bpf: Implement stack_map_get_build_id_offset_sleepable()
Subsystems (4)
Ihor Solodrai
reviewerv3 cover bpf: Implement stack_map_get_build_id_offset_sleepable()
I noticed a following deadlock splat on bpf-next, when running a subset of selftests/bpf:
#526 uprobe_multi_test:OK
#28 bpf_obj_id:OK
[ 27.561465]
[ 27.561603] ======================================================
[ 27.561899] WARNING: possible circular locking dependency detected
[ 27.562193] 7.0.0-rc5-g0eeb0094ba03 #3 Tainted: G OE
[ 27.562523] ------------------------------------------------------
[ 27.562820] uprobe_multi/561 is trying to acquire lock:
[ 27.563071] ff1100010768ba18 (&sb->s_type->i_mutex_key#8){++++}-{4:4}, at: netfs_start_io_read+0x24/0x110
[ 27.563556]
[ 27.563556] but task is already holding lock:
[ 27.563845] ff1100010b12a640 (&mm->mmap_lock){++++}-{4:4}, at: stack_map_get_build_id_offset+0xfa/0x650
[ 27.564296]
[ 27.564296] which lock already depends on the new lock.
[ 27.564296]
[ 27.564696]
[ 27.564696] the existing dependency chain (in reverse order) is:
[ 27.565044]
[ 27.565044] -> #1 (&mm->mmap_lock){++++}-{4:4}:
[ 27.565340] __might_fault+0xa3/0x100
[ 27.565572] _copy_to_iter+0xdf/0x1520
[ 27.565786] copy_page_to_iter+0x1c6/0x2d0
[ 27.566011] filemap_read+0x5f7/0xbf0
[ 27.566218] netfs_file_read_iter+0x1ca/0x240
[ 27.566453] vfs_read+0x482/0x9e0
[ 27.566658] ksys_read+0x115/0x1f0
[ 27.566852] do_syscall_64+0xde/0x390
[ 27.567057] entry_SYSCALL_64_after_hwframe+0x76/0x7e
[ 27.567322]
[ 27.567322] -> #0 (&sb->s_type->i_mutex_key#8){++++}-{4:4}:
[ 27.567701] __lock_acquire+0x1529/0x2a10
[ 27.567923] lock_acquire+0xd7/0x280
[ 27.568124] down_read_interruptible+0x55/0x340
[ 27.568368] netfs_start_io_read+0x24/0x110
[ 27.568622] netfs_file_read_iter+0x191/0x240
[ 27.568858] __kernel_read+0x401/0x850
[ 27.569068] freader_fetch+0x19b/0x790
[ 27.569279] __build_id_parse+0xde/0x580
[ 27.569528] stack_map_get_build_id_offset+0x248/0x650
[ 27.569802] __bpf_get_stack+0x653/0x890
[ 27.570019] bpf_get_stack_sleepable+0x1d/0x30
[ 27.570257] bpf_prog_51bb4c0f31ae471b_uprobe_sleepable+0x2d/0x42
[ 27.570594] uprobe_prog_run+0x425/0x870
[ 27.570816] uprobe_multi_link_handler+0x58/0xb0
[ 27.571062] handler_chain+0x234/0x1270
[ 27.571275] uprobe_notify_resume+0x4de/0xd60
[ 27.571542] exit_to_user_mode_loop+0xe0/0x480
[ 27.571785] exc_int3+0x1bd/0x260
[ 27.571973] asm_exc_int3+0x39/0x40
[ 27.572169]
[ 27.572169] other info that might help us debug this:
[ 27.572169]
[ 27.572571] Possible unsafe locking scenario:
[ 27.572571]
[ 27.572854] CPU0 CPU1
[ 27.573074] ---- ----
[ 27.573293] rlock(&mm->mmap_lock);
[ 27.573502] lock(&sb->s_type->i_mutex_key#8);
[ 27.573846] lock(&mm->mmap_lock);
[ 27.574135] rlock(&sb->s_type->i_mutex_key#8);
[ 27.574364]
[ 27.574364] *** DEADLOCK ***
[ 27.574364]
[ 27.574671] 3 locks held by uprobe_multi/561:
[ 27.574884] #0: ffffffff852e0038 (rcu_tasks_trace_srcu_struct){....}-{0:0}, at: uprobe_notify_resume+0x229/0xd60
[ 27.575371] #1: ffffffff852e0038 (rcu_tasks_trace_srcu_struct){....}-{0:0}, at: uprobe_prog_run+0x239/0x870
[ 27.575874] #2: ff1100010b12a640 (&mm->mmap_lock){++++}-{4:4}, at: stack_map_get_build_id_offset+0xfa/0x650
[ 27.576342]
[ 27.576342] stack backtrace:
[ 27.576578] CPU: 7 UID: 0 PID: 561 Comm: uprobe_multi Tainted: G OE 7.0.0-rc5-g0eeb0094ba03 #3 PREEMPT(full)
[ 27.576586] Tainted: [O]=OOT_MODULE, [E]=UNSIGNED_MODULE
[ 27.576588] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.3-5.el9 11/05/2023
[ 27.576591] Call Trace:
[ 27.576595] <TASK>
[ 27.576598] dump_stack_lvl+0x54/0x70
[ 27.576606] print_circular_bug+0x2e8/0x300
[ 27.576616] check_noncircular+0x12e/0x150
[ 27.576628] __lock_acquire+0x1529/0x2a10
[ 27.576642] ? __pfx_walk_pgd_range+0x10/0x10
[ 27.576651] ? srso_return_thunk+0x5/0x5f
[ 27.576656] ? lock_acquire+0xd7/0x280
[ 27.576661] ? avc_has_perm_noaudit+0x38/0x1f0
[ 27.576672] lock_acquire+0xd7/0x280
[ 27.576677] ? netfs_start_io_read+0x24/0x110
[ 27.576685] ? srso_return_thunk+0x5/0x5f
[ 27.576693] ? netfs_start_io_read+0x24/0x110
[ 27.576698] down_read_interruptible+0x55/0x340
[ 27.576702] ? netfs_start_io_read+0x24/0x110
[ 27.576707] ? __pfx_cmp_ex_search+0x10/0x10
[ 27.576716] netfs_start_io_read+0x24/0x110
[ 27.576723] netfs_file_read_iter+0x191/0x240
[ 27.576731] __kernel_read+0x401/0x850
[ 27.576741] ? __pfx___kernel_read+0x10/0x10
[ 27.576752] ? __pfx_fixup_exception+0x10/0x10
[ 27.576761] ? srso_return_thunk+0x5/0x5f
[ 27.576766] ? lock_is_held_type+0x76/0x100
[ 27.576773] ? srso_return_thunk+0x5/0x5f
[ 27.576777] ? mtree_range_walk+0x421/0x6c0
[ 27.576785] freader_fetch+0x19b/0x790
[ 27.576793] ? mt_find+0x14b/0x340
[ 27.576801] ? __pfx_freader_fetch+0x10/0x10
[ 27.576805] ? mt_find+0x29a/0x340
[ 27.576810] ? mt_find+0x14b/0x340
[ 27.576820] __build_id_parse+0xde/0x580
[ 27.576832] ? __pfx___build_id_parse+0x10/0x10
[ 27.576850] stack_map_get_build_id_offset+0x248/0x650
[ 27.576863] __bpf_get_stack+0x653/0x890
[ 27.576867] ? __pfx_stack_trace_consume_entry+0x10/0x10
[ 27.576880] ? __pfx___bpf_get_stack+0x10/0x10
[ 27.576884] ? lock_acquire+0xd7/0x280
[ 27.576889] ? uprobe_prog_run+0x239/0x870
[ 27.576895] ? srso_return_thunk+0x5/0x5f
[ 27.576899] ? stack_trace_save+0xae/0x100
[ 27.576908] bpf_get_stack_sleepable+0x1d/0x30
[ 27.576913] bpf_prog_51bb4c0f31ae471b_uprobe_sleepable+0x2d/0x42
[ 27.576919] uprobe_prog_run+0x425/0x870
[ 27.576927] ? srso_return_thunk+0x5/0x5f
[ 27.576933] ? __pfx_uprobe_prog_run+0x10/0x10
[ 27.576937] ? uprobe_notify_resume+0x413/0xd60
[ 27.576952] uprobe_multi_link_handler+0x58/0xb0
