Age | Commit message (Collapse) | Author | Files | Lines |
|
Patch series "Memory allocation profiling", v6.
Overview:
Low overhead [1] per-callsite memory allocation profiling. Not just for
debug kernels, overhead low enough to be deployed in production.
Example output:
root@moria-kvm:~# sort -rn /proc/allocinfo
127664128 31168 mm/page_ext.c:270 func:alloc_page_ext
56373248 4737 mm/slub.c:2259 func:alloc_slab_page
14880768 3633 mm/readahead.c:247 func:page_cache_ra_unbounded
14417920 3520 mm/mm_init.c:2530 func:alloc_large_system_hash
13377536 234 block/blk-mq.c:3421 func:blk_mq_alloc_rqs
11718656 2861 mm/filemap.c:1919 func:__filemap_get_folio
9192960 2800 kernel/fork.c:307 func:alloc_thread_stack_node
4206592 4 net/netfilter/nf_conntrack_core.c:2567 func:nf_ct_alloc_hashtable
4136960 1010 drivers/staging/ctagmod/ctagmod.c:20 [ctagmod] func:ctagmod_start
3940352 962 mm/memory.c:4214 func:alloc_anon_folio
2894464 22613 fs/kernfs/dir.c:615 func:__kernfs_new_node
...
Usage:
kconfig options:
- CONFIG_MEM_ALLOC_PROFILING
- CONFIG_MEM_ALLOC_PROFILING_ENABLED_BY_DEFAULT
- CONFIG_MEM_ALLOC_PROFILING_DEBUG
adds warnings for allocations that weren't accounted because of a
missing annotation
sysctl:
/proc/sys/vm/mem_profiling
Runtime info:
/proc/allocinfo
Notes:
[1]: Overhead
To measure the overhead we are comparing the following configurations:
(1) Baseline with CONFIG_MEMCG_KMEM=n
(2) Disabled by default (CONFIG_MEM_ALLOC_PROFILING=y &&
CONFIG_MEM_ALLOC_PROFILING_BY_DEFAULT=n)
(3) Enabled by default (CONFIG_MEM_ALLOC_PROFILING=y &&
CONFIG_MEM_ALLOC_PROFILING_BY_DEFAULT=y)
(4) Enabled at runtime (CONFIG_MEM_ALLOC_PROFILING=y &&
CONFIG_MEM_ALLOC_PROFILING_BY_DEFAULT=n && /proc/sys/vm/mem_profiling=1)
(5) Baseline with CONFIG_MEMCG_KMEM=y && allocating with __GFP_ACCOUNT
(6) Disabled by default (CONFIG_MEM_ALLOC_PROFILING=y &&
CONFIG_MEM_ALLOC_PROFILING_BY_DEFAULT=n) && CONFIG_MEMCG_KMEM=y
(7) Enabled by default (CONFIG_MEM_ALLOC_PROFILING=y &&
CONFIG_MEM_ALLOC_PROFILING_BY_DEFAULT=y) && CONFIG_MEMCG_KMEM=y
Performance overhead:
To evaluate performance we implemented an in-kernel test executing
multiple get_free_page/free_page and kmalloc/kfree calls with allocation
sizes growing from 8 to 240 bytes with CPU frequency set to max and CPU
affinity set to a specific CPU to minimize the noise. Below are results
from running the test on Ubuntu 22.04.2 LTS with 6.8.0-rc1 kernel on
56 core Intel Xeon:
kmalloc pgalloc
(1 baseline) 6.764s 16.902s
(2 default disabled) 6.793s (+0.43%) 17.007s (+0.62%)
(3 default enabled) 7.197s (+6.40%) 23.666s (+40.02%)
(4 runtime enabled) 7.405s (+9.48%) 23.901s (+41.41%)
(5 memcg) 13.388s (+97.94%) 48.460s (+186.71%)
(6 def disabled+memcg) 13.332s (+97.10%) 48.105s (+184.61%)
(7 def enabled+memcg) 13.446s (+98.78%) 54.963s (+225.18%)
Memory overhead:
Kernel size:
text data bss dec diff
(1) 26515311 18890222 17018880 62424413
(2) 26524728 19423818 16740352 62688898 264485
(3) 26524724 19423818 16740352 62688894 264481
(4) 26524728 19423818 16740352 62688898 264485
(5) 26541782 18964374 16957440 62463596 39183
Memory consumption on a 56 core Intel CPU with 125GB of memory:
Code tags: 192 kB
PageExts: 262144 kB (256MB)
SlabExts: 9876 kB (9.6MB)
PcpuExts: 512 kB (0.5MB)
Total overhead is 0.2% of total memory.
