Re: [PATCH bpf-next v5 2/3] bpf: Avoid faultable build ID reads under mm locks

[Andrii Nakryiko]

On Thu, May 14, 2026 at 5:53 PM Ihor Solodrai <ihor.solodrai@xxxxxxxxx> wrote:
>
> Sleepable 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().
>
> We have to use mmap_read_trylock() (and give up on failure) in this
> context because taking mmap_read_lock() is generally unsafe on code
> paths reachable from BPF programs [3], and may lead to deadlocks.
>
> [1] https://lore.kernel.org/all/20251218005818.614819-1-shakeel.butt@xxxxxxxxx/
> [2] https://lore.kernel.org/all/20260128183232.2854138-1-andrii@xxxxxxxxxx/
> [3] https://lore.kernel.org/bpf/2895ecd8-df1e-4cc0-b9f9-aef893dc2360@xxxxxxxxx/
>
> Fixes: d4dd9775ec24 ("bpf: wire up sleepable bpf_get_stack() and bpf_get_task_stack() helpers")
> Suggested-by: Puranjay Mohan <puranjay@xxxxxxxxxx>
> Signed-off-by: Ihor Solodrai <ihor.solodrai@xxxxxxxxx>
> ---
>  kernel/bpf/stackmap.c | 109 ++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 109 insertions(+)
>
> diff --git a/kernel/bpf/stackmap.c b/kernel/bpf/stackmap.c
> index 4ef0fd06cea5..08f7659505d1 100644
> --- a/kernel/bpf/stackmap.c
> +++ b/kernel/bpf/stackmap.c
> @@ -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 +159,109 @@ 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;
> +};
> +

tbh, it feels like this should be provided as some sort of primitive
by vma/mm API given how common it becomes when one tries to work with
VMAs efficiently (in terms of avoiding unnecessary mmap lock)... but
that would be a question to Suren maybe

> +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;
> +
> +       /*
> +        * Taking mmap_read_lock() is unsafe here, because the caller
> +        * BPF program might already hold it, causing a deadlock.
> +        */
> +       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;
> +}
> +

cc'ing Suren to help check we didn't miss any of the per-VMA/mmap
locking gotchas

> +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)

why is this only sleepable case? the only difference between sleepable
and non-sleepable is the use of build_id_parse[_file] vs
build_id_parse_nofault to fetch build ID, no? Other than that the
algorithm of locking VMAs is the same, no?

> +{
> +       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;

nit: we can calculate offset here instead of carrying over pgoff and
start, offset formula is cheap, no big deal


> +               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 +282,11 @@ 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.
> --
> 2.54.0
>