[ 27.576960] handler_chain+0x234/0x1270
[ 27.576976] ? __pfx_handler_chain+0x10/0x10
[ 27.576988] ? srso_return_thunk+0x5/0x5f
[ 27.576992] ? __kasan_kmalloc+0x72/0x90
[ 27.576998] ? __pfx_ri_timer+0x10/0x10
[ 27.577003] ? timer_init_key+0xf5/0x200
[ 27.577013] uprobe_notify_resume+0x4de/0xd60
[ 27.577025] ? uprobe_notify_resume+0x229/0xd60
[ 27.577030] ? __pfx_uprobe_notify_resume+0x10/0x10
[ 27.577035] ? srso_return_thunk+0x5/0x5f
[ 27.577039] ? atomic_notifier_call_chain+0xfc/0x110
[ 27.577047] ? srso_return_thunk+0x5/0x5f
[ 27.577051] ? notify_die+0xe5/0x130
[ 27.577056] ? __pfx_notify_die+0x10/0x10
[ 27.577063] ? srso_return_thunk+0x5/0x5f
[ 27.577071] exit_to_user_mode_loop+0xe0/0x480
[ 27.577076] ? asm_exc_int3+0x39/0x40
[ 27.577083] exc_int3+0x1bd/0x260
[ 27.577090] asm_exc_int3+0x39/0x40
[ 27.577095] RIP: 0033:0x6c4180
[ 27.577100] Code: c6 05 0b 6f 32 00 01 5d c3 90 c3 66 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 40 00 f3 0f 1e fa eb 8a 66 2e 0f 1f 84 00 00 00 00 00 <cc> 48 89 e5 31 c0 5d c3 0f 1f 84 00 00 00 00 00 55 48 89 e5 31 c0
[ 27.577104] RSP: 002b:00007ffcbf71ada8 EFLAGS: 00000246
[ 27.577109] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 00007fec9163db7b
[ 27.577112] RDX: 00007ffcbf71adbf RSI: 0000000000001000 RDI: 00000000007f0000
[ 27.577115] RBP: 00007ffcbf71add0 R08: 00007fec91731330 R09: 00007fec9178d060
[ 27.577118] R10: 00007fec9153ad98 R11: 0000000000000202 R12: 00007ffcbf71af08
[ 27.577120] R13: 0000000000787760 R14: 00000000009eade8 R15: 00007fec917bf000
[ 27.577135] </TASK>
#54/1 build_id/nofault-paged-out:OKIt was reproducable, which allowed me to bisect to commit 777a8560fd29 ("lib/buildid: use __kernel_read() for sleepable context")
Then I used Codex for analysis of the splat, patch and relevant lore discussions, and then asked it to produce a patch draft, which was a starting point for this series.
I verified that the patch series fixes the deadlock splat.
---
v2->v3:
* Split patch #2 in two: stack_map_get_build_id_offset_sleepable() implementation, and then introduce caching
* Drop taking mmap_lock if CONFIG_PER_VMA_LOCK=n, fall back to raw IPs instead
* Cache vm_{start,end} in addition to prev_file (Mykyta)
v2: https://lore.kernel.org/bpf/[email protected]/
v1->v2:
* Addressed feedback from Puranjay:
* split out a small refactoring patch
* use mmap_read_trylock()
* take into account CONFIG_PER_VMA_LOCK
* replace find_vma() with vma_lookup()
* cache prev_build_id to avoid re-parsing the same file
* Snapshot vm_pgoff and vm_start before unlocking (AI)
* To avoid repetitive unlocking statements, introduce struct stack_map_vma_lock to hold relevant lock state info and add an unlock helper
v1: https://lore.kernel.org/bpf/[email protected]/
---
Ihor Solodrai (3):
bpf: Factor out stack_map_build_id_set_ip() in stackmap.c
bpf: Avoid faultable build ID reads under mm locks
bpf: Cache build IDs in sleepable stackmap path
kernel/bpf/stackmap.c | 171 +++++++++++++++++++++++++++++++++++++++--- 1 file changed, 162 insertions(+), 9 deletions(-)
Signature · 1 line
2.54.0
Ihor Solodrai
reviewerPATCH v3 1/3 bpf: Factor out stack_map_build_id_set_ip() in stackmap.c
Factor out a small helper from stack_map_get_build_id_offset() in preparation for adding a sleepable build ID resolution path.
No functional changes.
--- kernel/bpf/stackmap.c | 21 ++++++++++++--------- 1 file changed, 12 insertions(+), 9 deletions(-)
kernel/bpf/stackmap.c +12 -9
@@ -152,6 +152,12 @@ static int fetch_build_id(struct vm_area_struct *vma, unsigned char *build_id, b : build_id_parse_nofault(vma, build_id, NULL); } +static inline void stack_map_build_id_set_ip (struct bpf_stack_build_id *id) +{ + id->status =BPF_STACK_BUILD_ID_IP ; +memset (id->build_id,0 ,BUILD_ID_SIZE_MAX ); +} +/* * Expects all id_offs[i].ip values to be set to correct initial IPs. * They will be subsequently:
@@ -165,23 +171,21 @@ static int fetch_build_id(struct vm_area_struct *vma, unsigned char *build_id, bstatic void stack_map_get_build_id_offset (struct bpf_stack_build_id *id_offs,u32 trace_nr,bool user,bool may_fault) { -int i;struct mmap_unlock_irq_work *work =NULL ;bool irq_work_busy =bpf_mmap_unlock_get_irq_work (&work); +bool has_user_ctx = user && current && current->mm;struct vm_area_struct *vma, *prev_vma =NULL ;const char *prev_build_id; +int i;/* If the irq_work is in use, fall back to report ips. Same * fallback is used for kernel stack (!user) on a stackmap with * build_id. */ -if (!user || !current || !current->mm || irq_work_busy || - !mmap_read_trylock (current->mm)) { +if (!has_user_ctx || irq_work_busy || !mmap_read_trylock (current->mm)) {/* cannot access current->mm, fall back to ips */ -for (i =0 ; i < trace_nr; i++) { - id_offs[i].status =BPF_STACK_BUILD_ID_IP ; -memset (id_offs[i].build_id,0 ,BUILD_ID_SIZE_MAX ); - } +for (i =0 ; i < trace_nr; i++) +stack_map_build_id_set_ip (&id_offs[i]);return ; }
@@ -196,8 +200,7 @@ static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs, vma =find_vma (current->mm, ip);if (!vma ||fetch_build_id (vma, id_offs[i].build_id, may_fault)) {/* per entry fall back to ips */ - id_offs[i].status =BPF_STACK_BUILD_ID_IP ; -memset (id_offs[i].build_id,0 ,BUILD_ID_SIZE_MAX ); +stack_map_build_id_set_ip (&id_offs[i]);continue ; } build_id_valid:
Signature · 1 line
2.54.0
Ihor Solodrai
reviewerPATCH v3 2/3 bpf: Avoid faultable build ID reads under mm locks
In reply to Ihor SolodraiSleepable build ID parsing can block in __kernel_read() [1], so the stackmap sleepable path must not call it while holding mmap_lock or a per-VMA read lock.
