Hi Abhishek
On 2026/10/8 06:40, Abhishek Bapat wrote: > Introduce a new IOCTL (`ALLOCINFO_IOC_TOGGLE_TRACE`) to selectively > toggle tracing on allocation call sites. Userspace tools can use the > existing filtering mechanism to specify the set of tags to toggle > tracing for. > > To facilitate low overhead execution for non-targeted call sites, add a > new `CODETAG_FLAG_TRACE_ON` flag to `struct codetag` to track per-site > activation. Protect these conditional branch evaluations using the global > `alloc_tag_trace_key` static branch and an inline static key check > pattern in the allocator hooks (`alloc_tag_add`, `alloc_tag_sub`, etc). > This ensures that the trace events are entirely skipped when no > allocation call sites are being actively traced, leaving only a NOP on > the allocation fast path. > > Signed-off-by: Abhishek Bapat <[email protected]> > --- > Documentation/mm/allocation-profiling.rst | 20 ++++ > include/linux/alloc_tag.h | 47 +++++--- > include/linux/codetag.h | 5 +- > include/uapi/linux/alloc_tag.h | 13 +++ > mm/alloc_tag.c | 128 ++++++++++++++++++++-- > 5 files changed, 187 insertions(+), 26 deletions(-) > > diff --git a/Documentation/mm/allocation-profiling.rst > b/Documentation/mm/allocation-profiling.rst > index b241450394df..63b668466342 100644 > --- a/Documentation/mm/allocation-profiling.rst > +++ b/Documentation/mm/allocation-profiling.rst > @@ -149,6 +149,26 @@ expose the full call stack and the lifetime of > individual allocations: > - slab and percpu: the address of the object's `codetag_ref` > - page allocator: the `struct page` pointer of the head page > > +Events are only emitted for tags that have tracing turned on and while > profiling > +is enabled. When no tag is traced, the hooks are behind a static branch. > + > +Enabling Tracing > +---------------- > + > +Tracing is toggled on and off per tag with the `ALLOCINFO_IOC_TOGGLE_TRACE` > +ioctl on `/proc/allocinfo`. It takes the same filter as used by > +`ALLOCINFO_IOC_GET_AT`; an empty mask selects all tags. Size limits in the > +filter are checked once, against the tag's size at the time of the call. Tags > +already in the requested state are skipped and a -ENOENT is returned if no > +tag matched. > + > +Tracing is turned off for a module's tags when the module is unloaded. > + > +Memory allocated before a tag was enabled has no alloc event but will still > +produce a free event. Likewise, memory alive when tag is disabled will not > +produce a free event. Tools should expect unmatched frees after enabling and > +missing frees after disabling. > + > Correlating the events > ---------------------- > > diff --git a/include/linux/alloc_tag.h b/include/linux/alloc_tag.h > index 49825177d8c8..94697b28053a 100644 > --- a/include/linux/alloc_tag.h > +++ b/include/linux/alloc_tag.h > @@ -136,9 +136,36 @@ static inline bool mem_alloc_profiling_enabled(void) > &mem_alloc_profiling_key); > } > > -static inline bool alloc_tag_trace_enabled(void) > +static inline void alloc_tag_set_inaccurate(struct alloc_tag *tag) > { > - return static_branch_unlikely(&alloc_tag_trace_key); > + atomic_or(CODETAG_FLAG_INACCURATE, &tag->ct.flags); > +} > + > +static inline bool alloc_tag_is_inaccurate(struct alloc_tag *tag) > +{ > + return !!