Message ID | 1478955748-14819-4-git-send-email-ard.biesheuvel@linaro.org |
---|---|
State | Accepted |
Commit | df8c2668d7277030cf98e7b218549e0466fa5f6b |
Headers | show |
On Sat, Nov 12, 2016 at 02:02:27PM +0100, Ard Biesheuvel wrote: > In preparation of adding support to ArmDmalib for DMA bus masters whose > view of memory is offset by a constant compared to the CPU's view, clean > up some abuse of the device address. > > The device address is not defined in terms of the CPU's address space, > and so it should not be used in CopyMem () or cache maintenance operations > that require a valid mapping. This not only applies to the above use case, > but also to the DebugUncachedMemoryAllocationLib that unmaps the > primary, cached mapping of an allocation, and returns a host address > which is an uncached alias offset by a constant. > > Since we should never access the device address from the CPU, there is > no need to record it in the MAPINFO struct. Instead, record the buffer > address in case of double buffering, since we do need to copy the contents > (in case of a bus master write) and free the buffer (in all cases) when > DmaUnmap() is called. > > Contributed-under: TianoCore Contribution Agreement 1.0 > Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org> For the fix itself: Reviewed-by: Leif Lindholm <leif.lindholm@linaro.org> However, can we wait for a few Tested-by:s to ensure this fix does not reveal any companion bugs? > --- > ArmPkg/Library/ArmDmaLib/ArmDmaLib.c | 15 ++++++++------- > 1 file changed, 8 insertions(+), 7 deletions(-) > > diff --git a/ArmPkg/Library/ArmDmaLib/ArmDmaLib.c b/ArmPkg/Library/ArmDmaLib/ArmDmaLib.c > index c2a44398d25a..7321388de63e 100644 > --- a/ArmPkg/Library/ArmDmaLib/ArmDmaLib.c > +++ b/ArmPkg/Library/ArmDmaLib/ArmDmaLib.c > @@ -27,7 +27,7 @@ > > typedef struct { > EFI_PHYSICAL_ADDRESS HostAddress; > - EFI_PHYSICAL_ADDRESS DeviceAddress; > + VOID *BufferAddress; > UINTN NumberOfBytes; > DMA_MAP_OPERATION Operation; > BOOLEAN DoubleBuffer; > @@ -94,7 +94,7 @@ DmaMap ( > ((*NumberOfBytes & (mCpu->DmaBufferAlignment - 1)) != 0)) { > > // Get the cacheability of the region > - Status = gDS->GetMemorySpaceDescriptor (*DeviceAddress, &GcdDescriptor); > + Status = gDS->GetMemorySpaceDescriptor ((UINTN)HostAddress, &GcdDescriptor); > if (EFI_ERROR(Status)) { > return Status; > } > @@ -128,6 +128,7 @@ DmaMap ( > } > > *DeviceAddress = ConvertToPhysicalAddress ((UINTN)Buffer); > + Map->BufferAddress = Buffer; > } else { > Map->DoubleBuffer = FALSE; > } > @@ -143,7 +144,7 @@ DmaMap ( > // So duplicate the check here when running in DEBUG mode, just to assert > // that we are not trying to create a consistent mapping for cached memory. > // > - Status = gDS->GetMemorySpaceDescriptor (*DeviceAddress, &GcdDescriptor); > + Status = gDS->GetMemorySpaceDescriptor ((UINTN)HostAddress, &GcdDescriptor); > ASSERT_EFI_ERROR(Status); > > ASSERT (Operation != MapOperationBusMasterCommonBuffer || > @@ -152,12 +153,11 @@ DmaMap ( > DEBUG_CODE_END (); > > // Flush the Data Cache (should not have any effect if the memory region is uncached) > - mCpu->FlushDataCache (mCpu, *DeviceAddress, *NumberOfBytes, > + mCpu->FlushDataCache (mCpu, (UINTN)HostAddress, *NumberOfBytes, > EfiCpuFlushTypeWriteBackInvalidate); > } > > Map->HostAddress = (UINTN)HostAddress; > - Map->DeviceAddress = *DeviceAddress; > Map->NumberOfBytes = *NumberOfBytes; > Map->Operation = Operation; > > @@ -200,10 +200,11 @@ DmaUnmap ( > if (Map->Operation == MapOperationBusMasterCommonBuffer) { > Status = EFI_INVALID_PARAMETER; > } else if (Map->Operation == MapOperationBusMasterWrite) { > - CopyMem ((VOID *)(UINTN)Map->HostAddress, (VOID *)(UINTN)Map->DeviceAddress, Map->NumberOfBytes); > + CopyMem ((VOID *)(UINTN)Map->HostAddress, Map->BufferAddress, > + Map->NumberOfBytes); > } > > - DmaFreeBuffer (EFI_SIZE_TO_PAGES (Map->NumberOfBytes), (VOID *)(UINTN)Map->DeviceAddress); > + DmaFreeBuffer (EFI_SIZE_TO_PAGES (Map->NumberOfBytes), Map->BufferAddress); > > } else { > if (Map->Operation == MapOperationBusMasterWrite) { > -- > 2.7.4 > _______________________________________________ edk2-devel mailing list edk2-devel@lists.01.org https://lists.01.org/mailman/listinfo/edk2-devel
