Add '-a' commandline option to show the old complete table that shows
dmabuf x process grid.
Bug: 135668399
Test: dmabuf_dump
Test: dmabuf_dump -h
Test: dmabuf_dump -a
Test: dmabuf_dump -a 123
Test: dmabuf_dump 123 456
Test: dmabuf_dump <pid>
Change-Id: If228c97baa375c515d94baf3b5be674cfcc09919
Signed-off-by: Sandeep Patil <sspatil@google.com>
- Get rid of the PSS number that didn't make sense.
- Fix double counting in per-process size when buffers had file
descriptors and mmap'ed
- Make the output much more readable
- Print totals where they make sense.
- Dont print per-process fd and map refs separately, total ref count per
process is sufficient.
Bug: 135668399
Test: dmabuf_dump (on AOSP blueline)
Change-Id: I10dc0a332d49110dd68f0eaa94a326679361deea
Signed-off-by: Sandeep Patil <sspatil@google.com>
The scudo memory allocator will allocate a large virtual memory address
that causes procrank to allocate a huge amount of data. Modify so it
caches the data a bit at a time.
Remove some allocations that are not necessary since the data is only used
in the first loop.
Fix a few pread64 checks against the actual bytes read, rather than the
read is < 0.
Fix sort by rss in procrank.
Bug: 135694447
Test: Ran unit tests (only one unit test actually runs the modified code).
Test: Ran the old version of procrank using a static libmeminfo and verified
Test: that the same data is found as the new version.
Change-Id: I8adc169b5607ec994ff13a9e161a479350d84c4d
procrank's usage in its output shows 20x increase from
last year. This is because it is keeping the process
maps around until termination. Fix that by getting rid
of ProcMemInfo objects when done parsing /proc/<pid>/maps,pagemap.
Note that the total allocations do not change and have not necessarily
regressed from Pie.
Bug: 130672819
Test: adb shell procrank | grep 'procrank\|cmdline'
Change-Id: Ib7bf960ed1d053347fcfc0c8aee9019607a1eb01
Signed-off-by: Sandeep Patil <sspatil@google.com>
procrank currently fails if a process gets killed while it is reading
the stats. This behavior is a regression from the previous version of
procrank and is often undesired.
Change procrank to silently ignore the process if it detects that it had
been killed while reading the stats. If the process is still around,
then print a warning about it and continue to read stats for other
processes in the system.
Bug: 130177765
Test: Tested by deliberately killing specific process in ProcessRecord()
constructor
Change-Id: I701808c3226bb9b3a350ccf8e67fb29b59b0d4e0
Signed-off-by: Sandeep Patil <sspatil@google.com>
librank was calling std::vector::find_if in nested loops when iterating
through processes and maps.
Changed the implementation to use std::map::insert, this resulted in
a significant improvement to execution speed.
Test: time librank on crosshatch
Before: 1m22.57s real 1m15.87s user 0m05.79s system
After: 0m03.85s real 0m00.88s user 0m02.92s system
Bug: 124523194
Change-Id: I2b519d0bfd7f7929c6c4c6c2374794cb1a744585
Signed-off-by: Erick Reyes <erickreyes@google.com>
The vendor partition does not give dmabuf_dump enough permissions to
retrieve shared buffer information for all processes. Move it to product
partition where it can be invoked from platform's dumpstate.
Bug: 63860998
Change-Id: I73957f59943d1e49fcb59abb723aad4eeeaa4172
Signed-off-by: Erick Reyes <erickreyes@google.com>
Adjust all callbacks to include the inode number as well.
Fixes: 123532375
Test: libprocinfo_test
Test: libdmabufinfo_test
Test: libmeminfo_test
Change-Id: I71fd75d90d5ee44585011c5ae654a1dd7e185458
Signed-off-by: Sandeep Patil <sspatil@google.com>
Test: build with master-art manifest
Exempt-From-Owner-Approval: Required for unbreaking ART development.
