Skip to contentSkip to conversation

bpf: sockmap: Remove preempt_disable in sock_map_sk_acquire

PATCH Applied [email protected] 8 participants 8 messages Jul 28, 2023 – Aug 1, 2023

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bpf: sockmap: Remove preempt_disable in sock_map_sk_acquire

From: Tomas Glozar <[email protected]>

Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC allocation later in sk_psock_init_link on PREEMPT_RT kernels, since GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist patchset notes for details).

This causes calling bpf_map_update_elem on BPF_MAP_TYPE_SOCKMAP maps to BUG (sleeping function called from invalid context) on RT kernels.

preempt_disable was introduced together with lock_sk and rcu_read_lock in commit 99ba2b5aba24e ("bpf: sockhash, disallow bpf_tcp_close and update in parallel"), probably to match disabled migration of BPF programs, and is no longer necessary.

Remove preempt_disable to fix BUG in sock_map_update_common on RT.

---
 net/core/sock_map.c | 2 --
 1 file changed, 2 deletions(-)
net/core/sock_map.c +0 -2
index 19538d628714..08ab108206bf 100644
@@ -115,7 +115,6 @@ static void sock_map_sk_acquire(struct sock *sk)
 	__acquires(&sk->sk_lock.slock)
 {
 	lock_sock(sk);
-	preempt_disable();
 	rcu_read_lock();
 }
 
@@ -123,7 +122,6 @@ static void sock_map_sk_release(struct sock *sk)
 	__releases(&sk->sk_lock.slock)
 {
 	rcu_read_unlock();
-	preempt_enable();
 	release_sock(sk);
 }
 
Signature · 1 line
2.39.3
#1 lore raw
Quoted 10 lines · On Fri, Jul 28, 2023 at 08:44:11AM +0200, [email protected] wrote:

From: Tomas Glozar <[email protected]>

Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC allocation later in sk_psock_init_link on PREEMPT_RT kernels, since GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist patchset notes for details).

This causes calling bpf_map_update_elem on BPF_MAP_TYPE_SOCKMAP maps to BUG (sleeping function called from invalid context) on RT kernels.

Quoted 6 lines

preempt_disable was introduced together with lock_sk and rcu_read_lock in commit 99ba2b5aba24e ("bpf: sockhash, disallow bpf_tcp_close and update in parallel"), probably to match disabled migration of BPF programs, and is no longer necessary.

Remove preempt_disable to fix BUG in sock_map_update_common on RT.

FYI, I'm not sure it's related but I started to see following splat recently:

[  189.360689][  T658] =============================
[  189.361149][  T658] [ BUG: Invalid wait context ]
[  189.361588][  T658] 6.5.0-rc2+ #589 Tainted: G           OE     
[  189.362174][  T658] -----------------------------
[  189.362660][  T658] test_progs/658 is trying to lock:
[  189.363176][  T658] ffff8881702652b8 (&psock->link_lock){....}-{3:3}, at: sock_map_update_common+0x1c4/0x340
[  189.364152][  T658] other info that might help us debug this:
[  189.364689][  T658] context-{5:5}
[  189.365021][  T658] 3 locks held by test_progs/658:
[  189.365508][  T658]  #0: ffff888177611a80 (sk_lock-AF_INET){+.+.}-{0:0}, at: sock_map_update_elem_sys+0x82/0x260
[  189.366503][  T658]  #1: ffffffff835a3180 (rcu_read_lock){....}-{1:3}, at: sock_map_update_elem_sys+0x78/0x260
[  189.367470][  T658]  #2: ffff88816cf19240 (&stab->lock){+...}-{2:2}, at: sock_map_update_common+0x12a/0x340
[  189.368420][  T658] stack backtrace:
[  189.368806][  T658] CPU: 0 PID: 658 Comm: test_progs Tainted: G           OE      6.5.0-rc2+ #589 98af30b3c42d747b51da05f1d0e4899e394be6c9
[  189.369889][  T658] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.2-1.fc38 04/01/2014
[  189.370736][  T658] Call Trace:
[  189.371063][  T658]  <TASK>
[  189.371365][  T658]  dump_stack_lvl+0xb2/0x120
[  189.371798][  T658]  __lock_acquire+0x9ad/0x2470
[  189.372243][  T658]  ? lock_acquire+0x104/0x350
[  189.372680][  T658]  lock_acquire+0x104/0x350
[  189.373104][  T658]  ? sock_map_update_common+0x1c4/0x340
[  189.373615][  T658]  ? find_held_lock+0x32/0x90
[  189.374074][  T658]  ? sock_map_update_common+0x12a/0x340
[  189.374587][  T658]  _raw_spin_lock_bh+0x38/0x80
[  189.375060][  T658]  ? sock_map_update_common+0x1c4/0x340
[  189.375571][  T658]  sock_map_update_common+0x1c4/0x340
[  189.376118][  T658]  sock_map_update_elem_sys+0x184/0x260
[  189.376704][  T658]  __sys_bpf+0x181f/0x2840
[  189.377147][  T658]  __x64_sys_bpf+0x1a/0x30
[  189.377556][  T658]  do_syscall_64+0x38/0x90
[  189.377980][  T658]  entry_SYSCALL_64_after_hwframe+0x6e/0xd8
[  189.378473][  T658] RIP: 0033:0x7fe52f47ab5d

