Hi All,

This patch series converts s390's PCI support from its platform specific DMA
API implementation in arch/s390/pci/pci_dma.c to the common DMA IOMMU layer.
The conversion itself is done in patches 3-4 with patch 2 providing the final
necessary IOMMU driver improvement to handle s390's special IOTLB flush
out-of-resource indication in virtualized environments. The conversion
itself only touches the s390 IOMMU driver and s390 arch code moving over
remaining functions from the s390 DMA API implementation. No changes to
common code are necessary.

After patch 4 the basic conversion is done and on our partitioning
machine hypervisor LPAR performance matches the previous implementation.
When running under z/VM or KVM however, performance plummets to about
half of the existing code due to a much higher rate of IOTLB flushes for
unmapped pages. Due to the hypervisors use of IOTLB flushes to
synchronize their shadow tables these are very expensive and minimizing
them is key for regaining the performance loss.

To this end patches 5-6 add a new, single queue, IOTLB flushing scheme
as an alternative to the existing per-CPU flush queues. Introducing an
alternative scheme was suggested by Robin Murphy[1]. The single queue
mode is introduced in patch 4 together with a new .shadow_on_flush flag
bit in struct dev_iommu. This allows IOMMU drivers to indicate that
their IOTLB flushes do the extra work of shadowing. This then lets the
dma-iommu code use a single queue.

Then patch 6 enables variable queue sizes using power of 2 values and
shift/mask to keep performance as close to the fixed size queue code as
possible. A larger queue size and timeout is used by dma-iommu when
shadow_on_flush is set. This same scheme may also be used by other IOMMU
drivers with similar requirements. Particularly virtio-iommu may be
a candidate.

I tested this code on s390x with LPAR, z/VM and KVM hypervisors on an
AMD Ryzen x86 system with native IOMMU and a guest with a modified
virtio-iommu[4] that set .shadow_on_flush = true.

This code is also available in the b4/dma_iommu topic branch of my
git.kernel.org repository[3] with tags matching the version sent.

NOTE: Due to the large drop in performance I think we should not merge
the DMA API conversion (patch 4) until we have a more suited IOVA
flushing scheme with similar improvements as the proposed changes.

Best regards,
Niklas

[0] https://lore.kernel.org/linux-iommu/20221109142903.4080275-1-schnelle@linux.ibm.com/
[1] https://lore.kernel.org/linux-iommu/3e402947-61f9-b7e8-1414-fde006257b6f@arm.com/
[2] https://lore.kernel.org/linux-iommu/a8e778da-7b41-a6ba-83c3-c366a426c3da@arm.com/
[3] https://git.kernel.org/pub/scm/linux/kernel/git/niks/linux.git/
[4] https://lore.kernel.org/lkml/20230726111433.1105665-1-schnelle@linux.ibm.com/

---
Changes in v12:
- Rebased on v6.5-rc7
- Changed queue type flag to an enum
- Incorporated feedback from Robin Murphy
  - Set options centrally and only once in iommu_dma_init_domain() with
    new helper iommu_dma_init_options()
  - Do not reset options of failing to init FQ
  - Fixed rebase mishap that partially rolled back patch 2
  - Simplified patch 4 by simply no claiming the deferred flush
    capability for ISM
  - Inlined and removed fq_flush_percpu()
  - Changed vzalloc() to vmalloc() for queue
- Added Acked-by's from Robin Murphy
- Link to v11: https://lore.kernel.org/r/20230717-dma_iommu-v11-0-a7a0b83c355c@linux.ibm.com

Changes in v11:
- Rebased on v6.5-rc2
- Added patch to force IOMMU_DOMAIN_DMA on s390 specific ISM devices
- Dropped the patch to properly set DMA mask on ISM devices which went upstream separately.
- s390 IOMMU driver now uses IOMMU_CAP_DEFERRED_FLUSH to enable DMA-FQ
  leaving no uses of IOMMU_DOMAIN_DMA_FQ in the driver.
- Link to v10: https://lore.kernel.org/r/20230310-dma_iommu-v10-0-f1fbd8310854@linux.ibm.com

