Setting memory.soft_limit_in_bytes on high-end devices with large memory
reserves affects performance of memory-hungry applications that have
large workingsets and keep thrashing because of the memory limits imposed.
Limit the usage of memory.soft_limit_in_bytes to low-memory devices only.
Add debug messages for future troubleshooting to capture cases when
vmpressure events are being ignored.
Bug: 78916015
Test: collect vmstat while running a heavy app
Change-Id: Ib4434b96d2be802ef89960b573486eae8d12f198
Merged-In: Ib4434b96d2be802ef89960b573486eae8d12f198
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Set F() capability set and 'drop' lmkd from AID_ROOT to AID_LMKD uid
and from AID_ROOT to AID_LMKD and AID_SYSTEM gid.
/dev/memcg/memory.pressure defaults to root.root mode 0000, set it up
as root.system mode 0040 to allow lmkd read access.
Instrument failure to set SCHED_FIFO.
Annotate access points that require elevated capabilities.
Test: check /proc/`pidof lmkd`/status for capability set
Test: lmkd_unit_test
Bug: 77650566
Change-Id: I986081a0434cf6e842b63a55726380205b30a3ea
Add ability to switch to the algorithm used by lowmemorykiller driver
for determining when to kill. It uses minfree levels to decide at which
levels of free memory and file cache to kill a process. oom_adj_score
is also determined by comparing current memory resources against minfree
levels.
ro.lmk.use_minfree_levels property is introduces for switching into this
mode. By default it is disabled.
Bug: 77299493
Bug: 75322373
Merged-In: I6b51972951026854a079fcda33d6786b7ed035e4
Change-Id: I6b51972951026854a079fcda33d6786b7ed035e4
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Current mechanism of getting system memory state by using sysinfo()
is not enough because it does not provide information about file cache
size which is needed to correctly assess memory state. Switch to using
data from /proc/meminfo that includes information provided by sysinfo()
and file cache size in addition to that.
Bug: 77299493
Bug: 75322373
Merged-In: I16106fc4f9254f17f776803e60502b7b6474e1b7
Change-Id: I16106fc4f9254f17f776803e60502b7b6474e1b7
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
/proc/zoneinfo and /proc/meminfo contain information necessary for lmkd
to assess system memory state. Add routines to parse these files.
Bug: 77299493
Bug: 75322373
Merged-In: Ie7d80bbb81fd0d2fc0629f6f678e6afc97fed1f6
Change-Id: Ie7d80bbb81fd0d2fc0629f6f678e6afc97fed1f6
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
To check system memory state lmkd is using same files every time vmpressure
event is received. Instead of opening and closing these files every time
we store the file descriptor and use pread() to reread the file from
the beginning.
Bug: 77299493
Bug: 75322373
Merged-In: I8e27f8b9526e82e3cc313a02fce215c2e4dd3d29
Change-Id: I8e27f8b9526e82e3cc313a02fce215c2e4dd3d29
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
lmkd should use ro.lmk.kill_heaviest_task property to select between
algorithms for victim selection. Set ro.lmk.kill_heaviest_task default
value to false in order to keep it compatible with previous versions
of lmkd (killing the heaviest task is a new mechanism).
Change-Id: I78d2dc79d9c54e636c26665605518d9c87b535b3
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Rename is_go_device variable to low_ram_device to better reflect
its meaning and relation to ro.config.low_ram variable.
Change-Id: I6e2eaebe79cf2e6edf861f7c602e52a5b573ad0a
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Move process initialization, mlockall() and sched_setscheduler() to
only occur if initialization succeeds and is enabled. This
conserves mlock'd memory and a FIFO processing slot if (deprecated)
kernel lowmemorykiller is active.
Cleanup: Android coding standard compliance.
Test: lmkd_unit_test
Bug: 33808187
Bug: 72838192
Change-Id: I51af0235c8eca4ef958ef2d83a8071ff58b49d78
lmkd_unit_test gets an selinux violation when run unprivileged:
lmkd_unit_test: type=1400 audit(0.0:???): avc: denied { search } for
name="lowmemorykiller" dev="sysfs" ino=??? scontext=u:r:shell:s0
tcontext=u:object_r:sysfs_lowmemorykiller:s0 tclass=dir
permissive=0
resulting in log noise. Check for INKERNEL_MINFREE_PATH existence
after known to be escalated as root user to suppress noise.
Test: lmkd_unit_test as unprivileged user, there should be no audits.
Change-Id: Ia5c7824e7070c98ee6eea20ca53097e4e56d61b6
Current kills message prints min_score_adj marked as "oom_adj" which
is confusing. This change marks it as "min_oom_adj" to be more clear.
Change-Id: If52f0fb25fe3208c928c5d6adce77c04b9f15a0e
system_server needs to register with lmkd, however it has no memory
cgroup under /dev/memcg/apps. This change detects if the process being
registered is system_server and suppresses the error message when
/dev/memcg/apps/uid_%d/pid_%d/memory.soft_limit_in_bytes file can't
be accessed.