The issue and the fix are conceptually similar to a recent procfs patch [2].
Resolve each covered VMA with a stable read-side reference, preferring lock_vma_under_rcu() and falling back to mmap_read_trylock() only long enough to acquire the VMA read lock. Take a reference to the backing file, drop the VMA lock, and then parse the build ID through (sleepable) build_id_parse_file().
[1]: https://lore.kernel.org/all/[email protected]/
[2]: https://lore.kernel.org/all/[email protected]/
--- kernel/bpf/stackmap.c | 105 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 105 insertions(+)
kernel/bpf/stackmap.c +105 -0
@@ -9,6 +9,7 @@#include <linux/perf_event.h> #include <linux/btf_ids.h> #include <linux/buildid.h> +#include <linux/mmap_lock.h> #include "percpu_freelist.h" #include "mmap_unlock_work.h"
@@ -158,6 +159,105 @@ static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id)memset (id->build_id,0 ,BUILD_ID_SIZE_MAX ); } +struct stack_map_vma_lock { +bool vma_locked; +struct vm_area_struct *vma; +struct mm_struct *mm; +}; + +static struct vm_area_struct * +stack_map_lock_vma (struct stack_map_vma_lock *lock,unsigned long ip) +{ +struct mm_struct *mm = lock->mm; +struct vm_area_struct *vma; + +if (WARN_ON_ONCE (!mm)) +return NULL ; + + vma =lock_vma_under_rcu (mm, ip); +if (vma) + goto vma_locked; + +if (!mmap_read_trylock (mm)) +return NULL ; + + vma =vma_lookup (mm, ip); +if (!vma) { +mmap_read_unlock (mm); +return NULL ; + } + +#ifdef CONFIG_PER_VMA_LOCK +if (!vma_start_read_locked (vma)) { +mmap_read_unlock (mm); +return NULL ; + } +mmap_read_unlock (mm); +#else +mmap_read_unlock (mm); +return NULL ; +#endif +vma_locked: + lock->vma_locked = true; + lock->vma = vma; +return vma; +} + +static void stack_map_unlock_vma (struct stack_map_vma_lock *lock) +{ +struct vm_area_struct *vma = lock->vma; + +if (lock->vma_locked) { +if (WARN_ON_ONCE (!vma)) + goto out; +vma_end_read (vma); + } +out: + lock->vma_locked = false; + lock->vma =NULL ; +} + +static void stack_map_get_build_id_offset_sleepable (struct bpf_stack_build_id *id_offs, +u32 trace_nr) +{ +struct mm_struct *mm = current->mm; +struct stack_map_vma_lock lock = { + .vma_locked = false, + .vma =NULL , + .mm = mm, + }; +unsigned long vm_pgoff, vm_start; +struct vm_area_struct *vma; +struct file *file; +u64 ip; + +for (u32 i =0 ; i < trace_nr; i++) { + ip =READ_ONCE (id_offs[i].ip); + vma =stack_map_lock_vma (&lock, ip); +if (!vma || !vma->vm_file) { +stack_map_build_id_set_ip (&id_offs[i]); +stack_map_unlock_vma (&lock); +continue ; + } + + file =get_file (vma->vm_file); + vm_pgoff = vma->vm_pgoff; + vm_start = vma->vm_start; +stack_map_unlock_vma (&lock); + +/* build_id_parse_file() may block on filesystem reads */ +if (build_id_parse_file (file, id_offs[i].build_id,NULL )) { +stack_map_build_id_set_ip (&id_offs[i]); +fput (file); +continue ; + } +fput (file); + + id_offs[i].offset = (vm_pgoff <<PAGE_SHIFT ) + ip - vm_start; + id_offs[i].status =BPF_STACK_BUILD_ID_VALID ; + } +} +/* * Expects all id_offs[i].ip values to be set to correct initial IPs. * They will be subsequently:
@@ -178,6 +278,11 @@ static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,const char *prev_build_id;int i; +if (may_fault && has_user_ctx) { +stack_map_get_build_id_offset_sleepable (id_offs, trace_nr); +return ; + } +/* If the irq_work is in use, fall back to report ips. Same * fallback is used for kernel stack (!user) on a stackmap with * build_id.
Signature · 1 line
2.54.0
[email protected]
AI reviewRe: [PATCH bpf v3 2/3] bpf: Avoid faultable build ID reads under mm locks
A subsystem pattern flags this as potentially concerning: the code uses !vma->vm_file to classify the VMA as anonymous. Per mm-vma.md's VMA classification rules, the correct test for anonymous VMAs…
Quoted 4 lines
kernel/bpf/stackmap.c +0
-0
[ ... ]
Quoted 70 lines
kernel/bpf/stackmap.c +66
-0@@ -158,6 +159,105 @@ static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id)memset (id->build_id,0 ,BUILD_ID_SIZE_MAX ); } +struct stack_map_vma_lock { +bool vma_locked; +struct vm_area_struct *vma; +struct mm_struct *mm; +}; + +static struct vm_area_struct * +stack_map_lock_vma (struct stack_map_vma_lock *lock,unsigned long ip) +{ +struct mm_struct *mm = lock->mm; +struct vm_area_struct *vma; + +if (WARN_ON_ONCE (!mm)) +return NULL ; + + vma =lock_vma_under_rcu (mm, ip); +if (vma) + goto vma_locked; + +if (!mmap_read_trylock (mm)) +return NULL ; + + vma =vma_lookup (mm, ip); +if (!vma) { +mmap_read_unlock (mm); +return NULL ; + } + +#ifdef CONFIG_PER_VMA_LOCK +if (!vma_start_read_locked (vma)) { +mmap_read_unlock (mm); +return NULL ; + } +mmap_read_unlock (mm); +#else +mmap_read_unlock (mm); +return NULL ; +#endif +vma_locked: + lock->vma_locked = true; + lock->vma = vma; +return vma; +} + +static void stack_map_unlock_vma (struct stack_map_vma_lock *lock) +{ +struct vm_area_struct *vma = lock->vma; + +if (lock->vma_locked) { +if (WARN_ON_ONCE (!vma)) + goto out; +vma_end_read (vma); + } +out: + lock->vma_locked = false; + lock->vma =NULL ; +} + +static void stack_map_get_build_id_offset_sleepable (struct bpf_stack_build_id *id_offs, +u32 trace_nr) +{ +struct mm_struct *mm = current->mm; +struct stack_map_vma_lock lock = { + .vma_locked = false, + .vma =NULL , + .mm = mm,
kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 237
Quoted 10 lines
kernel/bpf/stackmap.c +10
-0+ }; +unsigned long vm_pgoff, vm_start; +struct vm_area_struct *vma; +struct file *file; +u64 ip; + +for (u32 i =0 ; i < trace_nr; i++) { + ip =READ_ONCE (id_offs[i].ip); + vma =stack_map_lock_vma (&lock, ip); +if (!vma || !vma->vm_file) {
^^^^^^^^
A subsystem pattern flags this as potentially concerning: the code uses !vma->vm_file to classify the VMA as anonymous. Per mm-vma.md's VMA classification rules, the correct test for anonymous VMAs is vma_is_anonymous() (i.e. !vma->vm_ops). Some anonymous VMAs carry a vm_file reference (e.g. private mappings of /dev/zero set via vma_set_anonymous() while leaving vm_file populated).
