7be3fd486c | ||
---|---|---|
.. | ||
include | ||
tests | ||
Android.bp | ||
OWNERS | ||
README.md | ||
libllkd.cpp | ||
llkd-debuggable.rc | ||
llkd.cpp | ||
llkd.rc |
README.md
Android Live-LocK Daemon (llkd)
Android 10 includes the Android Live-LocK Daemon
(llkd
), which is designed to catch and mitigate kernel deadlocks. The llkd
component provides a default standalone implementation, but you can
alternatively integrate the llkd
code into another service, either as part of
the main loop or as a separate thread.
Detection scenarios
The llkd
has two detection scenarios: Persistent D or Z state, and persistent
stack signature.
Persistent D or Z state
If a thread is in D (uninterruptible sleep) or Z (zombie) state with no forward
progress for longer than ro.llk.timeout_ms or ro.llk.[D|Z].timeout_ms
, the
llkd
kills the process (or parent process). If a subsequent scan shows the
same process continues to exist, the llkd
confirms a live-lock condition and
panics the kernel in a manner that provides the most detailed bug report for the
condition.
The llkd
includes a self watchdog that alarms if llkd
locks up; watchdog is
double the expected time to flow through the mainloop and sampling is every
ro.llk_sample_ms
.
Persistent stack signature
For userdebug releases, the llkd
can detect kernel live-locks using persistent
stack signature checking. If a thread in any state except Z has a persistent
listed ro.llk.stack
kernel symbol that is reported for longer than
ro.llk.timeout_ms
or ro.llk.stack.timeout_ms
, the llkd
kills the process
(even if there is forward scheduling progress). If a subsequent scan shows the
same process continues to exist, the llkd
confirms a live-lock condition and
panics the kernel in a manner that provides the most detailed bug report for the
condition.
Note: Because forward scheduling progress is allowed, the llkd
does not
perform ABA detection{:.external}.
The lldk
check persists continuously when the live lock condition exists and
looks for the composed strings " symbol+0x"
or " symbol.cfi+0x"
in the
/proc/pid/stack
file on Linux. The list of symbols is in ro.llk.stack
and
defaults to the comma-separated list of
"cma_alloc,__get_user_pages,bit_wait_io,wait_on_page_bit_killable
".
Symbols should be rare and short-lived enough that on a typical system the
function is seen only once in a sample over the timeout period of
ro.llk.stack.timeout_ms
(samples occur every ro.llk.check_ms
). Due to lack
of ABA protection, this is the only way to prevent a false trigger. The symbol
function must appear below the function calling the lock that could contend. If
the lock is below or in the symbol function, the symbol appears in all affected
processes, not just the one that caused the lockup.
Coverage
The default implementation of llkd
does not monitor init
, [kthreadd]
, or
[kthreadd]
spawns. For the llkd
to cover [kthreadd]
-spawned threads:
- Drivers must not remain in a persistent D state,
OR
- Drivers must have mechanisms to recover the thread should it be killed
externally. For example, use
wait_event_interruptible()
instead ofwait_event()
.
If one of the above conditions is met, the llkd
ignorelist can be adjusted to
cover kernel components. Stack symbol checking involves an additional process
ignore list to prevent sepolicy violations on services that block ptrace
operations.
Android properties
The llkd
responds to several Android properties (listed below).
- Properties named
prop_ms
are in milliseconds. - Properties that use comma (,) separator for lists use a leading separator to preserve the default entry, then add or subtract entries with optional plus (+) and minus (-) prefixes respectively. For these lists, the string "false" is synonymous with an empty list, and blank or missing entries resort to the specified default value.
ro.config.low_ram
Device is configured with limited memory.
ro.debuggable
Device is configured for userdebug or eng build.
ro.llk.sysrq_t
If property is "eng", the default is not ro.config.low_ram
or ro.debuggable
.
