Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Ben, what was the driver which did not set "read" and caused EEH?
---
arch/powerpc/platforms/powernv/pci.c | 6 ++++++
1 file changed, 6 insertions(+)
@@ -601,6 +601,9 @@ int pnv_tce_build(struct iommu_table *tbl, long index, long npages,u64rpn=__pa(uaddr)>>tbl->it_page_shift;longi;+if(proto_tce&TCE_PCI_WRITE)+proto_tce|=TCE_PCI_READ;+for(i=0;i<npages;i++){unsignedlongnewtce=proto_tce|((rpn+i)<<tbl->it_page_shift);
@@ -622,6 +625,9 @@ int pnv_tce_xchg(struct iommu_table *tbl, long index,BUG_ON(*hpa&~IOMMU_PAGE_MASK(tbl));+if(newtce&TCE_PCI_WRITE)+newtce|=TCE_PCI_READ;+oldtce=xchg(pnv_tce(tbl,idx),cpu_to_be64(newtce));*hpa=be64_to_cpu(oldtce)&~(TCE_PCI_READ|TCE_PCI_WRITE);*direction=iommu_tce_direction(oldtce);
From: Benjamin Herrenschmidt <benh@kernel.crashing.org> Date: 2016-01-12 23:07:45
On Tue, 2016-01-12 at 15:40 +1100, Alexey Kardashevskiy wrote:
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty
platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Ben, what was the driver which did not set "read" and caused EEH?
From: Douglas Miller <hidden> Date: 2016-01-13 02:24:08
On 01/12/2016 05:07 PM, Benjamin Herrenschmidt wrote:
On Tue, 2016-01-12 at 15:40 +1100, Alexey Kardashevskiy wrote:
quoted
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty
platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Ben, what was the driver which did not set "read" and caused EEH?
aacraid
Cheers,
Ben.
Just to be precise, the driver wasn't responsible for setting READ. The
driver called scsi_dma_map() and the scsicmd was set (by scsi layer) as
DMA_FROM_DEVICE so the current code would set the permissions to
WRITE-ONLY. Previously, and in other architectures, this scsicmd would
have resulted in READ+WRITE permissions on the DMA map.
On 01/12/2016 05:07 PM, Benjamin Herrenschmidt wrote:
quoted
On Tue, 2016-01-12 at 15:40 +1100, Alexey Kardashevskiy wrote:
quoted
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty
platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Ben, what was the driver which did not set "read" and caused EEH?
aacraid
Cheers,
Ben.
Just to be precise, the driver wasn't responsible for setting READ. The
driver called scsi_dma_map() and the scsicmd was set (by scsi layer) as
DMA_FROM_DEVICE so the current code would set the permissions to
WRITE-ONLY. Previously, and in other architectures, this scsicmd would have
resulted in READ+WRITE permissions on the DMA map.
From: Douglas Miller <hidden> Date: 2016-01-19 19:01:29
On 01/18/2016 09:52 PM, Alexey Kardashevskiy wrote:
On 01/13/2016 01:24 PM, Douglas Miller wrote:
quoted
On 01/12/2016 05:07 PM, Benjamin Herrenschmidt wrote:
quoted
On Tue, 2016-01-12 at 15:40 +1100, Alexey Kardashevskiy wrote:
quoted
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty
platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Ben, what was the driver which did not set "read" and caused EEH?
aacraid
Cheers,
Ben.
Just to be precise, the driver wasn't responsible for setting READ. The
driver called scsi_dma_map() and the scsicmd was set (by scsi layer) as
DMA_FROM_DEVICE so the current code would set the permissions to
WRITE-ONLY. Previously, and in other architectures, this scsicmd
would have
resulted in READ+WRITE permissions on the DMA map.
I am still working on getting a machine to try this on. From code
inspection, it looks like it should work. The problem is shortage of
machines and machines tied-up by Test.
Thanks,
Doug
On 01/18/2016 09:52 PM, Alexey Kardashevskiy wrote:
quoted
On 01/13/2016 01:24 PM, Douglas Miller wrote:
quoted
On 01/12/2016 05:07 PM, Benjamin Herrenschmidt wrote:
quoted
On Tue, 2016-01-12 at 15:40 +1100, Alexey Kardashevskiy wrote:
quoted
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty
platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Ben, what was the driver which did not set "read" and caused EEH?
aacraid
Cheers,
Ben.
Just to be precise, the driver wasn't responsible for setting READ. The
driver called scsi_dma_map() and the scsicmd was set (by scsi layer) as
DMA_FROM_DEVICE so the current code would set the permissions to
WRITE-ONLY. Previously, and in other architectures, this scsicmd would have
resulted in READ+WRITE permissions on the DMA map.
I am still working on getting a machine to try this on. From code
inspection, it looks like it should work. The problem is shortage of
machines and machines tied-up by Test.
From: Douglas Miller <hidden> Date: 2016-02-09 14:28:25
We finally got the chance to test it end of last week. I forgot to
update everyone Monday. B all appearances, the patch fixes the problem.
We did not see any new issues with the patch (vs. same test scenarios
without).
I'll also update the bugzilla.