Changes in v10:
- Rebased on v6.4-rc3
- Removed the .tune_dma_iommu() op in favor of a .shadow_on_flush flag
  in struct dev_iommu which then let's the dma-iommu choose a single
  queue and larger timeouts and IOVA counts. This leaves the dma-iommu
  with full responsibility for the settings.
- The above change affects patches 5 and 6 and lead to a new subject for
  patch 6 since the flush queue size and timeout is no longer driver
  controlled
- Link to v9: https://lore.kernel.org/r/20230310-dma_iommu-v9-0-65bb8edd2beb@linux.ibm.com

Changes in v9:
- Rebased on v6.4-rc2
- Re-ordered iommu_group_store_type() to allow passing the device to
  iommu_dma_init_fq()
- Link to v8: https://lore.kernel.org/r/20230310-dma_iommu-v8-0-2347dfbed7af@linux.ibm.com

---
Niklas Schnelle (6):
      iommu: Allow .iotlb_sync_map to fail and handle s390's -ENOMEM return
      s390/pci: prepare is_passed_through() for dma-iommu
      s390/pci: Use dma-iommu layer
      iommu/s390: Disable deferred flush for ISM devices
      iommu/dma: Allow a single FQ in addition to per-CPU FQs
      iommu/dma: Use a large flush queue and timeout for shadow_on_flush

 Documentation/admin-guide/kernel-parameters.txt |   9 +-
 arch/s390/include/asm/pci.h                     |   7 -
 arch/s390/include/asm/pci_clp.h                 |   3 +
 arch/s390/include/asm/pci_dma.h                 | 119 +---
 arch/s390/pci/Makefile                          |   2 +-
 arch/s390/pci/pci.c                             |  22 +-
 arch/s390/pci/pci_bus.c                         |   5 -
 arch/s390/pci/pci_debug.c                       |  12 +-
 arch/s390/pci/pci_dma.c                         | 735 ------------------------
 arch/s390/pci/pci_event.c                       |  17 +-
 arch/s390/pci/pci_sysfs.c                       |  19 +-
 drivers/iommu/Kconfig                           |   4 +-
 drivers/iommu/amd/iommu.c                       |   5 +-
 drivers/iommu/apple-dart.c                      |   5 +-
 drivers/iommu/dma-iommu.c                       | 200 +++++--
 drivers/iommu/intel/iommu.c                     |   5 +-
 drivers/iommu/iommu.c                           |  20 +-
 drivers/iommu/msm_iommu.c                       |   5 +-
 drivers/iommu/mtk_iommu.c                       |   5 +-
 drivers/iommu/s390-iommu.c                      | 425 ++++++++++++--
 drivers/iommu/sprd-iommu.c                      |   5 +-
 drivers/iommu/sun50i-iommu.c                    |   6 +-
 drivers/iommu/tegra-gart.c                      |   5 +-
 include/linux/iommu.h                           |   6 +-
 24 files changed, 643 insertions(+), 1003 deletions(-)
---
base-commit: 706a741595047797872e669b3101429ab8d378ef
change-id: 20230310-dma_iommu-5e048c538647

Best regards,
-- 
Niklas Schnelle
Linux on Z Development

IBM Deutschland Research & Development GmbH
Vorsitzender des Aufsichtsrats: Gregor Pillen
Geschäftsführung: David Faller
Sitz der Gesellschaft: Böblingen / Registergericht: Amtsgericht Stuttgart, HRB 243294
IBM Data Privacy Statement - https://www.ibm.com/privacy 

On 8/25/23 6:11 AM, Niklas Schnelle wrote:
> Hi All,
> 
> This patch series converts s390's PCI support from its platform specific DMA
> API implementation in arch/s390/pci/pci_dma.c to the common DMA IOMMU layer.
> The conversion itself is done in patches 3-4 with patch 2 providing the final
> necessary IOMMU driver improvement to handle s390's special IOTLB flush
> out-of-resource indication in virtualized environments. The conversion
> itself only touches the s390 IOMMU driver and s390 arch code moving over
> remaining functions from the s390 DMA API implementation. No changes to
> common code are necessary.
> 

I also picked up this latest version and ran various tests with ISM, mlx5 and some NVMe drives.  FWIW, I have been including versions of this series in my s390 dev environments for a number of months now and have also been building my s390 pci iommufd nested translation series on top of this, so it's seen quite a bit of testing from me at least.

So as far as I'm concerned anyway, this series is ready for -next (after the merge window). 

