On 12/08/2014 03:38 PM, Eric W. Biederman wrote: > Brian King <brking at linux.vnet.ibm.com> writes: > >> On 12/02/2014 12:49 PM, Brian King wrote: >>> On 12/02/2014 12:47 PM, Brian King wrote: >>>> A kexec boot for some ipr SAS adapters was seen to take ~20 seconds >>>> just doing ipr adapter initialization. This is due to the fact that >>>> a kexec invokes the driver's shutdown handler which places the adapter >>>> into a state that requires a hard reset and resulting firmware initialization >>>> to be usable again, which takes significant time. By exporting kexec_in_progress, >>>> this process can be optimized significantly in the driver by essentially >>>> placing the adapter into a state where this hardware reset and re-initialization >>>> can be bypassed, eliminating this delay in kexec boot. >>>> >>>> Cc: Eric Biederman <ebiederm at xmission.com> >>>> Cc: kexec <kexec at lists.infradead.org> >>>> Signed-off-by: Brian King <brking at linux.vnet.ibm.com> >>>> --- >>>> >>>> kernel/kexec.c | 2 ++ >>>> 1 file changed, 2 insertions(+) >>>> >>>> diff -puN kernel/kexec.c~kexec_export_in_prog kernel/kexec.c >>>> --- scsi-queue/kernel/kexec.c~kexec_export_in_prog 2014-12-02 12:13:22.820325731 -0600 >>>> +++ scsi-queue-bjking1/kernel/kexec.c 2014-12-02 12:13:22.825325687 -0600 >>>> @@ -2768,3 +2768,5 @@ int kernel_kexec(void) >>>> mutex_unlock(&kexec_mutex); >>>> return error; >>>> } >>>> + >>>> +EXPORT_SYMBOL_GPL(kexec_in_progress); >>>> _ >>>> >>> >>> Eric, >>> >>> Can I get your ack on this so this can go through the scsi tree? The last time I sent this >>> out you had some issues with the patch description. >> >> Eric - can I get your ack here? > > You can have my nack. > > We really shouldn't need to special case things for kexec here. Why do > you do the full shutdown that is so slow to get out of? Does it really > make reboots more reliable? If you need it for a reboot why don't you > need it for a kexec reboot? You certainly need to stop all dma in > both cases. > > If it really matters to device drivers we need to refactor and make the > callbacks make sense, be maintainable. Right now caring that it is > kexec vs reboot seems broken. > > Now maybe I am wrong here, and there are some good really good reasons > for wanting to do a much stronger shutdown on reboot but at least > skimming through the change longs of the associated patches I am not > seeing the reason for making all of these changes conditional on kexec. > > What I am seeing is driver code getting really complicated possibly > beyond the ability of anyone to sanely test. The difference between an actual reboot and a kexec reboot is that for an actual reboot, the system / platform will end up hard resetting the adapter across the reboot. Due to this impending hard reset, some drivers may want to take additional actions before a full system reboot occurs. In the case of ipr, prior to a full system reset the key thing I want to ensure happens is to tell the adapter to quiesce any outstanding I/O to the disks. I/O that can be occurring at this time is either data being destaged from the non volatile write cache, RAID rebuilds, etc. The only firmware interface available to initiate this quiesce currently places the adapter into a state where the only way back out of it to a state where the host can begin to issue I/O again is to hard reset the adapter, as it is expected to only occur prior to system shutdown or system reboot. This reset and then the firmware initialization that follows can take a bit of time. If this pending I/O is not quiesced prior to a hard reset of the adapter the subsequent boot can then take even longer and, in some cases, may require additional actions to recover from. I can look at implementing this in a driver unique way so that we don't need this kexec patch, but this seemed the most straight forward. Thanks, Brian -- Brian King Power Linux I/O IBM Linux Technology Center