Change-Id: I6afc0e5444dfa21532a4c802f8c463091cab2b11
Fixes:
system/core/libmeminfo/libdmabufinfo/tools/dmabuf_dump.cpp:109:39: error: format specifies type 'unsigned long' but the argument has type 'uint64_t' (aka 'unsigned long long') [-Werror,-Wformat]
Test: lunch aosp_arm-eng && m dmabuf_dump
Change-Id: Ie605d9b94f5eff888f6a64801512216253a6babb
dmabuf_dump had a hard dependency on debugfs to gather the initial list
of buffers for all proceses, make it optional.
Change-Id: I1d271297c0ad6124b321a1ee8aa01d3e88ca9fed
Signed-off-by: Erick Reyes <erickreyes@google.com>
Removed dependency on libc++fs and changed liblog to shared linkage to
cope with build errors derived from the change.
Removed dmabuf_dump unnecessary dependencies and made it soc_specific.
Removed reference to std::filesystem::is_symlink to workaround the
following error when linking with the library:
ld.lld: error: undefined symbol:
std::__1::__fs::filesystem::__symlink_status(std::__1::__fs::filesystem::path
const&, std::__1::error_code*)
>>> referenced by filesystem:1743
(external/libcxx/include/filesystem:1743)
>>>
dmabufinfo.o:(android::dmabufinfo::AppendDmaBufInfo(int,
std::__1::vector<android::dmabufinfo::DmaBuffer,
std::__1::allocator<android::dmabufinfo::DmaBuffer> >*)) in archive
Bug: 63860998
Change-Id: Ieafdc87c64f153625df9e21fc8299292b2447aef
Signed-off-by: Erick Reyes <erickreyes@google.com>
Adding tool to dump the dmabuf information using the dmabufinfo library.
Test: run "dmabuf_dump" and "dmabuf_dump <pid>" on crosshatch
Bug: 63860998
Change-Id: I16fe0a701deb5086ca0e1a8ceecb70dbc17934e6
Signed-off-by: Erick Reyes <erickreyes@google.com>
The number of references to a buffer total_refs was not accounting for
a process holding multiple references to the same buffer.
Also, clarify AppendDmaBufInfo functionality in comments
Change-Id: I044a18acad1570492a925d51b9464d5b0e60ca35
Signed-off-by: Erick Reyes <erickreyes@google.com>
Add some test to verify the refcount and fd reference is correct when
the dma_buf is shared between processes.
Bug; 63860998
Test: libdmabufinfo_test
Change-Id: Id22e68e7a65820f19847b2faab11c78e6d942d92
Change the dmabufinfo library to handle multiple references to the same
buffer across PIDs by doing these changes:
- Add function AppendDmaBufInfo: appends the information for a specific
PID to an existing DmaBuffer vector
- Change the way fd/map references to a buffer belonging to a PID are
stored in DmaBuffer from vector<pid_t> to unorderedmap<pid_t, int>
- When parsing proc/%pid/{maps|fds}, look up the buffer by inode and
update the data if found
Test: builds
Bug: 63860998
Change-Id: If22feea772f182086ae3446b46d626a750674b1f
Signed-off-by: Erick Reyes <erickreyes@google.com>
Remove all tests that validated libmeminfo against the libpagemap
implementation and make libmeminfo_test self sufficient.
Bug: 111694435
Bug: 113035067
Test: libmeminfo_test
Change-Id: I21b9a0e2c71ef3f1f672a3b1de8080c0c36b4bfe
Signed-off-by: Sandeep Patil <sspatil@google.com>
This is the userspace component for the improved shared memory buffer
tracking.
A new static library called libdmabufinfo will provide APIs to parse dmabuf
debugfs entries. The library can also map and read the file descriptors
per process to cross-reference the global data per process.
This change only adds a global API to read dmabuf stats from debugfs.
