Re: aacraid: Regression in 4.14.56 with *genirq/affinity: assign vectors to all possible CPUs*
From: Ming Lei
Date: Sun Aug 12 2018 - 23:32:45 EST
On Sat, Aug 11, 2018 at 03:50:21PM +0200, Greg Kroah-Hartman wrote:
> On Sat, Aug 11, 2018 at 10:14:18AM +0200, Paul Menzel wrote:
> > Dear Greg,
> >
> >
> > Am 10.08.2018 um 17:55 schrieb Greg Kroah-Hartman:
> > > On Fri, Aug 10, 2018 at 04:11:23PM +0200, Paul Menzel wrote:
> >
> > > > On 08/10/18 15:36, Greg Kroah-Hartman wrote:
> > > > > 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 doesnât 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.
> > > >
> > > > Ah, I didnât see that [1] yet. Also I canât find the original message, and a
> > > > way to reply to that thread. Therefore, here is my reply.
> > > >
> > > > > Can you try the patches listed there?
> > > >
> > > > I tried some of these already without success.
> > > >
> > > > b5b6e8c8d3b4 scsi: virtio_scsi: fix IO hang caused by automatic irq vector affinity
> > > > 2f31115e940c scsi: core: introduce force_blk_mq
> > > > adbe552349f2 scsi: megaraid_sas: fix selection of reply queue
> > > >
> > > > The commit above is already in v4.14.56.
> > > >
> > > > 8b834bff1b73 scsi: hpsa: fix selection of reply queue
> > > >
> > > > The problem persists.
> > > >
> > > > The problem also persists with the state below.
> > > >
> > > > 3528f73a4e5d scsi: core: introduce force_blk_mq
> > > > 16dc4d8215f3 scsi: hpsa: fix selection of reply queue
> > > > f0a7ab12232d scsi: virtio_scsi: fix IO hang caused by automatic irq vector affinity
> > > > 6aba81b5a2f5 genirq/affinity: Don't return with empty affinity masks on error
> > > > 1aa1166eface (tag: v4.14.62, stable/linux-4.14.y) Linux 4.14.62
> > > >
> > > > So, some more commits are necessary.
> > >
> > > Or I revert the original patch here, and the follow-on ones that were
> > > added to "fix" this issue. I think that might be the better thing
> > > overall here, right? Have you tried that?
> >
> > Yes, reverting the commit fixed the issue for us. If Christoph or Ming do
> > not have another suggestion for a commit, that would be the way to go.
>
> Christoph or Ming, any ideas here?
>
> In looking at the aacraid code, I don't see anywhere that this is using
> a specific cpu number for queues or anything, but I could be wrong.
> Ideally this should also be failing in 4.17 or 4.18-rc right now as well
> as I don't see anything that would have "fixed" this recently. Unless
> I'm missing something here?
>From both aac_fib_vector_assign() and aac_src_deliver_message(), looks
every msix vector is used inside driver, so causes this trouble.
>From 84676c1f21e8ff54b (genirq/affinity: assign vectors to all possible CPUs),
actually it should have been since 9a0ef98e18 (genirq/affinity: Assign vectors
to all present CPUs), it isn't guaranteed that one vector can be mapped to
at least one online CPU.
So it should be fixed in driver by similar approach taken in the
following patches:
scsi: hpsa: fix selection of reply queue
scsi: megaraid_sas: fix selection of reply queue
And the reason this issue won't be observed is that the following
patches(merged to v4.17) may avoid this issue, but in theory, the issue
won't be avoided completely by these patches:
d3056812e7df genirq/affinity: Spread irq vectors among present CPUs as far as possible
1a2d0914e23a genirq/affinity: Allow irq spreading from a given starting point
b3e6aaa8d94d genirq/affinity: Move actual irq vector spreading into a helper function
47778f33dcba genirq/affinity: Rename *node_to_possible_cpumask as *node_to_cpumask
0211e12dd0a5 genirq/affinity: Don't return with empty affinity masks on error
In short, this issue should have been fixed in upstream aacraid, and the idea
is simple, use the current CPU id to retrieve the msix vector offset, this
way will make sure that the msix vector can be mapped from one online
CPU.
Thanks,
Ming