Re: [PATCH net-next v4 0/3] kernel: add support to collect hardware logs in crash recovery kernel

[Rahul Lakkireddy]

On Friday, April 04/20/18, 2018 at 19:06:09 +0530, Eric W. Biederman wrote:
> Rahul Lakkireddy <rahul.lakkireddy@xxxxxxxxxxx> writes:
> 
> > On Thursday, April 04/19/18, 2018 at 20:23:37 +0530, Eric W. Biederman wrote:
> >> Rahul Lakkireddy <rahul.lakkireddy@xxxxxxxxxxx> writes:
> >> 
> >> > On Thursday, April 04/19/18, 2018 at 07:10:30 +0530, Dave Young wrote:
> >> >> On 04/18/18 at 06:01pm, Rahul Lakkireddy wrote:
> >> >> > On Wednesday, April 04/18/18, 2018 at 11:45:46 +0530, Dave Young wrote:
> >> >> > > Hi Rahul,
> >> >> > > On 04/17/18 at 01:14pm, Rahul Lakkireddy wrote:
> >> >> > > > On production servers running variety of workloads over time, kernel
> >> >> > > > panic can happen sporadically after days or even months. It is
> >> >> > > > important to collect as much debug logs as possible to root cause
> >> >> > > > and fix the problem, that may not be easy to reproduce. Snapshot of
> >> >> > > > underlying hardware/firmware state (like register dump, firmware
> >> >> > > > logs, adapter memory, etc.), at the time of kernel panic will be very
> >> >> > > > helpful while debugging the culprit device driver.
> >> >> > > > 
> >> >> > > > This series of patches add new generic framework that enable device
> >> >> > > > drivers to collect device specific snapshot of the hardware/firmware
> >> >> > > > state of the underlying device in the crash recovery kernel. In crash
> >> >> > > > recovery kernel, the collected logs are added as elf notes to
> >> >> > > > /proc/vmcore, which is copied by user space scripts for post-analysis.
> >> >> > > > 
> >> >> > > > The sequence of actions done by device drivers to append their device
> >> >> > > > specific hardware/firmware logs to /proc/vmcore are as follows:
> >> >> > > > 
> >> >> > > > 1. During probe (before hardware is initialized), device drivers
> >> >> > > > register to the vmcore module (via vmcore_add_device_dump()), with
> >> >> > > > callback function, along with buffer size and log name needed for
> >> >> > > > firmware/hardware log collection.
> >> >> > > 
> >> >> > > I assumed the elf notes info should be prepared while kexec_[file_]load
> >> >> > > phase. But I did not read the old comment, not sure if it has been discussed
> >> >> > > or not.
> >> >> > > 
> >> >> > 
> >> >> > We must not collect dumps in crashing kernel. Adding more things in
> >> >> > crash dump path risks not collecting vmcore at all. Eric had
> >> >> > discussed this in more detail at:
> >> >> > 
> >> >> > https://lkml.org/lkml/2018/3/24/319
> >> >> > 
> >> >> > We are safe to collect dumps in the second kernel. Each device dump
> >> >> > will be exported as an elf note in /proc/vmcore.
> >> >> 
> >> >> I understand that we should avoid adding anything in crash path.  And I also
> >> >> agree to collect device dump in second kernel.  I just assumed device
> >> >> dump use some memory area to store the debug info and the memory
> >> >> is persistent so that this can be done in 2 steps, first register the
> >> >> address in elf header in kexec_load, then collect the dump in 2nd
> >> >> kernel.  But it seems the driver is doing some other logic to collect
> >> >> the info instead of just that simple like I thought. 
> >> >> 
> >> >
> >> > It seems simpler, but I'm concerned with waste of memory area, if
> >> > there are no device dumps being collected in second kernel. In
> >> > approach proposed in these series, we dynamically allocate memory
> >> > for the device dumps from second kernel's available memory.
> >> 
> >> Don't count that kernel having more than about 128MiB.
> >> 
> >
> > If large dump is expected, Administrator can increase the memory
> > allocated to the second kernel (using crashkernel boot param), to
> > ensure device dumps get collected.
> 
> Except 128MiB is already a already a huge amount to reserve.  I
> typically have run crash dumps with 16MiB of memory and thought it was
> overkill.  Looking below 32MiB seems a bit high but it is small enough
> that it is still doable.  I am baffled at how 2GiB can be guaranteed to fit
> in 32MiB (sparse register space?) but if it works reliably.
> 

Yes, we skip portions in on-chip memory dump that do not add to debug
value (such as the large regions reserved for holding Payload data
going through the device) and hence the overall dump size reduces
significantly.

> >> For that reason if for no other it would be nice if it was possible to
> >> have the driver to not initialize the device and just stand there
> >> handing out the data a piece at a time as it is read from /proc/vmcore.
> >> 
> >
> > Since cxgb4 is a network driver, it can be used to transfer the dumps
> > over the network. So we must ensure the dumps get collected and
> > stored, before device gets initialized to transfer dumps over
> > the network.
> 
> Good point.  For some reason I was thinking it was an infiniband and not
> an 10GiB ethernet device.
> 
> >> The 2GiB number I read earlier concerns me for working in a limited
> >> environment.
> >> 
> >
> > All dumps, including the 2GB on-chip memory dump, is compressed by
> > the cxgb4 driver as they are collected. The overall compressed dump
> > comes out at max 32 MB.
> >
> >> It might even make sense to separate this into a completely separate
> >> module (depended upon the main driver if it makes sense to share
> >> the functionality) so that people performing crash dumps would not
> >> hesitate to include the code in their initramfs images.
> >> 
> >> I can see splitting a device up into a portion only to be used in case
> >> of a crash dump and a normal portion like we do for main memory but I
> >> doubt that makes sense in practice.
> >> 
> >
> > This is not required, especially in case of network drivers, which
> > must collect underlying device dump and initialize the device to
> > transfer dumps over the network.
> 
> I have a practical concern.  What happens if the previous kernel left
> the device in such a bad stat the driver can not successfully initialize
> it.
> 
> Does failure to initialize cxgb4 after a crash now mean that you can not
> capture the crash dump to see the crazy state the device was in?
> 
> Typically the initramfs for a crash dump does not include unnecessary
> drivers so that hardware in states the drivers can't handle won't
> prevent taking a crash dump.
> 
> I understand the issue if you are taking a dump over your 10GiB ethernet
> it is a moot point.  But if you are writing your dump to disk, or
> writing it over a management gigabit ethernet then it is still an issue.
> 
> Is there a decoupling so that a totally b0rked device can't prevent
> taking it's own dump?
> 

As long as we can safely map and access the BAR registers, we
can collect the dumps regardless of whatever state the device and
firmware were left in and store it as part of /proc/vmcore. After
that, we attempt to re-initialize the device and restart the
firmware. So, even if driver initialization fails at this point,
we still have the dumps as part of vmcore.

Thanks,
Rahul