfs/libxfs/xfs_trans_space.h +0 -0
index 88221c7a04cc..c6df01a2a158 100644

fs/xfs/libxfs/xfs_trans_space.h line 61

Quoted 6 lines
fs/xfs/libxfs/xfs_trans_space.h +1 -1
@@ -57,7 +57,7 @@
 	XFS_DAREMOVE_SPACE_RES(mp, XFS_DATA_FORK)
 #define	XFS_IALLOC_SPACE_RES(mp)	\
 	(M_IGEO(mp)->ialloc_blks + \
-	 (xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1 * \
+	 ((xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1) * \
 	  (M_IGEO(mp)->inobt_maxlevels - 1)))

Ugh. THese macros really need rewriting as static inline functions. This would not have happened if it were written as:

static inline int
xfs_ialloc_space_res(struct xfs_mount *mp)

{
	int	res = M_IGEO(mp)->ialloc_blks;

	res += M_IGEO(mp)->inobt_maxlevels - 1;
	if (xfs_sb_version_hasfinobt(&mp->m_sb))
		res += M_IGEO(mp)->inobt_maxlevels - 1;
	return res;
}
Quoted 8 lines
Quoted 4 lines

Next question: why is this even a macro that is calculated on demand instead of a read-only constant held in inode geometry calculated at mount time? Then it doesn't even need to be an inline function and can just be rolled into xfs_ialloc_setup_geometry()....

Yeah, I hate those macros too. Fixing all that sounds like a <cough> cleanup series for someone, but in the meantime this is easy enough to backport to stable kernels.

Well, I'm not suggesting that we have to fix all of them at once. Just converting this specific one to a IGEO variable is probably only 20 lines of code, which is still an "easy to backport" fix.

i.e. XFS_IALLOC_SPACE_RES() is used in just 7 places in the code, 4 of them are in that same header file, so it's a simple, standalone patch that fixes the bug by addressing the underlying cause of the problem (i.e. nasty macro!).

Historically speaking , we have cleaned up stuff like this to fix the bug, not done a one liner and then left fixing the root cause to some larger chunk of future work. The "one-liner" approach is largely a recent invention. I look at this sort of thing as being similar to cleaning up typedefs: we remove typedefs as we change surrounding code, thereby slowly remove them over time. We could just remove them all as one big patchset, but we don't do that because of all the outstanding work it would cause conflicts in.

Perhaps we've lost sight of the fact that doing things in little chunks on demand actually results in a lot of good cleanup change over time. We really don't have to do cleanups as one huge chunk of work all at once....

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#4 lore raw
Quoted 55 lines · Dave Chinner
Quoted 45 lines · Darrick J. Wong
Quoted 44 lines · Dave Chinner
Quoted 22 lines · Brian Foster

The block reservation calculation for inode allocation is supposed to consist of the blocks required for the inode chunk plus (maxlevels-1) of the inode btree multiplied by the number of inode btrees in the fs (2 when finobt is enabled, 1 otherwise).

Instead, the macro returns (ialloc_blocks + 2) due to a precedence error in the calculation logic. This leads to block reservation overruns via generic/531 on small block filesystems with finobt enabled. Add braces to fix the calculation and reserve the appropriate number of blocks.

Fixes 9d43b180af67 ("xfs: update inode allocation/free transaction reservations for finobt")
---
 fs/xfs/libxfs/xfs_trans_space.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/xfs/libxfs/xfs_trans_space.h +0 -0
index 88221c7a04cc..c6df01a2a158 100644

fs/xfs/libxfs/xfs_trans_space.h line 61

Quoted 6 lines
fs/xfs/libxfs/xfs_trans_space.h +1 -1
@@ -57,7 +57,7 @@
 	XFS_DAREMOVE_SPACE_RES(mp, XFS_DATA_FORK)
 #define	XFS_IALLOC_SPACE_RES(mp)	\
 	(M_IGEO(mp)->ialloc_blks + \
-	 (xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1 * \
+	 ((xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1) * \
 	  (M_IGEO(mp)->inobt_maxlevels - 1)))

