platform_system_core/llkd
Mark Salyzyn 22e05fb5a4 llkd: report stack signature matched
Adjusted debugging messaging to add clarity.  Report _which_ stack
signature matched that triggered the kernel panic.  Reduce the noise
associated with missing /stack to VERBOSE as that is for development
debugging only.

Test: observe during unit test we see something like following logs:

livelock: Found SyS_openat in stack for pid XXX
livelock: S 120.000s XXX->YYY port-bridge [kill]
livelock: Killing '/vendor/bin/port-bridge' (XXX) to check forward\
          scheduling progress in S state for\
          '/vendor/bin/port-bridge' (YYY)
. . .
livelock: Found SyS_openat in stack for pid XXXXX
livelock: S 120.000s XXXXX->XXXXX llkd_unit_test [kill]
livelock: Killing '/data/nativetest64/llkd_unit_test/llkd_unit_test\
          (XXXXX) to check forward scheduling progress in S state

Test: llkd_unit_test
Bug: 33808187
Change-Id: Ifac7dd9a656208563bb20e28739abb741358d964
2019-01-04 10:46:05 -08:00
..
include llkd: adbd is allowed to be a bad parent on userdebug 2018-12-14 08:44:33 -08:00
tests llkd: Add stack symbol checking 2018-08-30 13:53:19 -07:00
Android.bp llkd: Add stack symbol checking 2018-08-30 13:53:19 -07:00
libllkd.cpp llkd: report stack signature matched 2019-01-04 10:46:05 -08:00
llkd-debuggable.rc llkd: Add stack symbol checking 2018-08-30 13:53:19 -07:00
llkd.cpp llkd: clear PR_SET_DUMPABLE 2018-08-28 13:13:50 -07:00
llkd.rc llkd: Add stack symbol checking 2018-08-30 13:53:19 -07:00
OWNERS llkd: add live-lock daemon 2018-04-18 14:01:56 -07:00
README.md llkd: adbd is allowed to be a bad parent on userdebug 2018-12-14 08:44:33 -08:00

Android Live-LocK Daemon

Introduction

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.

Operations

There are two detection scenarios. Persistent D or Z state, and persistent stack signature.

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.

For usedebug releases only, persistent stack signature checking is enabled. If a thread in any state but Z, has a persistent listed ro.llk.stack kernel symbol always being reported, even if there is forward scheduling progress, for longer than ro.llk.timeout_ms, or ro.llk.stack.timeout_ms, then issue a kill to the process. If another scan shows the same process continues to exist, then have a confirmed live-lock condition and need to panic. There is no ABA detection since forward scheduling progress is allowed, thus the condition for the symbols are:

  • Check is looking for " symbol+0x" or " symbol.cfi+0x" in /proc/pid/stack.
  • The symbol should be rare and short lived enough that on a typical system the function is seen at most only once in a sample over the timeout period of ro.llk.stack.timeout_ms, samples occur every ro.llk.check_ms. This can be the only way to prevent a false trigger as there is no ABA protection.
  • Persistent continuously when the live lock condition exists.
  • Should be just below the function that is calling the lock that could contend, because if the lock is below or in the symbol function, the symbol will show in all affected processes, not just the one that caused the lockup.

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. For the stack symbol checking, there is an additional process blacklist so that we do not incide sepolicy violations on services that block ptrace operations.

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

Android Properties llkd respond to (prop_ms parms are in milliseconds):

ro.config.low_ram

device is configured with limited memory.

ro.debuggable

device is configured for userdebug or eng build.

ro.llk.sysrq_t

default not ro.config.low_ram, or ro.debuggable if property is "eng". if true do sysrq t (dump all threads).

ro.llk.enable

default false, allow live-lock daemon to be enabled.

llk.enable

default ro.llk.enable, and evaluated for eng.

ro.khungtask.enable

default false, allow [khungtask] daemon to be enabled.

khungtask.enable

default ro.khungtask.enable and evaluated for eng.

ro.llk.mlockall

default false, enable call to mlockall().

ro.khungtask.timeout

default value 12 minutes, [khungtask] maximum timelimit.

ro.llk.timeout_ms

default 10 minutes, D or Z maximum timelimit, double this value and it sets the alarm watchdog for llkd.

ro.llk.D.timeout_ms

default ro.llk.timeout_ms, D maximum timelimit.

ro.llk.Z.timeout_ms

default ro.llk.timeout_ms, Z maximum timelimit.

ro.llk.stack.timeout_ms

default ro.llk.timeout_ms, checking for persistent stack symbols maximum timelimit. Only active on userdebug or eng builds.

ro.llk.check_ms

default 2 minutes samples of threads for D or Z.

ro.llk.stack

default cma_alloc,__get_user_pages,bit_wait_io comma separated list of kernel symbols. The string "false" is the equivalent to an empty list. Look for kernel stack symbols that if ever persistently present can indicate a subsystem is locked up. Beware, check does not on purpose do forward scheduling ABA except by polling every ro.llk_check_ms over the period ro.llk.stack.timeout_ms, so stack symbol should be exceptionally rare and fleeting. One must be convinced that it is virtually impossible for symbol to show up persistently in all samples of the stack. Again, looks for a match for either " symbol+0x" or " symbol.cfi+0x" in stack expansion. Only available on userdebug or eng builds, limited privileges due to security concerns on user builds prevents this checking.

ro.llk.blacklist.process

default 0,1,2 (kernel, init and [kthreadd]) plus process names init,[kthreadd],[khungtaskd],lmkd,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.

ro.llk.blacklist.parent

default 0,2,adbd (kernel, [kthreadd] and adbd). 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.

ro.llk.blacklist.uid

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.

ro.llk.blacklist.process.stack

default process names init,lmkd.llkd,llkd,keystore,ueventd,apexd,logd. The string "false" is the equivalent to an empty list. This subset of processes are not monitored for live lock stack signatures. Also prevents the sepolicy violation associated with processes that block ptrace, as these can not be checked anyways. Only active on userdebug and eng builds.

Architectural Concerns

  • built-in [khungtask] daemon is too generic and trips on driver code that sits around in D state too much. To switch to S instead makes the task(s) killable, so the drivers should be able to resurrect them if needed.
  • Properties are limited to 92 characters.
  • Create kernel module and associated gTest to actually test panic.
  • Create gTest to test out blacklist (ro.llk.blacklist.properties generally not be inputs). Could require more test-only interfaces to libllkd.
  • Speed up gTest using something else than ro.llk.properties, which should not be inputs as they should be baked into the product.