selftests/bpf: retry bpf_map_update_elem() when E2BIG is returned
Subsystems (3)
PATCH 1/1 selftests/bpf: retry bpf_map_update_elem() when E2BIG is returned
Executing the test_maps binary on platforms with extremely high core counts may cause intermittent assertion failures in test_update_delete() (called via test_map_parallel()). This can occur because bpf_map_update_elem() under some circumstances (specifically in this case while performing bpf_map_update_elem() with BPF_NOEXIST on a BPF_MAP_TYPE_HASH with its map_flags set to BPF_F_NO_PREALLOC) can return an E2BIG error code i.e.
error -7 7
tools/testing/selftests/bpf/test_maps.c:#: void test_update_delete(unsigned int, void *): Assertion `err == 0' failed. tools/testing/selftests/bpf/test_maps.c:#: void__run_parallel(unsigned int, void (*)(unsigned int, void *), void *): Assertion `status == 0' failed.
As it turns out, is_map_full() which is called from alloc_htab_elem() can take on a conservative approach when htab->use_percpu_counter is true (which is the case here because the percpu_counter is used when a BPF_MAP_TYPE_HASH is created with its map_flags set to BPF_F_NO_PREALLOC). This conservative approach approach prioritizes preventing over-allocation and potential issues that could arise from possibly exceeding htab->map.max_entries in highly concurrent environments, even if it means slightly under-utilizing the htab map's capacity.
Given that bpf_map_update_elem() from test_update_delete() can return E2BIG, update can_retry() such that it also accounts for the E2BIG error code (specifically only when running with map_flags being set to BPF_F_NO_PREALLOC). The retry loop will allow the global count belonging to the percpu_counter to become synchronized and better reflect the current htab map's capacity.
--- tools/testing/selftests/bpf/test_maps.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
tools/testing/selftests/bpf/test_maps.c +2 -1
@@ -1399,7 +1399,8 @@ static void test_map_stress(void)static bool can_retry (int err) {return (err ==EAGAIN || err ==EBUSY || - (err ==ENOMEM && map_opts.map_flags ==BPF_F_NO_PREALLOC )); + ((err ==ENOMEM || err ==E2BIG ) && + map_opts.map_flags ==BPF_F_NO_PREALLOC )); } int map_update_retriable(int map_fd, const void *key, const void *value, int flags, int attempts,
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
Quoted 14 lines · Matt Bobrowski
Executing the
test_mapsbinary on platforms with extremely high core counts may cause intermittent assertion failures intest_update_delete()(called viatest_map_parallel()). This can occur becausebpf_map_update_elem()under some circumstances (specifically in this case while performingbpf_map_update_elem()withBPF_NOEXISTon aBPF_MAP_TYPE_HASHwith itsmap_flagsset toBPF_F_NO_PREALLOC) can return an E2BIG error code i.e.error -7 7
tools/testing/selftests/bpf/test_maps.c:#: voidtest_update_delete(unsigned int, void *): Assertion `err == 0' failed. tools/testing/selftests/bpf/test_maps.c:#: void__run_parallel(unsigned int, void (*)(unsigned int, void *), void *): Assertion `status == 0' failed.
Quoted 5 lines
As it turns out,
is_map_full()which is called fromalloc_htab_elem()can take on a conservative approach whenhtab->use_percpu_counteris true (which is the case here because thepercpu_counteris used when aBPF_MAP_TYPE_HASHis created with itsmap_flagsset toBPF_F_NO_PREALLOC). This conservative approach approach prioritizes
s/approach approach/approach
AFAICT checkpatch.pl also warns double "approach", as well as line exceed 75 character above.
Quoted 5 lines
preventing over-allocation and potential issues that could arise from possibly exceeding
htab->map.max_entriesin highly concurrent environments, even if it means slightly under-utilizing the htab map's capacity.