On such VMAs, this code would proceed down the file-backed path and hand /dev/zero (or any other vma_set_anonymous()-style driver file) to build_id_parse_file() instead of immediately falling back to reporting the raw IP. build_id_parse_file() would fail the ELF-magic check and return an error, so the practical effect is a wasted pinned-file reference and a pointless sleepable read attempt. However, the mm subsystem guide directs that this pattern be reported.
The existing non-sleepable stack_map_get_build_id_offset()/build_id_parse() in this same subsystem also uses !vma->vm_file, so the new code is consistent with pre-existing style.
kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 241
Quoted 4 lines
kernel/bpf/stackmap.c +4
-0+stack_map_build_id_set_ip (&id_offs[i]); +stack_map_unlock_vma (&lock); +continue ; + }
Mykyta Yatsenko suggested an optimization in v2 that has not been
implemented: cache vm_start and vm_end, and if the next IP is in range, reuse the previous build ID without locking the VMA. This would avoid repeated lock/unlock cycles for consecutive IPs in the same VMA.
The current v3 code only caches the file reference for same-file optimization, not the VMA range. See the original discussion at
https://lore.kernel.org/bpf/[email protected]/
Quoted 24 lines
kernel/bpf/stackmap.c +19
-0+ + file =get_file (vma->vm_file); + vm_pgoff = vma->vm_pgoff; + vm_start = vma->vm_start; +stack_map_unlock_vma (&lock); + +/* build_id_parse_file() may block on filesystem reads */ +if (build_id_parse_file (file, id_offs[i].build_id,NULL )) { +stack_map_build_id_set_ip (&id_offs[i]); +fput (file); +continue ; + } +fput (file); + + id_offs[i].offset = (vm_pgoff <<PAGE_SHIFT ) + ip - vm_start; + id_offs[i].status =BPF_STACK_BUILD_ID_VALID ; + } +} +/* * Expects all id_offs[i].ip values to be set to correct initial IPs. * They will be subsequently: @@ -178,6 +278,11 @@ static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,const char *prev_build_id;
kernel/bpf/stackmap.c static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs, · line 288
Quoted 10 lines
kernel/bpf/stackmap.c +5
-0int i; +if (may_fault && has_user_ctx) { +stack_map_get_build_id_offset_sleepable (id_offs, trace_nr); +return ; + } +/* If the irq_work is in use, fall back to report ips. Same * fallback is used for kernel stack (!user) on a stackmap with * build_id.
Commit message:
Quoted 14 lines
bpf: Avoid faultable build ID reads under mm locks
Sleepable build ID parsing can block in
__kernel_read()[1], so the stackmap sleepable path must not call it while holdingmmap_lockor a per-VMA read lock.
The issue and the fix are conceptually similar to a recent procfs patch [2].
Resolve each covered VMA with a stable read-side reference, preferringlock_vma_under_rcu()and falling back tommap_read_trylock()only long enough to acquire the VMA read lock. Take a reference to the backing file, drop the VMA lock, and then parse the build ID through (sleepable)build_id_parse_file().
[1]: https://lore.kernel.org/all/[email protected]/
[2]: https://lore.kernel.org/all/[email protected]/
Quoted 1 line
Fixes 777a8560fd29 ("lib/buildid: use __kernel_read() for sleepable context")
The existing Fixes tag points to commit 777a8560fd29 ("lib/buildid: use __kernel_read() for sleepable context"), but the bug was actually introduced by commit d4dd9775ec24 ("bpf: wire up sleepable bpf_get_stack() and bpf_get_task_stack() helpers").
The d4dd9775ec24 commit introduced the pattern of calling potentially sleepable build_id_parse() while holding mm locks via find_vma(). Commit 777a8560fd29 only made the latent blocking behavior more severe by using __kernel_read(), but the architectural issue of calling sleepable functions under locks was introduced earlier.
Should this be:
---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md
CI run summary: https://github.com/kernel-patches/bpf/actions/runs/25711733137
Ihor Solodrai
reviewerRe: [PATCH bpf v3 2/3] bpf: Avoid faultable build ID reads under mm locks
Quoted 104 lines · On 5/11/26 9:16 PM, [email protected] wrote:
Quoted 4 lines
kernel/bpf/stackmap.c +0
-0[ ... ]
Quoted 70 lines
kernel/bpf/stackmap.c +66
-0@@ -158,6 +159,105 @@ static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id)memset (id->build_id,0 ,BUILD_ID_SIZE_MAX ); } +struct stack_map_vma_lock { +bool vma_locked; +struct vm_area_struct *vma; +struct mm_struct *mm; +}; + +static struct vm_area_struct * +stack_map_lock_vma (struct stack_map_vma_lock *lock,unsigned long ip) +{ +struct mm_struct *mm = lock->mm; +struct vm_area_struct *vma; + +if (WARN_ON_ONCE (!mm)) +return NULL ; + + vma =lock_vma_under_rcu (mm, ip); +if (vma) + goto vma_locked; + +if (!mmap_read_trylock (mm)) +return NULL ; + + vma =vma_lookup (mm, ip); +if (!vma) { +mmap_read_unlock (mm); +return NULL ; + } + +#ifdef CONFIG_PER_VMA_LOCK +if (!vma_start_read_locked (vma)) { +mmap_read_unlock (mm); +return NULL ; + } +mmap_read_unlock (mm); +#else +mmap_read_unlock (mm); +return NULL ; +#endif +vma_locked: + lock->vma_locked = true; + lock->vma = vma; +return vma; +} + +static void stack_map_unlock_vma (struct stack_map_vma_lock *lock) +{ +struct vm_area_struct *vma = lock->vma; + +if (lock->vma_locked) { +if (WARN_ON_ONCE (!vma)) + goto out; +vma_end_read (vma); + } +out: + lock->vma_locked = false; + lock->vma =NULL ; +} + +static void stack_map_get_build_id_offset_sleepable (struct bpf_stack_build_id *id_offs, +u32 trace_nr) +{ +struct mm_struct *mm = current->mm; +struct stack_map_vma_lock lock = { + .vma_locked = false, + .vma =NULL , + .mm = mm,kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 237
Quoted 10 lines
kernel/bpf/stackmap.c +10
-0+ }; +unsigned long vm_pgoff, vm_start; +struct vm_area_struct *vma; +struct file *file; +u64 ip; + +for (u32 i =0 ; i < trace_nr; i++) { + ip =READ_ONCE (id_offs[i].ip); + vma =stack_map_lock_vma (&lock, ip); +if (!vma || !vma->vm_file) {^^^^^^^^A subsystem pattern flags this as potentially concerning: the code uses !vma->vm_file to classify the VMA as anonymous. Per mm-vma.md's VMA classification rules, the correct test for anonymous VMAs is
vma_is_anonymous()(i.e. !vma->vm_ops). Some anonymous VMAs carry avm_filereference (e.g. private mappings of/dev/zeroset viavma_set_anonymous()while leavingvm_filepopulated).On such VMAs, this code would proceed down the file-backed path and hand
/dev/zero(or any other vma_set_anonymous()-style driver file) tobuild_id_parse_file()instead of immediately falling back to reporting the raw IP.build_id_parse_file()would fail the ELF-magic check and return an error, so the practical effect is a wasted pinned-file reference and a pointless sleepable read attempt. However, the mm subsystem guide directs that this pattern be reported.
Quoted 8 lines
The existing non-sleepable stack_map_get_build_id_offset()/build_id_parse() in this same subsystem also uses !vma->vm_file, so the new code is consistent with pre-existing style.
kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 241
Quoted 4 lines
kernel/bpf/stackmap.c +4
-0+stack_map_build_id_set_ip (&id_offs[i]); +stack_map_unlock_vma (&lock); +continue ; + }
This is a minor edge case, resulting in unnecessary sleepable read. The fix is simple, but would be inconsistent with the surrounding code.
