On Fri, Aug 10, 2018 at 03:21:52PM +0200, Paul Menzel wrote:
> Dear Greg,
>
>
> Commit ef86f3a7 (genirq/affinity: assign vectors to all possible CPUs) added
> for Linux 4.14.56 causes the aacraid module to not detect the attached devices
> anymore on a Dell PowerEdge R720 with two six core 24x E5-2630 @ 2.30GHz.
>
> ```
> $ dmesg | grep raid
> [ 0.269768] raid6: sse2x1 gen() 7179 MB/s
> [ 0.290069] raid6: sse2x1 xor() 5636 MB/s
> [ 0.311068] raid6: sse2x2 gen() 9160 MB/s
> [ 0.332076] raid6: sse2x2 xor() 6375 MB/s
> [ 0.353075] raid6: sse2x4 gen() 11164 MB/s
> [ 0.374064] raid6: sse2x4 xor() 7429 MB/s
> [ 0.379001] raid6: using algorithm sse2x4 gen() 11164 MB/s
> [ 0.386001] raid6: .... xor() 7429 MB/s, rmw enabled
> [ 0.391008] raid6: using ssse3x2 recovery algorithm
> [ 3.559682] megaraid cmm: 2.20.2.7 (Release Date: Sun Jul 16 00:01:03 EST 2006)
> [ 3.570061] megaraid: 2.20.5.1 (Release Date: Thu Nov 16 15:32:35 EST 2006)
> [ 10.725767] Adaptec aacraid driver 1.2.1[50834]-custom
> [ 10.731724] aacraid 0000:04:00.0: can't disable ASPM; OS doesn't have ASPM control
> [ 10.743295] aacraid: Comm Interface type3 enabled
> $ lspci -nn | grep Adaptec
> 04:00.0 Serial Attached SCSI controller [0107]: Adaptec Series 8 12G SAS/PCIe 3 [9005:028d] (rev 01)
> 42:00.0 Serial Attached SCSI controller [0107]: Adaptec Smart Storage PQI 12G SAS/PCIe 3 [9005:028f] (rev 01)
> ```
>
> But, it still works with a Dell PowerEdge R715 with two eight core AMD
> Opteron 6136, the card below.
>
> ```
> $ lspci -nn | grep Adaptec
> 22:00.0 Serial Attached SCSI controller [0107]: Adaptec Series 8 12G SAS/PCIe 3 [9005:028d] (rev 01)
> ```
>
> Reverting the commit fixes the issue.
>
> commit ef86f3a72adb8a7931f67335560740a7ad696d1d
> Author: Christoph Hellwig <hch@xxxxxx>
> Date: Fri Jan 12 10:53:05 2018 +0800
>
> genirq/affinity: assign vectors to all possible CPUs
>
> commit 84676c1f21e8ff54befe985f4f14dc1edc10046b upstream.
>
> Currently we assign managed interrupt vectors to all present CPUs. This
> works fine for systems were we only online/offline CPUs. But in case of
> systems that support physical CPU hotplug (or the virtualized version of
> it) this means the additional CPUs covered for in the ACPI tables or on
> the command line are not catered for. To fix this we'd either need to
> introduce new hotplug CPU states just for this case, or we can start
> assining vectors to possible but not present CPUs.
>
> Reported-by: Christian Borntraeger <borntraeger@xxxxxxxxxx>
> Tested-by: Christian Borntraeger <borntraeger@xxxxxxxxxx>
> Tested-by: Stefan Haberland <sth@xxxxxxxxxxxxxxxxxx>
> Fixes: 4b855ad37194 ("blk-mq: Create hctx for each present CPU")
> Cc: linux-kernel@xxxxxxxxxxxxxxx
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
>
> The problem doesn’t happen with Linux 4.17.11, so there are commits in
> Linux master fixing this. Unfortunately, my attempts to find out failed.
>
> I was able to cherry-pick the three commits below on top of 4.14.62,
> but the problem persists.
>
> 6aba81b5a2f5 genirq/affinity: Don't return with empty affinity masks on error
> 355d7ecdea35 scsi: hpsa: fix selection of reply queue
> e944e9615741 scsi: virtio_scsi: fix IO hang caused by automatic irq vector affinity
>
> Trying to cherry-pick the commits below, referencing the commit
> in question, gave conflicts.
>
> 1. adbe552349f2 scsi: megaraid_sas: fix selection of reply queue
> 2. d3056812e7df genirq/affinity: Spread irq vectors among present CPUs as far as possible
>
> To avoid further trial and error with the server with a slow firmware,
> do you know what commits should fix the issue?
Look at the email on the stable mailing list:
Subject: Re: Fix for 84676c1f (b5b6e8c8) missing in 4.14.y
it should help you out here. Can you try the patches listed there?
> PS: I couldn’t find, who suggested this for stable, that means how
> it was picked to be added to stable. Is there an easy way to find
> that out?
Dig through the archives is usually the best way. Looks like it came in
through a suggestion from Sasha.
thanks,
greg k-h
