Messages in this thread | | | Date | Wed, 10 Aug 2011 14:12:35 +0100 | From | "Jan Beulich" <> | Subject | Re: [PATCH] x86, efi: Don't recursively acquire rtc_lock |
| |
>>> On 10.08.11 at 11:51, Ingo Molnar <mingo@elte.hu> wrote:
> * Jan Beulich <JBeulich@novell.com> wrote: > >> >>> On 10.08.11 at 11:03, Ingo Molnar <mingo@elte.hu> wrote: >> >> > * Jan Beulich <JBeulich@novell.com> wrote: >> > >> >> >>> On 08.08.11 at 15:40, Matt Fleming <matt@console-pimps.org> wrote: >> >> > On Fri, 2011-08-05 at 18:04 +0100, Jan Beulich wrote: >> >> >> >> >> >> Virtual platforms will have to take care of the serialization in the >> >> >> host anyway, so the guest side implementation of getwallclock et al >> >> >> is entirely unaffected. >> >> > >> >> > Ah, OK, that's the important part. I didn't realise that rtc_lock isn't >> >> > actually required by any other code. In which case, yes, it completely >> >> > makes sense to push the locking of rtc_lock down into the >> >> > implementations that actually need it. >> >> > >> >> > It'd be great if I could get some ACK's from the virtualization guys. >> >> > >> >> > --------8<-------- >> >> > >> >> > From a0a39dbb69f6ac675846bf00f30ad153506a4567 Mon Sep 17 00:00:00 2001 >> >> > From: Matt Fleming <matt.fleming@intel.com> >> >> > Date: Mon, 8 Aug 2011 12:59:35 +0100 >> >> > Subject: [PATCH] x86, efi: Don't recursively acquire rtc_lock >> >> > >> >> > A deadlock was introduced on x86 in commit ef68c8f87ed1 ("x86: >> >> > Serialize EFI time accesses on rtc_lock") because efi_get_time() and >> >> > friends can be called with rtc_lock already held by >> >> > read_persistent_time(), e.g. >> >> > >> >> > timekeeping_init() >> >> > read_persistent_clock() <-- acquire rtc_lock >> >> > efi_get_time() >> >> > phys_efi_get_time() <-- acquire rtc_lock <DEADLOCK> >> >> > >> >> > To fix this let's push the locking down into the get_wallclock() and >> >> > set_wallclock() implementations. Only the clock implementations that >> >> > access the x86 RTC directly need to acquire rtc_lock, so it makes >> >> > sense to push the locking down into the rtc, vrtc and efi code. >> >> > >> >> > The virtualization implementations don't require rtc_lock to be held >> >> > because they provide their own serialization. >> >> > >> >> > Signed-off-by: Matt Fleming <matt.fleming@intel.com> >> >> >> >> Acked-by: Jan Beulich <jbeulich@novell.com> >> > >> > How urgent is thix fix - can it wait until Thomas comes back and >> > starts processing patches again? >> >> Without it booting on EFI can't be expected to work (and while I >> was really sure I tested this before submitting, I apparently must >> have used a stale kernel then - I'm really sorry for that), so >> getting it applied soon would be rather desirable. >> >> But then again I don't seem to be able to boot 3.0.1 with both >> patches applied on my only EFI box, so there must be something else >> that broke, and hence maybe I really did test it (on 3.0-rc7) and >> it worked then. Debugging... > > Ok, once you figure it out we can merge all fixes upstream ASAP.
So this is a combination of two recent events: In my patch set, I recently removed efi.c's X86_32 conditional around setting x86_platform.[gs]et_wallclock. Without that removal, the deadlock wouldn't occur (because mach_get_cmos_time() would get called instead) on 64-bits, which is the only thing I run.
With that conditional removed, Matthew's NX consolidation patch gets into the way: For the physical call path it is simply wrong (correct only for the phys_efi_set_virtual_address_map() call, as there the memory map got virtual addresses assigned already) - it switches from early_runtime_code_mapping_set_exec()'s using of md->phys_addr to efi_set_executable() using md->virt_addr. But the latter isn't set at that point yet. And it's unclear to me why this works on 32-bit.
Bottom line - the fix here is needed for 32-bit booting only (and hence isn't - afaic - as critical as I first thought). But the adjustment I'm holding is a necessary one (and I don't even understand why, when 32- and 64-bit EFI code got merged, this was left 32-bit only), and the NX consolidation change needs to be fixed too (basically as a prerequisite).
Jan
| |