On 14 November 2016 at 16:16, Leif Lindholm <leif.lindholm@linaro.org> wrote: > On Sat, Nov 12, 2016 at 02:02:27PM +0100, Ard Biesheuvel wrote: >> In preparation of adding support to ArmDmalib for DMA bus masters whose >> view of memory is offset by a constant compared to the CPU's view, clean >> up some abuse of the device address. >> >> The device address is not defined in terms of the CPU's address space, >> and so it should not be used in CopyMem () or cache maintenance operations >> that require a valid mapping. This not only applies to the above use case, >> but also to the DebugUncachedMemoryAllocationLib that unmaps the >> primary, cached mapping of an allocation, and returns a host address >> which is an uncached alias offset by a constant. >> >> Since we should never access the device address from the CPU, there is >> no need to record it in the MAPINFO struct. Instead, record the buffer >> address in case of double buffering, since we do need to copy the contents >> (in case of a bus master write) and free the buffer (in all cases) when >> DmaUnmap() is called. >> >> Contributed-under: TianoCore Contribution Agreement 1.0 >> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org> > > For the fix itself: > Reviewed-by: Leif Lindholm <leif.lindholm@linaro.org> > > However, can we wait for a few Tested-by:s to ensure this fix does not > reveal any companion bugs? > Perhaps, yes. In case anyone is up to doing that, please find the branch here https://git.linaro.org/people/ard.biesheuvel/uefi-next.git/log/?h=armdmalib-offset However, given that the split CPU/bus master view is introduced in the next patch, the only use case where the device address differs from the host address is when using the DebugUncachedMemoryAllocationLib, which is currently broken AFAICT (it attempts to unmap the linear mapping of the allocation by setting the memory attributes to '0', which triggers an assert in the ArmPkg MMU code) _______________________________________________ edk2-devel mailing list edk2-devel@lists.01.org https://lists.01.org/mailman/listinfo/edk2-devel
On 15 November 2016 at 09:19, Ard Biesheuvel <ard.biesheuvel@linaro.org> wrote: > On 14 November 2016 at 16:16, Leif Lindholm <leif.lindholm@linaro.org> wrote: >> On Sat, Nov 12, 2016 at 02:02:27PM +0100, Ard Biesheuvel wrote: >>> In preparation of adding support to ArmDmalib for DMA bus masters whose >>> view of memory is offset by a constant compared to the CPU's view, clean >>> up some abuse of the device address. >>> >>> The device address is not defined in terms of the CPU's address space, >>> and so it should not be used in CopyMem () or cache maintenance operations >>> that require a valid mapping. This not only applies to the above use case, >>> but also to the DebugUncachedMemoryAllocationLib that unmaps the >>> primary, cached mapping of an allocation, and returns a host address >>> which is an uncached alias offset by a constant. >>> >>> Since we should never access the device address from the CPU, there is >>> no need to record it in the MAPINFO struct. Instead, record the buffer >>> address in case of double buffering, since we do need to copy the contents >>> (in case of a bus master write) and free the buffer (in all cases) when >>> DmaUnmap() is called. >>> >>> Contributed-under: TianoCore Contribution Agreement 1.0 >>> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org> >> >> For the fix itself: >> Reviewed-by: Leif Lindholm <leif.lindholm@linaro.org> >> >> However, can we wait for a few Tested-by:s to ensure this fix does not >> reveal any companion bugs? >> > > Perhaps, yes. > > In case anyone is up to doing that, please find the branch here > https://git.linaro.org/people/ard.biesheuvel/uefi-next.git/log/?h=armdmalib-offset > I tested your branch on the usual victims (R0/1/2, FVP Foundation & AEMv8 and TC2) and they all work fine for me. Tested-by: Ryan Harkin <ryan.harkin@linaro.org> > However, given that the split CPU/bus master view is introduced in the > next patch, the only use case where the device address differs from > the host address is when using the DebugUncachedMemoryAllocationLib, > which is currently broken AFAICT (it attempts to unmap the linear > mapping of the allocation by setting the memory attributes to '0', > which triggers an assert in the ArmPkg MMU code) _______________________________________________ edk2-devel mailing list edk2-devel@lists.01.org https://lists.01.org/mailman/listinfo/edk2-devel