the patch did not help with that

jirka

Quoted 30 lines
---
 net/core/sock_map.c | 2 --
 1 file changed, 2 deletions(-)
net/core/sock_map.c +0 -2
index 19538d628714..08ab108206bf 100644
@@ -115,7 +115,6 @@ static void sock_map_sk_acquire(struct sock *sk)
 	__acquires(&sk->sk_lock.slock)
 {
 	lock_sock(sk);
-	preempt_disable();
 	rcu_read_lock();
 }
 
@@ -123,7 +122,6 @@ static void sock_map_sk_release(struct sock *sk)
 	__releases(&sk->sk_lock.slock)
 {
 	rcu_read_unlock();
-	preempt_enable();
 	release_sock(sk);
 }
 
Signature · 1 line
2.39.3
reviewer
#2 lore raw
Quoted 54 lines · Jiri Olsa
Quoted 10 lines · On Fri, Jul 28, 2023 at 08:44:11AM +0200, [email protected] wrote:

From: Tomas Glozar <[email protected]>

Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC allocation later in sk_psock_init_link on PREEMPT_RT kernels, since GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist patchset notes for details).

This causes calling bpf_map_update_elem on BPF_MAP_TYPE_SOCKMAP maps to BUG (sleeping function called from invalid context) on RT kernels.

Quoted 6 lines

preempt_disable was introduced together with lock_sk and rcu_read_lock in commit 99ba2b5aba24e ("bpf: sockhash, disallow bpf_tcp_close and update in parallel"), probably to match disabled migration of BPF programs, and is no longer necessary.

Remove preempt_disable to fix BUG in sock_map_update_common on RT.

FYI, I'm not sure it's related but I started to see following splat recently:

[  189.360689][  T658] =============================
[  189.361149][  T658] [ BUG: Invalid wait context ]
[  189.361588][  T658] 6.5.0-rc2+ #589 Tainted: G           OE
[  189.362174][  T658] -----------------------------
[  189.362660][  T658] test_progs/658 is trying to lock:
[  189.363176][  T658] ffff8881702652b8 (&psock->link_lock){....}-{3:3}, at: sock_map_update_common+0x1c4/0x340
[  189.364152][  T658] other info that might help us debug this:
[  189.364689][  T658] context-{5:5}
[  189.365021][  T658] 3 locks held by test_progs/658:
[  189.365508][  T658]  #0: ffff888177611a80 (sk_lock-AF_INET){+.+.}-{0:0}, at: sock_map_update_elem_sys+0x82/0x260
[  189.366503][  T658]  #1: ffffffff835a3180 (rcu_read_lock){....}-{1:3}, at: sock_map_update_elem_sys+0x78/0x260
[  189.367470][  T658]  #2: ffff88816cf19240 (&stab->lock){+...}-{2:2}, at: sock_map_update_common+0x12a/0x340
[  189.368420][  T658] stack backtrace:
[  189.368806][  T658] CPU: 0 PID: 658 Comm: test_progs Tainted: G           OE      6.5.0-rc2+ #589 98af30b3c42d747b51da05f1d0e4899e394be6c9
[  189.369889][  T658] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.2-1.fc38 04/01/2014
[  189.370736][  T658] Call Trace:
[  189.371063][  T658]  <TASK>
[  189.371365][  T658]  dump_stack_lvl+0xb2/0x120
[  189.371798][  T658]  __lock_acquire+0x9ad/0x2470
[  189.372243][  T658]  ? lock_acquire+0x104/0x350
[  189.372680][  T658]  lock_acquire+0x104/0x350
[  189.373104][  T658]  ? sock_map_update_common+0x1c4/0x340
[  189.373615][  T658]  ? find_held_lock+0x32/0x90
[  189.374074][  T658]  ? sock_map_update_common+0x12a/0x340
[  189.374587][  T658]  _raw_spin_lock_bh+0x38/0x80
[  189.375060][  T658]  ? sock_map_update_common+0x1c4/0x340
[  189.375571][  T658]  sock_map_update_common+0x1c4/0x340
[  189.376118][  T658]  sock_map_update_elem_sys+0x184/0x260
[  189.376704][  T658]  __sys_bpf+0x181f/0x2840
[  189.377147][  T658]  __x64_sys_bpf+0x1a/0x30
[  189.377556][  T658]  do_syscall_64+0x38/0x90
[  189.377980][  T658]  entry_SYSCALL_64_after_hwframe+0x6e/0xd8
[  189.378473][  T658] RIP: 0033:0x7fe52f47ab5d
Quoted 1 line