Bug: 63860998
Test: dmabufinfo_test
Change-Id: Ib0deedc73337a70875e03e07d3e9a692c4c33e63
Signed-off-by: Greg Hackmann <ghackmann@google.com>
Signed-off-by: Erick Reyes <erickreyes@google.com>
Signed-off-by: Sandeep Patil <sspatil@google.com>
Also adds a tool to make use of this if it exists call 'wsstop'
Bug: 111694435
Test: wsstop -d 3 -n 100 1
Change-Id: I50415f0bdc09c09b5b414cf0e4fff8f2907c5823
Signed-off-by: Sandeep Patil <sspatil@google.com>
There is no need for the vmalloc reader to be in the sysmeminfo object.
The caller has to create the SysMemInfo object for no reason right now.
Bug: 111694435
Test: libmeminfo_test 1
Change-Id: I263c08c169ea6caf2221bfb915b26ef3c0b1c20c
Signed-off-by: Sandeep Patil <sspatil@google.com>
.. and make sure we have some helper methods exposed to the clients
to do some basic checks for the pagemap values. For example, to check
if the page is present or swapped.
Bug: 111694435
Test: libmeminfo_test 1 --gtest_filter=ValidateProcMemInfo.TestPageMap
Test: libmeminfo_test 1
Change-Id: Ic6ae91f4214b42346f3d0b54164a43ac79d5ade1
Signed-off-by: Sandeep Patil <sspatil@google.com>
Enables procrank, procmem, librank and showmap.
Only showmap gets built into the default builds.
Bug: 111694435
Test: builds
Test: showmap exists in final builds
Change-Id: Ib5901aa2bd4ccc6aa3d02915fda92f8320651315
Signed-off-by: Sandeep Patil <sspatil@google.com>
The test has become invalid due to addition of IsSmapsRollupSupported()
and its internal usage by SmapsOrRollup() method. The method now returns
success and returns statistics even if 'smaps_rollup' doesn't exist. It
does that by internally checking for it's existence and falls back to
using /proc/<pid>/smaps
Bug: 111694435
Test: mma -j
Change-Id: I1983a23f1f617aee126bc66a4c1fbd4abb50f565
Signed-off-by: Sandeep Patil <sspatil@google.com>
Consolidate the checking of /proc/<pid>/smaps_rollup support
in libmeminfo and do it in a thread safe way.
Use the API in ProcMemInfo as well to eliminate the extra
parameters passed to SmapsOrRollup* methods.
Bug: 111694435
Test: libmeminfo_test 1 --gtest_filter=TestProcMemInfo.IsSmapsSupportedTest
Test: Tested with and without the smaps_rollup support in kernel.
Change-Id: I992057f06b54569025fa0cdade9618da2675d1de
Merged-In: I992057f06b54569025fa0cdade9618da2675d1de
Signed-off-by: Sandeep Patil <sspatil@google.com>
The Vma and ProcMemInfo objects do not need separate stats
objects for storing working set. The Vma either has working set
or memory usage information and never both.
Bug: 111694435
Test: libmeminfo_test 1
Change-Id: I2df05f7e750bbba4325474633e705d6d68dd2ccb
Merged-In: I2df05f7e750bbba4325474633e705d6d68dd2ccb
Signed-off-by: Sandeep Patil <sspatil@google.com>
Caused by passing invalid parameters to getline(3) and the test
failure went unnoticed.
Bug: 111694435
Test: libmeminfo_test 1 --gtest_filter=TestProcMemInfo.ForEachVmaFromFileTest
Change-Id: Ideb39604c58f89237b05d2f7c8edb67c5ae65768
Merged-In: Ideb39604c58f89237b05d2f7c8edb67c5ae65768
Signed-off-by: Sandeep Patil <sspatil@google.com>
Needed by showmap and also android_s_Debug to classify each allocation
into multiple heaps.
The APIs added are:
ForEachVmaFromFile - Global API to parse a file expected to be in the
same format as /proc/<pid>/smaps and make a callback for each VMA found.
ProcMemInfo::ForEachVma - Same as 'ForEachVmaFromFile' but for a
ProcMemInfo object corresponding to a process(pid).