I'd rather leave it as is for now, and if there is a strong reason to switch to vma_is_anonymous() - then fix it everywhere in stackmap.c
Quoted 9 lines
Mykyta Yatsenko suggested an optimization in v2 that has not been
implemented: cachevm_startandvm_end, and if the next IP is in range, reuse the previous build ID without locking the VMA. This would avoid repeated lock/unlock cycles for consecutive IPs in the same VMA.The current v3 code only caches the file reference for same-file optimization, not the VMA range. See the original discussion at
https://lore.kernel.org/bpf/[email protected]/
The bot failed to notice that the optimization suggested by Mykyta is implemented in the patch #3.
Quoted 65 lines
Quoted 25 lines
kernel/bpf/stackmap.c +19
-0+ + file =get_file (vma->vm_file); + vm_pgoff = vma->vm_pgoff; + vm_start = vma->vm_start; +stack_map_unlock_vma (&lock); + +/* build_id_parse_file() may block on filesystem reads */ +if (build_id_parse_file (file, id_offs[i].build_id,NULL )) { +stack_map_build_id_set_ip (&id_offs[i]); +fput (file); +continue ; + } +fput (file); + + id_offs[i].offset = (vm_pgoff <<PAGE_SHIFT ) + ip - vm_start; + id_offs[i].status =BPF_STACK_BUILD_ID_VALID ; + } +} +/* * Expects all id_offs[i].ip values to be set to correct initial IPs. * They will be subsequently: @@ -178,6 +278,11 @@ static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,const char *prev_build_id;kernel/bpf/stackmap.c static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs, · line 288
Quoted 10 lines
kernel/bpf/stackmap.c +5
-0int i; +if (may_fault && has_user_ctx) { +stack_map_get_build_id_offset_sleepable (id_offs, trace_nr); +return ; + } +/* If the irq_work is in use, fall back to report ips. Same * fallback is used for kernel stack (!user) on a stackmap with * build_id. Commit message:
Quoted 14 lines
bpf: Avoid faultable build ID reads under mm locks
Sleepable build ID parsing can block in
__kernel_read()[1], so the stackmap sleepable path must not call it while holdingmmap_lockor a per-VMA read lock.
The issue and the fix are conceptually similar to a recent procfs patch [2].
Resolve each covered VMA with a stable read-side reference, preferringlock_vma_under_rcu()and falling back tommap_read_trylock()only long enough to acquire the VMA read lock. Take a reference to the backing file, drop the VMA lock, and then parse the build ID through (sleepable)build_id_parse_file().
[1]: https://lore.kernel.org/all/[email protected]/
[2]: https://lore.kernel.org/all/[email protected]/Quoted 1 line
Fixes 777a8560fd29 ("lib/buildid: use __kernel_read() for sleepable context")The existing Fixes tag points to commit
777a8560fd29("lib/buildid: use__kernel_read()for sleepable context"), but the bug was actually introduced by commitd4dd9775ec24("bpf: wire up sleepablebpf_get_stack()andbpf_get_task_stack()helpers").The
d4dd9775ec24commit introduced the pattern of calling potentially sleepablebuild_id_parse()while holding mm locks viafind_vma(). Commit777a8560fd29only made the latent blocking behavior more severe by using__kernel_read(), but the architectural issue of calling sleepable functions under locks was introduced earlier.
Quoted 3 lines
Should this be:
Fixes d4dd9775ec24 ("bpf: wire up sleepable bpf_get_stack() and bpf_get_task_stack() helpers")
This is reasonable. Will change Fixes tag in v4.
Going to wait a bit in case there is more feedback from reviewers.
Quoted 7 lines
---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.mdCI run summary: https://github.com/kernel-patches/bpf/actions/runs/25711733137
[email protected]
AI reviewRe: [PATCH bpf v3 2/3] bpf: Avoid faultable build ID reads under mm locks
In reply to Ihor SolodraiThank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider: - [High] Sleepable build ID parsing unconditionally fails on systems where CONFIG_PER_VMA_LOCK is disabled.…
Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:
- [High] Sleepable build ID parsing unconditionally fails on systems where CONFIG_PER_VMA_LOCK is disabled.
- [High] Incorrect VMA classification using !vma->vm_file violates MM API and risks unsafe driver reads.
- [High] Missing caching causes a severe performance regression due to redundant file I/O for IPs within the same VMA. --
commit 8402e095f9b6a82fa06a32af10007e005d6c5e76
Author: Ihor Solodrai <[email protected]>
bpf: Avoid faultable build ID reads under mm locks
This commit updates the stackmap sleepable path to avoid calling build_id_parse_file() while holding mmap_lock or a per-VMA read lock. It achieves this by acquiring a stable read-side reference to the backing file, dropping the VMA lock, and then parsing the build ID.
Quoted 4 lines
kernel/bpf/stackmap.c +0
-0
[ ... ]
Quoted 13 lines
kernel/bpf/stackmap.c +9
-0@@ -158,6 +159,105 @@ static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id)memset (id->build_id,0 ,BUILD_ID_SIZE_MAX ); } +struct stack_map_vma_lock { +bool vma_locked; +struct vm_area_struct *vma; +struct mm_struct *mm; +}; + +static struct vm_area_struct * +stack_map_lock_vma (struct stack_map_vma_lock *lock,unsigned long ip) +{
[ ... ]
Quoted 6 lines
kernel/bpf/stackmap.c +6
-0+ vma =vma_lookup (mm, ip); +if (!vma) { +mmap_read_unlock (mm); +return NULL ; + } +
kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 199
Quoted 10 lines
kernel/bpf/stackmap.c +10
-0+#ifdef CONFIG_PER_VMA_LOCK +if (!vma_start_read_locked (vma)) { +mmap_read_unlock (mm); +return NULL ; + } +mmap_read_unlock (mm); +#else +mmap_read_unlock (mm); +return NULL ; +#endif
Does this unconditionally fail on systems where CONFIG_PER_VMA_LOCK is disabled?
Looking at the fallback block, it unlocks the mmap_read_lock and returns NULL. Will this cause stack_map_get_build_id_offset_sleepable() to always receive a NULL VMA and fail to extract any build IDs on architectures lacking per-VMA lock support?
[ ... ]
Quoted 5 lines
kernel/bpf/stackmap.c +5
-0+static void stack_map_get_build_id_offset_sleepable (struct bpf_stack_build_id *id_offs, +u32 trace_nr) +{ +struct mm_struct *mm = current->mm; +struct stack_map_vma_lock lock = {
kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 234
Quoted 10 lines
kernel/bpf/stackmap.c +10
-0+ .vma_locked = false, + .vma = NULL, + .mm = mm, + }; +unsigned long vm_pgoff, vm_start; +struct vm_area_struct *vma; +struct file *file; +u64 ip; + +for (u32 i =0 ; i < trace_nr; i++) {
I see this is addressed in a later commit in the series ("bpf: Cache build IDs in sleepable stackmap path"), but is there a risk of a performance regression here?
The original non-sleepable path cached the previous VMA and build ID. Since that caching logic is missing here, will adjacent stack frames from the same binary repeatedly trigger file reference acquisitions and blocking reads for every single IP?
kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 237
Quoted 3 lines
kernel/bpf/stackmap.c +3
-0+ ip =READ_ONCE (id_offs[i].ip); + vma =stack_map_lock_vma (&lock, ip); +if (!vma || !vma->vm_file) {
Is checking vma->vm_file sufficient to determine if the mapping is file-backed?