Thanks,
Matt

On 2023-08-25 19:26, Matthew Rosato wrote:
> On 8/25/23 6:11 AM, Niklas Schnelle wrote:
>> Hi All,
>>
>> This patch series converts s390's PCI support from its platform specific DMA
>> API implementation in arch/s390/pci/pci_dma.c to the common DMA IOMMU layer.
>> The conversion itself is done in patches 3-4 with patch 2 providing the final
>> necessary IOMMU driver improvement to handle s390's special IOTLB flush
>> out-of-resource indication in virtualized environments. The conversion
>> itself only touches the s390 IOMMU driver and s390 arch code moving over
>> remaining functions from the s390 DMA API implementation. No changes to
>> common code are necessary.
>>
> 
> I also picked up this latest version and ran various tests with ISM, mlx5 and some NVMe drives.  FWIW, I have been including versions of this series in my s390 dev environments for a number of months now and have also been building my s390 pci iommufd nested translation series on top of this, so it's seen quite a bit of testing from me at least.
> 
> So as far as I'm concerned anyway, this series is ready for -next (after the merge window).

Agreed; I'll trust your reviews for the s390-specific parts, so indeed 
it looks like this should have all it needs now and is ready for a nice 
long soak in -next once Joerg opens the tree for 6.7 material.

Cheers,
Robin.

On Fri, Aug 25, 2023 at 12:11:15PM +0200, Niklas Schnelle wrote:
> Niklas Schnelle (6):
>       iommu: Allow .iotlb_sync_map to fail and handle s390's -ENOMEM return
>       s390/pci: prepare is_passed_through() for dma-iommu
>       s390/pci: Use dma-iommu layer
>       iommu/s390: Disable deferred flush for ISM devices
>       iommu/dma: Allow a single FQ in addition to per-CPU FQs
>       iommu/dma: Use a large flush queue and timeout for shadow_on_flush

Applied, thanks.

Hi Niklas,

On Fri, Aug 25, 2023 at 12:11:15PM +0200, Niklas Schnelle wrote:
> Niklas Schnelle (6):
>       iommu: Allow .iotlb_sync_map to fail and handle s390's -ENOMEM return
>       s390/pci: prepare is_passed_through() for dma-iommu
>       s390/pci: Use dma-iommu layer
>       iommu/s390: Disable deferred flush for ISM devices
>       iommu/dma: Allow a single FQ in addition to per-CPU FQs
>       iommu/dma: Use a large flush queue and timeout for shadow_on_flush

Turned out this series has non-trivial conflicts with Jasons
default-domain work so I had to remove it from the IOMMU tree for now.
Can you please rebase it to the latest iommu/core branch and re-send? I
will take it into the tree again then.

Thanks,

	Joerg

On Tue, Sep 26, 2023 at 05:04:28PM +0200, Joerg Roedel wrote:
> Hi Niklas,
> 
> On Fri, Aug 25, 2023 at 12:11:15PM +0200, Niklas Schnelle wrote:
> > Niklas Schnelle (6):
> >       iommu: Allow .iotlb_sync_map to fail and handle s390's -ENOMEM return
> >       s390/pci: prepare is_passed_through() for dma-iommu
> >       s390/pci: Use dma-iommu layer
> >       iommu/s390: Disable deferred flush for ISM devices
> >       iommu/dma: Allow a single FQ in addition to per-CPU FQs
> >       iommu/dma: Use a large flush queue and timeout for shadow_on_flush
> 
> Turned out this series has non-trivial conflicts with Jasons
> default-domain work so I had to remove it from the IOMMU tree for now.
> Can you please rebase it to the latest iommu/core branch and re-send? I
> will take it into the tree again then.

