| Message ID | 20170912043859.32473-1-bsingharora@gmail.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Revisit MCE handling for UE Errors | expand |
Hi Balbir, Very cool. How are you testing it? Is it failing memory pages and poisoning them out properly? Looks like you have a printk in the machine_check_early path, which you shouldn't. I guess because we don't mark that context as an NMI. Which we could... but I think you want to put as little as possible in that path, so avoiding the print would be preferable. Perhaps you could mark the mce event somehow that the failure can be reported during processing it? Firmware logging is a good question, I could not really see where this all gets plumbed through. If this is expected to be a common problem for some types of attached memory, then we really need to build up a log of these errors that can be used to exclude the memory after a reboot too. Do we have anything like this capability in firmware? Thanks, Nick On Tue, 12 Sep 2017 14:38:55 +1000 Balbir Singh <bsingharora@gmail.com> wrote: > This patch series is designed to hook up memory_failure on > UE errors, this is specially helpful for user_mode UE errors. > > The first patch is a cleanup patch, it removes dead code. > I could not find any users of get_mce_fault_addr(). > The second patch walks kernel/user mode page tables in > real mode to extract the effective address of the instruction > that caused the UE error and the effective address it was > trying to access (for load/store). The third patch hooks > up the pfn for instruction UE errors (ierror). > > The fourth patch hooks up memory_failure to the MCE patch. > > TODO: > Log the address in NVRAM, so that we can recover from > bad pages at boot and keep the blacklist persistent. > > Changelog v2: > - address review comments from Nick and Mahesh > (initialization of pfn and more comments on failure > when addr_to_pfn() or anaylse_instr() fail) > - Hookup ierrors to the framework as well > (comments from Mahesh) > > Balbir Singh (4): > powerpc/mce.c: Remove unused function get_mce_fault_addr() > powerpc/mce: Hookup derror (load/store) UE errors > powerpc/mce: Hookup ierror (instruction) UE errors > powerpc/mce: hookup memory_failure for UE errors > > arch/powerpc/include/asm/mce.h | 4 +- > arch/powerpc/kernel/mce.c | 108 ++++++++++++++++++++++++---------------- > arch/powerpc/kernel/mce_power.c | 105 +++++++++++++++++++++++++++++++++++--- > 3 files changed, 163 insertions(+), 54 deletions(-) >
On Tue, Sep 12, 2017 at 3:03 PM, Nicholas Piggin <npiggin@gmail.com> wrote: > Hi Balbir, > > Very cool. How are you testing it? Is it failing memory pages > and poisoning them out properly? > Yep, I tested it and it seems to work correctly so far. I am testing this on a simulator with injected MCE UE errors for both the data and instruction side. > Looks like you have a printk in the machine_check_early path, > which you shouldn't. I guess because we don't mark that context > as an NMI. Which we could... but I think you want to put as > little as possible in that path, so avoiding the print would > be preferable. Perhaps you could mark the mce event somehow that > the failure can be reported during processing it? > Good point, I did see that printk handles stuff via printk_nmi_enter/exit, but its best avoided. Will spin v2 > Firmware logging is a good question, I could not really see > where this all gets plumbed through. If this is expected to be > a common problem for some types of attached memory, then we > really need to build up a log of these errors that can be used > to exclude the memory after a reboot too. Do we have anything > like this capability in firmware? It's to be built, we should log these to NVRAM and revisit at every boot to isolate these pages Balbir Singh.