[PATCH -next 0/1] audit: ignore userspace log when backlog limit exceeded and backlog_wait_time is 0
12:50 a.m.
The audit backlog overflow is an actual problem in our system. We set
the backlog_wait_time to 0 for improving OS performance. Therefore,
under the impact of a large number of audit logs produced by sudo, the
backlog in buffer will greatly exceed backlog_limit.
When the backlog exceed the backlog_limit and backlog_wait_time is set
to 0, the process will only sleep for a very short time (jiffies). The
backlog may still exceed backlog_limit in extreme cases.
The more reasonable way to fix this problem is:
1. If backlog_wait_time is set to zero, ignore the log;
2. If backlog_wait_time is set to non-zero, let process sleep for
backlog_wait_time.
The above log limit logic is also the same as that in the existing
audit_log_start function.
This may not be a perfect solution, a further discussion maybe
necessary.
luhuaxin (1):
audit: ignore userspace log when backlog limit exceeded and
backlog_wait_time is 0
kernel/audit.c | 14 +++++++++++---
1 file changed, 11 insertions(+), 3 deletions(-)
--
2.23.0
12:50 a.m.
New subject: [PATCH -next 1/1] audit: ignore userspace log when backlog limit exceeded and backlog_wait_time is 0
When the backlog exceed the backlog_limit and backlog_wait_time is set
to 0, the process will only sleep for a very short time (jiffies). The
backlog may still exceed backlog_limit in extreme cases.
The more reasonable way to fix this problem is:
1. If backlog_wait_time is set to zero, ignore the log;
2. If backlog_wait_time is set to non-zero, let process sleep for
backlog_wait_time.
The above log limit logic is also the same as that in the existing
audit_log_start function.
Fixes: 8f110f530635 ("[PATCH] audit: ensure userspace is penalized the
same as the kernel when under pressure")
Signed-off-by: luhuaxin <luhuaxin1(a)huawei.com>
---
kernel/audit.c | 14 +++++++++++---
1 file changed, 11 insertions(+), 3 deletions(-)
diff --git a/kernel/audit.c b/kernel/audit.c
index 249e11628..70450f70a 100644
--- a/kernel/audit.c
+++ b/kernel/audit.c
@@ -1545,7 +1545,8 @@ static void audit_receive(struct sk_buff *skb)
/* can't block with the ctrl lock, so penalize the sender now */
if (audit_backlog_limit &&
- (skb_queue_len(&audit_queue) > audit_backlog_limit)) {
+ (skb_queue_len(&audit_queue) > audit_backlog_limit) &&
+ audit_backlog_wait_time) {
DECLARE_WAITQUEUE(wait, current);
/* wake kauditd to try and flush the queue */
@@ -1842,9 +1843,8 @@ struct audit_buffer *audit_log_start(struct audit_context *ctx,
gfp_t gfp_mask,
* while holding the mutex, although we do penalize the sender
* later in audit_receive() when it is safe to block
*/
+ long stime = audit_backlog_wait_time;
if (!(auditd_test_task(current) || audit_ctl_owner_current())) {
- long stime = audit_backlog_wait_time;
-
while (audit_backlog_limit &&
(skb_queue_len(&audit_queue) > audit_backlog_limit)) {
/* wake kauditd to try and flush the queue */
@@ -1872,6 +1872,14 @@ struct audit_buffer *audit_log_start(struct audit_context *ctx,
gfp_t gfp_mask,
return NULL;
}
}
+ } else if (!stime && audit_backlog_limit &&
+ (skb_queue_len(&audit_queue) > audit_backlog_limit)) {
+ if (audit_rate_check() && printk_ratelimit())
+ pr_warn("audit_backlog=%d > audit_backlog_limit=%d\n",
+ skb_queue_len(&audit_queue),
+ audit_backlog_limit);
+ audit_log_lost("backlog limit exceeded");
+ return NULL;
}
ab = audit_buffer_alloc(ctx, gfp_mask, type);
--
2.23.0
7:20 p.m.
New subject: [PATCH -next 1/1] audit: ignore userspace log when backlog limit exceeded and backlog_wait_time is 0
On Tue, Dec 28, 2021 at 1:50 AM luhuaxin <luhuaxin1(a)huawei.com> wrote:
When the backlog exceed the backlog_limit and backlog_wait_time is set
to 0, the process will only sleep for a very short time (jiffies). The
backlog may still exceed backlog_limit in extreme cases.
The more reasonable way to fix this problem is:
1. If backlog_wait_time is set to zero, ignore the log;
2. If backlog_wait_time is set to non-zero, let process sleep for
backlog_wait_time.
The above log limit logic is also the same as that in the existing
audit_log_start function.
Fixes: 8f110f530635 ("[PATCH] audit: ensure userspace is penalized the
same as the kernel when under pressure")
One quick comment on the "Fixes" tag above: you shouldn't add the
"[PATCH]" string to the commit's subject, you should use the commit
subject that you would see if you typed `git log --oneline`. It also
shouldn't be word-wrapped, it should be all on one line in your
patch/email.
Regardless of the above, I don't think this is a patch we want to
merge upstream. I can understand the desire to improve performance,
but this doesn't seem appropriate to me; adjusting the
backlog_wait_time to very low values just so you can drop audit
records is not an approach we want to advocate, or support, upstream.
Signed-off-by: luhuaxin <luhuaxin1(a)huawei.com>
---
kernel/audit.c | 14 +++++++++++---
1 file changed, 11 insertions(+), 3 deletions(-)
diff --git a/kernel/audit.c b/kernel/audit.c
index 249e11628..70450f70a 100644
--- a/kernel/audit.c
+++ b/kernel/audit.c
@@ -1545,7 +1545,8 @@ static void audit_receive(struct sk_buff *skb)
/* can't block with the ctrl lock, so penalize the sender now */
if (audit_backlog_limit &&
- (skb_queue_len(&audit_queue) > audit_backlog_limit)) {
+ (skb_queue_len(&audit_queue) > audit_backlog_limit) &&
+ audit_backlog_wait_time) {
DECLARE_WAITQUEUE(wait, current);
/* wake kauditd to try and flush the queue */
@@ -1842,9 +1843,8 @@ struct audit_buffer *audit_log_start(struct audit_context *ctx,
gfp_t gfp_mask,
* while holding the mutex, although we do penalize the sender
* later in audit_receive() when it is safe to block
*/
+ long stime = audit_backlog_wait_time;
if (!(auditd_test_task(current) || audit_ctl_owner_current())) {
- long stime = audit_backlog_wait_time;
-
while (audit_backlog_limit &&
(skb_queue_len(&audit_queue) > audit_backlog_limit)) {
/* wake kauditd to try and flush the queue */
@@ -1872,6 +1872,14 @@ struct audit_buffer *audit_log_start(struct audit_context *ctx,
gfp_t gfp_mask,
return NULL;
}
}
+ } else if (!stime && audit_backlog_limit &&
+ (skb_queue_len(&audit_queue) > audit_backlog_limit)) {
+ if (audit_rate_check() && printk_ratelimit())
+ pr_warn("audit_backlog=%d >
audit_backlog_limit=%d\n",
+ skb_queue_len(&audit_queue),
+ audit_backlog_limit);
+ audit_log_lost("backlog limit exceeded");
+ return NULL;
}
ab = audit_buffer_alloc(ctx, gfp_mask, type);
--
2.23.0
--
paul moore
www.paul-moore.com
1036
days inactive
1037
days old
linux-audit@lists.linux-audit.osci.io
2 comments
2 participants
participants (2)
-
luhuaxin -
Paul Moore