Messages in this thread | | | Date | Wed, 28 Sep 2011 16:45:45 +0200 | From | Takashi Iwai <> | Subject | Re: S4 resume broken since 2.6.39 (3.1, too) |
| |
At Wed, 28 Sep 2011 16:45:48 +0200, Rafael J. Wysocki wrote: > > On Wednesday, September 28, 2011, Takashi Iwai wrote: > > At Wed, 28 Sep 2011 15:28:12 +0200, > > Takashi Iwai wrote: > > > > > > At Wed, 28 Sep 2011 15:26:19 +0200, > > > Rafael J. Wysocki wrote: > > > > > > > > On Wednesday, September 28, 2011, Takashi Iwai wrote: > > > > > At Tue, 27 Sep 2011 18:56:44 +0200, > > > > > Rafael J. Wysocki wrote: > > > > > > > > > > > > On Tuesday, September 27, 2011, Takashi Iwai wrote: > > > > > > > At Mon, 26 Sep 2011 19:26:44 -0700, > > > > > > > Yinghai Lu wrote: > > > > > > > > > > > > > > > > On 09/22/2011 11:11 AM, Takashi Iwai wrote: > > > > > > > > > > > > > > > > > At Thu, 22 Sep 2011 07:33:17 -0700, > > > > > > > > > Yinghai Lu wrote: > > > > > > > > >> > > > > > > > > >> On Wed, Sep 21, 2011 at 11:48 AM, Rafael J. Wysocki <rjw@sisk.pl> wrote: > > > > > > > > >>> It looks like init_memory_mapping() is sometimes called with "end" > > > > > > > > >>> beyond the last mapped PFN and it explodes when we try to write stuff to > > > > > > > > >>> that address during image restoration. > > > > > > > > >>> > > > > > > > > >>> IOW, the Yinghai's assumption that init_memory_mapping() would always be > > > > > > > > >>> called with a "good end" on x86_64 was overomptimistic. > > > > > > > > >> > > > > > > > > >> for 64bit x86, kernel_physical_mapping_init() will use > > > > > > > > >> map_low_page()/call early_memmap() to access ram for page_table that is above > > > > > > > > >> rather last mapped PFN. > > > > > > > > >> > > > > > > > > >> the point is: > > > > > > > > >> on system with 64g, usable ram will be [0,2048m), [4g, 64g) > > > > > > > > >> init_memory_mapping will be called two times for them. > > > > > > > > >> before putting page_table high, > > > > > > > > >> page table will be two parts: one is just below 512M, and one below 2048m. > > > > > > > > >> after putting page_table high, > > > > > > > > >> page table will be two parts: one is just below 2048M, and one below 64G. > > > > > > > > > > > > > > > > > > So, how can this change break S4 resume? > > > > > > > > > > > > > > > > > > > > > > > > not sure. > > > > > > > > > > > > > > > > seems resume has it's own page table during transition... > > > > > > > > > > > > > > > > > > > > > > > > > Any hint for further debugging? > > > > > > > > > > > > > > > > > > > > > > > > you may try to insert dead loop in arch/x86/power/hibernate_asm_64.S::restore_image or core_restore_code > > > > > > > > > > > > > > > > to see which part cause reset. > > > > > > > > > > > > > > That's the answer I was afraid of :) > > > > > > > > > > > > I wonder. Does hibernation work on the machine in question with > > > > > > acpi_sleep=nonvs and if so, is the problem reproducible with this > > > > > > setting? > > > > > > > > > > It still shows the same problem. Reboots after a few S4 cycles. > > > > > I'll test kernel S4. > > > > > > > > This most likely means that NVS is not involved. > > > > > > Indeed, kernel S4 also failed both with and without acpi_sleep=nonvs. > > > > > > > Have you tried to revert commit d1ee433 (x86, trampoline: Use the unified > > > > trampoline setup for ACPI wakeup)? > > > > > > Not yet. I'll check it later. > > > > Unfortunately the commit can't be reverted cleanly. > > The commits around it seem mixed up together. > > > > If you have a patch applicable to 3.0 or 3.1 kernel, let me know. > > No, I don't, but can you simply go back to that commit and see > if the issue is reproducible with HEAD = d1ee433, and if it is, > see if it still will be reproducible after going one commit back from there?
Ah, that should be easy. I'll give it a try.
Takashi
| |