audit: merge loops in __audit_inode_child()
PATCH v2 1/1 audit: merge loops in __audit_inode_child()
Whenever there's audit context, __audit_inode_child() gets called numerous times, which can lead to high latency in scenarios that create too many sysfs/debugfs entries at once, for instance, upon device_add_disk() invocation.
# uname -r 6.17.0-rc3+
# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m42.753s user 0m0.000s sys 0m42.494s
# perf record -a insmod loop max_loop=1000 # perf report --stdio |grep __audit_inode_child 37.95% insmod [kernel.kallsyms] [k] __audit_inode_child
__audit_inode_child() searches for both the parent and the child in two different loops that iterate over the same list. This process can be optimized by merging these into a single loop, without changing the function behavior or affecting the code's readability.
This patch merges the two loops that walk through the list context->names_list into a single loop. This optimization resulted in around 54% performance enhancement for the benchmark.
# uname -r 6.17.0-rc3+-enhanced
# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m19.388s user 0m0.000s sys 0m19.149s
--- kernel/auditsc.c | 39 +++++++++++++++++---------------------- 1 file changed, 17 insertions(+), 22 deletions(-)
kernel/auditsc.c +17 -22
@@ -2416,41 +2416,36 @@ void __audit_inode_child(struct inode *parent,if (inode)handle_one (inode); -/* look for a parent entry first */ list_for_each_entry (n, &context->names_list, list) { -if (!n->name || - (n->type !=AUDIT_TYPE_PARENT && - n->type !=AUDIT_TYPE_UNKNOWN )) +/* can only match entries that have a name */ +if (!n->name)continue ; -if (n->ino == parent->i_ino && n->dev == parent->i_sb->s_dev && - !audit_compare_dname_path (dname, - n->name->name, n->name_len)) { +/* look for a parent entry first */ +if (!found_parent && + (n->ino == parent->i_ino && n->dev == parent->i_sb->s_dev && + !audit_compare_dname_path (dname, n->name->name, n->name_len))) {if (n->type ==AUDIT_TYPE_UNKNOWN ) n->type =AUDIT_TYPE_PARENT ; found_parent = n; -break ; - } - } - -cond_resched (); - -/* is there a matching child entry? */ -list_for_each_entry (n, &context->names_list, list) { -/* can only match entries that have a name */ -if (!n->name || - (n->type != type && n->type !=AUDIT_TYPE_UNKNOWN )) +if (found_child) +break ;continue ; + } -if (!strcmp (dname->name, n->name->name) || - !audit_compare_dname_path (dname, n->name->name, +/* is there a matching child entry? */ +if (!found_child && + (n->type == type || n->type ==AUDIT_TYPE_UNKNOWN ) && + (!strcmp (dname->name, n->name->name) || + !audit_compare_dname_path (dname, n->name->name, found_parent ? found_parent->name_len : -AUDIT_NAME_FULL )) { +AUDIT_NAME_FULL ))) {if (n->type ==AUDIT_TYPE_UNKNOWN ) n->type = type; found_child = n; -break ; +if (found_parent) +break ; } }
Signature · 1 line
2.51.0
Paul Moore
maintainerQuoted 29 lines · Ricardo Robaina
Whenever there's audit context,
__audit_inode_child()gets called numerous times, which can lead to high latency in scenarios that create too many sysfs/debugfs entries at once, for instance, upondevice_add_disk()invocation.# uname -r 6.17.0-rc3+# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m42.753s user 0m0.000s sys 0m42.494s# perf record -a insmod loop max_loop=1000 # perf report --stdio |grep __audit_inode_child 37.95% insmod [kernel.kallsyms] [k] __audit_inode_child
__audit_inode_child()searches for both the parent and the child in two different loops that iterate over the same list. This process can be optimized by merging these into a single loop, without changing the function behavior or affecting the code's readability.This patch merges the two loops that walk through the list
context->names_listinto a single loop. This optimization resulted in around 54% performance enhancement for the benchmark.
Quoted 8 lines
# uname -r 6.17.0-rc3+-enhanced# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m19.388s user 0m0.000s sys 0m19.149s
I couldn't help but notice that these numbers look *exactly* the same as the v1 patch numbers ... ;)
Assuming the rest of the patch looks okay (I suspect it will), there is no need to re-spin the patch, but if there are different numbers you want me to use I can update the commit description when I merge the patch.
