Messages in this thread | | | Date | Wed, 10 Dec 2014 11:41:52 -0800 | From | Shaohua Li <> | Subject | Re: [PATCH 2/3] X86: add a generic API to let vdso code detect context switch |
| |
On Wed, Dec 10, 2014 at 11:11:27AM -0800, Andy Lutomirski wrote: > On Wed, Dec 10, 2014 at 10:51 AM, Shaohua Li <shli@fb.com> wrote: > > On Wed, Dec 10, 2014 at 10:38:41AM -0800, Andy Lutomirski wrote: > >> On Sun, Dec 7, 2014 at 7:03 PM, Shaohua Li <shli@fb.com> wrote: > >> > vdso code can't disable preempt, so it can be preempted at any time. > >> > This makes a challenge to implement specific features. This patch adds a > >> > generic API to let vdso code detect context switch. > >> > > >> > With this patch, every cpu maintains a context switch count. The upper > >> > bits of the count is the logical cpu id, so the count can't be identical > >> > for any cpu. The low bits of the count will be increased for each > >> > context switch. For a x86_64 cpu with 4096 cpus, the context switch will > >> > be overflowed for 2^(64 - 12) context switch, which is a long time and can be > >> > ignored. The change of the count in giving time can be used to detect if > >> > context switch occurs. > >> > >> Why do you need those high bits? I don't understand how you could > >> possibly confuse one cpu's count with another's unless you fail to > >> make sure that you're reading the same address both times. > >> > >> That being said, I don't like this patch. I'm not sure I have a much > >> better idea, though. More thoughts in the 0/0 email to follow. > > the vdso code doesn't disable preemption, so it can be migrated between > > cpus at any time, the usage is: > > > > get_countext_switch_count (in cpu A) > > do_something (in cpu B) > > get_countext_switch_count (in cpu C) > > > > The cpu A, B, C could be completely different. We want to make sure > > there is no preemption here and we use the context switch count to judge > > this. If the high bits is ignored, the context switch count could be > > identical even A != C, then our judgement using the switch count is > > wrong. > > Sure, but you could compare the cpu numbers, too.
Aha, makes sense.
Thanks, Shaohua
| |