Niklas, I think you just 'take yours' to resolve this. All the
IOMMU_DOMAIN_PLATFORM related and .default_domain = parts should be
removed. Let me know if you need anything

Thanks,
Jason

On Tue, 2023-09-26 at 13:08 -0300, Jason Gunthorpe wrote:
> On Tue, Sep 26, 2023 at 05:04:28PM +0200, Joerg Roedel wrote:
> > Hi Niklas,
> > 
> > On Fri, Aug 25, 2023 at 12:11:15PM +0200, Niklas Schnelle wrote:
> > > Niklas Schnelle (6):
> > >       iommu: Allow .iotlb_sync_map to fail and handle s390's -ENOMEM return
> > >       s390/pci: prepare is_passed_through() for dma-iommu
> > >       s390/pci: Use dma-iommu layer
> > >       iommu/s390: Disable deferred flush for ISM devices
> > >       iommu/dma: Allow a single FQ in addition to per-CPU FQs
> > >       iommu/dma: Use a large flush queue and timeout for shadow_on_flush
> > 
> > Turned out this series has non-trivial conflicts with Jasons
> > default-domain work so I had to remove it from the IOMMU tree for now.
> > Can you please rebase it to the latest iommu/core branch and re-send? I
> > will take it into the tree again then.
> 
> Niklas, I think you just 'take yours' to resolve this. All the
> IOMMU_DOMAIN_PLATFORM related and .default_domain = parts should be
> removed. Let me know if you need anything
> 
> Thanks,
> Jason

Hi Joerg, Hi Jason,

I've run into an unfortunate problem, not with the rebase itself but
with the iommu/core branch. 

Jason is right, I basically need to just remove the platform ops and
.default_domain ops. This seems to work fine for an NVMe both in the
host and also when using the IOMMU with vfio-pci + KVM. I've already
pushed the result of that to my git.kernel.org:
https://git.kernel.org/pub/scm/linux/kernel/git/niks/linux.git/log/?h=b4/dma_iommu

The problem is that something seems to  be broken in the iommu/core
branch. Regardless of whether I have my DMA API conversion on top or
with the base iommu/core branch I can not use ConnectX-4 VFs.

# lspci
111a:00:00.0 Ethernet controller: Mellanox Technologies MT27710 Family [ConnectX-4 Lx Virtual Function]
# dmesg | grep mlx
[    3.189749] mlx5_core 111a:00:00.0: mlx5_mdev_init:1802:(pid 464): Failed initializing cmdif SW structs, aborting
[    3.189783] mlx5_core: probe of 111a:00:00.0 failed with error -12

This same card works on v6.6-rc3 both with and without my DMA API
conversion patch series applied. Looking at mlx5_mdev_init() -> 
mlx5_cmd_init(). The -ENOMEM seems to come from the following
dma_pool_create():

cmd->pool = dma_pool_create("mlx5_cmd", mlx5_core_dma_dev(dev), size, align, 0);

I'll try to debug this further but wanted to let you know already in
case you have some ideas. Either way as it doesn't seem to be related
to the DMA API conversion I can sent that out again regardless if you
want, really don't want to miss another cycle.

Thanks,
Niklas

Hi Niklas,

On Wed, Sep 27, 2023 at 10:55:23AM +0200, Niklas Schnelle wrote:
> The problem is that something seems to  be broken in the iommu/core
> branch. Regardless of whether I have my DMA API conversion on top or
> with the base iommu/core branch I can not use ConnectX-4 VFs.

Have you already tried to bisect the issue in the iommu/core branch?
The result might sched some light on the issue.

Regards,

	Joerg

On Wed, 2023-09-27 at 11:55 +0200, Joerg Roedel wrote:
> Hi Niklas,
> 
> On Wed, Sep 27, 2023 at 10:55:23AM +0200, Niklas Schnelle wrote:
> > The problem is that something seems to  be broken in the iommu/core
> > branch. Regardless of whether I have my DMA API conversion on top or
> > with the base iommu/core branch I can not use ConnectX-4 VFs.
> 
> Have you already tried to bisect the issue in the iommu/core branch?
> The result might sched some light on the issue.
> 
> Regards,
> 
> 	Joerg

Hi Joerg,

Working on it, somehow I must have messed up earlier. It now looks like
it might in fact be caused by my DMA API conversion rebase and the
"s390/pci: Use dma-iommu layer" commit. Maybe there is some interaction
with Jason's patches that I haven't thought about. So sorry for any
wrong blame.

