Skip to contentSkip to conversation

bpf: avoid holding freeze_mutex during mmap operation

PATCH In review Andrii Nakryiko 4 participants 5 messages Jan 24, 2025 – Jan 28, 2025

Subsystems (2)
lore
Share & tools
lore

Andrii Nakryiko

author maintainer
#0 lore raw

PATCH 1/1 bpf: avoid holding freeze_mutex during mmap operation

We use map->freeze_mutex to prevent races between map_freeze() and memory mapping BPF map contents with writable permissions. The way we naively do this means we'll hold freeze_mutex for entire duration of all the mm and VMA manipulations, which is completely unnecessary. This can potentially also lead to deadlocks, as reported by syzbot in [0].

So, instead, hold freeze_mutex only during writeability checks, bump (proactively) "write active" count for the map, unlock the mutex and proceed with mmap logic. And only if something went wrong during mmap logic, then undo that "write active" counter increment.

Note, instead of checking VM_MAYWRITE we check VM_WRITE before and after mmaping, because we also have a logic that unsets VM_MAYWRITE forcefully, if VM_WRITE is not set. So VM_MAYWRITE could be set early on for read-only mmaping, but it won't be afterwards. VM_WRITE is a consistent way to detect writable mmaping in our implementation.

[0] https://lore.kernel.org/bpf/[email protected]/

Fixes fc9702273e2e ("bpf: Add mmap() support for BPF_MAP_TYPE_ARRAY")
Reported-by [email protected]
---
 kernel/bpf/syscall.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)
kernel/bpf/syscall.c +13 -7
index 0daf098e3207..0d5b39e99770 100644
@@ -1035,7 +1035,7 @@ static const struct vm_operations_struct bpf_map_default_vmops = {
 static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
 {
 	struct bpf_map *map = filp->private_data;
-	int err;
+	int err = 0;
 
 	if (!map->ops->map_mmap || !IS_ERR_OR_NULL(map->record))
 		return -ENOTSUPP;
@@ -1059,7 +1059,12 @@ static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
 			err = -EACCES;
 			goto out;
 		}
+		bpf_map_write_active_inc(map);
 	}
+out:
+	mutex_unlock(&map->freeze_mutex);
+	if (err)
+		return err;
 
 	/* set default open/close callbacks */
 	vma->vm_ops = &bpf_map_default_vmops;
@@ -1070,13 +1075,14 @@ static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
 		vm_flags_clear(vma, VM_MAYWRITE);
 
 	err = map->ops->map_mmap(map, vma);
-	if (err)
-		goto out;
+	if (err) {
+		if (vma->vm_flags & VM_WRITE) {
+			mutex_lock(&map->freeze_mutex);
+			bpf_map_write_active_dec(map);
+			mutex_unlock(&map->freeze_mutex);
+		}
+	}
 
-	if (vma->vm_flags & VM_MAYWRITE)
-		bpf_map_write_active_inc(map);
-out:
-	mutex_unlock(&map->freeze_mutex);
 	return err;
 }
 
Signature · 1 line
2.43.5
Acked-by on v1 1/1
#1 lore raw
Quoted 17 lines · Andrii Nakryiko

We use map->freeze_mutex to prevent races between map_freeze() and memory mapping BPF map contents with writable permissions. The way we naively do this means we'll hold freeze_mutex for entire duration of all the mm and VMA manipulations, which is completely unnecessary. This can potentially also lead to deadlocks, as reported by syzbot in [0].

So, instead, hold freeze_mutex only during writeability checks, bump (proactively) "write active" count for the map, unlock the mutex and proceed with mmap logic. And only if something went wrong during mmap logic, then undo that "write active" counter increment.

Note, instead of checking VM_MAYWRITE we check VM_WRITE before and after mmaping, because we also have a logic that unsets VM_MAYWRITE forcefully, if VM_WRITE is not set. So VM_MAYWRITE could be set early on for read-only mmaping, but it won't be afterwards. VM_WRITE is a consistent way to detect writable mmaping in our implementation.