Anonymous VMAs, such as private mappings of /dev/zero or driver mappings calling vma_set_anonymous(), can legitimately have vm_file set.
Could passing these anonymous VMAs to build_id_parse_file() invoke __kernel_read() on a driver's read handler, leading to unintended side effects or hangs?
Would it be safer to use !vma_is_anonymous(vma) instead?
Quoted 16 lines
kernel/bpf/stackmap.c +16
-0+stack_map_build_id_set_ip (&id_offs[i]); +stack_map_unlock_vma (&lock); +continue ; + } + + file =get_file (vma->vm_file); + vm_pgoff = vma->vm_pgoff; + vm_start = vma->vm_start; +stack_map_unlock_vma (&lock); + +/* build_id_parse_file() may block on filesystem reads */ +if (build_id_parse_file (file, id_offs[i].build_id,NULL )) { +stack_map_build_id_set_ip (&id_offs[i]); +fput (file); +continue ; + }
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2
Ihor Solodrai
reviewerRe: [PATCH bpf v3 2/3] bpf: Avoid faultable build ID reads under mm locks
Quoted 4 lines · On 5/12/26 9:03 PM, [email protected] wrote:
Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:
- [High] Sleepable build ID parsing unconditionally fails on systems whereCONFIG_PER_VMA_LOCKis disabled.
- [High] Incorrect VMA classification using!vma->vm_fileviolates MM API and risks unsafe driver reads.
- [High] Missing caching causes a severe performance regression due to redundant file I/O for IPs within the same VMA.
Sashiko essentially repeated the issues flagged by BPF CI bot, which I already replied to. Same in the second message.
Ignoring these for now.
Quoted 122 lines
--
commit
8402e095f9b6a82fa06a32af10007e005d6c5e76
Author: Ihor Solodrai <[email protected]>
bpf: Avoid faultable build ID reads under mm locksThis commit updates the stackmap sleepable path to avoid calling
build_id_parse_file()while holdingmmap_lockor a per-VMA read lock. It achieves this by acquiring a stable read-side reference to the backing file, dropping the VMA lock, and then parsing the build ID.Quoted 4 lines
kernel/bpf/stackmap.c +0
-0[ ... ]
Quoted 13 lines
kernel/bpf/stackmap.c +9
-0@@ -158,6 +159,105 @@ static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id)memset (id->build_id,0 ,BUILD_ID_SIZE_MAX ); } +struct stack_map_vma_lock { +bool vma_locked; +struct vm_area_struct *vma; +struct mm_struct *mm; +}; + +static struct vm_area_struct * +stack_map_lock_vma (struct stack_map_vma_lock *lock,unsigned long ip) +{[ ... ]
Quoted 6 lines
kernel/bpf/stackmap.c +6
-0+ vma =vma_lookup (mm, ip); +if (!vma) { +mmap_read_unlock (mm); +return NULL ; + } +kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 199
Quoted 10 lines
kernel/bpf/stackmap.c +10
-0+#ifdef CONFIG_PER_VMA_LOCK +if (!vma_start_read_locked (vma)) { +mmap_read_unlock (mm); +return NULL ; + } +mmap_read_unlock (mm); +#else +mmap_read_unlock (mm); +return NULL ; +#endif Does this unconditionally fail on systems where
CONFIG_PER_VMA_LOCKis disabled?Looking at the fallback block, it unlocks the
mmap_read_lockand returns NULL. Will this causestack_map_get_build_id_offset_sleepable()to always receive a NULL VMA and fail to extract any build IDs on architectures lacking per-VMA lock support?[ ... ]
Quoted 5 lines
kernel/bpf/stackmap.c +5
-0+static void stack_map_get_build_id_offset_sleepable (struct bpf_stack_build_id *id_offs, +u32 trace_nr) +{ +struct mm_struct *mm = current->mm; +struct stack_map_vma_lock lock = {kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 234
Quoted 10 lines
kernel/bpf/stackmap.c +10
-0+ .vma_locked = false, + .vma = NULL, + .mm = mm, + }; +unsigned long vm_pgoff, vm_start; +struct vm_area_struct *vma; +struct file *file; +u64 ip; + +for (u32 i =0 ; i < trace_nr; i++) {I see this is addressed in a later commit in the series ("bpf: Cache build IDs in sleepable stackmap path"), but is there a risk of a performance regression here?
The original non-sleepable path cached the previous VMA and build ID. Since that caching logic is missing here, will adjacent stack frames from the same binary repeatedly trigger file reference acquisitions and blocking reads for every single IP?
kernel/bpf/stackmap.c static inline void stack_map_build_id_set_ip(struct bpf_stack_build_id *id) · line 237
Quoted 3 lines
kernel/bpf/stackmap.c +3
-0+ ip =READ_ONCE (id_offs[i].ip); + vma =stack_map_lock_vma (&lock, ip); +if (!vma || !vma->vm_file) {Is checking
vma->vm_filesufficient to determine if the mapping is file-backed?Anonymous VMAs, such as private mappings of
/dev/zeroor driver mappings callingvma_set_anonymous(), can legitimately havevm_fileset.Could passing these anonymous VMAs to
build_id_parse_file()invoke__kernel_read()on a driver's read handler, leading to unintended side effects or hangs?Would it be safer to use !vma_is_anonymous(vma) instead?
Quoted 16 lines
kernel/bpf/stackmap.c +16
-0+stack_map_build_id_set_ip (&id_offs[i]); +stack_map_unlock_vma (&lock); +continue ; + } + + file =get_file (vma->vm_file); + vm_pgoff = vma->vm_pgoff; + vm_start = vma->vm_start; +stack_map_unlock_vma (&lock); + +/* build_id_parse_file() may block on filesystem reads */ +if (build_id_parse_file (file, id_offs[i].build_id,NULL )) { +stack_map_build_id_set_ip (&id_offs[i]); +fput (file); +continue ; + }
Ihor Solodrai
reviewerPATCH v3 3/3 bpf: Cache build IDs in sleepable stackmap path
In reply to Ihor SolodraiStack traces often contain adjacent IPs from the same VMA or from different VMAs backed by the same ELF file. Cache the last successfully parsed build ID together with the resolved VMA range and backing file so the sleepable build-ID path can avoid repeated VMA locking and file parsing in common cases.
