On Tue, Oct 24, 2023 at 11:29 AM Roman Gushchin
[off-list ref] wrote:

On Tue, Oct 24, 2023 at 06:45:57AM -0700, Suren Baghdasaryan wrote:

quoted

Updates since the last version [1]
- Simplified allocation tagging macros;
- Runtime enable/disable sysctl switch (/proc/sys/vm/mem_profiling)
instead of kernel command-line option;
- CONFIG_MEM_ALLOC_PROFILING_BY_DEFAULT to select default enable state;
- Changed the user-facing API from debugfs to procfs (/proc/allocinfo);
- Removed context capture support to make patch incremental;
- Renamed uninstrumented allocation functions to use _noprof suffix;
- Added __GFP_LAST_BIT to make the code cleaner;
- Removed lazy per-cpu counters; it turned out the memory savings was
minimal and not worth the performance impact;

Hello Suren,

quoted

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 is performance
comparison between the baseline kernel, profiling when enabled, profiling
when disabled and (for comparison purposes) baseline with
CONFIG_MEMCG_KMEM enabled and allocations using __GFP_ACCOUNT:

                        kmalloc                 pgalloc
(1 baseline)            12.041s                 49.190s
(2 default disabled)    14.970s (+24.33%)       49.684s (+1.00%)
(3 default enabled)     16.859s (+40.01%)       56.287s (+14.43%)
(4 runtime enabled)     16.983s (+41.04%)       55.760s (+13.36%)
(5 memcg)               33.831s (+180.96%)      51.433s (+4.56%)

some recent changes [1] to the kmem accounting should have made it quite a bit
faster. Would be great if you can provide new numbers for the comparison.
Maybe with the next revision?

And btw thank you (and Kent): your numbers inspired me to do this kmemcg
performance work. I expect it still to be ~twice more expensive than your
stuff because on the memcg side we handle separately charge and statistics,
but hopefully the difference will be lower.

Yes, I saw them! Well done! I'll definitely update my numbers once the
patches land in their final form.

Thank you!

Thank you for the optimizations!

[1]:
  patches from next tree, so no stable hashes:
    mm: kmem: reimplement get_obj_cgroup_from_current()
    percpu: scoped objcg protection
    mm: kmem: scoped objcg protection
    mm: kmem: make memcg keep a reference to the original objcg
    mm: kmem: add direct objcg pointer to task_struct
    mm: kmem: optimize get_obj_cgroup_from_current()