Ugh. THese macros really need rewriting as static inline functions. This would not have happened if it were written as:

static inline int
xfs_ialloc_space_res(struct xfs_mount *mp)

{
	int	res = M_IGEO(mp)->ialloc_blks;

	res += M_IGEO(mp)->inobt_maxlevels - 1;
	if (xfs_sb_version_hasfinobt(&mp->m_sb))
		res += M_IGEO(mp)->inobt_maxlevels - 1;
	return res;
}
Quoted 8 lines
Quoted 4 lines

Next question: why is this even a macro that is calculated on demand instead of a read-only constant held in inode geometry calculated at mount time? Then it doesn't even need to be an inline function and can just be rolled into xfs_ialloc_setup_geometry()....

Yeah, I hate those macros too. Fixing all that sounds like a <cough> cleanup series for someone, but in the meantime this is easy enough to backport to stable kernels.

Quoted 9 lines

Well, I'm not suggesting that we have to fix all of them at once. Just converting this specific one to a IGEO variable is probably only 20 lines of code, which is still an "easy to backport" fix.

i.e. XFS_IALLOC_SPACE_RES() is used in just 7 places in the code, 4 of them are in that same header file, so it's a simple, standalone patch that fixes the bug by addressing the underlying cause of the problem (i.e. nasty macro!).

I agree that the inline is nicer than the macro, but a transaction reservation value seems misplaced to me in the IGEO. Perhaps having something analogous to struct xfs_trans_resv might be more appropriate.

Regardless, I agree with Darrick on the backporting situation. The original patch needs to be backportable to however many upstream stable releases back to v3.16 and similarly for distro kernels. Either patch might not be complex overall, but for somebody who might be processing hundreds of backports across various subsystems refactoring things as such in the same patch is clearly not equivalent to a one line change to an otherwise unchanged line since the original commit. I'll post a patch on top of this one to rework into an inline if people view that as preferable to the macro.

Brian

Quoted 21 lines

Historically speaking , we have cleaned up stuff like this to fix the bug, not done a one liner and then left fixing the root cause to some larger chunk of future work. The "one-liner" approach is largely a recent invention. I look at this sort of thing as being similar to cleaning up typedefs: we remove typedefs as we change surrounding code, thereby slowly remove them over time. We could just remove them all as one big patchset, but we don't do that because of all the outstanding work it would cause conflicts in.

Perhaps we've lost sight of the fact that doing things in little chunks on demand actually results in a lot of good cleanup change over time. We really don't have to do cleanups as one huge chunk of work all at once....

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#5 lore raw
Quoted 1 line · Brian Foster

On Thu, Jul 16, 2020 at 12:02:09PM +1000, Dave Chinner wrote:

...

Quoted 8 lines
Quoted 5 lines

i.e. XFS_IALLOC_SPACE_RES() is used in just 7 places in the code, 4 of them are in that same header file, so it's a simple, standalone patch that fixes the bug by addressing the underlying cause of the problem (i.e. nasty macro!).

I agree that the inline is nicer than the macro, but a transaction reservation value seems misplaced to me in the IGEO. Perhaps having something analogous to struct xfs_trans_resv might be more appropriate.

For whatever my opinion is worth these days, it seems like doing a survey to see how many of these reservations are static would be a good first step, and then decide where they should all go if they should move. I agree that IGEO might be a little odd, depending on what other static reservation types there are and what they're associated with.

I see both sides of the discussion re: how fixes like this move forward and what's easily backportable but in this case (and maybe I'm missing context) it seems like a wider survey would be wise before deciding to move this one value to IGEO in particular.

-Eric

#6 lore raw
Quoted 20 lines · Eric Sandeen
Quoted 1 line · Brian Foster

On Thu, Jul 16, 2020 at 12:02:09PM +1000, Dave Chinner wrote:

...

