[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH v11 5/5] migration: introduce 'userfaultfd-wrlat.py' script

From: Andrey Gruzdev
Subject: Re: [PATCH v11 5/5] migration: introduce 'userfaultfd-wrlat.py' script
Date: Thu, 21 Jan 2021 20:15:49 +0300
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101 Thunderbird/68.10.0

On 21.01.2021 18:37, Peter Xu wrote:
On Thu, Jan 21, 2021 at 04:12:23PM +0300, Andrey Gruzdev wrote:
+/* KRETPROBE for handle_userfault(). */
+int retprobe_handle_userfault(struct pt_regs *ctx)
+    u64 pid = (u32) bpf_get_current_pid_tgid();
+    u64 *addr_p;
+    /*
+     * Here we just ignore the return value. In case of spurious wakeup
+     * or pending signal we'll still get (at least for v5.8.0 kernel)
+     * Anyhow, handle_userfault() would be re-entered if such case happens,
+     * keeping initial timestamp unchanged for the faulting thread.
AFAIU this comment is not matching what the code does.  But I agree it's not a
big problem because we won't miss any long delays (because the one long delayed
sample will just be split into two or multiple delays, which will still be
reflected in the histogram at last).  Or am I wrong?
Mm, not really sure about comment.. I need to read kernel code again.
Not relevant to kernel; I was only talking about the last sentence where we
won't "keeping initial timestamp unchanged" but we'll do the statistic anyways.
Because exactly as you said we'll get VM_FAULT_RETRY unconditionally while we
won't be able to identify whether the page fault request is resolved or not.

Yep, agree. My point is also that trying to get timestamp of real PF resolution is
complicated and not very reasonable since a sequence of softbit and hardbit modifications
to the in-memory paging structures occur, not immediately reflected in particular TLB entry.
But for our statistics this level of accuracy is OK, I think.
Andrey Gruzdev, Principal Engineer
Virtuozzo GmbH  +7-903-247-6397

reply via email to

[Prev in Thread] Current Thread [Next in Thread]