[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <1BF0973C-E2E5-49E3-B3F9-80FF7D6727B2@fb.com>
Date: Wed, 22 Jul 2020 16:49:47 +0000
From: Song Liu <songliubraving@...com>
To: Peter Zijlstra <peterz@...radead.org>
CC: Alexei Starovoitov <alexei.starovoitov@...il.com>,
open list <linux-kernel@...r.kernel.org>,
bpf <bpf@...r.kernel.org>, Networking <netdev@...r.kernel.org>,
Alexei Starovoitov <ast@...nel.org>,
"Daniel Borkmann" <daniel@...earbox.net>,
Kernel Team <Kernel-team@...com>,
"john.fastabend@...il.com" <john.fastabend@...il.com>,
"kpsingh@...omium.org" <kpsingh@...omium.org>,
"brouer@...hat.com" <brouer@...hat.com>
Subject: Re: [PATCH v3 bpf-next 1/2] bpf: separate bpf_get_[stack|stackid] for
perf events BPF
> On Jul 22, 2020, at 8:40 AM, Peter Zijlstra <peterz@...radead.org> wrote:
>
> On Tue, Jul 21, 2020 at 10:40:19PM +0000, Song Liu wrote:
>
>> We only need to block precise_ip >= 2. precise_ip == 1 is OK.
>
> Uuuh, how? Anything PEBS would have the same problem. Sure, precise_ip
> == 1 will not correct the IP, but the stack will not match regardless.
>
> You need IP,SP(,BP) to be a consistent set _AND_ have it match the
> current stack, PEBS simply cannot do that, because the regs get recorded
> (much) earlier than the PMI and the stack can have changed in the
> meantime.
>
By "OK", I meant unwinder will not report error (in my tests). For
accurate stack, we should do the same for precise_ip == 1.
Thanks,
Song
Powered by blists - more mailing lists