Quoted 8 lines
Quoted 5 lines

i.e. XFS_IALLOC_SPACE_RES() is used in just 7 places in the code, 4 of them are in that same header file, so it's a simple, standalone patch that fixes the bug by addressing the underlying cause of the problem (i.e. nasty macro!).

I agree that the inline is nicer than the macro, but a transaction reservation value seems misplaced to me in the IGEO. Perhaps having something analogous to struct xfs_trans_resv might be more appropriate.

For whatever my opinion is worth these days, it seems like doing a survey to see how many of these reservations are static would be a good first step, and then decide where they should all go if they should move. I agree that IGEO might be a little odd, depending on what other static reservation types there are and what they're associated with.

Quoted 4 lines

I see both sides of the discussion re: how fixes like this move forward and what's easily backportable but in this case (and maybe I'm missing context) it seems like a wider survey would be wise before deciding to move this one value to IGEO in particular.

Agreed. AFAICT the first patch is a bug fix for broken functionality, so I will put it in the 5.9 branch update next week.

--D

Quoted 1 line

-Eric

#7 lore raw

PATCH 2/2 xfs: replace ialloc space res macro with inline helper

In reply to Brian Foster

Rewrite the macro as a static inline helper to clean up the logic and have one less macro.

---
 fs/xfs/libxfs/xfs_trans_space.h | 24 ++++++++++++++++--------
 fs/xfs/xfs_inode.c              |  4 ++--
 fs/xfs/xfs_symlink.c            |  2 +-
 3 files changed, 19 insertions(+), 11 deletions(-)
fs/xfs/libxfs/xfs_trans_space.h +16 -8
index c6df01a2a158..d08dfc8795c3 100644
@@ -55,10 +55,18 @@
 	 XFS_DIRENTER_MAX_SPLIT(mp,nl))
 #define	XFS_DIRREMOVE_SPACE_RES(mp)	\
 	XFS_DAREMOVE_SPACE_RES(mp, XFS_DATA_FORK)
-#define	XFS_IALLOC_SPACE_RES(mp)	\
-	(M_IGEO(mp)->ialloc_blks + \
-	 ((xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1) * \
-	  (M_IGEO(mp)->inobt_maxlevels - 1)))
+
+static inline int
+xfs_ialloc_space_res(
+	struct xfs_mount	*mp)
+{
+	int			res = M_IGEO(mp)->ialloc_blks;
+
+	res += M_IGEO(mp)->inobt_maxlevels - 1;
+	if (xfs_sb_version_hasfinobt(&mp->m_sb))
+		res += M_IGEO(mp)->inobt_maxlevels - 1;
+	return res;
+}
 
 /*
  * Space reservation values for various transactions.
@@ -71,7 +79,7 @@
 #define	XFS_ATTRSET_SPACE_RES(mp, v)	\
 	(XFS_DAENTER_SPACE_RES(mp, XFS_ATTR_FORK) + XFS_B_TO_FSB(mp, v))
 #define	XFS_CREATE_SPACE_RES(mp,nl)	\
-	(XFS_IALLOC_SPACE_RES(mp) + XFS_DIRENTER_SPACE_RES(mp,nl))
+	(xfs_ialloc_space_res(mp) + XFS_DIRENTER_SPACE_RES(mp,nl))
 #define	XFS_DIOSTRAT_SPACE_RES(mp, v)	\
 	(XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK) + (v))
 #define	XFS_GROWFS_SPACE_RES(mp)	\
@@ -81,18 +89,18 @@
 #define	XFS_LINK_SPACE_RES(mp,nl)	\
 	XFS_DIRENTER_SPACE_RES(mp,nl)
 #define	XFS_MKDIR_SPACE_RES(mp,nl)	\
-	(XFS_IALLOC_SPACE_RES(mp) + XFS_DIRENTER_SPACE_RES(mp,nl))
+	(xfs_ialloc_space_res(mp) + XFS_DIRENTER_SPACE_RES(mp,nl))
 #define	XFS_QM_DQALLOC_SPACE_RES(mp)	\
 	(XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK) + \
 	 XFS_DQUOT_CLUSTER_SIZE_FSB)
 #define	XFS_QM_QINOCREATE_SPACE_RES(mp)	\
-	XFS_IALLOC_SPACE_RES(mp)
+	xfs_ialloc_space_res(mp)
 #define	XFS_REMOVE_SPACE_RES(mp)	\
 	XFS_DIRREMOVE_SPACE_RES(mp)
 #define	XFS_RENAME_SPACE_RES(mp,nl)	\
 	(XFS_DIRREMOVE_SPACE_RES(mp) + XFS_DIRENTER_SPACE_RES(mp,nl))
 #define	XFS_SYMLINK_SPACE_RES(mp,nl,b)	\
-	(XFS_IALLOC_SPACE_RES(mp) + XFS_DIRENTER_SPACE_RES(mp,nl) + (b))
+	(xfs_ialloc_space_res(mp) + XFS_DIRENTER_SPACE_RES(mp,nl) + (b))
 #define XFS_IFREE_SPACE_RES(mp)		\
 	(xfs_sb_version_hasfinobt(&mp->m_sb) ? \
 			M_IGEO(mp)->inobt_maxlevels : 0)
fs/xfs/xfs_inode.c +2 -2
index 5c07bf491d9f..3420bc595e1b 100644
@@ -1195,7 +1195,7 @@ xfs_create(
 	unlock_dp_on_error = false;
 