the patch did not help with that

I think the above splat is not related to this patch. In function sock_map_update_common func we have
raw_spin_lock_bh(&stab->lock);

  sock_map_add_link(psock, link, map, &stab->sks[idx]);
    spin_lock_bh(&psock->link_lock);
    ...
    spin_unlock_bh(&psock->link_lock);

  raw_spin_unlock_bh(&stab->lock);

I think you probably have CONFIG_PROVE_RAW_LOCK_NESTING turned on in your config.

In the above case, for RT kernel, spin_lock_bh will become 'mutex' and it is sleepable, while raw_spin_lock_bh remains to be a spin lock. The warning is about potential locking violation with RT kernel.

To fix the issue, you can convert spin_lock_bh to raw_spin_lock_bh to silence the warning.

Quoted 34 lines

jirka

Quoted 30 lines
---
  net/core/sock_map.c | 2 --
  1 file changed, 2 deletions(-)
net/core/sock_map.c +0 -2
index 19538d628714..08ab108206bf 100644
@@ -115,7 +115,6 @@ static void sock_map_sk_acquire(struct sock *sk)
  	__acquires(&sk->sk_lock.slock)
  {
  	lock_sock(sk);
-	preempt_disable();
  	rcu_read_lock();
  }
 
@@ -123,7 +122,6 @@ static void sock_map_sk_release(struct sock *sk)
  	__releases(&sk->sk_lock.slock)
  {
  	rcu_read_unlock();
-	preempt_enable();
  	release_sock(sk);
  }
 
Signature · 1 line
2.39.3
maintainer Reviewed-by on v1 1/1
#3 lore raw
In reply to [email protected]
Quoted 15 lines · On Fri, Jul 28, 2023 at 08:44 AM +02, [email protected] wrote:

From: Tomas Glozar <[email protected]>

Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC allocation later in sk_psock_init_link on PREEMPT_RT kernels, since GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist patchset notes for details).

This causes calling bpf_map_update_elem on BPF_MAP_TYPE_SOCKMAP maps to BUG (sleeping function called from invalid context) on RT kernels.

preempt_disable was introduced together with lock_sk and rcu_read_lock in commit 99ba2b5aba24e ("bpf: sockhash, disallow bpf_tcp_close and update in parallel"), probably to match disabled migration of BPF programs, and is no longer necessary.

Quoted 4 lines

Remove preempt_disable to fix BUG in sock_map_update_common on RT.

---

We disable softirq and hold a spin lock when modifying the map/hash in sock_{map,hash}_update_common so this LGTM:

You might want some extra tags:

Fixes 99ba2b5aba24 ("bpf: sockhash, disallow bpf_tcp_close and update in parallel")
maintainer Reviewed-by
#4 lore raw
Quoted 17 lines · Jakub Sitnicki
Quoted 15 lines · On Fri, Jul 28, 2023 at 08:44 AM +02, [email protected] wrote:

From: Tomas Glozar <[email protected]>

Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC allocation later in sk_psock_init_link on PREEMPT_RT kernels, since GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist patchset notes for details).

This causes calling bpf_map_update_elem on BPF_MAP_TYPE_SOCKMAP maps to BUG (sleeping function called from invalid context) on RT kernels.

preempt_disable was introduced together with lock_sk and rcu_read_lock in commit 99ba2b5aba24e ("bpf: sockhash, disallow bpf_tcp_close and update in parallel"), probably to match disabled migration of BPF programs, and is no longer necessary.