--- kernel/bpf/stackmap.c | 51 ++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 48 insertions(+), 3 deletions(-)
kernel/bpf/stackmap.c +48 -3
@@ -226,13 +226,34 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i .vma = NULL, .mm = mm, }; -unsigned long vm_pgoff, vm_start; +struct { +struct file *file; +const char *build_id; +unsigned long vm_start; +unsigned long vm_end; +unsigned long vm_pgoff; + } cache = {}; +unsigned long vm_pgoff, vm_start, vm_end;struct vm_area_struct *vma;struct file *file;u64 ip;for (u32 i =0 ; i < trace_nr; i++) { ip =READ_ONCE (id_offs[i].ip); + +/* +* Range cache fast path: if ip falls within the previously +* resolved VMA range, reuse the cache build_id without +* re-acquiring the VMA lock. +*/ +if (cache.build_id && ip >= cache.vm_start && ip < cache.vm_end) { +memcpy (id_offs[i].build_id, cache.build_id,BUILD_ID_SIZE_MAX ); + vm_start = cache.vm_start; + vm_end = cache.vm_end; + vm_pgoff = cache.vm_pgoff; + goto build_id_valid; + } + vma =stack_map_lock_vma (&lock, ip);if (!vma || !vma->vm_file) {stack_map_build_id_set_ip (&id_offs[i]);
@@ -240,9 +261,22 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *icontinue ; } - file =get_file (vma->vm_file); + file = vma->vm_file; vm_pgoff = vma->vm_pgoff; vm_start = vma->vm_start; + vm_end = vma->vm_end; + +if (file == cache.file) { +/* +* Same backing file as previous (e.g. different VMAs +* of the same ELF binary). Reuse the cache build_id. +*/ +memcpy (id_offs[i].build_id, cache.build_id,BUILD_ID_SIZE_MAX ); +stack_map_unlock_vma (&lock); + goto build_id_valid; + } + + file =get_file (file);stack_map_unlock_vma (&lock);/* build_id_parse_file() may block on filesystem reads */
@@ -251,11 +285,22 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *ifput (file);continue ; } -fput (file); +if (cache.file) +fput (cache.file); + cache.file = file; + cache.build_id = id_offs[i].build_id; + +build_id_valid: + cache.vm_start = vm_start; + cache.vm_end = vm_end; + cache.vm_pgoff = vm_pgoff; id_offs[i].offset = (vm_pgoff <<PAGE_SHIFT ) + ip - vm_start; id_offs[i].status =BPF_STACK_BUILD_ID_VALID ; } + +if (cache.file) +fput (cache.file); }/*
Signature · 1 line
2.54.0
[email protected]
AI reviewRe: [PATCH bpf v3 3/3] bpf: Cache build IDs in sleepable stackmap path
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [High] Incorrect VMA classification using `!vma->vm_file` instead of `vma_is_anonymous()`. --
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] Incorrect VMA classification using !vma->vm_file instead of vma_is_anonymous(). --
commit ce8ba4f759b0570ea69e4efa358a2ec6ddcc2fba
Author: Ihor Solodrai <[email protected]>
bpf: Cache build IDs in sleepable stackmap path
This commit introduces a cache for the last successfully parsed build ID, VMA range, and backing file in the sleepable stackmap path. This avoids repeated VMA locking and file parsing for adjacent IPs.
Quoted 4 lines
kernel/bpf/stackmap.c +0
-0
[ ... ]
Quoted 1 line
kernel/bpf/stackmap.c +0
-0@@ -226,13 +226,34 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i
[ ... ]
Quoted 2 lines
kernel/bpf/stackmap.c +0
-0for (u32 i =0 ; i < trace_nr; i++) { ip =READ_ONCE (id_offs[i].ip);
[ ... ]
kernel/bpf/stackmap.c static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i · line 258
Quoted 2 lines
kernel/bpf/stackmap.c +0
-0vma =stack_map_lock_vma (&lock, ip);if (!vma || !vma->vm_file) {
While this specific check was not introduced by this patch, I notice the code relies on !vma->vm_file to determine if a VMA is file-backed.
Can this cause issues with anonymous VMAs that still have a backing file?
For example, private mappings of /dev/zero or driver memory can have a non-NULL vma->vm_file but are still considered anonymous.
If the stackmap path proceeds to parse the ELF build ID from these files, could it trigger unintended side effects, such as consuming FIFO data or altering hardware state when reading from character devices?
Would it be safer to check vma_is_anonymous(vma) here to properly classify the VMA?
Quoted 4 lines
kernel/bpf/stackmap.c +0
-0stack_map_build_id_set_ip (&id_offs[i]);stack_map_unlock_vma (&lock);continue ; }
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=3
Re: [PATCH bpf v3 3/3] bpf: Cache build IDs in sleepable stackmap path
In reply to Ihor SolodraiQuoted 81 lines · Ihor Solodrai
Stack traces often contain adjacent IPs from the same VMA or from different VMAs backed by the same ELF file. Cache the last successfully parsed build ID together with the resolved VMA range and backing file so the sleepable build-ID path can avoid repeated VMA locking and file parsing in common cases.
Suggested-by Mykyta Yatsenko <[email protected]>Signed-off-by Ihor Solodrai <[email protected]>--- kernel/bpf/stackmap.c | 51 ++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 48 insertions(+), 3 deletions(-)kernel/bpf/stackmap.c +36
-3@@ -226,13 +226,34 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i .vma = NULL, .mm = mm, }; -unsigned long vm_pgoff, vm_start; +struct { +struct file *file; +const char *build_id; +unsigned long vm_start; +unsigned long vm_end; +unsigned long vm_pgoff; + } cache = {}; +unsigned long vm_pgoff, vm_start, vm_end;struct vm_area_struct *vma;struct file *file;u64 ip;for (u32 i =0 ; i < trace_nr; i++) { ip =READ_ONCE (id_offs[i].ip); + +/* +* Range cache fast path: if ip falls within the previously +* resolved VMA range, reuse the cache build_id without +* re-acquiring the VMA lock. +*/ +if (cache.build_id && ip >= cache.vm_start && ip < cache.vm_end) { +memcpy (id_offs[i].build_id, cache.build_id,BUILD_ID_SIZE_MAX ); + vm_start = cache.vm_start; + vm_end = cache.vm_end; + vm_pgoff = cache.vm_pgoff; + goto build_id_valid; + } + vma =stack_map_lock_vma (&lock, ip);if (!vma || !vma->vm_file) {stack_map_build_id_set_ip (&id_offs[i]);@@ -240,9 +261,22 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *icontinue ; } - file =get_file (vma->vm_file); + file = vma->vm_file; vm_pgoff = vma->vm_pgoff; vm_start = vma->vm_start; + vm_end = vma->vm_end; + +if (file == cache.file) { +/* +* Same backing file as previous (e.g. different VMAs +* of the same ELF binary). Reuse the cache build_id. +*/ +memcpy (id_offs[i].build_id, cache.build_id,BUILD_ID_SIZE_MAX ); +stack_map_unlock_vma (&lock); + goto build_id_valid; + } + + file =get_file (file);stack_map_unlock_vma (&lock);/* build_id_parse_file() may block on filesystem reads */ @@ -251,11 +285,22 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *ifput (file);continue ; } -fput (file);
kernel/bpf/stackmap.c static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i · line 297
Quoted 10 lines
kernel/bpf/stackmap.c +9
-0+if (cache.file) +fput (cache.file); + cache.file = file; + cache.build_id = id_offs[i].build_id; + +build_id_valid: + cache.vm_start = vm_start; + cache.vm_end = vm_end; + cache.vm_pgoff = vm_pgoff;
nit: Maybe we can also put
if (cache.build_id != id_offs[i])
memcpy(id_offs[i].build_id, cache.build_id, BUILD_ID_SIZE_MAX);
here? So all writes to the id_offs[i] are at the same place for clearer structure?
Overall it looks correct. Did you try to run it to see if we hit this cache at all?
Quoted 9 lines
kernel/bpf/stackmap.c +3
-0id_offs[i].offset = (vm_pgoff <<PAGE_SHIFT ) + ip - vm_start; id_offs[i].status =BPF_STACK_BUILD_ID_VALID ; } + +if (cache.file) +fput (cache.file); }/*
Ihor Solodrai
reviewerRe: [PATCH bpf v3 3/3] bpf: Cache build IDs in sleepable stackmap path
Quoted 97 lines · Mykyta Yatsenko
Quoted 81 lines · Ihor Solodrai
Stack traces often contain adjacent IPs from the same VMA or from different VMAs backed by the same ELF file. Cache the last successfully parsed build ID together with the resolved VMA range and backing file so the sleepable build-ID path can avoid repeated VMA locking and file parsing in common cases.