Quoted 5 lines

[0] https://lore.kernel.org/bpf/[email protected]/

Fixes fc9702273e2e ("bpf: Add mmap() support for BPF_MAP_TYPE_ARRAY")
Reported-by [email protected]
#2 lore raw
In reply to Andrii Nakryiko
Quoted 12 lines · Andrii Nakryiko

We use map->freeze_mutex to prevent races between map_freeze() and memory mapping BPF map contents with writable permissions. The way we naively do this means we'll hold freeze_mutex for entire duration of all the mm and VMA manipulations, which is completely unnecessary. This can potentially also lead to deadlocks, as reported by syzbot in [0].

So, instead, hold freeze_mutex only during writeability checks, bump (proactively) "write active" count for the map, unlock the mutex and proceed with mmap logic. And only if something went wrong during mmap logic, then undo that "write active" counter increment.

Quoted 5 lines

Note, instead of checking VM_MAYWRITE we check VM_WRITE before and after mmaping, because we also have a logic that unsets VM_MAYWRITE forcefully, if VM_WRITE is not set. So VM_MAYWRITE could be set early on for read-only mmaping, but it won't be afterwards. VM_WRITE is a consistent way to detect writable mmaping in our implementation.

bpf_map_mmap_open/bpf_map_mmap_close use VM_MAYWRITE,

Do they need to change as well?

Quoted 36 lines

[0] https://lore.kernel.org/bpf/[email protected]/

Fixes fc9702273e2e ("bpf: Add mmap() support for BPF_MAP_TYPE_ARRAY")
Reported-by [email protected]
---
 kernel/bpf/syscall.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)
kernel/bpf/syscall.c +6 -1
index 0daf098e3207..0d5b39e99770 100644
@@ -1035,7 +1035,7 @@ static const struct vm_operations_struct bpf_map_default_vmops = {
 static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
 {
        struct bpf_map *map = filp->private_data;
-       int err;
+       int err = 0;
 
        if (!map->ops->map_mmap || !IS_ERR_OR_NULL(map->record))
                return -ENOTSUPP;
@@ -1059,7 +1059,12 @@ static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
                        err = -EACCES;
                        goto out;
                }
+               bpf_map_write_active_inc(map);
        }
+out:
+       mutex_unlock(&map->freeze_mutex);
+       if (err)
+               return err;
 
        /* set default open/close callbacks */
        vma->vm_ops = &bpf_map_default_vmops;

kernel/bpf/syscall.c static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma) · line 1082

Quoted 10 lines
kernel/bpf/syscall.c +5 -2
@@ -1070,13 +1075,14 @@ static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
                vm_flags_clear(vma, VM_MAYWRITE);
 
        err = map->ops->map_mmap(map, vma);
-       if (err)
-               goto out;
+       if (err) {
+               if (vma->vm_flags & VM_WRITE) {
+                       mutex_lock(&map->freeze_mutex);
+                       bpf_map_write_active_dec(map);
+                       mutex_unlock(&map->freeze_mutex);

Extra lock/unlock looks unnecessary.

This functiona and map_freeze() need to see frozen and write_active coherent, but write_active_dec looks like without mutex.
It's atomic64_dec.

#3 lore raw
Quoted 20 lines · Alexei Starovoitov
Quoted 12 lines · Andrii Nakryiko

We use map->freeze_mutex to prevent races between map_freeze() and memory mapping BPF map contents with writable permissions. The way we naively do this means we'll hold freeze_mutex for entire duration of all the mm and VMA manipulations, which is completely unnecessary. This can potentially also lead to deadlocks, as reported by syzbot in [0].

So, instead, hold freeze_mutex only during writeability checks, bump (proactively) "write active" count for the map, unlock the mutex and proceed with mmap logic. And only if something went wrong during mmap logic, then undo that "write active" counter increment.

Quoted 5 lines

Note, instead of checking VM_MAYWRITE we check VM_WRITE before and after mmaping, because we also have a logic that unsets VM_MAYWRITE forcefully, if VM_WRITE is not set. So VM_MAYWRITE could be set early on for read-only mmaping, but it won't be afterwards. VM_WRITE is a consistent way to detect writable mmaping in our implementation.