Bug: 73483785
Test: verified logcat output
Change-Id: I03df7831f41f512ac8d3ebc46330546d08a3cbc6
Occasionally processes are killed or crash while lmkd is trying to access
their oom_score_adj file. This is not necessarily an erroneous condition
and therefore should not be reported as an error. Demoting error to a
warning with a detailed message.
Bug: 72039129
Test: verified logcat output
Change-Id: I97444ba5198c02cb7f7ba03c3af12e4aad9d233e
This change allows lmkd-test to work correctly when it's started
using absolute as well as relative path.
Change-Id: Ida58d13d66a224119a363f59cc59289f0167d3c9
(cherry pick from commit 0f10051805)
A number of tools and tests require communication with lmkd.
In order to avoid code duplication liblmkd_utils implements
functions commonly used when interacting with lmkd process.
Isolate communication protocol details into lmkd.h
Bug: 63631020
Change-Id: Id840983d55b7db60013d52dee0c3187943811822
Merged-In: Id840983d55b7db60013d52dee0c3187943811822
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit 3cfb2c8b10)
lmkd testing requires communication with lmkd daemon to register new
native processes. New implementation allows more than one communication
channel to lmkd. Current max number of communication channels is set to
two - one for ActivityManager and another one for a test process.
Bug: 63631020
Change-Id: I736115938a3c5ad9253bce29a17cd5349af190eb
Merged-In: I736115938a3c5ad9253bce29a17cd5349af190eb
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit 979591b627)
Currently selection criteria for in-kernel vs userspace lmk is kernel
driver presence and device not being a Go device. This change removes
Go device check leaving kernel driver presence to be the only selection
criteria.
Bug: 71502948
Change-Id: I394a7920433a8d090e207ea86296356413a63fe7
Merged-In: I394a7920433a8d090e207ea86296356413a63fe7
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit caa2dc56fd)
New ro.lmk.kill_timeout_ms property defines timeout in ms after a
successful kill cycle for more kills to be considered. This is
necessary because memory pressure after a kill does not go down
instantly and system needs time to reflect new memory state. This
timeout prevents extra kills in the period immediately after a
kill cycle. By default it is set to 0 which disables this feature.
Bug: 63631020
Test: alloc-stress
Change-Id: Ia847118c8c4a659a7fc38cd5cd0042acb514ae28
Merged-In: Ia847118c8c4a659a7fc38cd5cd0042acb514ae28
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit 65f54a2665)
Record free memory at low vmpressure levels and whenever pressure
increases beyond low free up enough memory to downgrade memory pressure
to low. This is done by freeing enough memory to get to the max free
memory levels seen during low vmpressure.
The kill logic for Go devices is not changed as these devices are designed
to operate under high memory pressure.
Bug: 63631020
Test: alloc-stress
Change-Id: Ic8396eee08013b1c709072a13525601d5c8bf1f1
Merged-In: Ic8396eee08013b1c709072a13525601d5c8bf1f1
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit e82e15c242)
lmkd checks for vmpressure events using epoll_wait() with eventfds of
all registered events. It's possible that multiple events of different
priorities happen before epoll_wait() returns. For these cases we
use conservative approach by assuming that the system is under the
highest registered vmpressure levels. This speeds up lmkd response time
to high memory pressure by not responding to possibly stale low pressure
levels when vmpressure rises quickly.
Bug: 63631020
Test: alloc-stress
Change-Id: I79a85c3342e7e1b3a3be82945266b2cc60b437cf
Merged-In: I79a85c3342e7e1b3a3be82945266b2cc60b437cf
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit 1bd2fc4fb6)
After events are specified by writing into cgroup.event_control file
the file should be closed.
Change-Id: Id015e6a7bac2b74bbc8d8793c85f529ee00bdf55
Merged-In: Id015e6a7bac2b74bbc8d8793c85f529ee00bdf55
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit c71355991d)
For tracing lmkd kills inside kernel it is useful to have traces
indicating when and which process lmkd is killing. By default the
tracing is disabled.
Bug: 63631020
Test: alloc-stress
Change-Id: I3ceb2bde0c292eec55855cb4535927f3b4c5d08b
Merged-In: I3ceb2bde0c292eec55855cb4535927f3b4c5d08b
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit 662492ab1d)
Killing the most memory-demanding process from the set of eligible
processes yields better results on high-performance devices than
killing the first one we could find. This is in line with how in-kernel
lowmemorykiller driver chooses its victims.
Bug: 63631020
Test: alloc-stress
Change-Id: Ie1ef7f33f3e79698a9b4120c14490386d6129f9b
Merged-In: Ie1ef7f33f3e79698a9b4120c14490386d6129f9b
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit ad2fd9150b)
vmpressure upgrade/downgrade logic based on swap utilization works well
for low memory devices because of a small swap size, however for high
performance devices this measure is not a good indication of the memory
pressure because of large swap resources. This change sets the default
levels to disable upgrade/downgrade logic by default and each device
can set these properties appropriately.