If true, dump all threads (sysrq t
).
ro.llk.enable
Allow live-lock daemon to be enabled. Default is false.
llk.enable
Evaluated for eng builds. Default is ro.llk.enable
.
ro.khungtask.enable
Allow [khungtask]
daemon to be enabled. Default is false.
khungtask.enable
Evaluated for eng builds. Default is ro.khungtask.enable
.
ro.llk.mlockall
Enable call to mlockall()
. Default is false.
ro.khungtask.timeout
[khungtask]
maximum time limit. Default is 12 minutes.
ro.llk.timeout_ms
D or Z maximum time limit. Default is 10 minutes. Double this value to set the
alarm watchdog for llkd
.
ro.llk.D.timeout_ms
D maximum time limit. Default is ro.llk.timeout_ms
.
ro.llk.Z.timeout_ms
Z maximum time limit. Default is ro.llk.timeout_ms
.
ro.llk.stack.timeout_ms
Checks for persistent stack symbols maximum time limit. Default is
ro.llk.timeout_ms
. Active only on userdebug or eng builds.
ro.llk.check_ms
Samples of threads for D or Z. Default is two minutes.
ro.llk.stack
Checks for kernel stack symbols that if persistently present can indicate a
subsystem is locked up. Default is
cma_alloc,__get_user_pages,bit_wait_io,wait_on_page_bit_killable
comma-separated list of kernel symbols. The check doesn't do forward scheduling
ABA except by polling every ro.llk_check_ms
over the period
ro.llk.stack.timeout_ms
, so stack symbols should be exceptionally rare and
fleeting (it is highly unlikely for a symbol to show up persistently in all
samples of the stack). Checks for a match for " symbol+0x"
or
" symbol.cfi+0x"
in stack expansion. Available only on userdebug or eng
builds; security concerns on user builds result in limited privileges that
prevent this check.
ro.llk.ignorelist.process
The llkd
does not watch the specified processes. Default is 0,1,2
(kernel
,
init
, and [kthreadd]
) plus process names
init,[kthreadd],[khungtaskd],lmkd,llkd,watchdogd, [watchdogd],[watchdogd/0],...,[watchdogd/get_nprocs-1]
.
A process can be a comm
, cmdline
, or pid
reference. An automated default
can be larger than the current maximum property size of 92.
Note: false
is an extremely unlikely process to want to ignore.
ro.llk.ignorelist.parent
The llkd
does not watch processes that have the specified parent(s). Default
is 0,2,adbd&[setsid]
(kernel
, [kthreadd]
, and adbd
only for zombie
setsid
). An ampersand (&) separator specifies that the parent is ignored only
in combination with the target child process. Ampersand was selected because it
is never part of a process name; however, a setprop
in the shell requires the
ampersand to be escaped or quoted, although the init rc
file where this is
normally specified does not have this issue. A parent or target process can be a
comm
, cmdline
, or pid
reference.
ro.llk.ignorelist.uid
The llkd
does not watch processes that match the specified uid(s).
Comma-separated list of uid numbers or names. Default is empty or false.
ro.llk.ignorelist.process.stack
The llkd
does not monitor the specified subset of processes for live lock stack
signatures. Default is process names
init,lmkd.llkd,llkd,keystore,keystore2,ueventd,apexd,logd
. Prevents the sepolicy
violation associated with processes that block ptrace
(as these can't be
checked). Active only on userdebug and eng builds. For details on build
types, refer to Building Android.
Architectural concerns
- Properties are limited to 92 characters. However, this is not limited for
defaults defined in the
include/llkd.h
file in the sources. - The built-in
[khungtask]
daemon is too generic and trips on driver code that sits around in D state too much. Switching drivers to sleep, or S state, would make task(s) killable, and need to be resurrectable by drivers on an as-need basis.
Library interface (optional)
You can optionally incorporate the llkd
into another privileged daemon using
the following C interface from the libllkd
component:
#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 automatically spawns, otherwise the caller
must call llkCheckMilliseconds
in its main loop. The function returns the
period of time before the next expected call to this handler.