Re: [PATCH v3 8/8] selftests/dmabuf-heaps: Add dma-buf memcg accounting tests

Thank you for your contribution! Sashiko AI review found 5 potential issue(s) to consider:
- [Medium] selftests/dmabuf-heaps: check_deps() runs before handle_build() in vmtest.sh
- [Medium] selftests/dmabuf-heaps: fail instead of skip on EOPNOTSUPP in XFER_CHARGE test
- [Medium] selftests/dmabuf-heaps: pidfd_alloc_supported() treats EOPNOTSUPP as supported
- [Medium] selftests: dmabuf-heaps: no memory controller check in open_self_cgroupfd()
- [Medium] selftests: cgroup: MEMCG_DMABUF_WAIT_RETRIES is shorter than 2s rstat FLUSH_TIME
--

commit 764a620a8574ac6b7e6c57e6f71e4c743c373656
Author: Albert Esteve [off-list ref]

selftests/dmabuf-heaps: Add dma-buf memcg accounting tests

This commit adds selftests for dma-buf memcg accounting via charge_fd in
DMA_HEAP_IOCTL_ALLOC and DMA_BUF_IOCTL_XFER_CHARGE across the dmabuf-heaps
and cgroup test suites, along with a vmtest.sh runner script.

[Severity: Medium]
Is MEMCG_DMABUF_WAIT_RETRIES (15 retries at 100ms intervals, or 1.4 seconds
before the final read) long enough to avoid races with the periodic rstat
flush in test_memcg_dmabuf() and test_memcg_dmabuf_xfer_at_limit()?

In mm/memcontrol.c, reading memory.stat calls mem_cgroup_flush_stats(), which
only flushes rstat synchronously when stats_updates exceeds
MEMCG_CHARGE_BATCH * num_online_cpus() (64 * num_online_cpus()):

mm/memcontrol.c:mem_cgroup_flush_stats()
    if (atomic_long_read(&vmstats->stats_updates) >
        MEMCG_CHARGE_BATCH * num_online_cpus())
        ...

Charging or uncharging a ONE_MEG (1 MB) dma-buf in a fresh child cgroup
updates MEMCG_DMABUF by 256 pages on 4KB-page kernels (or 16 pages on
64KB-page kernels). When num_online_cpus() >= 4 (where 256 > 64 * 4 is false)
or on >4KB-page kernels, reading memory.stat skips synchronous flushing and
relies on the periodic deferred work stats_flush_dwork, which runs every
FLUSH_TIME (2UL * HZ, or 2.0 seconds).

In test_memcg_dmabuf(), even if the first poll for ONE_MEG catches a periodic
flush, that flush resets stats_updates to 0 and schedules the next
stats_flush_dwork 2.0 seconds later. Immediately closing dmabuf_fd adds only
256 to stats_updates, so the subsequent 1.4-second poll for dmabuf == 0 can
time out before the next 2.0-second periodic flush runs. Should
MEMCG_DMABUF_WAIT_RETRIES be at least 30, matching
MEMCG_SOCKSTAT_WAIT_RETRIES?

[ ... ]

[ ... ]

[Severity: Medium]
What happens in pidfd_alloc_supported() when the kernel is built with
CONFIG_MEMCG=n?

In drivers/dma-buf/dma-heap.c, dma_heap_ioctl_allocate() passes a non-zero
charge_fd to mem_cgroup_get_from_fd():

drivers/dma-buf/dma-heap.c:dma_heap_ioctl_allocate()
    if (heap_allocation->charge_fd) {
        memcg = mem_cgroup_get_from_fd(heap_allocation->charge_fd);
        if (IS_ERR(memcg))
            return ERR_CAST(memcg);
    }

When CONFIG_MEMCG is disabled, the stub in include/linux/memcontrol.h returns
ERR_PTR(-EOPNOTSUPP):

include/linux/memcontrol.h:mem_cgroup_get_from_fd()
    static inline struct mem_cgroup *mem_cgroup_get_from_fd(int fd)
    {
        return ERR_PTR(-EOPNOTSUPP);
    }