ProcMemInfo::Smaps - Wrapper to ProcMemInfo::ForEachVma, except the
function collects 'struct Vma' in a member vector and returns the
reference for the same.
Added showmap2 using the APIs and the corresponding tests the same time.
Bug: 111694435
Test: showmap_test.sh
Test: libmeminfo_test 1
Change-Id: I3065809cf94ecf3da88529809701035c47a8ce34
Signed-off-by: Sandeep Patil <sspatil@google.com>
This adds the tests and SmapsRollup() parsing function in
ProcMemInfo. Adds tests to check the return value as well as
the correctness.
Bug: 111694435
Test: libmeminfo_test 1 --gtest_filter=TestProcMemInfo.*
Test: libmeminfo_benchmark --benchmark_filter=BM_SmapsRollup_
Result:
----------------------------------------------------------
Benchmark Time CPU Iterations
----------------------------------------------------------
BM_SmapsRollup_old 4751 ns 4730 ns 149458
BM_SmapsRollup_new 4858 ns 4837 ns 144636
----------------------------------------------------------
Change-Id: Ia051fe53a7622e3091502ff7166efafae35e7935
Signed-off-by: Sandeep Patil <sspatil@google.com>
This is to replace occurrences of get_allocated_vmalloc_memory().
Splitting into libmeminfo already found a bug with current code which
failed to account for memory allocated by modules due to addition of
the extra [%module_name%] in __builtin_return_address().
See: https://elixir.bootlin.com/linux/latest/source/kernel/kallsyms.c#L373
Also improves the performance a bit in the process.
Bug: 119639955
Bug: 111694435
Test: libmeminfo_test 1 --gtest_filter=SysMemInfoParser.TestVmallocInfo
Test: libmeminfo_benchmark --benchmark_filter=BM_VmallocInfo_*
Result:
----------------------------------------------------------------
Benchmark Time CPU Iterations
----------------------------------------------------------------
BM_VmallocInfo_old_fixed 459239 ns 457268 ns 1532
BM_VmallocInfo_new 386032 ns 384353 ns 1821
----------------------------------------------------------------
Change-Id: I1b6606ac73b5cc2dac31d24487b462ec9abfb2ef
Signed-off-by: Sandeep Patil <sspatil@google.com>
The new variant is primarily used in framework. See: go/ag/5780400
for usage. Also add tests, benchmarks and fix several issues found in SysMemInfo
class.
New benchmark results are:
--------------------------------------------------------------
Benchmark Time CPU Iterations
--------------------------------------------------------------
BM_ReadMemInfo_old 7726 ns 7696 ns 90201
BM_ReadMemInfo_new 7554 ns 7525 ns 90358
BM_ZramTotal_old 6446 ns 6406 ns 108361
BM_ZramTotal_new 6529 ns 6488 ns 106545
BM_MemInfoWithZram_old 14485 ns 14412 ns 48492
BM_MemInfoWithZram_new 20572 ns 20459 ns 33438
--------------------------------------------------------------
The reason for BM_MemInfoWithZram_new shows worse numbers is because
the new API also tries to find more than 1 zram device (if it exists).
The old implementation hard coded everything to "/sys/block/zram0/"
Test: libmeminfo_test 1
Bug: 114325007
Bug: 111694435
Change-Id: I246d9e9a54986ee9b2542d1eaac79ecf7310b23a
Signed-off-by: Sandeep Patil <sspatil@google.com>
In the same time replace iomanipulators with
::android::base::StringPrintf which makes the output much more readable
and the overall program simpler.
Bug: 114325007
Bug: 111694435
Test: procmem 1
Test: procmem -W 1
Test: procmem -u 1
Test: procmem -p 1
Test: procmem -w 1
Test: procmem -w -u 1
Test: procmem -w -p 1
Change-Id: Ibf7ef59b24dfcb851c78a3b2fad672d96a708b98
Signed-off-by: Sandeep Patil <sspatil@google.com>