Thanks,
Niklas

On Wed, 2023-09-27 at 13:24 +0200, Niklas Schnelle wrote:
> On Wed, 2023-09-27 at 11:55 +0200, Joerg Roedel wrote:
> > Hi Niklas,
> > 
> > On Wed, Sep 27, 2023 at 10:55:23AM +0200, Niklas Schnelle wrote:
> > > The problem is that something seems to  be broken in the iommu/core
> > > branch. Regardless of whether I have my DMA API conversion on top or
> > > with the base iommu/core branch I can not use ConnectX-4 VFs.
> > 
> > Have you already tried to bisect the issue in the iommu/core branch?
> > The result might sched some light on the issue.
> > 
> > Regards,
> > 
> > 	Joerg
> 
> Hi Joerg,
> 
> Working on it, somehow I must have messed up earlier. It now looks like
> it might in fact be caused by my DMA API conversion rebase and the
> "s390/pci: Use dma-iommu layer" commit. Maybe there is some interaction
> with Jason's patches that I haven't thought about. So sorry for any
> wrong blame.
> 
> Thanks,
> Niklas

Hi,

I tracked the problem down from mlx5_core's alloc_cmd_page() via
dma_alloc_coherent(), ops->alloc, iommu_dma_alloc_remap(), and
__iommu_dma_alloc_noncontiguous() to a failed iommu_dma_alloc_iova().
The allocation here is for 4K so nothing crazy.

On second look I also noticed:

nvme 2007:00:00.0: Using 42-bit DMA addresses

for the NVMe that is working. The problem here seems to be that we set
iommu_dma_forcedac = true in s390_iommu_probe_finalize() because we
have currently have a reserved region over the first 4 GiB anyway so
will always use IOVAs larger than that. That however is too late since
iommu_dma_set_pci_32bit_workaround() is already checked in
__iommu_probe_device() which is called just before ops-
>probe_finalize(). So I moved setting iommu_dma_forcedac = true to
zpci_init_iommu() and that gets rid of the notice for the NVMe but I
still get a failure of iommu_dma_alloc_iova() in
__iommu_dma_alloc_noncontiguous(). So I'll keep digging.

Thanks,
Niklas

On Wed, 2023-09-27 at 15:20 +0200, Niklas Schnelle wrote:
> On Wed, 2023-09-27 at 13:24 +0200, Niklas Schnelle wrote:
> > On Wed, 2023-09-27 at 11:55 +0200, Joerg Roedel wrote:
> > > Hi Niklas,
> > > 
> > > On Wed, Sep 27, 2023 at 10:55:23AM +0200, Niklas Schnelle wrote:
> > > > The problem is that something seems to  be broken in the iommu/core
> > > > branch. Regardless of whether I have my DMA API conversion on top or
> > > > with the base iommu/core branch I can not use ConnectX-4 VFs.
> > > 
> > > Have you already tried to bisect the issue in the iommu/core branch?
> > > The result might sched some light on the issue.
> > > 
> > > Regards,
> > > 
> > > 	Joerg
> > 
> > Hi Joerg,
> > 
> > Working on it, somehow I must have messed up earlier. It now looks like
> > it might in fact be caused by my DMA API conversion rebase and the
> > "s390/pci: Use dma-iommu layer" commit. Maybe there is some interaction
> > with Jason's patches that I haven't thought about. So sorry for any
> > wrong blame.
> > 
> > Thanks,
> > Niklas
> 
> Hi,
> 
> I tracked the problem down from mlx5_core's alloc_cmd_page() via
> dma_alloc_coherent(), ops->alloc, iommu_dma_alloc_remap(), and
> __iommu_dma_alloc_noncontiguous() to a failed iommu_dma_alloc_iova().
> The allocation here is for 4K so nothing crazy.
> 
> On second look I also noticed:
> 
> nvme 2007:00:00.0: Using 42-bit DMA addresses
> 
> for the NVMe that is working. The problem here seems to be that we set
> iommu_dma_forcedac = true in s390_iommu_probe_finalize() because we
> have currently have a reserved region over the first 4 GiB anyway so
> will always use IOVAs larger than that. That however is too late since
> iommu_dma_set_pci_32bit_workaround() is already checked in
> __iommu_probe_device() which is called just before ops-
> > probe_finalize(). So I moved setting iommu_dma_forcedac = true to
> zpci_init_iommu() and that gets rid of the notice for the NVMe but I
> still get a failure of iommu_dma_alloc_iova() in
> __iommu_dma_alloc_noncontiguous(). So I'll keep digging.
> 
> Thanks,
> Niklas


Ok I think I got it and this doesn't seem strictly s390x specific but
I'd think should happen with iommu.forcedac=1 everywhere.