Benchmarks:
Hackbench tests run 100 times:
hackbench -s 512 -l 200 -g 15 -f 25 -P
baseline disabled profiling enabled profiling
avg 0.3543 0.3559 (+0.0016) 0.3566 (+0.0023)
stdev 0.0137 0.0188 0.0077
hackbench -l 10000
baseline disabled profiling enabled profiling
avg 6.4218 6.4306 (+0.0088) 6.5077 (+0.0859)
stdev 0.0933 0.0286 0.0489
stress-ng tests:
stress-ng --class memory --seq 4 -t 60
stress-ng --class cpu --seq 4 -t 60
Results posted at: https://evilpiepirate.org/~kent/memalloc_prof_v4_stress-ng/
[2] https://lore.kernel.org/all/20240306182440.2003814-1-surenb@google.com/
This patch (of 37):
The next patch drops vmalloc.h from a system header in order to fix a
circular dependency; this adds it to all the files that were pulling it in
implicitly.
[kent.overstreet@linux.dev: fix arch/alpha/lib/memcpy.c]
Link: https://lkml.kernel.org/r/20240327002152.3339937-1-kent.overstreet@linux.dev
[surenb@google.com: fix arch/x86/mm/numa_32.c]
Link: https://lkml.kernel.org/r/20240402180933.1663992-1-surenb@google.com
[kent.overstreet@linux.dev: a few places were depending on sizes.h]
Link: https://lkml.kernel.org/r/20240404034744.1664840-1-kent.overstreet@linux.dev
[arnd@arndb.de: fix mm/kasan/hw_tags.c]
Link: https://lkml.kernel.org/r/20240404124435.3121534-1-arnd@kernel.org
[surenb@google.com: fix arc build]
Link: https://lkml.kernel.org/r/20240405225115.431056-1-surenb@google.com
Link: https://lkml.kernel.org/r/20240321163705.3067592-1-surenb@google.com
Link: https://lkml.kernel.org/r/20240321163705.3067592-2-surenb@google.com
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Pasha Tatashin <pasha.tatashin@soleen.com>
Tested-by: Kees Cook <keescook@chromium.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Alex Gaynor <alex.gaynor@gmail.com>
Cc: Alice Ryhl <aliceryhl@google.com>
Cc: Andreas Hindborg <a.hindborg@samsung.com>
Cc: Benno Lossin <benno.lossin@proton.me>
Cc: "Björn Roy Baron" <bjorn3_gh@protonmail.com>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Gary Guo <gary@garyguo.net>
Cc: Miguel Ojeda <ojeda@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wedson Almeida Filho <wedsonaf@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
|
|
The current logic for handling resets is more complicated than it needs
to be. The deferred_reset flag is used to indicate a reset is needed
and the deferred_reset_state is the requested, post-reset, state.
Also, the deferred_reset logic was added to vfio migration drivers to
prevent a circular locking dependency with respect to mm_lock and state
mutex. This is mainly because of the copy_to/from_user() functions(which
takes mm_lock) invoked under state mutex.
Remove all of the deferred reset logic and just pass the requested
next state to pds_vfio_reset() so it can be used for VMM and DSC
initiated resets.
This removes the need for pds_vfio_state_mutex_lock(), so remove that
and replace its use with a simple mutex_unlock().
Also, remove the reset_mutex as it's no longer needed since the
state_mutex can be the driver's primary protector.
