Android Live-LocK Daemon (llkd) is used to catch kernel deadlocks and mitigate.
Code is structured to allow integration into another service as either as part of the main loop, or spun off as a thread should that be necessary. A default standalone implementation is provided by llkd component.
The 'C' interface from libllkd component is thus:
#include "llkd.h" bool llkInit(const char* threadname) /* return true if enabled */ unsigned llkCheckMillseconds(void) /* ms to sleep for next check */
If a threadname is provided, a thread will be automatically spawned, otherwise caller must call llkCheckMilliseconds in its main loop. Function will return the period of time before the next expected call to this handler.
If a thread is in D or Z state with no forward progress for longer than ro.llk.timeout_ms, or ro.llk.[D|Z].timeout_ms, kill the process or parent process respectively. If another scan shows the same process continues to exist, then have a confirmed live-lock condition and need to panic. Panic the kernel in a manner to provide the greatest bugreporting details as to the condition. Add a alarm self watchdog should llkd ever get locked up that is double the expected time to flow through the mainloop. Sampling is every ro.llk_sample_ms.
Default will not monitor init, or [kthreadd] and all that [kthreadd] spawns. This reduces the effectiveness of llkd by limiting its coverage. If there is value in covering [kthreadd] spawned threads, the requirement will be that the drivers not remain in a persistent 'D' state, or that they have mechanisms to recover the thread should it be killed externally (this is good driver coding hygiene, a common request to add such to publicly reviewed kernel.org maintained drivers). For instance use wait_event_interruptible() instead of wait_event(). The blacklists can be adjusted accordingly if these conditions are met to cover kernel components.
An accompanying gTest set have been added, and will setup a persistent D or Z process, with and without forward progress, but not in a live-lock state because that would require a buggy kernel, or a module or kernel modification to stimulate. The test will check that llkd will mitigate first by killing the appropriate process. D state is setup by vfork() waiting for exec() in child process. Z state is setup by fork() and an un-waited for child process. Should be noted that both of these conditions should never happen on Android on purpose, and llkd effectively sweeps up processes that create these conditions. If the test can, it will reconfigure llkd to expedite the test duration by adjusting the ro.llk.* Android properties. Tests run the D state with some scheduling progress to ensure that ABA checking prevents false triggers. If 100% reliable ABA on platform, then ro.llk.killtest can be set to false; however this will result in some of the unit tests to panic kernel instead of deal with more graceful kill operation.
Android Properties llkd respond to (prop_ms parms are in milliseconds):
default false, if true do not sysrq t (dump all threads).
default false, allow live-lock daemon to be enabled.
default ro.llk.enable, and evaluated for eng.
default false, allow [khungtask] daemon to be enabled.
default ro.khungtask.enable and evaluated for eng.
default false, enable call to mlockall().
default value 12 minutes, [khungtask] maximum timelimit.
default 10 minutes, D or Z maximum timelimit, double this value and it sets the alarm watchdog for llkd.
default ro.llk.timeout_ms, D maximum timelimit.
default ro.llk.timeout_ms, Z maximum timelimit.
default 2 minutes samples of threads for D or Z.
default 0,1,2 (kernel, init and [kthreadd]) plus process names init,[kthreadd],[khungtaskd],lmkd,lmkd.llkd,llkd,watchdogd, [watchdogd],[watchdogd/0],...,[watchdogd/get_nprocs-1]. The string false is the equivalent to an empty list. Do not watch these processes. A process can be comm, cmdline or pid reference. NB: automated default here can be larger than the current maximum property size of 92. NB: false is a very very very unlikely process to want to blacklist.
default 0,2 (kernel and [kthreadd]). The string false is the equivalent to an empty list. Do not watch processes that have this parent. A parent process can be comm, cmdline or pid reference.
default empty or false, comma separated list of uid numbers or names. The string false is the equivalent to an empty list. Do not watch processes that match this uid.