[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20251006221043.07cdb0fd@gandalf.local.home>
Date: Mon, 6 Oct 2025 22:10:43 -0400
From: Steven Rostedt <rostedt@...dmis.org>
To: Runping Lai <runpinglai@...gle.com>
Cc: Masami Hiramatsu <mhiramat@...nel.org>, Mathieu Desnoyers
<mathieu.desnoyers@...icios.com>, Wattson CI <wattson-external@...gle.com>,
kernel-team@...roid.com, linux-kernel@...r.kernel.org,
linux-trace-kernel@...r.kernel.org, Luo Gengkun
<luogengkun@...weicloud.com>, Linus Torvalds
<torvalds@...ux-foundation.org>
Subject: Re: [PATCH v1] Revert "tracing: Fix tracing_marker may trigger page
fault during preempt_disable"
On Tue, 7 Oct 2025 00:34:17 +0000
Runping Lai <runpinglai@...gle.com> wrote:
> This reverts commit 3d62ab32df065e4a7797204a918f6489ddb8a237.
>
> It's observed on Pixel 6 that this commit causes a severe functional
> regression: all user-space writes to trace_marker now fail. The write
> does not goes through at all. The error is observed in the shell as
> 'printf: write: Bad address'. This breaks a primary ftrace interface
> for user-space debugging and profiling. In kernel trace file, it's
> logged as 'tracing_mark_write: <faulted>'. After reverting this commit,
> functionality is restored.
This is very interesting. The copy is being done in an atomic context. If
the fault had to do anything other than update a page table, it is likely
not to do anything and return a fault.
What preemption model is Pixel 6 running in? CONFIG_PREEMPT_NONE?
The original code is buggy, but if this is causing a regression, then we
likely need to do something else, like copy in a pre-allocated buffer?
-- Steve
>
> Signed-off-by: Runping Lai <runpinglai@...gle.com>
> Reported-by: Wattson CI <wattson-external@...gle.com>
> ---
> kernel/trace/trace.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> index 156e7e0bf559..bb9a6284a629 100644
> --- a/kernel/trace/trace.c
> +++ b/kernel/trace/trace.c
> @@ -7213,7 +7213,7 @@ static ssize_t write_marker_to_buffer(struct trace_array *tr, const char __user
> entry = ring_buffer_event_data(event);
> entry->ip = ip;
>
> - len = copy_from_user_nofault(&entry->buf, ubuf, cnt);
> + len = __copy_from_user_inatomic(&entry->buf, ubuf, cnt);
> if (len) {
> memcpy(&entry->buf, FAULTED_STR, FAULTED_SIZE);
> cnt = FAULTED_SIZE;
> @@ -7310,7 +7310,7 @@ static ssize_t write_raw_marker_to_buffer(struct trace_array *tr,
>
> entry = ring_buffer_event_data(event);
>
> - len = copy_from_user_nofault(&entry->id, ubuf, cnt);
> + len = __copy_from_user_inatomic(&entry->id, ubuf, cnt);
> if (len) {
> entry->id = -1;
> memcpy(&entry->buf, FAULTED_STR, FAULTED_SIZE);
Powered by blists - more mailing lists