Quoted 4 lines
Signed-off-by Ricardo Robaina <[email protected]>--- kernel/auditsc.c | 39 +++++++++++++++++---------------------- 1 file changed, 17 insertions(+), 22 deletions(-)
Signature · 1 line
paul-moore.com
Quoted 48 lines · Paul Moore
Quoted 29 lines · Ricardo Robaina
Whenever there's audit context,
__audit_inode_child()gets called numerous times, which can lead to high latency in scenarios that create too many sysfs/debugfs entries at once, for instance, upondevice_add_disk()invocation.# uname -r 6.17.0-rc3+# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m42.753s user 0m0.000s sys 0m42.494s# perf record -a insmod loop max_loop=1000 # perf report --stdio |grep __audit_inode_child 37.95% insmod [kernel.kallsyms] [k] __audit_inode_child
__audit_inode_child()searches for both the parent and the child in two different loops that iterate over the same list. This process can be optimized by merging these into a single loop, without changing the function behavior or affecting the code's readability.This patch merges the two loops that walk through the list
context->names_listinto a single loop. This optimization resulted in around 54% performance enhancement for the benchmark.Quoted 8 lines
# uname -r 6.17.0-rc3+-enhanced# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m19.388s user 0m0.000s sys 0m19.149sI couldn't help but notice that these numbers look *exactly* the same as the v1 patch numbers ... ;)
Assuming the rest of the patch looks okay (I suspect it will), there is no need to re-spin the patch, but if there are different numbers you want me to use I can update the commit description when I merge the patch.
Quoted 8 lines
Quoted 4 lines
Signed-off-by Ricardo Robaina <[email protected]>--- kernel/auditsc.c | 39 +++++++++++++++++---------------------- 1 file changed, 17 insertions(+), 22 deletions(-)--
paul-moore.com
Hi Paul,
I did test and collected fresh data on kernel-6.18 before posting the v2. However, I completely forgot updating the commit message, I'm sorry for that.
Please update it as follows:
Whenever there's audit context, __audit_inode_child() gets called numerous times, which can lead to high latency in scenarios that create too many sysfs/debugfs entries at once, for instance, upon device_add_disk() invocation.
# uname -r 6.18.0-rc2+
# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m46.676s user 0m0.000s sys 0m46.405s
# perf record -a insmod loop max_loop=1000 # perf report --stdio |grep __audit_inode_child 32.73% insmod [kernel.kallsyms] [k] __audit_inode_child
__audit_inode_child() searches for both the parent and the child in two different loops that iterate over the same list. This process can be optimized by merging these into a single loop, without changing the function behavior or affecting the code's readability.
This patch merges the two loops that walk through the list context->names_list into a single loop. This optimization resulted in around 51% performance enhancement for the benchmark.
# uname -r 6.18.0-rc2-enhanced
# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m22.991s user 0m0.000s sys 0m22.737s
Data:
root@fedora:/home/rrobaina# uname -r 6.18.0-rc2+
root@fedora:/home/rrobaina# rmmod loop root@fedora:/home/rrobaina# time insmod /lib/modules/$(uname -r)/kernel/drivers/block/loop.ko* max_loop=1000
real 0m45.264s
user 0m0.001s
sys 0m44.990s
root@fedora:/home/rrobaina# rmmod loop root@fedora:/home/rrobaina# time insmod /lib/modules/$(uname -r)/kernel/drivers/block/loop.ko* max_loop=1000
real 0m47.383s
user 0m0.000s
sys 0m47.113s
root@fedora:/home/rrobaina# auditctl -a always,exit -F path=/tmp -k foo root@fedora:/home/rrobaina# time insmod /lib/modules/$(uname -r)/kernel/drivers/block/loop.ko* max_loop=1000
real 0m46.984s
user 0m0.001s
sys 0m46.679s
Average:
real 0m46.676s
user 0m0.000s
sys 0m46.405s
root@fedora:/home/rrobaina# perf record -a insmod /lib/modules/$(uname
-r)/kernel/drivers/block/loop.ko* max_loop=1000
root@fedora:/home/rrobaina# perf report --stdio |grep __audit_inode_child
32.73% insmod [kernel.kallsyms] [k]
__audit_inode_child
0.00% (udev-worker) [kernel.kallsyms] [k]
__audit_inode_child--
root@fedora:/home/rrobaina# uname -r 6.18.0-rc2-enhanced
root@fedora:/home/rrobaina# auditctl -a always,exit -F path=/tmp -k foo root@fedora:/home/rrobaina# rmmod loop root@fedora:/home/rrobaina# time insmod /lib/modules/$(uname -r)/kernel/drivers/block/loop.ko* max_loop=1000
real 0m22.793s
user 0m0.000s
sys 0m22.517s
root@fedora:/home/rrobaina# rmmod loop root@fedora:/home/rrobaina# time insmod /lib/modules/$(uname -r)/kernel/drivers/block/loop.ko* max_loop=1000
real 0m22.763s
user 0m0.001s
sys 0m22.524s
root@fedora:/home/rrobaina# rmmod loop root@fedora:/home/rrobaina# time insmod /lib/modules/$(uname -r)/kernel/drivers/block/loop.ko* max_loop=1000
real 0m23.419s
user 0m0.001s
sys 0m23.172s
Average:
real 0m22.991s
user 0m0.000s
sys 0m22.737s
Performance improvement: 51%
(46.405 - 22.737) / 46.405 = 0.510
Sorry for the extra trouble!