Quoted 3 lines

bpf_map_mmap_open/bpf_map_mmap_open use VM_MAYWRITE,

Do they need to change as well?

So I didn't want to elaborate too much on this (because of verboseness), but it is indeed non-obvious (I was confused by this for a bit while working on the patch).

We have this piece of logic in the middle of bpf_map_mmap():

if (!(vma->vm_flags & VM_WRITE))

    vm_flags_clear(vma, VM_MAYWRITE);

After this point, VM_WRITE and VM_MAYWRITE are equivalent (when it comes to BPF maps mmap-ing). I.e., if we have writable mapping, we'll have both VM_WRITE and VM_MAYWRITE; if we have read-only mapping, we won't have either. We can't have any other mix of those two.

bpf_map_write_active_inc() used to happen after this point, and so we were checking VM_MAYWRITE, but I had to move bpf_map_write_active_inc() before that point, so I switched to VM_WRITE check.

bpf_map_mmap_open/bpf_map_mmap_close happen after this vm_flags_clear(vma, VM_MAYWRITE), so whether they use VM_MAYWRITE or VM_WRITE doesn't matter. So they should be fine as is.

It is confusing, though, I agree. So maybe we should just normalize all the checks to VM_WRITE and leave a comment that MAYWRITE and WRITE are coupled with our custom mmaping logic?

Quoted 50 lines
Quoted 37 lines

[0] https://lore.kernel.org/bpf/[email protected]/

Fixes fc9702273e2e ("bpf: Add mmap() support for BPF_MAP_TYPE_ARRAY")
Reported-by [email protected]
---
 kernel/bpf/syscall.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)
kernel/bpf/syscall.c +6 -1
index 0daf098e3207..0d5b39e99770 100644
@@ -1035,7 +1035,7 @@ static const struct vm_operations_struct bpf_map_default_vmops = {
 static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
 {
        struct bpf_map *map = filp->private_data;
-       int err;
+       int err = 0;
 
        if (!map->ops->map_mmap || !IS_ERR_OR_NULL(map->record))
                return -ENOTSUPP;
@@ -1059,7 +1059,12 @@ static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
                        err = -EACCES;
                        goto out;
                }
+               bpf_map_write_active_inc(map);
        }
+out:
+       mutex_unlock(&map->freeze_mutex);
+       if (err)
+               return err;
 
        /* set default open/close callbacks */
        vma->vm_ops = &bpf_map_default_vmops;

kernel/bpf/syscall.c static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma) · line 1082

Quoted 10 lines
kernel/bpf/syscall.c +5 -2
@@ -1070,13 +1075,14 @@ static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
                vm_flags_clear(vma, VM_MAYWRITE);
 
        err = map->ops->map_mmap(map, vma);
-       if (err)
-               goto out;
+       if (err) {
+               if (vma->vm_flags & VM_WRITE) {
+                       mutex_lock(&map->freeze_mutex);
+                       bpf_map_write_active_dec(map);
+                       mutex_unlock(&map->freeze_mutex);

Extra lock/unlock looks unnecessary.

Quoted 3 lines

This functiona and map_freeze() need to see frozen and write_active coherent, but write_active_dec looks like without mutex.
It's atomic64_dec.

Yep, I think you are right. I wanted a no-brainer change and not having to think about any memory ordering effects or anything like that. But seeing bpf_map_is_rdonly() checks this without any lock anyways, I think we should be fine. I can drop this lock/unlock for v2.

#4 lore raw
Quoted 41 lines · Andrii Nakryiko
Quoted 20 lines · Alexei Starovoitov
Quoted 12 lines · Andrii Nakryiko

We use map->freeze_mutex to prevent races between map_freeze() and memory mapping BPF map contents with writable permissions. The way we naively do this means we'll hold freeze_mutex for entire duration of all the mm and VMA manipulations, which is completely unnecessary. This can potentially also lead to deadlocks, as reported by syzbot in [0].

So, instead, hold freeze_mutex only during writeability checks, bump (proactively) "write active" count for the map, unlock the mutex and proceed with mmap logic. And only if something went wrong during mmap logic, then undo that "write active" counter increment.