The reason iommu_dma_alloc_iova() fails seems to be that mlx5_core does
dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64)) in 
mlx5_pci_init()->set_dma_caps() which happens after it already called
mlx5_mdev_init()->mlx5_cmd_init()->alloc_cmd_page() so for the
dma_alloc_coherent() in there the dev->coherent_dma_mask is still
DMA_BIT_MASK(32) for which we can't find an IOVA because well we don't
have IOVAs below 4 GiB. Not entirely sure what caused this not to be
enforced before.

Thanks,
Niklas

On Wed, 2023-09-27 at 16:31 +0200, Niklas Schnelle wrote:
> On Wed, 2023-09-27 at 15:20 +0200, Niklas Schnelle wrote:
> > On Wed, 2023-09-27 at 13:24 +0200, Niklas Schnelle wrote:
> > > On Wed, 2023-09-27 at 11:55 +0200, Joerg Roedel wrote:
> > > > Hi Niklas,
> > > > 
> > > > On Wed, Sep 27, 2023 at 10:55:23AM +0200, Niklas Schnelle wrote:
> > > > > The problem is that something seems to  be broken in the iommu/core
> > > > > branch. Regardless of whether I have my DMA API conversion on top or
> > > > > with the base iommu/core branch I can not use ConnectX-4 VFs.
> > > > 
> > > > Have you already tried to bisect the issue in the iommu/core branch?
> > > > The result might sched some light on the issue.
> > > > 
> > > > Regards,
> > > > 
> > > > 	Joerg
> > > 
> > > Hi Joerg,
> > > 
> > > Working on it, somehow I must have messed up earlier. It now looks like
> > > it might in fact be caused by my DMA API conversion rebase and the
> > > "s390/pci: Use dma-iommu layer" commit. Maybe there is some interaction
> > > with Jason's patches that I haven't thought about. So sorry for any
> > > wrong blame.
> > > 
> > > Thanks,
> > > Niklas
> > 
> > Hi,
> > 
> > I tracked the problem down from mlx5_core's alloc_cmd_page() via
> > dma_alloc_coherent(), ops->alloc, iommu_dma_alloc_remap(), and
> > __iommu_dma_alloc_noncontiguous() to a failed iommu_dma_alloc_iova().
> > The allocation here is for 4K so nothing crazy.
> > 
> > On second look I also noticed:
> > 
> > nvme 2007:00:00.0: Using 42-bit DMA addresses
> > 
> > for the NVMe that is working. The problem here seems to be that we set
> > iommu_dma_forcedac = true in s390_iommu_probe_finalize() because we
> > have currently have a reserved region over the first 4 GiB anyway so
> > will always use IOVAs larger than that. That however is too late since
> > iommu_dma_set_pci_32bit_workaround() is already checked in
> > __iommu_probe_device() which is called just before ops-
> > > probe_finalize(). So I moved setting iommu_dma_forcedac = true to
> > zpci_init_iommu() and that gets rid of the notice for the NVMe but I
> > still get a failure of iommu_dma_alloc_iova() in
> > __iommu_dma_alloc_noncontiguous(). So I'll keep digging.
> > 
> > Thanks,
> > Niklas
> 
> 
> Ok I think I got it and this doesn't seem strictly s390x specific but
> I'd think should happen with iommu.forcedac=1 everywhere.
> 
> The reason iommu_dma_alloc_iova() fails seems to be that mlx5_core does
> dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64)) in 
> mlx5_pci_init()->set_dma_caps() which happens after it already called
> mlx5_mdev_init()->mlx5_cmd_init()->alloc_cmd_page() so for the
> dma_alloc_coherent() in there the dev->coherent_dma_mask is still
> DMA_BIT_MASK(32) for which we can't find an IOVA because well we don't
> have IOVAs below 4 GiB. Not entirely sure what caused this not to be
> enforced before.
> 
> Thanks,
> Niklas
> 

Ok, another update. On trying it out again this problem actually also
occurs when applying this v12 on top of v6.6-rc3 too. Also I guess
unlike my prior thinking it probably doesn't occur with
iommu.forcedac=1 since that still allows IOVAs below 4 GiB and we might
be the only ones who don't support those. From my point of view this
sounds like a mlx5_core issue they really should call
dma_set_mask_and_coherent() before their first call to
dma_alloc_coherent() not after. So I guess I'll send a v13 of this
series rebased on iommu/core and with an additional mlx5 patch and then
let's hope we can get that merged in a way that doesn't leave us with
broken ConnectX VFs for too long.