(atomic_read(&tag->ct.flags) & CODETAG_FLAG_INACCURATE); > +} > + > +static inline void alloc_tag_set_traced(struct alloc_tag *tag) > +{ > + atomic_or(CODETAG_FLAG_TRACE_ON, &tag->ct.flags); > +} > + > +static inline void alloc_tag_clear_traced(struct alloc_tag *tag) > +{ > + atomic_andnot(CODETAG_FLAG_TRACE_ON, &tag->ct.flags); > +} > + > +static inline bool alloc_tag_is_traced(const struct alloc_tag *tag) > +{ > + return !!(atomic_read(&tag->ct.flags) & CODETAG_FLAG_TRACE_ON); > +} > + > +static inline bool alloc_tag_trace_enabled(const struct alloc_tag *tag) > +{ > + if (static_branch_unlikely(&alloc_tag_trace_key)) > + return tag && alloc_tag_is_traced(tag); > + return false; > } > > void alloc_tag_trace_mem_alloc(const void *ptr, struct alloc_tag *tag, > @@ -151,7 +178,7 @@ void __alloc_tag_trace_hit(struct alloc_tag *tag); > > static __always_inline void alloc_tag_trace_hit(struct alloc_tag *tag) > { > - if (alloc_tag_trace_enabled()) > + if (alloc_tag_trace_enabled(tag)) > __alloc_tag_trace_hit(tag); > } > > @@ -225,7 +252,7 @@ static inline void alloc_tag_add(union codetag_ref *ref, > struct alloc_tag *tag, > if (likely(alloc_tag_ref_set(ref, tag))) { > this_cpu_add(tag->counters->bytes, bytes); > > - if (alloc_tag_trace_enabled()) > + if (alloc_tag_trace_enabled(tag)) > /* Trace successful allocs with their unique ptr */ > alloc_tag_trace_mem_alloc(ptr, tag, bytes); > } > @@ -249,23 +276,13 @@ static inline void alloc_tag_sub(union codetag_ref > *ref, size_t bytes, const voi > this_cpu_sub(tag->counters->bytes, bytes); > this_cpu_dec(tag->counters->calls); > > - if (alloc_tag_trace_enabled()) > + if (alloc_tag_trace_enabled(tag)) > /* Trace frees with their unique ptr */ > alloc_tag_trace_mem_free(ptr, tag, bytes); > > ref->ct = NULL; > } > > -static inline void alloc_tag_set_inaccurate(struct alloc_tag *tag) > -{ > - tag->ct.flags |= CODETAG_FLAG_INACCURATE; > -} > - > -static inline bool alloc_tag_is_inaccurate(struct alloc_tag *tag) > -{ > - return !!(tag->ct.flags & CODETAG_FLAG_INACCURATE); > -} > - > #define alloc_tag_record(p) ((p) = current->alloc_tag) > > #else /* CONFIG_MEM_ALLOC_PROFILING */ > diff --git a/include/linux/codetag.h b/include/linux/codetag.h > index a25a085c2df1..f728295d50c0 100644 > --- a/include/linux/codetag.h > +++ b/include/linux/codetag.h > @@ -18,6 +18,7 @@ struct module; > > /* codetag flags */ > #define CODETAG_FLAG_INACCURATE (1 << 0) > +#define CODETAG_FLAG_TRACE_ON (1 << 1) > > /* > * An instance of this structure is created in a special ELF section at every > @@ -25,7 +26,7 @@ struct module; > * an array of these. > */ > struct codetag { > - unsigned int flags; > + atomic_t flags; > unsigned int lineno; > const char *modname; > const char *function; > @@ -71,7 +72,7 @@ struct codetag_iterator { > .function = __func__, \ > .filename = __FILE__, \ > .lineno = __LINE__, \ > - .flags = 0, \ > + .flags = ATOMIC_INIT(0), \ > } > > void codetag_lock_module_list(struct codetag_type *cttype); > diff --git a/include/uapi/linux/alloc_tag.h b/include/uapi/linux/alloc_tag.h > index 7d4618bea043..dafcba2f7e2c 100644 > --- a/include/uapi/linux/alloc_tag.h > +++ b/include/uapi/linux/alloc_tag.h > @@ -85,9 +85,20 @@ struct allocinfo_get_at { > struct allocinfo_tag_data data; > }; > > +struct allocinfo_toggle_traces { > + /* inputs */ > + /* > + * Tags matching the filter are toggled; an empty mask selects all tags. > + * Size criteria are checked once, against the tag's size at call time. > + */ > + struct allocinfo_filter filter; > + __u64 enable; > +}; > + > #define _ALLOCINFO_IOC_CONTENT_ID 0 > #define _ALLOCINFO_IOC_GET_AT 1 > #define _ALLOCINFO_IOC_GET_NEXT 2 > +#define _ALLOCINFO_IOC_TOGGLE_TRACE 3 > > #define ALLOCINFO_IOC_BASE 0xA6 > #define ALLOCINFO_IOC_CONTENT_ID _IOR(ALLOCINFO_IOC_BASE, > _ALLOCINFO_IOC_CONTENT_ID, \ > @@ -96,5 +107,7 @@ struct allocinfo_get_at { > struct allocinfo_get_at) > #define ALLOCINFO_IOC_GET_NEXT _IOR(ALLOCINFO_IOC_BASE, > _ALLOCINFO_IOC_GET_NEXT, \ > struct allocinfo_tag_data) > +#define ALLOCINFO_IOC_TOGGLE_TRACE _IOW(ALLOCINFO_IOC_BASE, > _ALLOCINFO_IOC_TOGGLE_TRACE, \ > + struct allocinfo_toggle_traces) > > #endif /* _UAPI_ALLOC_TAG_H */ > diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c > index ca2412a67312..262f00f2bba9 100644 > --- a/mm/alloc_tag.c > +++ b/mm/alloc_tag.c > @@ -60,6 +60,15 @@ DEFINE_STATIC_KEY_FALSE(mem_profiling_compressed); > DEFINE_STATIC_KEY_FALSE(alloc_tag_trace_key); > EXPORT_SYMBOL(alloc_tag_trace_key); > > +static atomic_t alloc_tag_trace_cnt = ATOMIC_INIT(0); > + > +/* > + * As `codetag_lock_module_list` is a read lock, we need an additional mutex > + * to protect against the race conditions involved in the alloc tag trace > + * toggle path. > + */ > +static DEFINE_MUTEX(alloc_tag_trace_mutex); > + > struct alloc_tag_kernel_section kernel_tags = { NULL, 0 }; > unsigned long alloc_tag_ref_mask; > int alloc_tag_ref_offs; > @@ -296,7 +305,7 @@ static bool matches_filter(struct codetag *ct, struct > allocinfo_filter *filter, > return false; > > if (filter->mask & ALLOCINFO_FILTER_MASK_INACCURATE) { > - inaccurate = !!(ct->flags & CODETAG_FLAG_INACCURATE); > + inaccurate = alloc_tag_is_inaccurate(ct_to_alloc_tag(ct)); > if (inaccurate != !!(filter->inaccurate)) > return false; > } > @@ -317,6 +326,23 @@ static bool matches_filter(struct codetag *ct, struct > allocinfo_filter *filter, > return true; > } > > +/* > + * Checks that a user supplied filter only uses known fields and that its > size > + * range, if any, is not inverted. > + */ > +static bool allocinfo_filter_valid(const struct allocinfo_filter *filter) > +{ > + if (filter->mask & ~ALLOCINFO_FILTER_MASKS) > + return false; > + > + if ((filter->mask & ALLOCINFO_FILTER_MASK_MIN_SIZE) && > + (filter->mask & ALLOCINFO_FILTER_MASK_MAX_SIZE) && > + filter->min_size > filter->max_size) > + return false; > + > + return true; > +} > + > /* > * Seeks the ioctl iterator to the specified 0-indexed tag position, reads > its > * profiling data and returns it to userspace. > @@ -333,12 +359,7 @@ static int allocinfo_ioctl_get_at(struct seq_file *m, > void __user *arg) > if (copy_from_user(¶ms, arg, sizeof(params))) > return -EFAULT; > > - if (params.filter.mask & ~ALLOCINFO_FILTER_MASKS) > - return -EINVAL; > - > - if ((params.filter.mask & ALLOCINFO_FILTER_MASK_MIN_SIZE) && > - (params.filter.mask & ALLOCINFO_FILTER_MASK_MAX_SIZE) && > - params.filter.min_size > params.filter.max_size) > + if (!allocinfo_filter_valid(¶ms.filter)) > return -EINVAL; > > priv = m->private; > @@ -443,6 +464,75 @@ static int allocinfo_ioctl_get_next(struct seq_file *m, > void __user *arg) > return ret; > } > > +static void alloc_tag_trace_toggle(struct alloc_tag *tag, bool enable) > +{ > + if (enable) { > + if (alloc_tag_is_traced(tag)) > + return; > + > + alloc_tag_set_traced(tag); > + if (atomic_fetch_inc(&alloc_tag_trace_cnt) == 0) > + static_branch_enable(&alloc_tag_trace_key); > + } else { > + if (!alloc_tag_is_traced(tag)) > + return; > + > + alloc_tag_clear_traced(tag); > + if (atomic_dec_and_test(&alloc_tag_trace_cnt)) > + static_branch_disable(&alloc_tag_trace_key); > + } > +} > + > +/* > + * Toggles tracing on every allocation tag that matches the user supplied > + * filter. An empty filter mask selects all tags, same as for > + * ALLOCINFO_IOC_GET_AT. Tags already in the requested state are not an > error. > + */ > +static int allocinfo_ioctl_toggle_trace(struct seq_file *m, void __user *arg) > +{ > + struct allocinfo_toggle_traces params; > + struct codetag_iterator iter; > + struct codetag *ct; > + struct alloc_tag_counters counters; > + bool fetched_counters; > + int matches = 0, ret; > + > + if (!capable(CAP_SYS_ADMIN)) > + return -EPERM; > + > + if (copy_from_user(¶ms, arg, sizeof(params))) > + return -EFAULT; > + > + if (!allocinfo_filter_valid(¶ms.filter)) > + return -EINVAL; > + > + codetag_lock_module_list(alloc_tag_cttype); > + > + iter = codetag_get_ct_iter(alloc_tag_cttype); > + > + /* Toggle tracing on all codetags that match */ > + while ((ct = codetag_next_ct(&iter))) { > + fetched_counters = false; > + if (matches_filter(ct, ¶ms.filter, &counters, > &fetched_counters)) { > + matches++; > + > + mutex_lock(&alloc_tag_trace_mutex); Could we grab and drop alloc_tag_trace_mutex outside the loop? > + alloc_tag_trace_toggle(ct_to_alloc_tag(ct), > !!params.enable); > + mutex_unlock(&alloc_tag_trace_mutex); > + } > + } > + > + if (matches == 0) > + /* Nothing matched the filter */ > + ret = -ENOENT; > + else > + ret = 0; > + > + codetag_unlock_module_list(alloc_tag_cttype); > + > + return ret; > +} > + > /* > * Entry point ioctl function for /proc/allocinfo routing requests to fetch > the > * layout content ID, seek to a specific tag, or read sequential tags. > @@ -463,6 +553,9 @@ static long allocinfo_ioctl(struct file *file, unsigned > int cmd, > case ALLOCINFO_IOC_GET_NEXT: > ret = allocinfo_ioctl_get_next(file->private_data, arg); > break; > + case ALLOCINFO_IOC_TOGGLE_TRACE: > + ret = allocinfo_ioctl_toggle_trace(file->private_data, arg); > + break; > default: > ret = -ENOIOCTLCMD; > break; > @@ -492,8 +585,6 @@ static const struct proc_ops allocinfo_proc_ops = { > > void __alloc_tag_trace_hit(struct alloc_tag *tag) > { > - if (unlikely(!tag)) > - return; > trace_alloc_tag_hit(tag); > } > EXPORT_SYMBOL(__alloc_tag_trace_hit); > @@ -1031,6 +1122,24 @@ static int load_module(struct module *mod, struct > codetag *start, struct codetag > return 0; > } > > +static void unload_module(struct module *mod, struct codetag *start, struct > codetag *stop) > +{ > + struct alloc_tag *start_tag = ct_to_alloc_tag(start); > + struct alloc_tag *stop_tag = ct_to_alloc_tag(stop); > + struct alloc_tag *tag; > + > + /* > + * Turn tracing off for the tags of the module being unloaded. Without > + * this, `alloc_tag_trace_cnt` would never reach zero and tracing would > + * stay enabled forever. > + * > + * `alloc_tag_trace_mutex` is not needed here as this code path is > + * protected by a `down_write(&cttype->mod_lock)`. > + */ > + for (tag = start_tag; tag < stop_tag; tag++) > + alloc_tag_trace_toggle(tag, false); > +} > + > static void replace_module(struct module *mod, struct module *new_mod) > { > MA_STATE(mas, &mod_area_mt, 0, module_tags.size); > @@ -1357,6 +1466,7 @@ static int __init alloc_tag_init(void) > .alloc_section_mem = reserve_module_tags, > .free_section_mem = release_module_tags, > .module_load = load_module, > + .module_unload = unload_module, > .module_replaced = replace_module, > #endif > };
