Thanks for details information! For the switch question, I get your point. logd.auditd is the switch of whether logd can record selinux audit log. But I'm looking for the switch of ratelimit. I mean, removing ratelimit is really helpful for selinux debugging, especially in bringup stage. But meanwhile, removing it just opens the gate for potential DOS. So should we have a simple command that can disable ratelimit during bringup debugging and enable it for release. Just like we can switch to permissive mode by setenforce 0. -----Original Message----- From: Stephen Smalley [mailto:sds@xxxxxxxxxxxxx] Sent: 2015年5月7日 21:00 To: Zhi Xin; selinux@xxxxxxxxxxxxx Cc: seandroid-list@xxxxxxxxxxxxx Subject: Re: Give out all the avc logs in ome time On 05/06/2015 11:18 PM, Zhi Xin wrote: > Thanks a lot~ > > I have another question: in last email, you mentioned two ways to eliminate the ratelimit. > > (1) disable the audit ratelimit in logd by removing the AUDIT_STATUS_RATE_LIMIT flag from status.mask. > (2) use the kernel > https://bitbucket.org/seandroid/kernel-msm/commits/0388e1630648c481e42 > 929135babb1dbba272e27 > > I can understand how the kernel part work. But why do you still need to set a user space process like logd ? I mean the avc log is given by kernel. Why a user space process can affect its output ? Is is because the kernel log brought to user by this logd and if I want to disable audit ratelimit, I have to apply both of them ? > > I also wonder do we have a more simple mechanism like a open/off switch during runtime ? Until logd (or auditd in Linux distributions like Fedora) starts running and registers itself to receive audit messages, the kernel logs the audit records directly via printk, and those calls are wrapped with printk_ratelimit checks to avoid a potential DOS. Once logd (or auditd) is registered to receive audit messages, the kernel sends audit records via the netlink audit socket to logd, and logd then writes them to /dev/kmsg (so that they are still available via dmesg; this differs from auditd behavior in Linux distributions) and also writes them to the logcat events and main log buffers (and this is obviously Android-specific). logd also grabs any previously logged audit records from dmesg and writes them to the logcat buffers so that even the audit records generated before logd started should be included in the logcat buffers, not just the ones generated afterward. The printk ratelimit is only relevant for the audit messages generated before logd is started, and for any warning/error messages logged via printk (e.g. printk limit exceeded, backlog limit exceeded, rate limit exceeded, out of memory...). It is not configurable and is always applied by the audit system when logging messages via printk. That's why we need a kernel patch to disable it. The audit ratelimit on the other hand is relevant for all audit messages, whether logged via printk or sent over the netlink audit socket, but is only set if logd explicitly sets it to something. And since logd only sets it if it is also registering itself to receive audit messages, then effectively the audit ratelimit only applies to audit messages sent over the netlink audit socket to logd. Not to messages logged before logd starts. I hadn't noticed this before, but in just looking at the logd code, I see that there are also a couple of system properties that you can set to alter its behavior wrt to audit: name type default description logd.auditd bool true Enable selinux audit daemon logd.auditd.dmesg bool true selinux audit messages duplicated and sent on to dmesg log So you could possibly disable logd's audit functionality altogether by setting logd.auditd to false and rebooting. At which point you would only be concerned with the printk ratelimit. _______________________________________________ Selinux mailing list Selinux@xxxxxxxxxxxxx To unsubscribe, send email to Selinux-leave@xxxxxxxxxxxxx. To get help, send an email containing "help" to Selinux-request@xxxxxxxxxxxxx.