Thanks,
Niklas

On Wed, Sep 27, 2023 at 05:24:20PM +0200, Niklas Schnelle wrote:

> Ok, another update. On trying it out again this problem actually also
> occurs when applying this v12 on top of v6.6-rc3 too. Also I guess
> unlike my prior thinking it probably doesn't occur with
> iommu.forcedac=1 since that still allows IOVAs below 4 GiB and we might
> be the only ones who don't support those. From my point of view this
> sounds like a mlx5_core issue they really should call
> dma_set_mask_and_coherent() before their first call to
> dma_alloc_coherent() not after. So I guess I'll send a v13 of this
> series rebased on iommu/core and with an additional mlx5 patch and then
> let's hope we can get that merged in a way that doesn't leave us with
> broken ConnectX VFs for too long.

Yes, OK. It definitely sounds wrong that mlx5 is doing dma allocations before
setting it's dma_set_mask_and_coherent(). Please link to this thread
and we can get Leon or Saeed to ack it for Joerg.

(though wondering why s390 is the only case that ever hit this?)

Jason

On 9/27/23 11:40 AM, Jason Gunthorpe wrote:
> On Wed, Sep 27, 2023 at 05:24:20PM +0200, Niklas Schnelle wrote:
> 
>> Ok, another update. On trying it out again this problem actually also
>> occurs when applying this v12 on top of v6.6-rc3 too. Also I guess
>> unlike my prior thinking it probably doesn't occur with
>> iommu.forcedac=1 since that still allows IOVAs below 4 GiB and we might
>> be the only ones who don't support those. From my point of view this
>> sounds like a mlx5_core issue they really should call
>> dma_set_mask_and_coherent() before their first call to
>> dma_alloc_coherent() not after. So I guess I'll send a v13 of this
>> series rebased on iommu/core and with an additional mlx5 patch and then
>> let's hope we can get that merged in a way that doesn't leave us with
>> broken ConnectX VFs for too long.
> 
> Yes, OK. It definitely sounds wrong that mlx5 is doing dma allocations before
> setting it's dma_set_mask_and_coherent(). Please link to this thread
> and we can get Leon or Saeed to ack it for Joerg.
> 

Hi Niklas,

I bisected the start of this issue to the following commit (only noticeable on s390 when you apply this subject series on top):

06cd555f73caec515a14d42ef052221fa2587ff9 ("net/mlx5: split mlx5_cmd_init() to probe and reload routines")

Which went in during the merge window.  Please include with your fix and/or report to the mlx5 maintainers.  Looks like the changes in this patch match what you and Jason describe; it splits up mlx5_cmd_init() and moves part of the call earlier.  The net result is we first call mlx5_mdev_init>mlx5_cmd_init->alloc_cmd_page->dma_alloc_coherent and then sometime later call mlx5_pci_init->set_dma_caps->dma_set_mask_and_coherent. 

Prior to this patch, we would not drive mlx5_cmd_init (and thus that first dma_alloc_coherent) until mlx5_init_one which happens _after_ mlx5_pci_init->set_dma_caps->dma_set_mask_and_coherent.

Thanks,
Matt

On 27/09/2023 4:40 pm, Jason Gunthorpe wrote:
> On Wed, Sep 27, 2023 at 05:24:20PM +0200, Niklas Schnelle wrote:
> 
>> Ok, another update. On trying it out again this problem actually also
>> occurs when applying this v12 on top of v6.6-rc3 too. Also I guess
>> unlike my prior thinking it probably doesn't occur with
>> iommu.forcedac=1 since that still allows IOVAs below 4 GiB and we might
>> be the only ones who don't support those. From my point of view this
>> sounds like a mlx5_core issue they really should call
>> dma_set_mask_and_coherent() before their first call to
>> dma_alloc_coherent() not after. So I guess I'll send a v13 of this
>> series rebased on iommu/core and with an additional mlx5 patch and then
>> let's hope we can get that merged in a way that doesn't leave us with
>> broken ConnectX VFs for too long.
> 
> Yes, OK. It definitely sounds wrong that mlx5 is doing dma allocations before
> setting it's dma_set_mask_and_coherent(). Please link to this thread
> and we can get Leon or Saeed to ack it for Joerg.
> 
> (though wondering why s390 is the only case that ever hit this?)

