Re: [PATCH] Revert "perf: Fix perf_event_validate_size()"

From: Mark Rutland
Date: Fri Dec 15 2023 - 05:48:33 EST


On Thu, Dec 14, 2023 at 03:11:25PM +0000, Mark Rutland wrote:
> On Wed, Dec 13, 2023 at 04:06:20PM -0800, Lucas De Marchi wrote:
> > This reverts commit 382c27f4ed28f803b1f1473ac2d8db0afc795a1b.
> >
> > When using for_each_sibling_event() lockdep complains on
> >
> > WARN_ON_ONCE(__lockdep_enabled && \
> > (this_cpu_read(hardirqs_enabled) && \
> > lockdep_is_held(&(event)->ctx->mutex) != LOCK_STATE_HELD))
> >
> > after a simple `perf stat -e instructions sleep 1`:
> >
> > WARNING: CPU: 8 PID: 1471 at kernel/events/core.c:1950 __do_sys_perf_event_open+0xf37/0x1080
> > Modules linked in: x86_pkg_temp_thermal coretemp kvm_intel kvm e1000e igc irqbypass mei_pxp mei_hdcp crct10dif_pclmul wmi_bmof crc32_pclmul ghash_clmulni_intel i2c_i801 ptp mei_me i2c_smbus pps_core mei intel_lpss_pci video wmi fuse
> > CPU: 8 PID: 1471 Comm: perf Not tainted 6.7.0-rc5-linus-demarchi+ #8
> > Hardware name: Intel Corporation Rocket Lake Client Platform/RocketLake S UDIMM 4L ERB, BIOS RKLSFWI1.R00.2022.A00.2101052151 01/05/2021
> > RIP: 0010:__do_sys_perf_event_open+0xf37/0x1080
> > Code: ff e8 ed 30 ab 00 48 8b 81 28 02 00 00 83 ce ff 48 8d b8 a8 00 00 00 e8 67 38 ab 00 48 8b 4c 24 08 83 e8 01 0f 84 56 f9 ff ff <0f> 0b e9 4f f9 ff ff be 03 00 00 00 e8 d8 74 40 00 e9 54 fa ff ff
> > RSP: 0018:ffffc90001d6be40 EFLAGS: 00010297
> > RAX: 00000000ffffffff RBX: ffff888112bc8040 RCX: ffff888100c34568
> > RDX: 0000000080000000 RSI: ffffffff823e2806 RDI: ffffffff824023cb
> > RBP: 0000000000000000 R08: 0000000000000001 R09: 0000000000000001
> > R10: 0000000000000001 R11: 0000000000000000 R12: ffff888106560800
> > R13: 0000000000000000 R14: 0000000000080002 R15: ffff888100c34568
> > FS: 00007f723d196c00(0000) GS:ffff888450800000(0000) knlGS:0000000000000000
> > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > CR2: 00005624c0efdaa0 CR3: 00000001199de002 CR4: 0000000000770ef0
> > PKRU: 55555554
> > Call Trace:
> > <TASK>
> > ? __do_sys_perf_event_open+0xf37/0x1080
> > ? __warn+0x80/0x170
> > ? __do_sys_perf_event_open+0xf37/0x1080
> > ? report_bug+0x191/0x1c0
> > ? handle_bug+0x3c/0x70
> > ? exc_invalid_op+0x17/0x70
> > ? asm_exc_invalid_op+0x1a/0x20
> > ? __do_sys_perf_event_open+0xf37/0x1080
> > do_syscall_64+0x42/0xf0
> > entry_SYSCALL_64_after_hwframe+0x6e/0x76
> > RIP: 0033:0x7f723bf1ea7d
> > Code: 5b 41 5c c3 66 0f 1f 84 00 00 00 00 00 f3 0f 1e fa 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 83 a3 0f 00 f7 d8 64 89 01 48
> > RSP: 002b:00007ffc7118ba68 EFLAGS: 00000246 ORIG_RAX: 000000000000012a
> > RAX: ffffffffffffffda RBX: 00000000ffffffff RCX: 00007f723bf1ea7d
> > RDX: 00000000ffffffff RSI: 00000000000005c0 RDI: 00005624c0efb0c0
> > RBP: 00007ffc7118bb20 R08: 0000000000000008 R09: 0000000000000000
> > R10: 00000000ffffffff R11: 0000000000000246 R12: 0000000000000000
> > R13: 0000000000000008 R14: 00005624c0efb0b0 R15: 00005624c0efb0b0
> > </TASK>
> > irq event stamp: 45697
> > hardirqs last enabled at (45703): [<ffffffff81173dca>] console_unlock+0x10a/0x160
> > hardirqs last disabled at (45708): [<ffffffff81173daf>] console_unlock+0xef/0x160
> > softirqs last enabled at (45276): [<ffffffff810dab0e>] irq_exit_rcu+0x8e/0xf0
> > softirqs last disabled at (45267): [<ffffffff810dab0e>] irq_exit_rcu+0x8e/0xf0
> >
> > References: https://gitlab.freedesktop.org/drm/intel/-/issues/9847
> > Signed-off-by: Lucas De Marchi <lucas.demarchi@xxxxxxxxx>
> > ---
> >
> > Not sure about a proper fix for this. At first I thought it was indeed
> > because previously it was not validating the the size correctly. However
> > then I noticed the warning is actually due to for_each_sibling_event().
>
> I think what's happening here is that at the point we call
> perf_event_validate_size(), we've found and locked the relevant
> perf_event_context, but haven't set event->ctx yet via
> perf_install_in_context().
>
> Thanks to earlier checks, we know that group_leader->ctx == ctx, so we have
> actually locked the relevant context, but for_each_sibling_event() doesn't have
> access to the context to discover that.
>
> The diff below worked for me, though it's a bit ugly.

Thinking a bit more, it'd be cleaner to just special-case this within
perf_event_validate_size(), as below.

I'll go turn that into a proper patch and send it out shortly.

Mark.
---->8----
diff --git a/kernel/events/core.c b/kernel/events/core.c
index c9d123e13b57..9efd0d7775e7 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -1947,6 +1947,16 @@ static bool perf_event_validate_size(struct perf_event *event)
group_leader->nr_siblings + 1) > 16*1024)
return false;

+ /*
+ * When creating a new group leader, group_leader->ctx is initialized
+ * after the size has been validated, but we cannot safely use
+ * for_each_sibling_event() until group_leader->ctx is set. A new group
+ * leader cannot have any siblings yet, so we can safely skip checking
+ * the non-existent siblings.
+ */
+ if (event == group_leader)
+ return true;
+
for_each_sibling_event(sibling, group_leader) {
if (__perf_event_read_size(sibling->attr.read_format,
group_leader->nr_siblings + 1) > 16*1024)