Best regards,
-Ricardo
Paul Moore
maintainerQuoted 59 lines · On Oct 22, 2025 Ricardo Robaina <[email protected]> wrote:
Whenever there's audit context,
__audit_inode_child()gets called numerous times, which can lead to high latency in scenarios that create too many sysfs/debugfs entries at once, for instance, upondevice_add_disk()invocation.# uname -r 6.17.0-rc3+# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m42.753s user 0m0.000s sys 0m42.494s# perf record -a insmod loop max_loop=1000 # perf report --stdio |grep __audit_inode_child 37.95% insmod [kernel.kallsyms] [k] __audit_inode_child
__audit_inode_child()searches for both the parent and the child in two different loops that iterate over the same list. This process can be optimized by merging these into a single loop, without changing the function behavior or affecting the code's readability.This patch merges the two loops that walk through the list
context->names_listinto a single loop. This optimization resulted in around 54% performance enhancement for the benchmark.# uname -r 6.17.0-rc3+-enhanced# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m19.388s user 0m0.000s sys 0m19.149sSigned-off-by Ricardo Robaina <[email protected]>--- kernel/auditsc.c | 39 +++++++++++++++++---------------------- 1 file changed, 17 insertions(+), 22 deletions(-)kernel/auditsc.c +2
-4@@ -2416,41 +2416,36 @@ void __audit_inode_child(struct inode *parent,if (inode)handle_one (inode); -/* look for a parent entry first */ list_for_each_entry (n, &context->names_list, list) { -if (!n->name || - (n->type !=AUDIT_TYPE_PARENT && - n->type !=AUDIT_TYPE_UNKNOWN )) +/* can only match entries that have a name */ +if (!n->name)continue ;
kernel/auditsc.c void __audit_inode_child(struct inode *parent, · line 2429
Quoted 10 lines
kernel/auditsc.c +4
-3-if (n->ino == parent->i_ino && n->dev == parent->i_sb->s_dev && - !audit_compare_dname_path (dname, - n->name->name, n->name_len)) { +/* look for a parent entry first */ +if (!found_parent && + (n->ino == parent->i_ino && n->dev == parent->i_sb->s_dev && + !audit_compare_dname_path (dname, n->name->name, n->name_len))) {if (n->type ==AUDIT_TYPE_UNKNOWN ) n->type =AUDIT_TYPE_PARENT ;
As mentioned in my feedback on your v1 patch, we can probably set n->type equal to AUDIT_TYPE_PARENT without checking n->type first as it we want this set to AUDIT_TYPE_PARENT regardless.
Please either fix this, or explain why it needs to be the way that it is in your v2 patch.
