[PATCH OLK-6.6 v3 0/2] mm/mem_sampling: fix some mem sampling issues
This patch fix issues in mem_sampling: - Prevent mem_sampling from enabling if SPE initialization fails. - Fix inaccurate sampling during NUMA balancing and DAMON. These changes improve memory sampling accuracy and stability on ARM systems. changes since v2: -- remove the feature patch. -- fix bugzilla number. Ze Zuo (2): mm/mem_sampling: Prevent mem_sampling from being enabled if SPE init failed mm/mem_sampling: Fix inaccurate sampling for NUMA balancing and DAMON drivers/arm/mm_monitor/mm_spe.c | 6 +++--- include/linux/mem_sampling.h | 7 ++++--- mm/mem_sampling.c | 11 +++++++++-- 3 files changed, 16 insertions(+), 8 deletions(-) -- 2.25.1
hulk inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/ICH7BF CVE: NA -------------------------------- When the ARM Statistical Profiling Extension (SPE) PMU driver is not properly initialized or has already been unloaded, it is not safe to enable the `mem_sampling` feature. Add a check to ensure that `mem_sampling` can only be enabled if `spe_probe_status == SPE_INIT_SUCC`. This prevents enabling memory sampling in an invalid state, which may otherwise lead to undefined behavior or system instability. Fixs: 5e73787bd539 ("mm/numa: Use mem_sampling framework for NUMA balancing") Signed-off-by: Ze Zuo <zuoze1@huawei.com> --- drivers/arm/mm_monitor/mm_spe.c | 2 +- mm/mem_sampling.c | 11 +++++++++-- 2 files changed, 10 insertions(+), 3 deletions(-) diff --git a/drivers/arm/mm_monitor/mm_spe.c b/drivers/arm/mm_monitor/mm_spe.c index 192f1f9c6adb..3c0438ab54e4 100644 --- a/drivers/arm/mm_monitor/mm_spe.c +++ b/drivers/arm/mm_monitor/mm_spe.c @@ -114,8 +114,8 @@ void mm_spe_buffer_free(void) continue; mm_spe_percpu_buffer_free(cpu); } - spe_probe_status -= 1; set_mem_sampling_state(false); + spe_probe_status -= 1; } EXPORT_SYMBOL_GPL(mm_spe_buffer_free); diff --git a/mm/mem_sampling.c b/mm/mem_sampling.c index 8d79e83e64f0..74f95e4611fe 100644 --- a/mm/mem_sampling.c +++ b/mm/mem_sampling.c @@ -278,6 +278,9 @@ static void numa_balancing_mem_sampling_cb_unregister(void) } static void set_numabalancing_mem_sampling_state(bool enabled) { + if (!mem_sampling_ops.sampling_start || !mm_spe_enabled()) + return; + if (enabled) { numa_balancing_mem_sampling_cb_register(); static_branch_enable(&sched_numabalancing_mem_sampling); @@ -342,6 +345,9 @@ static void damon_mem_sampling_record_cb_unregister(void) static void set_damon_mem_sampling_state(bool enabled) { + if (!mem_sampling_ops.sampling_start || !mm_spe_enabled()) + return; + if (enabled) { damon_mem_sampling_record_cb_register(); static_branch_enable(&mm_damon_mem_sampling); @@ -412,14 +418,15 @@ static void __set_mem_sampling_state(bool enabled) void set_mem_sampling_state(bool enabled) { + if (!mem_sampling_ops.sampling_start || !mm_spe_enabled()) + return; + if (mem_sampling_saved_state != MEM_SAMPLING_STATE_EMPTY) { mem_sampling_saved_state = enabled ? MEM_SAMPLING_STATE_ENABLE : MEM_SAMPLING_STATE_DISABLE; return; } - if (!mem_sampling_ops.sampling_start || !mm_spe_enabled()) - return; if (enabled) sysctl_mem_sampling_mode = MEM_SAMPLING_NORMAL; else -- 2.25.1