Suggested-by Mykyta Yatsenko <[email protected]>Signed-off-by Ihor Solodrai <[email protected]>--- kernel/bpf/stackmap.c | 51 ++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 48 insertions(+), 3 deletions(-)kernel/bpf/stackmap.c +36
-3@@ -226,13 +226,34 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i .vma = NULL, .mm = mm, }; -unsigned long vm_pgoff, vm_start; +struct { +struct file *file; +const char *build_id; +unsigned long vm_start; +unsigned long vm_end; +unsigned long vm_pgoff; + } cache = {}; +unsigned long vm_pgoff, vm_start, vm_end;struct vm_area_struct *vma;struct file *file;u64 ip;for (u32 i =0 ; i < trace_nr; i++) { ip =READ_ONCE (id_offs[i].ip); + +/* +* Range cache fast path: if ip falls within the previously +* resolved VMA range, reuse the cache build_id without +* re-acquiring the VMA lock. +*/ +if (cache.build_id && ip >= cache.vm_start && ip < cache.vm_end) { +memcpy (id_offs[i].build_id, cache.build_id,BUILD_ID_SIZE_MAX ); + vm_start = cache.vm_start; + vm_end = cache.vm_end; + vm_pgoff = cache.vm_pgoff; + goto build_id_valid; + } + vma =stack_map_lock_vma (&lock, ip);if (!vma || !vma->vm_file) {stack_map_build_id_set_ip (&id_offs[i]);@@ -240,9 +261,22 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *icontinue ; } - file =get_file (vma->vm_file); + file = vma->vm_file; vm_pgoff = vma->vm_pgoff; vm_start = vma->vm_start; + vm_end = vma->vm_end; + +if (file == cache.file) { +/* +* Same backing file as previous (e.g. different VMAs +* of the same ELF binary). Reuse the cache build_id. +*/ +memcpy (id_offs[i].build_id, cache.build_id,BUILD_ID_SIZE_MAX ); +stack_map_unlock_vma (&lock); + goto build_id_valid; + } + + file =get_file (file);stack_map_unlock_vma (&lock);/* build_id_parse_file() may block on filesystem reads */ @@ -251,11 +285,22 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *ifput (file);continue ; } -fput (file);kernel/bpf/stackmap.c static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i · line 297
Quoted 10 lines
kernel/bpf/stackmap.c +9
-0+if (cache.file) +fput (cache.file); + cache.file = file; + cache.build_id = id_offs[i].build_id; + +build_id_valid: + cache.vm_start = vm_start; + cache.vm_end = vm_end; + cache.vm_pgoff = vm_pgoff;nit: Maybe we can also put
if (cache.build_id != id_offs[i])
Quoted 2 lines
memcpy(id_offs[i].build_id, cache.build_id, BUILD_ID_SIZE_MAX);here? So all writes to the id_offs[i] are at the same place for clearer structure?
Looks reasonable to me, will do this in v4.
Quoted 2 lines
Overall it looks correct. Did you try to run it to see if we hit this cache at all?
I vibe coded a bit of instrumentation to check. Works on a trivial example. I think real-world deep stacks should have a high cache hit rate here.
kernel/bpf/stackmap.c +39 -0
@@ -10,9 +10,19 @@#include <linux/btf_ids.h> #include <linux/buildid.h> #include <linux/mmap_lock.h> +#include <linux/debugfs.h> +#include <linux/seq_file.h> #include "percpu_freelist.h" #include "mmap_unlock_work.h" +#define DEBUG_BUILD_ID_CACHE + +#ifdef DEBUG_BUILD_ID_CACHE +static atomic64_t bid_cache_range_hits; +static atomic64_t bid_cache_file_hits; +static atomic64_t bid_cache_misses; +#endif +#define STACK_CREATE_FLAG_MASK \ (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY | \ BPF_F_STACK_BUILD_ID)
@@ -247,6 +257,9 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i* re-acquiring the VMA lock. */ if (cache.build_id && ip >= cache.vm_start && ip < cache.vm_end) { +#ifdef DEBUG_BUILD_ID_CACHE +atomic64_inc (&bid_cache_range_hits); +#endif memcpy (id_offs[i].build_id, cache.build_id,BUILD_ID_SIZE_MAX ); vm_start = cache.vm_start; vm_end = cache.vm_end;
@@ -271,6 +284,9 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i* Same backing file as previous (e.g. different VMAs * of the same ELF binary). Reuse the cache build_id. */ +#ifdef DEBUG_BUILD_ID_CACHE +atomic64_inc (&bid_cache_file_hits); +#endif memcpy (id_offs[i].build_id, cache.build_id,BUILD_ID_SIZE_MAX );stack_map_unlock_vma (&lock); goto build_id_valid;
@@ -279,6 +295,9 @@ static void stack_map_get_build_id_offset_sleepable(struct bpf_stack_build_id *i file =get_file (file);stack_map_unlock_vma (&lock); +#ifdef DEBUG_BUILD_ID_CACHE +atomic64_inc (&bid_cache_misses); +#endif /* build_id_parse_file() may block on filesystem reads */ if (build_id_parse_file (file, id_offs[i].build_id,NULL )) {stack_map_build_id_set_ip (&id_offs[i]);
@@ -943,3 +962,23 @@ const struct bpf_map_ops stack_trace_map_ops = { .map_mem_usage = stack_map_mem_usage, .map_btf_id = &stack_trace_map_btf_ids[0], }; + +#ifdef DEBUG_BUILD_ID_CACHE +static int bid_cache_stats_show (struct seq_file *m,void *v) +{ +seq_printf (m,"range_hits %lld\nfile_hits %lld\nmisses %lld\n" , + (long long )atomic64_read (&bid_cache_range_hits), + (long long )atomic64_read (&bid_cache_file_hits), + (long long )atomic64_read (&bid_cache_misses)); +return 0 ; +} +DEFINE_SHOW_ATTRIBUTE (bid_cache_stats); + +static int __init bid_cache_debugfs_init (void ) +{ +debugfs_create_file ("bpf_stackmap_buildid_cache" ,0444 ,NULL ,NULL , + &bid_cache_stats_fops); +return 0 ; +} +late_initcall (bid_cache_debugfs_init); +#endif /* DEBUG_BUILD_ID_CACHE */
tools/testing/selftests/bpf/uprobe_multi.c +19 -0
@@ -95,6 +95,23 @@ static int usdt(void)extern char build_id_start[];extern char build_id_end[]; +/* Deep-recursion driver to give the sleepable build_id cache a stack with +* many adjacent frames in the same VMA. Used by uprobe-deep mode. +*/ +static int __attribute__((noinline))recurse (int depth) +{ +if (depth <=0 ) { + (void )uprobe (); +return 0 ; + } +return recurse (depth -1 ) +1 ; +} + +static int trigger_uprobe_deep (int depth) +{ +return recurse (depth); +} +int __attribute__((weak))trigger_uprobe (bool build_id_resident) {int page_sz =sysconf (_SC_PAGESIZE);
@@ -142,6 +159,8 @@ int main(int argc, char **argv)return trigger_uprobe (false/* page-out build ID */ );if (!strcmp ("uprobe-paged-in" , argv[1 ]))return trigger_uprobe (true/* page-in build ID */ ); +if (!strcmp ("uprobe-deep" , argv[1 ])) +return trigger_uprobe_deep (32 ); error:fprintf (stderr,"usage: %s <bench|usdt>\n" , argv[0 ]);
Quoted 12 lines
Acked-by Mykyta Yatsenko <[email protected]>Quoted 9 lines
kernel/bpf/stackmap.c +3
-0id_offs[i].offset = (vm_pgoff <<PAGE_SHIFT ) + ip - vm_start; id_offs[i].status =BPF_STACK_BUILD_ID_VALID ; } + +if (cache.file) +fput (cache.file); }/*