Thanks,
Doug
On 02/08/2016 07:37 PM, Alexey Kardashevskiy wrote:
On 01/20/2016 06:01 AM, Douglas Miller wrote:
quoted
On 01/18/2016 09:52 PM, Alexey Kardashevskiy wrote:
quoted
On 01/13/2016 01:24 PM, Douglas Miller wrote:
quoted
On 01/12/2016 05:07 PM, Benjamin Herrenschmidt wrote:
quoted
On Tue, 2016-01-12 at 15:40 +1100, Alexey Kardashevskiy wrote:
quoted
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty
platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit
when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Ben, what was the driver which did not set "read" and caused EEH?
aacraid
Cheers,
Ben.
Just to be precise, the driver wasn't responsible for setting READ.
The
driver called scsi_dma_map() and the scsicmd was set (by scsi
layer) as
DMA_FROM_DEVICE so the current code would set the permissions to
WRITE-ONLY. Previously, and in other architectures, this scsicmd
would have
resulted in READ+WRITE permissions on the DMA map.
I am still working on getting a machine to try this on. From code
inspection, it looks like it should work. The problem is shortage of
machines and machines tied-up by Test.
We finally got the chance to test it end of last week. I forgot to update
everyone Monday. B all appearances, the patch fixes the problem. We did not
see any new issues with the patch (vs. same test scenarios without).
I'll also update the bugzilla.
Thanks. Care to add "Tested-by"?
Thanks,
Doug
On 02/08/2016 07:37 PM, Alexey Kardashevskiy wrote:
quoted
On 01/20/2016 06:01 AM, Douglas Miller wrote:
quoted
On 01/18/2016 09:52 PM, Alexey Kardashevskiy wrote:
quoted
On 01/13/2016 01:24 PM, Douglas Miller wrote:
quoted
On 01/12/2016 05:07 PM, Benjamin Herrenschmidt wrote:
quoted
On Tue, 2016-01-12 at 15:40 +1100, Alexey Kardashevskiy wrote:
quoted
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty
platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Ben, what was the driver which did not set "read" and caused EEH?
aacraid
Cheers,
Ben.
Just to be precise, the driver wasn't responsible for setting READ. The
driver called scsi_dma_map() and the scsicmd was set (by scsi layer) as
DMA_FROM_DEVICE so the current code would set the permissions to
WRITE-ONLY. Previously, and in other architectures, this scsicmd would
have
resulted in READ+WRITE permissions on the DMA map.
I am still working on getting a machine to try this on. From code
inspection, it looks like it should work. The problem is shortage of
machines and machines tied-up by Test.
From: Douglas Miller <hidden> Date: 2016-02-10 12:26:17
Tested-by: Douglas Miller <redacted>
On 01/11/2016 10:40 PM, Alexey Kardashevskiy wrote:
quoted hunk
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
---
Ben, what was the driver which did not set "read" and caused EEH?
---
arch/powerpc/platforms/powernv/pci.c | 6 ++++++
1 file changed, 6 insertions(+)
@@ -601,6 +601,9 @@ int pnv_tce_build(struct iommu_table *tbl, long index, long npages,u64rpn=__pa(uaddr)>>tbl->it_page_shift;longi;+if(proto_tce&TCE_PCI_WRITE)+proto_tce|=TCE_PCI_READ;+for(i=0;i<npages;i++){unsignedlongnewtce=proto_tce|((rpn+i)<<tbl->it_page_shift);
@@ -622,6 +625,9 @@ int pnv_tce_xchg(struct iommu_table *tbl, long index,BUG_ON(*hpa&~IOMMU_PAGE_MASK(tbl));+if(newtce&TCE_PCI_WRITE)+newtce|=TCE_PCI_READ;+oldtce=xchg(pnv_tce(tbl,idx),cpu_to_be64(newtce));*hpa=be64_to_cpu(oldtce)&~(TCE_PCI_READ|TCE_PCI_WRITE);*direction=iommu_tce_direction(oldtce);
From: Michael Ellerman <mpe@ellerman.id.au> Date: 2016-02-16 03:18:25
On Tue, 2016-12-01 at 04:40:20 UTC, Alexey Kardashevskiy wrote:
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
Tested-by: Douglas Miller <redacted>
Are you planning on sending a non-RFC version of this?
If so is it an urgent fix I should send upstream now? And if so should it
also be CC'ed to stable?
cheers
From: Douglas Miller <hidden> Date: 2016-02-16 13:20:41
On 02/15/2016 09:18 PM, Michael Ellerman wrote:
On Tue, 2016-12-01 at 04:40:20 UTC, Alexey Kardashevskiy wrote:
quoted
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
Tested-by: Douglas Miller <redacted>
Are you planning on sending a non-RFC version of this?
If so is it an urgent fix I should send upstream now? And if so should it
also be CC'ed to stable?
cheers
I vote for the more-urgent case. This is the result of a fairly recent
change in behavior and we may not have found all the adapters that will
break yet. If we get this upstream sooner, it means less exposure.
On Tue, 2016-12-01 at 04:40:20 UTC, Alexey Kardashevskiy wrote:
quoted
Quite often drivers set only "write" permission assuming that this
includes "read" permission as well and this works on plenty platforms.
However IODA2 is strict about this and produces an EEH when "read"
permission is not and reading happens.
This adds a workaround in IODA code to always add the "read" bit when
the "write" bit is set.
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Alexey Kardashevskiy <redacted>
Tested-by: Douglas Miller <redacted>
Are you planning on sending a non-RFC version of this?
Just posted.
If so is it an urgent fix I should send upstream now?
Yes.
> And if so should it also be CC'ed to stable?
Ben suggested that yes, it should. Thanks. Sorry about breaking things.
--
Alexey