Because dmabuf_heap_alloc_pidfd() returns -1 with errno == EOPNOTSUPP,
returning ret < 0 causes pidfd_alloc_supported() to return true when
CONFIG_MEMCG=n, and test_alloc_pidfd_self(), test_alloc_pidfd_child(),
test_alloc_cgroupfd_self(), and test_xfer_charge_cgroupfd_self() then proceed
and report test failures instead of skipping. Should pidfd_alloc_supported()
also verify that errno != EOPNOTSUPP (or errno == EINVAL)?

[ ... ]

[Severity: Medium]
Can test_alloc_cgroupfd_self() and test_xfer_charge_cgroupfd_self() fail with
ENOENT instead of skipping when the test binary is run outside vmtest.sh (for
example via make run_tests or vmtest.sh -s) in a cgroup where the memory
controller is not enabled?

Unlike the pidfd path in mem_cgroup_get_from_fd(), which uses
get_mem_cgroup_from_mm() and walks up to an online ancestor, the cgroupfd
path requires the memory controller CSS to be directly enabled on the target
cgroup itself:

mm/memcontrol.c:mem_cgroup_get_from_fd()
    cgrp = cgroup_get_from_fd(fd);
    if (!IS_ERR(cgrp)) {
        rcu_read_lock();
        css = cgroup_css(cgrp, &memory_cgrp_subsys);
        if (!css || !css_tryget_online(css)) {
            rcu_read_unlock();
            cgroup_put(cgrp);
            return ERR_PTR(-ENOENT);
        }

When dmabuf-heap runs in a leaf cgroup (such as a standard systemd user or
service scope) whose parent has not enabled +memory in
cgroup.subtree_control, or when booted with cgroup_disable=memory,
open_self_cgroupfd() still succeeds in opening the directory, and the ioctl
fails with ENOENT. Should open_self_cgroupfd() check that memory is active in
that cgroup (or should these two tests skip when the ioctl fails with
errno == ENOENT)?

[Severity: Medium]
Does test_xfer_charge_cgroupfd_self() report a failure instead of skipping
when main() iterates over non-system heaps in /dev/dma_heap (such as CMA
heaps when CONFIG_DMABUF_HEAPS_CMA=y)?

In drivers/dma-buf/dma-heap.c, dma_heap_ioctl_allocate() validates charge_fd
before calling heap->ops->allocate(), so pidfd_alloc_supported() returns true
for all heaps. However, in drivers/dma-buf/dma-buf.c, dma_buf_xfer_charge()
checks whether the exporter enabled memcg accounting on the buffer:

drivers/dma-buf/dma-buf.c:dma_buf_xfer_charge()
    if (!dmabuf->memcg_account)
        return -EOPNOTSUPP;

Because only system_heap sets exp_info.memcg_account = true (while exporters
such as cma_heap leave it false), ioctl(dmabuf_fd, DMA_BUF_IOCTL_XFER_CHARGE,
&xfer) returns -1 with errno == EOPNOTSUPP on those heaps. Should
test_xfer_charge_cgroupfd_self() call ksft_test_result_skip() when ret < 0
and errno == EOPNOTSUPP?

[ ... ]

[ ... ]

[Severity: Medium]
If vmtest.sh is invoked with -b on a clean tree where the selftest binaries
have not been built yet, won't check_deps() exit with KSFT_SKIP before
handle_build() has a chance to build them?

In the main flow, check_deps() runs before handle_build(), and inside
check_deps() only SHELL_MODE returns early before checking for
DMABUF_HEAP_TEST and MEMCONTROL_TEST. Should check_deps() also skip the
binary existence checks when BUILD is true (or run after handle_build())?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261027-v2_20230123_tjmercier_google_com-v3-0-78d2c4cb8647@redhat.com?part=8