[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <11ae956a-9d0c-ca80-c3a7-a16b2c53e737@huawei.com>
Date: Thu, 20 Jun 2024 19:27:16 +0800
From: "Liao, Chang" <liaochang1@...wei.com>
To: Jiri Olsa <olsajiri@...il.com>, Oleg Nesterov <oleg@...hat.com>
CC: <rostedt@...dmis.org>, <mhiramat@...nel.org>, <ast@...nel.org>,
<daniel@...earbox.net>, <andrii@...nel.org>, <nathan@...nel.org>,
<peterz@...radead.org>, <mingo@...hat.com>, <mark.rutland@....com>,
<linux-perf-users@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
<bpf@...r.kernel.org>
Subject: Re: [PATCH bpf-next] uprobes: Fix the xol slots reserved for
uretprobe trampoline
在 2024/6/20 17:06, Jiri Olsa 写道:
> On Thu, Jun 20, 2024 at 10:36:02AM +0200, Oleg Nesterov wrote:
>> On 06/20, Liao, Chang wrote:
>>>
>>> However, when i asm porting uretprobe trampoline to arm64
>>> to explore its benefits on that architecture, i discovered the problem that
>>> single slot is not large enought for trampoline code.
>
> ah ok, makes sense now.. x86_64 has the slot big enough for the trampoline,
> but arm64 does not
>
>>
>> Ah, but then I'd suggest to make the changelog more clear. It looks as
>> if the problem was introduced by the patch from Jiri. Note that we was
>> confused as well ;)
>>
>> And,
>>
>> + /* Reserve enough slots for the uretprobe trampoline */
>> + for (slot_nr = 0;
>> + slot_nr < max((insns_size / UPROBE_XOL_SLOT_BYTES), 1);
>> + slot_nr++)
>>
>> this doesn't look right. Just suppose that insns_size = UPROBE_XOL_SLOT_BYTES + 1.
>> I'd suggest DIV_ROUND_UP(insns_size, UPROBE_XOL_SLOT_BYTES).
>>
>> And perhaps it would be better to send this change along with
>> uretprobe_trampoline_for_arm64 ?
>
> +1, also I'm curious what's the gain on arm64?
I am currently finalizing the uretprobe trampoline and syscall implementation on arm64.
While i have addressed most of issues, there are stiil a few bugs that reguire more effort.
Once these are fixed, i will use Redis to evaluate the performance gains on arm64. In the
next revision, i will submit a patchset that includes all relevant code changs, testcases
and benchmark data, which will allows a comprehensive review and dicussion.
>
> thanks,
> jirka
--
BR
Liao, Chang
Powered by blists - more mailing lists