Bug: 63631020
Test: alloc-stress
Change-Id: Ifd4fbd4d6bb3e82f0f87b029df94934f1e7b1c9c
Merged-In: Ifd4fbd4d6bb3e82f0f87b029df94934f1e7b1c9c
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(cherry pick from commit 96bf3a600c)
Ability to monitor all available vmpressure event levels is needed
to accommodate systems with different memory resources. Low memory
systems can rely on medium and critical level events because working
under memory pressure is usual mode of operation. High performance
systems with more memory need to react earlier using also low
vmpressure level events to free memory early and prevent low memory
condition affecting its performance.
Bug: 63631020
Test: alloc-stress
Change-Id: I0cef1bd4c97d32c005045ae47f0ce3464ed98899
Merged-In: I0cef1bd4c97d32c005045ae47f0ce3464ed98899
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Remove a use-after free reference of procp->pid, using the already
captured pid variable.
Test: lmkd_unit_tests
Bug: 33808187
Change-Id: I3f5f8dd9acab2e28c81465d6195b73ae47e0a3c4
This way, we don't fault in the entirety of our DSOs immediately;
instead, used pages are "sticky" in memory. Works only on kernel 4.4
and up: downlevel, we ignore the mlockall failure.
Once we get statically-linked lmkd in better shape, we'll just switch
to that.
Change-Id: I07a75ee3bc1264a1db41635c2acf611fede99b91
We pin lmkd in memory so that we don't take page faults (and thus
requisition memory) while we're in the process of responding to a
low-memory condition. mlockall(2) is the right primitive for this
pinning. Previously, we used the MCL_FUTURE flag to mlockall: used
this way, mlockall doesn't actually pin all pages in memory, since
MCL_FUTURE affects only the default flags for future mappings and
doesn't affect mapping already in existence at the time of the
mlockall call --- like the lmkd executable itself.
This patch adds the MCL_CURRENT flag, which also pins all pages
already mapped.
Test: code inspection
Change-Id: I4563959367a2f0a9cadc3ea41731b7f311326685
Get min adj score when needed, instead to propagate from caller funct.
Bug:65642829
Test: tested on gobo_512
Change-Id: I87f15c5c6206d471002828a24d0462b0abb0d196
If the system has enough memory the pressure events are ignored or
downgraded from critical to medium.
Bug: 65642829
Test: tested on gobo
Change-Id: I44e66d2e35508aceee5c1719313db217b80d582e
For Go devices fallback to memcg/memory pressure events for LMK.
Go devices will use memcg pressure events while inkernel module
is disabled.
Bug: 64852905
Test: tested on gobo
Change-Id: I267ab00be85e324331f6c91551ba013184de817e
When close to oom the system tend to become very thrashy
and keeps paging. This change looks at the current working sent and
checks it against swap.
Test: tested on gobo
Bug: 64721547
Change-Id: I93d42def93cbc03a01a54988fd5286ec9f124e36
- Don't show misleading messages, add the correct description for
the reason of process killing
- Don't kill Launcher3 upgrade the score to perceptible (200)
- Limit the verbose logging
Test: tested on gobo
Bug: 64316273
Change-Id: Ic7c13853d939afe4f20cdaa5b488b6e1810f5ddf
By doing so different platforms can customize them accordingly.
For instance a low mem, 512MB device can do
ro.lmk.medium=400, while a device with more memory can keep
the default(800).
Test: tested on gobo.
Bug: 64316084
Change-Id: Ifc4f3853bc06f30488adb25883ccd9aaf683ba9b
Point to log/log.h where necessary, define LOG_TAG where necessary.
Accept that private/android_logger.h is suitable replacement for
log/logger.h and android/log.h.
Correct liblog/README
Effectively a cleanup and controlled select revert of
'system/core: drop or replace log/logger.h' and
'system/core: Replace log/log.h with android/log.h'.
Test: compile
Bug: 30465923
Change-Id: Ic2ad157bad6f5efe2c6af293a73bb753300b17a2
Should use android/log.h instead of log/log.h as a good example
to all others. Adjust header order to comply with Android Coding
standards.
Test: Compile
Bug: 26552300
Bug: 31289077
Change-Id: I33a8fb4e754d2dc4754d335660c450e0a67190fc
android_get_control_socket has a warning from the implicit cast from
long to int. The warning was being hidden because cutils/sockets.h was
included with -isystem. Move android_get_control_socket to sockets.cpp,
since we don't want header only dependencies anyways, and fix the
warning with a range check and a static_cast.
Bug: 31492149
Test: m -j <module that uses sockets.h and -Wall>
Change-Id: I1f394ab26d4ec8a7dd0e7907c10416d7f8647624
This will make sure we get invoked in time,
and combined with a kernel change to increase
the window size (and thus rate-limit the
notifications) based on the amount of memory
on the device, this should provide a nice
balance between receiving too many notifications
vs getting notified too late.
Change-Id: I2f5e92390b94e582aaa06d88763a232e39673496