Quoted 8 lines
Quoted 4 lines

Remove preempt_disable to fix BUG in sock_map_update_common on RT.

---

We disable softirq and hold a spin lock when modifying the map/hash in sock_{map,hash}_update_common so this LGTM:

Agree.

Quoted 1 line
Quoted 5 lines

You might want some extra tags:

Fixes 99ba2b5aba24 ("bpf: sockhash, disallow bpf_tcp_close and update in parallel")
maintainer
#5 lore raw
Quoted 27 lines · John Fastabend
Quoted 17 lines · Jakub Sitnicki
Quoted 15 lines · On Fri, Jul 28, 2023 at 08:44 AM +02, [email protected] wrote:

From: Tomas Glozar <[email protected]>

Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC allocation later in sk_psock_init_link on PREEMPT_RT kernels, since GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist patchset notes for details).

This causes calling bpf_map_update_elem on BPF_MAP_TYPE_SOCKMAP maps to BUG (sleeping function called from invalid context) on RT kernels.

preempt_disable was introduced together with lock_sk and rcu_read_lock in commit 99ba2b5aba24e ("bpf: sockhash, disallow bpf_tcp_close and update in parallel"), probably to match disabled migration of BPF programs, and is no longer necessary.

Quoted 8 lines
Quoted 4 lines

Remove preempt_disable to fix BUG in sock_map_update_common on RT.

---

We disable softirq and hold a spin lock when modifying the map/hash in sock_{map,hash}_update_common so this LGTM:

Quoted 11 lines

Agree.

Quoted 1 line
Quoted 5 lines

You might want some extra tags:

Fixes 99ba2b5aba24 ("bpf: sockhash, disallow bpf_tcp_close and update in parallel")

ENOCOFFEE here (which is never an excuse, but at least today is really true), but I considered you may want this patch via the ebpf tree only after applying it to net.

Hopefully should not be tragic, but please let me know if you prefer the change reverted from net and going via the other path.

Thanks!

Paolo

#6 lore raw
Quoted 41 lines · Paolo Abeni
Quoted 27 lines · John Fastabend
Quoted 17 lines · Jakub Sitnicki
Quoted 15 lines · On Fri, Jul 28, 2023 at 08:44 AM +02, [email protected] wrote:

From: Tomas Glozar <[email protected]>

Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC allocation later in sk_psock_init_link on PREEMPT_RT kernels, since GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist patchset notes for details).

This causes calling bpf_map_update_elem on BPF_MAP_TYPE_SOCKMAP maps to BUG (sleeping function called from invalid context) on RT kernels.

preempt_disable was introduced together with lock_sk and rcu_read_lock in commit 99ba2b5aba24e ("bpf: sockhash, disallow bpf_tcp_close and update in parallel"), probably to match disabled migration of BPF programs, and is no longer necessary.

Quoted 8 lines
Quoted 4 lines

Remove preempt_disable to fix BUG in sock_map_update_common on RT.

---

We disable softirq and hold a spin lock when modifying the map/hash in sock_{map,hash}_update_common so this LGTM:

Quoted 11 lines

Agree.

Quoted 1 line
Quoted 5 lines

You might want some extra tags:

Fixes 99ba2b5aba24 ("bpf: sockhash, disallow bpf_tcp_close and update in parallel")
Quoted 6 lines

ENOCOFFEE here (which is never an excuse, but at least today is really true), but I considered you may want this patch via the ebpf tree only after applying it to net.

Hopefully should not be tragic, but please let me know if you prefer the change reverted from net and going via the other path.

It's fine. Probably it shouldn't conflict with other sockmap fixes.

#7 lore raw
In reply to [email protected]
This patch was applied to netdev/net.git (main) by Paolo Abeni <[email protected]>:

Hello:

This patch was applied to netdev/net.git (main)
by Paolo Abeni <[email protected]>:

Quoted 8 lines · On Fri, 28 Jul 2023 08:44:11 +0200 you wrote:

From: Tomas Glozar <[email protected]>

Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC allocation later in sk_psock_init_link on PREEMPT_RT kernels, since GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist patchset notes for details).

[...]

Here is the summary with links:
- [net] bpf: sockmap: Remove preempt_disable in sock_map_sk_acquire

    https://git.kernel.org/netdev/net/c/13d2618b48f1

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html