 	error = xfs_dir_createname(tp, dp, name, ip->i_ino,
-					resblks - XFS_IALLOC_SPACE_RES(mp));
+					resblks - xfs_ialloc_space_res(mp));
 	if (error) {
 		ASSERT(error != -ENOSPC);
 		goto out_trans_cancel;
@@ -1290,7 +1290,7 @@ xfs_create_tmpfile(
 	if (error)
 		return error;
 
-	resblks = XFS_IALLOC_SPACE_RES(mp);
+	resblks = xfs_ialloc_space_res(mp);
 	tres = &M_RES(mp)->tr_create_tmpfile;
 
 	error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
fs/xfs/xfs_symlink.c +1 -1
index 8e88a7ca387e..b0a58cdd4c78 100644
@@ -243,7 +243,7 @@ xfs_symlink(
 	 */
 	xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
 
-	resblks -= XFS_IALLOC_SPACE_RES(mp);
+	resblks -= xfs_ialloc_space_res(mp);
 	/*
 	 * If the symlink will fit into the inode, write it inline.
 	 */
Signature · 1 line
2.21.3
#8 lore raw

Re: [PATCH 2/2] xfs: replace ialloc space res macro with inline helper

Quoted 24 lines · Brian Foster

Rewrite the macro as a static inline helper to clean up the logic and have one less macro.

---
 fs/xfs/libxfs/xfs_trans_space.h | 24 ++++++++++++++++--------
 fs/xfs/xfs_inode.c              |  4 ++--
 fs/xfs/xfs_symlink.c            |  2 +-
 3 files changed, 19 insertions(+), 11 deletions(-)
fs/xfs/libxfs/xfs_trans_space.h +2 -4
index c6df01a2a158..d08dfc8795c3 100644
@@ -55,10 +55,18 @@
 	 XFS_DIRENTER_MAX_SPLIT(mp,nl))
 #define	XFS_DIRREMOVE_SPACE_RES(mp)	\
 	XFS_DAREMOVE_SPACE_RES(mp, XFS_DATA_FORK)
-#define	XFS_IALLOC_SPACE_RES(mp)	\
-	(M_IGEO(mp)->ialloc_blks + \
-	 ((xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1) * \
-	  (M_IGEO(mp)->inobt_maxlevels - 1)))
+
+static inline int

fs/xfs/libxfs/xfs_trans_space.h line 69

Quoted 10 lines
fs/xfs/libxfs/xfs_trans_space.h +10 -0
+xfs_ialloc_space_res(
+	struct xfs_mount	*mp)
+{
+	int			res = M_IGEO(mp)->ialloc_blks;
+
+	res += M_IGEO(mp)->inobt_maxlevels - 1;
+	if (xfs_sb_version_hasfinobt(&mp->m_sb))
+		res += M_IGEO(mp)->inobt_maxlevels - 1;
+	return res;
+}

This misses the point I made. i.e. that the space reservation is constant and never changes, yet we calculate it -twice- per inode create. That means we can be calculating it hundreds of thousands of times a second instead of just reading a variable that is likely hot in cache.

