[PATCH 2/2] fs: track the inode having file locks with a flag in ->i_opflags

[Mateusz Guzik <mjguzik@gmail.com>]

Opening and closing an inode dirties the ->i_readcount field.

There is false-sharing with both of these operations, extent of which
depends on how misaligned the inode happens to be.

This notably concerns the ->i_flctx field. Since most inodes don't have
the field populated, this bit can be managed with a flag in ->i_opflags
instead which bypasses the problem.

Here are results I obtained while opening a file read-only in a loop
with 24 cores doing the work on Sapphire Rapids. Utilizing the flag as
opposed to reading ->i_flctx field was toggled at runtime as the benchmark
was running, to make sure both results come from the same alignment.

before: 3233740
after:  3373346 (+4%)

before: 3284313
after:  3518711 (+7%)

before: 3505545
after:  4092806 (+16%)

Or to put it differently, this varies wildly depending on how (un)lucky
you get.

The primary bottleneck before and after is the avoidable lockref trip in
do_dentry_open(), afterwards it's a toss up between apparmor and
->i_readcount.

Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
---

of course someone(tm) should sort out struct inode, but that's a longer
story and even then it may be beneficial to not load i_flctx.

bench code for will-it-scale below.

It differs from open3 in that it is not read-write. Opening a file
read-write in parallel suffers *more* false-sharing and is less common,
thus a more likely case of read-only.

# cat tests/openro3.c
#include <stdlib.h>
#include <unistd.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <assert.h>

static char tmpfile[] = "/tmp/willitscale.XXXXXX";

char *testcase_description = "Same file open/close read-only";

void testcase_prepare(unsigned long nr_tasks)
{
	int fd = mkstemp(tmpfile);

	assert(fd >= 0);
	close(fd);
}

void testcase(unsigned long long *iterations, unsigned long nr)
{
	while (1) {
		int fd = open(tmpfile, O_RDONLY);
		assert(fd >= 0);
		close(fd);

		(*iterations)++;
	}
}

void testcase_cleanup(void)
{
	unlink(tmpfile);
}

 fs/locks.c               | 14 ++++++++++++--
 include/linux/filelock.h | 15 +++++++++++----
 include/linux/fs.h       |  1 +
 3 files changed, 24 insertions(+), 6 deletions(-)

diff --git a/fs/locks.c b/fs/locks.c
index 04a3f0e20724..9cd23dd2a74e 100644
--- a/fs/locks.c
+++ b/fs/locks.c
@@ -178,7 +178,6 @@ locks_get_lock_context(struct inode *inode, int type)
 {
 	struct file_lock_context *ctx;
 
-	/* paired with cmpxchg() below */
 	ctx = locks_inode_context(inode);
 	if (likely(ctx) || type == F_UNLCK)
 		goto out;
@@ -196,7 +195,18 @@ locks_get_lock_context(struct inode *inode, int type)
 	 * Assign the pointer if it's not already assigned. If it is, then
 	 * free the context we just allocated.
 	 */
-	if (cmpxchg(&inode->i_flctx, NULL, ctx)) {
+	spin_lock(&inode->i_lock);
+	if (!(inode->i_opflags & IOP_FLCTX)) {
+		VFS_BUG_ON_INODE(inode->i_flctx, inode);
+		WRITE_ONCE(inode->i_flctx, ctx);
+		/*
+		 * Paired with locks_inode_context().
+		 */
+		smp_store_release(&inode->i_opflags, inode->i_opflags | IOP_FLCTX);
+		spin_unlock(&inode->i_lock);
+	} else {
+		VFS_BUG_ON_INODE(!inode->i_flctx, inode);
+		spin_unlock(&inode->i_lock);
 		kmem_cache_free(flctx_cache, ctx);
 		ctx = locks_inode_context(inode);
 	}
diff --git a/include/linux/filelock.h b/include/linux/filelock.h
index 37e1b33bd267..e23cbc562534 100644
--- a/include/linux/filelock.h
+++ b/include/linux/filelock.h
@@ -233,8 +233,12 @@ static inline struct file_lock_context *
 locks_inode_context(const struct inode *inode)
 {
 	/*
-	 * Paired with the fence in locks_get_lock_context().
+	 * Paired with smp_store_release in locks_get_lock_context().
+	 *
+	 * Ensures ->i_flctx will be visible if we spotted the flag.
 	 */
+	if (likely(!(smp_load_acquire(&inode->i_opflags) & IOP_FLCTX)))
+		return 0;
 	return READ_ONCE(inode->i_flctx);
 }
 
@@ -441,7 +445,7 @@ static inline int break_lease(struct inode *inode, unsigned int mode)
 	 * could end up racing with tasks trying to set a new lease on this
 	 * file.
 	 */
-	flctx = READ_ONCE(inode->i_flctx);
+	flctx = locks_inode_context(inode);
 	if (!flctx)
 		return 0;
 	smp_mb();
@@ -460,7 +464,7 @@ static inline int break_deleg(struct inode *inode, unsigned int mode)
 	 * could end up racing with tasks trying to set a new lease on this
 	 * file.
 	 */
-	flctx = READ_ONCE(inode->i_flctx);
+	flctx = locks_inode_context(inode);
 	if (!flctx)
 		return 0;
 	smp_mb();
@@ -493,8 +497,11 @@ static inline int break_deleg_wait(struct inode **delegated_inode)
 
 static inline int break_layout(struct inode *inode, bool wait)
 {
+	struct file_lock_context *flctx;
+
 	smp_mb();
-	if (inode->i_flctx && !list_empty_careful(&inode->i_flctx->flc_lease))
+	flctx = locks_inode_context(inode);
+	if (flctx && !list_empty_careful(&flctx->flc_lease))
 		return __break_lease(inode,
 				wait ? O_WRONLY : O_WRONLY | O_NONBLOCK,
 				FL_LAYOUT);
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 7c2aa286ef87..314a1349747b 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -655,6 +655,7 @@ is_uncached_acl(struct posix_acl *acl)
 #define IOP_MGTIME		0x0020
 #define IOP_CACHED_LINK		0x0040
 #define IOP_FASTPERM_MAY_EXEC	0x0080
+#define IOP_FLCTX		0x0100
 
 /*
  * Inode state bits.  Protected by inode->i_lock
-- 
2.48.1