Quoted 5 lines

Note, instead of checking VM_MAYWRITE we check VM_WRITE before and after mmaping, because we also have a logic that unsets VM_MAYWRITE forcefully, if VM_WRITE is not set. So VM_MAYWRITE could be set early on for read-only mmaping, but it won't be afterwards. VM_WRITE is a consistent way to detect writable mmaping in our implementation.

Quoted 3 lines

bpf_map_mmap_open/bpf_map_mmap_open use VM_MAYWRITE,

Do they need to change as well?

So I didn't want to elaborate too much on this (because of verboseness), but it is indeed non-obvious (I was confused by this for a bit while working on the patch).

We have this piece of logic in the middle of bpf_map_mmap():

if (!(vma->vm_flags & VM_WRITE))

    vm_flags_clear(vma, VM_MAYWRITE);

After this point, VM_WRITE and VM_MAYWRITE are equivalent (when it comes to BPF maps mmap-ing). I.e., if we have writable mapping, we'll have both VM_WRITE and VM_MAYWRITE; if we have read-only mapping, we won't have either. We can't have any other mix of those two.

Quoted 8 lines

bpf_map_write_active_inc() used to happen after this point, and so we were checking VM_MAYWRITE, but I had to move bpf_map_write_active_inc() before that point, so I switched to VM_WRITE check.

bpf_map_mmap_open/bpf_map_mmap_close happen after this vm_flags_clear(vma, VM_MAYWRITE), so whether they use VM_MAYWRITE or VM_WRITE doesn't matter. So they should be fine as is.

I see. Yeah. I think this analysis is correct.
It seems to be fine as-is.

Quoted 3 lines

It is confusing, though, I agree. So maybe we should just normalize all the checks to VM_WRITE and leave a comment that MAYWRITE and WRITE are coupled with our custom mmaping logic?

Yeah. That would be nice.

Quoted 54 lines
Quoted 50 lines
Quoted 37 lines

[0] https://lore.kernel.org/bpf/[email protected]/

Fixes fc9702273e2e ("bpf: Add mmap() support for BPF_MAP_TYPE_ARRAY")
Reported-by [email protected]
---
 kernel/bpf/syscall.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)
kernel/bpf/syscall.c +6 -1
index 0daf098e3207..0d5b39e99770 100644
@@ -1035,7 +1035,7 @@ static const struct vm_operations_struct bpf_map_default_vmops = {
 static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
 {
        struct bpf_map *map = filp->private_data;
-       int err;
+       int err = 0;
 
        if (!map->ops->map_mmap || !IS_ERR_OR_NULL(map->record))
                return -ENOTSUPP;
@@ -1059,7 +1059,12 @@ static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
                        err = -EACCES;
                        goto out;
                }
+               bpf_map_write_active_inc(map);
        }
+out:
+       mutex_unlock(&map->freeze_mutex);
+       if (err)
+               return err;
 
        /* set default open/close callbacks */
        vma->vm_ops = &bpf_map_default_vmops;

kernel/bpf/syscall.c static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma) · line 1082

Quoted 10 lines
kernel/bpf/syscall.c +5 -2
@@ -1070,13 +1075,14 @@ static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
                vm_flags_clear(vma, VM_MAYWRITE);
 
        err = map->ops->map_mmap(map, vma);
-       if (err)
-               goto out;
+       if (err) {
+               if (vma->vm_flags & VM_WRITE) {
+                       mutex_lock(&map->freeze_mutex);
+                       bpf_map_write_active_dec(map);
+                       mutex_unlock(&map->freeze_mutex);

Extra lock/unlock looks unnecessary.

Quoted 3 lines

This functiona and map_freeze() need to see frozen and write_active coherent, but write_active_dec looks like without mutex.
It's atomic64_dec.

Quoted 5 lines

Yep, I think you are right. I wanted a no-brainer change and not having to think about any memory ordering effects or anything like that. But seeing bpf_map_is_rdonly() checks this without any lock anyways, I think we should be fine. I can drop this lock/unlock for v2.

Not only that. map_update/delete do it as well. So extra mutex_lock provokes questions like mine :) So pls remove.