Suggested-by: Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>
Reviewed-by: Shannon Nelson <shannon.nelson@amd.com>
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Link: https://lore.kernel.org/r/20240308182149.22036-3-brett.creeley@amd.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
|
|
Only supporting a single region/range is limiting,
wasteful, and in some cases broken (i.e. when there
are large gaps in the iova memory ranges). Fix this
by adding support for multiple regions based on
what the device tells the driver it can support.
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Signed-off-by: Shannon Nelson <shannon.nelson@amd.com>
Link: https://lore.kernel.org/r/20231117001207.2793-7-brett.creeley@amd.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
|
|
Since the host seq/ack bitmaps are part of a region
move them into struct pds_vfio_region. Also, make use
of the bmp_bytes value for validation purposes.
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Signed-off-by: Shannon Nelson <shannon.nelson@amd.com>
Link: https://lore.kernel.org/r/20231117001207.2793-6-brett.creeley@amd.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
|
|
A later patch in the series implements multi-region
support. That will require specific regions to be
passed to relevant functions. Prepare for that change
by passing the region structure to relevant functions.
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Signed-off-by: Shannon Nelson <shannon.nelson@amd.com>
Link: https://lore.kernel.org/r/20231117001207.2793-5-brett.creeley@amd.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
|
|
An upcoming change in this series will add support
for multiple regions. To prepare for that, move
region specific information into struct pds_vfio_region
and rename the members for readability. This will
reduce the size of the patch that actually implements
multiple region support.
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Signed-off-by: Shannon Nelson <shannon.nelson@amd.com>
Link: https://lore.kernel.org/r/20231117001207.2793-4-brett.creeley@amd.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
|
|
Since the seq/ack operations never happen in parallel there
is no need for multiple scatter gather lists per region.
The current implementation is wasting memory. Fix this by
only using a single scatter-gather list for both the seq
and ack operations.
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Signed-off-by: Shannon Nelson <shannon.nelson@amd.com>
Link: https://lore.kernel.org/r/20231117001207.2793-3-brett.creeley@amd.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
|
|
The incorrect check is being done for comparing the
iova/length being requested to sync. This can cause
the dirty sync operation to fail. Fix this by making
sure the iova offset added to the requested sync
length doesn't exceed the region_size.
Also, the region_start is assumed to always be at 0.
This can cause dirty tracking to fail because the
device/driver bitmap offset always starts at 0,
however, the region_start/iova may not. Fix this by
determining the iova offset from region_start to
determine the bitmap offset.
Fixes: f232836a9152 ("vfio/pds: Add support for dirty page tracking")
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Signed-off-by: Shannon Nelson <shannon.nelson@amd.com>
Link: https://lore.kernel.org/r/20231117001207.2793-2-brett.creeley@amd.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
|
|
In order to support dirty page tracking, the driver has to implement
the VFIO subsystem's vfio_log_ops. This includes log_start, log_stop,
and log_read_and_clear.
All of the tracker resources are allocated and dirty tracking on the
device is started during log_start. The resources are cleaned up and
dirty tracking on the device is stopped during log_stop. The dirty
pages are determined and reported during log_read_and_clear.
In order to support these callbacks admin queue commands are used.
All of the adminq queue command structures and implementations
are included as part of this patch.
PDS_LM_CMD_DIRTY_STATUS is added to query the current status of
dirty tracking on the device. This includes if it's enabled (i.e.
number of regions being tracked from the device's perspective) and
the maximum number of regions supported from the device's perspective.
PDS_LM_CMD_DIRTY_ENABLE is added to enable dirty tracking on the
specified number of regions and their iova ranges.
PDS_LM_CMD_DIRTY_DISABLE is added to disable dirty tracking for all
regions on the device.
PDS_LM_CMD_READ_SEQ and PDS_LM_CMD_DIRTY_WRITE_ACK are added to
support reading and acknowledging the currently dirtied pages.
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Signed-off-by: Shannon Nelson <shannon.nelson@amd.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Reviewed-by: Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>
Link: https://lore.kernel.org/r/20230807205755.29579-7-brett.creeley@amd.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
|