Probably because most systems happen to be able to satisfy the 
allocation within the default 32-bit mask - the whole bottom 4GB of IOVA 
space being reserved is pretty atypical.

TBH it makes me wonder the opposite - how this ever worked on s390 
before? And I think the answer to that is "by pure chance", since upon 
inspection the existing s390_pci_dma_ops implementation appears to pay 
absolutely no attention to the device's DMA masks whatsoever :(

Robin.

On 2023-09-27 09:55, Niklas Schnelle wrote:
> On Tue, 2023-09-26 at 13:08 -0300, Jason Gunthorpe wrote:
>> On Tue, Sep 26, 2023 at 05:04:28PM +0200, Joerg Roedel wrote:
>>> Hi Niklas,
>>>
>>> On Fri, Aug 25, 2023 at 12:11:15PM +0200, Niklas Schnelle wrote:
>>>> Niklas Schnelle (6):
>>>>        iommu: Allow .iotlb_sync_map to fail and handle s390's -ENOMEM return
>>>>        s390/pci: prepare is_passed_through() for dma-iommu
>>>>        s390/pci: Use dma-iommu layer
>>>>        iommu/s390: Disable deferred flush for ISM devices
>>>>        iommu/dma: Allow a single FQ in addition to per-CPU FQs
>>>>        iommu/dma: Use a large flush queue and timeout for shadow_on_flush
>>>
>>> Turned out this series has non-trivial conflicts with Jasons
>>> default-domain work so I had to remove it from the IOMMU tree for now.
>>> Can you please rebase it to the latest iommu/core branch and re-send? I
>>> will take it into the tree again then.
>>
>> Niklas, I think you just 'take yours' to resolve this. All the
>> IOMMU_DOMAIN_PLATFORM related and .default_domain = parts should be
>> removed. Let me know if you need anything
>>
>> Thanks,
>> Jason
> 
> Hi Joerg, Hi Jason,
> 
> I've run into an unfortunate problem, not with the rebase itself but
> with the iommu/core branch.
> 
> Jason is right, I basically need to just remove the platform ops and
> .default_domain ops. This seems to work fine for an NVMe both in the
> host and also when using the IOMMU with vfio-pci + KVM. I've already
> pushed the result of that to my git.kernel.org:
> https://git.kernel.org/pub/scm/linux/kernel/git/niks/linux.git/log/?h=b4/dma_iommu
> 
> The problem is that something seems to  be broken in the iommu/core
> branch. Regardless of whether I have my DMA API conversion on top or
> with the base iommu/core branch I can not use ConnectX-4 VFs.
> 
> # lspci
> 111a:00:00.0 Ethernet controller: Mellanox Technologies MT27710 Family [ConnectX-4 Lx Virtual Function]
> # dmesg | grep mlx
> [    3.189749] mlx5_core 111a:00:00.0: mlx5_mdev_init:1802:(pid 464): Failed initializing cmdif SW structs, aborting
> [    3.189783] mlx5_core: probe of 111a:00:00.0 failed with error -12
> 
> This same card works on v6.6-rc3 both with and without my DMA API
> conversion patch series applied. Looking at mlx5_mdev_init() ->
> mlx5_cmd_init(). The -ENOMEM seems to come from the following
> dma_pool_create():
> 
> cmd->pool = dma_pool_create("mlx5_cmd", mlx5_core_dma_dev(dev), size, align, 0);
> 
> I'll try to debug this further but wanted to let you know already in
> case you have some ideas.

I could imagine that potentially something in the initial default domain 
conversion somehow interferes with the DMA ops in a way that ends up 
causing alloc_cmd_page() to fail (maybe calling zpci_dma_init_device() 
at the wrong point, or too many times?). FWIW I see nothing that would 
obviously affect dma_pool_create() itself.

Robin.

> Either way as it doesn't seem to be related
> to the DMA API conversion I can sent that out again regardless if you
> want, really don't want to miss another cycle.
> 
> Thanks,
> Niklas