Quoted 8 lines
Given that
bpf_map_update_elem()fromtest_update_delete()can return E2BIG, updatecan_retry()such that it also accounts for the E2BIG error code (specifically only when running withmap_flagsbeing set toBPF_F_NO_PREALLOC). The retry loop will allow the global count belonging to thepercpu_counterto become synchronized and better reflect the current htab map's capacity.Signed-off-by Matt Bobrowski <[email protected]>
Other than the nitpick above, this looks good to me.
Quoted 21 lines
--- tools/testing/selftests/bpf/test_maps.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)tools/testing/selftests/bpf/test_maps.c +2
-2@@ -1399,7 +1399,8 @@ static void test_map_stress(void)static bool can_retry (int err) {return (err ==EAGAIN || err ==EBUSY || - (err ==ENOMEM && map_opts.map_flags ==BPF_F_NO_PREALLOC )); + ((err ==ENOMEM || err ==E2BIG ) && + map_opts.map_flags ==BPF_F_NO_PREALLOC )); } int map_update_retriable(int map_fd, const void *key, const void *value, int flags, int attempts, --2.51.2.1041.gc1ab5b90ca-goog
Quoted 22 lines · Song Liu
Quoted 14 lines · Matt Bobrowski
Executing the
test_mapsbinary on platforms with extremely high core counts may cause intermittent assertion failures intest_update_delete()(called viatest_map_parallel()). This can occur becausebpf_map_update_elem()under some circumstances (specifically in this case while performingbpf_map_update_elem()withBPF_NOEXISTon aBPF_MAP_TYPE_HASHwith itsmap_flagsset toBPF_F_NO_PREALLOC) can return an E2BIG error code i.e.error -7 7
tools/testing/selftests/bpf/test_maps.c:#: voidtest_update_delete(unsigned int, void *): Assertion `err == 0' failed. tools/testing/selftests/bpf/test_maps.c:#: void__run_parallel(unsigned int, void (*)(unsigned int, void *), void *): Assertion `status == 0' failed.Quoted 5 lines
As it turns out,
is_map_full()which is called fromalloc_htab_elem()can take on a conservative approach whenhtab->use_percpu_counteris true (which is the case here because thepercpu_counteris used when aBPF_MAP_TYPE_HASHis created with itsmap_flagsset toBPF_F_NO_PREALLOC). This conservative approach approach prioritizes
Quoted 4 lines
s/approach approach/approach
AFAICT checkpatch.pl also warns double "approach", as well as line exceed 75 character above.
ACK. Will respin with nits addressed at some point today.
Quoted 40 lines
Quoted 5 lines
preventing over-allocation and potential issues that could arise from possibly exceeding
htab->map.max_entriesin highly concurrent environments, even if it means slightly under-utilizing the htab map's capacity.Quoted 8 lines
Given that
bpf_map_update_elem()fromtest_update_delete()can return E2BIG, updatecan_retry()such that it also accounts for the E2BIG error code (specifically only when running withmap_flagsbeing set toBPF_F_NO_PREALLOC). The retry loop will allow the global count belonging to thepercpu_counterto become synchronized and better reflect the current htab map's capacity.Signed-off-by Matt Bobrowski <[email protected]>Other than the nitpick above, this looks good to me.
Acked-by Song Liu <[email protected]>Quoted 21 lines
--- tools/testing/selftests/bpf/test_maps.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)tools/testing/selftests/bpf/test_maps.c +2
-2@@ -1399,7 +1399,8 @@ static void test_map_stress(void)static bool can_retry (int err) {return (err ==EAGAIN || err ==EBUSY || - (err ==ENOMEM && map_opts.map_flags ==BPF_F_NO_PREALLOC )); + ((err ==ENOMEM || err ==E2BIG ) && + map_opts.map_flags ==BPF_F_NO_PREALLOC )); } int map_update_retriable(int map_fd, const void *key, const void *value, int flags, int attempts, --2.51.2.1041.gc1ab5b90ca-goog