hulk inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/ICH7BF CVE: NA -------------------------------- SPE memory access records can be used by NUMA balancing and DAMON for accurate kernel-side memory profiling. However, the current implementation suffers from inaccuracies in kernel-side consumers due to two key issues: 1. The `mem_sampling_record` exposed to kernel consumers is not correctly synchronized with `struct arm_spe_record`. This leads to fields such as physical address, virtual address, or task-related metadata being outdated or inconsistent. 2. In dual-buffer configurations, the function `mm_spe_getbuf_addr()` fails to return the correct `record_base` when the backup buffer is in use. As a result, consumers may fetch stale or invalid sampling data. This patch addresses both problems by: - Explicitly copying all relevant fields from `struct arm_spe_record` into `mem_sampling_record` before it is passed to users. - Updating `mm_spe_getbuf_addr()` to check both active and backup buffer state and return the correct `record_base` accordingly. With this fix, NUMA auto-balancing and DAMON-based sampling can operate on accurate and up-to-date SPE sample data in all buffer modes. Fixs: 5e73787bd539 ("mm/numa: Use mem_sampling framework for NUMA balancing") Signed-off-by: Ze Zuo <zuoze1@huawei.com> --- drivers/arm/mm_monitor/mm_spe.c | 4 ++-- include/linux/mem_sampling.h | 7 ++++--- 2 files changed, 6 insertions(+), 5 deletions(-) diff --git a/drivers/arm/mm_monitor/mm_spe.c b/drivers/arm/mm_monitor/mm_spe.c index 3c0438ab54e4..73fa4acdc998 100644 --- a/drivers/arm/mm_monitor/mm_spe.c +++ b/drivers/arm/mm_monitor/mm_spe.c @@ -366,11 +366,11 @@ void mm_spe_decoding(void) arm_spe_decode_buf(spe_buf->cur, spe_buf->size); } -struct mm_spe_buf *mm_spe_getbuf_addr(void) +void *mm_spe_getbuf_addr(void) { struct mm_spe_buf *spe_buf = this_cpu_ptr(&per_cpu_spe_buf); - return spe_buf; + return spe_buf->record_base; } int mm_spe_getnum_record(void) diff --git a/include/linux/mem_sampling.h b/include/linux/mem_sampling.h index 71d098fd3c05..b325c9955d31 100644 --- a/include/linux/mem_sampling.h +++ b/include/linux/mem_sampling.h @@ -49,6 +49,7 @@ struct mem_sampling_record { u64 context_id; u64 boost_spe_addr[8]; u64 rem_addr; + u16 boost_spe_idx; u16 source; }; @@ -57,7 +58,7 @@ struct mem_sampling_ops_struct { void (*sampling_stop)(void); void (*sampling_continue)(void); void (*sampling_decoding)(void); - struct mm_spe_buf* (*mm_spe_getbuf_addr)(void); + void* (*mm_spe_getbuf_addr)(void); int (*mm_spe_getnum_record)(void); }; @@ -83,7 +84,7 @@ void mm_spe_stop(void); void mm_spe_continue(void); void mm_spe_decoding(void); int mm_spe_getnum_record(void); -struct mm_spe_buf *mm_spe_getbuf_addr(void); +void *mm_spe_getbuf_addr(void); int mm_spe_enabled(void); void arm_spe_set_probe_status(int status); #else @@ -93,7 +94,7 @@ static inline void mm_spe_decoding(void) { } static inline void arm_spe_set_probe_status(int status) { } static inline int mm_spe_start(void) { return 0; } static inline int mm_spe_getnum_record(void) { return 0; } -static inline struct mm_spe_buf *mm_spe_getbuf_addr(void) { return NULL; } +static inline void *mm_spe_getbuf_addr(void) { return NULL; } static inline int mm_spe_enabled(void) { return 0; } #endif /* CONFIG_ARM_SPE_MEM_SAMPLING */ -- 2.25.1
反馈： 您发送到kernel@openeuler.org的补丁/补丁集，已成功转换为PR！ PR链接地址： https://gitee.com/openeuler/kernel/pulls/16779 邮件列表地址：https://mailweb.openeuler.org/archives/list/kernel@openeuler.org/message/5AI... FeedBack: The patch(es) which you have sent to kernel@openeuler.org mailing list has been converted to a pull request successfully! Pull request link: https://gitee.com/openeuler/kernel/pulls/16779 Mailing list address: https://mailweb.openeuler.org/archives/list/kernel@openeuler.org/message/5AI...
participants (2)
-
patchwork bot -
Ze Zuo