IOWs, if we are going to improve this code, it should to be moved to a pre-calculated, read-only, per-mount variable so the repeated calculation goes away entirely.

Then the macro/function goes away entirely an is replaced simply by mp->m_ialloc_space_res or M_IGEO(mp)->alloc_space_res....

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#9 lore raw

Re: [PATCH 2/2] xfs: replace ialloc space res macro with inline helper

Quoted 33 lines · Dave Chinner
Quoted 32 lines · Brian Foster

Rewrite the macro as a static inline helper to clean up the logic and have one less macro.

---
 fs/xfs/libxfs/xfs_trans_space.h | 24 ++++++++++++++++--------
 fs/xfs/xfs_inode.c              |  4 ++--
 fs/xfs/xfs_symlink.c            |  2 +-
 3 files changed, 19 insertions(+), 11 deletions(-)
fs/xfs/libxfs/xfs_trans_space.h +10 -4
index c6df01a2a158..d08dfc8795c3 100644
@@ -55,10 +55,18 @@
 	 XFS_DIRENTER_MAX_SPLIT(mp,nl))
 #define	XFS_DIRREMOVE_SPACE_RES(mp)	\
 	XFS_DAREMOVE_SPACE_RES(mp, XFS_DATA_FORK)
-#define	XFS_IALLOC_SPACE_RES(mp)	\
-	(M_IGEO(mp)->ialloc_blks + \
-	 ((xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1) * \
-	  (M_IGEO(mp)->inobt_maxlevels - 1)))
+
+static inline int
+xfs_ialloc_space_res(
+	struct xfs_mount	*mp)
+{
+	int			res = M_IGEO(mp)->ialloc_blks;
+
+	res += M_IGEO(mp)->inobt_maxlevels - 1;
+	if (xfs_sb_version_hasfinobt(&mp->m_sb))
+		res += M_IGEO(mp)->inobt_maxlevels - 1;
Quoted 9 lines

fs/xfs/libxfs/xfs_trans_space.h line 69

Quoted 2 lines
fs/xfs/libxfs/xfs_trans_space.h +2 -0
+	return res;
+}

This misses the point I made. i.e. that the space reservation is constant and never changes, yet we calculate it -twice- per inode create. That means we can be calculating it hundreds of thousands of times a second instead of just reading a variable that is likely hot in cache.

Partly.. I mentioned in my earlier reply that the geometry structure doesn't seem like the right place to stuff transaction reservation related values. An alternative I mentioned would be a new (or update to the existing) structure that similarly precalculates log reservations, but I wasn't going to go do that without some discussion/feedback on the approach. Replacing the macro seemed broadly acceptable, so I sent that patch as an incremental improvement and to keep the bug fix isolated.

Also, while stuffing a new value in a pre-existing structure might lend itself to a one-off patch, I'm not sure that creating a new data structure does lest it fail to justify the existence of the structure itself. Therefore, it might be better to create a small series to convert over several values to start such a structure and perhaps do the rest over time to reduce the churn..

Quoted 4 lines

IOWs, if we are going to improve this code, it should to be moved to a pre-calculated, read-only, per-mount variable so the repeated calculation goes away entirely.

I figured replacing the macro was an incremental improvement independent from how precalculated transaction block reservations might be structured. TBH, I don't mind the macros as much as others seem to, so feel free to defer or discard this patch altogether..

Quoted 3 lines

Then the macro/function goes away entirely an is replaced simply by mp->m_ialloc_space_res or M_IGEO(mp)->alloc_space_res....

Feel free to comment on my feedback on this in the previous reply..

Brian

Quoted 7 lines

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]