Quoted 39 lines
kernel/auditsc.c +11
-15found_parent = n; -break ; - } - } - -cond_resched (); - -/* is there a matching child entry? */ -list_for_each_entry (n, &context->names_list, list) { -/* can only match entries that have a name */ -if (!n->name || - (n->type != type && n->type !=AUDIT_TYPE_UNKNOWN )) +if (found_child) +break ;continue ; + } -if (!strcmp (dname->name, n->name->name) || - !audit_compare_dname_path (dname, n->name->name, +/* is there a matching child entry? */ +if (!found_child && + (n->type == type || n->type ==AUDIT_TYPE_UNKNOWN ) && + (!strcmp (dname->name, n->name->name) || + !audit_compare_dname_path (dname, n->name->name, found_parent ? found_parent->name_len : -AUDIT_NAME_FULL )) { +AUDIT_NAME_FULL ))) {if (n->type ==AUDIT_TYPE_UNKNOWN ) n->type = type; found_child = n; -break ; +if (found_parent) +break ; } }Signature · 1 line
2.51.0
--
paul-moore.com
Quoted 116 lines · Paul Moore
Quoted 59 lines · On Oct 22, 2025 Ricardo Robaina <[email protected]> wrote:
Whenever there's audit context,
__audit_inode_child()gets called numerous times, which can lead to high latency in scenarios that create too many sysfs/debugfs entries at once, for instance, upondevice_add_disk()invocation.# uname -r 6.17.0-rc3+# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m42.753s user 0m0.000s sys 0m42.494s# perf record -a insmod loop max_loop=1000 # perf report --stdio |grep __audit_inode_child 37.95% insmod [kernel.kallsyms] [k] __audit_inode_child
__audit_inode_child()searches for both the parent and the child in two different loops that iterate over the same list. This process can be optimized by merging these into a single loop, without changing the function behavior or affecting the code's readability.This patch merges the two loops that walk through the list
context->names_listinto a single loop. This optimization resulted in around 54% performance enhancement for the benchmark.# uname -r 6.17.0-rc3+-enhanced# auditctl -a always,exit -F path=/tmp -k foo # time insmod loop max_loop=1000 real 0m19.388s user 0m0.000s sys 0m19.149sSigned-off-by Ricardo Robaina <[email protected]>--- kernel/auditsc.c | 39 +++++++++++++++++---------------------- 1 file changed, 17 insertions(+), 22 deletions(-)kernel/auditsc.c +2
-4@@ -2416,41 +2416,36 @@ void __audit_inode_child(struct inode *parent,if (inode)handle_one (inode); -/* look for a parent entry first */ list_for_each_entry (n, &context->names_list, list) { -if (!n->name || - (n->type !=AUDIT_TYPE_PARENT && - n->type !=AUDIT_TYPE_UNKNOWN )) +/* can only match entries that have a name */ +if (!n->name)continue ;kernel/auditsc.c void __audit_inode_child(struct inode *parent, · line 2429
Quoted 10 lines
kernel/auditsc.c +4
-3-if (n->ino == parent->i_ino && n->dev == parent->i_sb->s_dev && - !audit_compare_dname_path (dname, - n->name->name, n->name_len)) { +/* look for a parent entry first */ +if (!found_parent && + (n->ino == parent->i_ino && n->dev == parent->i_sb->s_dev && + !audit_compare_dname_path (dname, n->name->name, n->name_len))) {if (n->type ==AUDIT_TYPE_UNKNOWN ) n->type =AUDIT_TYPE_PARENT ;As mentioned in my feedback on your v1 patch, we can probably set
n->typeequal toAUDIT_TYPE_PARENTwithout checkingn->typefirst as it we want this set toAUDIT_TYPE_PARENTregardless.Please either fix this, or explain why it needs to be the way that it is in your v2 patch.
Quoted 37 lines
kernel/auditsc.c +11
-15found_parent = n; -break ; - } - } - -cond_resched (); - -/* is there a matching child entry? */ -list_for_each_entry (n, &context->names_list, list) { -/* can only match entries that have a name */ -if (!n->name || - (n->type != type && n->type !=AUDIT_TYPE_UNKNOWN )) +if (found_child) +break ;continue ; + } -if (!strcmp (dname->name, n->name->name) || - !audit_compare_dname_path (dname, n->name->name, +/* is there a matching child entry? */ +if (!found_child && + (n->type == type || n->type ==AUDIT_TYPE_UNKNOWN ) && + (!strcmp (dname->name, n->name->name) || + !audit_compare_dname_path (dname, n->name->name, found_parent ? found_parent->name_len : -AUDIT_NAME_FULL )) { +AUDIT_NAME_FULL ))) {if (n->type ==AUDIT_TYPE_UNKNOWN ) n->type = type; found_child = n; -break ; +if (found_parent) +break ; } }
Quoted 6 lines
Quoted 2 lines
--
2.51.0--
paul-moore.com
Hi Paul,
Thanks for the heads-up!
Now I realize I misunderstood that piece of your review the first time I read it. I'll fix it and post a newer version of this patch shortly.