On 15 November 2016 at 11:34, Ryan Harkin <ryan.harkin@linaro.org> wrote: > On 15 November 2016 at 09:19, Ard Biesheuvel <ard.biesheuvel@linaro.org> wrote: >> On 14 November 2016 at 16:16, Leif Lindholm <leif.lindholm@linaro.org> wrote: >>> On Sat, Nov 12, 2016 at 02:02:27PM +0100, Ard Biesheuvel wrote: >>>> In preparation of adding support to ArmDmalib for DMA bus masters whose >>>> view of memory is offset by a constant compared to the CPU's view, clean >>>> up some abuse of the device address. >>>> >>>> The device address is not defined in terms of the CPU's address space, >>>> and so it should not be used in CopyMem () or cache maintenance operations >>>> that require a valid mapping. This not only applies to the above use case, >>>> but also to the DebugUncachedMemoryAllocationLib that unmaps the >>>> primary, cached mapping of an allocation, and returns a host address >>>> which is an uncached alias offset by a constant. >>>> >>>> Since we should never access the device address from the CPU, there is >>>> no need to record it in the MAPINFO struct. Instead, record the buffer >>>> address in case of double buffering, since we do need to copy the contents >>>> (in case of a bus master write) and free the buffer (in all cases) when >>>> DmaUnmap() is called. >>>> >>>> Contributed-under: TianoCore Contribution Agreement 1.0 >>>> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org> >>> >>> For the fix itself: >>> Reviewed-by: Leif Lindholm <leif.lindholm@linaro.org> >>> >>> However, can we wait for a few Tested-by:s to ensure this fix does not >>> reveal any companion bugs? >>> >> >> Perhaps, yes. >> >> In case anyone is up to doing that, please find the branch here >> https://git.linaro.org/people/ard.biesheuvel/uefi-next.git/log/?h=armdmalib-offset >> > > I tested your branch on the usual victims (R0/1/2, FVP Foundation & > AEMv8 and TC2) and they all work fine for me. > > Tested-by: Ryan Harkin <ryan.harkin@linaro.org> > Thanks! _______________________________________________ edk2-devel mailing list edk2-devel@lists.01.org https://lists.01.org/mailman/listinfo/edk2-devel
On Tue, Nov 15, 2016 at 01:07:37PM +0000, Ard Biesheuvel wrote: > On 15 November 2016 at 11:34, Ryan Harkin <ryan.harkin@linaro.org> wrote: > >>>> Since we should never access the device address from the CPU, there is > >>>> no need to record it in the MAPINFO struct. Instead, record the buffer > >>>> address in case of double buffering, since we do need to copy the contents > >>>> (in case of a bus master write) and free the buffer (in all cases) when > >>>> DmaUnmap() is called. > >>>> > >>>> Contributed-under: TianoCore Contribution Agreement 1.0 > >>>> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org> > >>> > >>> For the fix itself: > >>> Reviewed-by: Leif Lindholm <leif.lindholm@linaro.org> > >>> > >>> However, can we wait for a few Tested-by:s to ensure this fix does not > >>> reveal any companion bugs? > >>> > >> > >> Perhaps, yes. > >> > >> In case anyone is up to doing that, please find the branch here > >> https://git.linaro.org/people/ard.biesheuvel/uefi-next.git/log/?h=armdmalib-offset > >> > > > > I tested your branch on the usual victims (R0/1/2, FVP Foundation & > > AEMv8 and TC2) and they all work fine for me. > > > > Tested-by: Ryan Harkin <ryan.harkin@linaro.org> OK, so on the off-hand chance that this does break some non-upstream drivers, I'll wait for further comments until I'm back from holiday, 28 November, before pushing this. Additional Tested-by:s would still be appreciated. Regards, Leif _______________________________________________ edk2-devel mailing list edk2-devel@lists.01.org https://lists.01.org/mailman/listinfo/edk2-devel
On 15 November 2016 at 18:01, Leif Lindholm <leif.lindholm@linaro.org> wrote: >> >> In case anyone is up to doing that, please find the branch here >> >> https://git.linaro.org/people/ard.biesheuvel/uefi-next.git/log/?h=armdmalib-offset >> > >> > I tested your branch on the usual victims (R0/1/2, FVP Foundation & >> > AEMv8 and TC2) and they all work fine for me. >> > >> > Tested-by: Ryan Harkin <ryan.harkin@linaro.org> > > OK, so on the off-hand chance that this does break some non-upstream > drivers, I'll wait for further comments until I'm back from holiday, > 28 November, before pushing this. Series pushed as 3571884..bfe3427 - thanks! / Leif _______________________________________________ edk2-devel mailing list edk2-devel@lists.01.org https://lists.01.org/mailman/listinfo/edk2-devel
diff --git a/ArmPkg/Library/ArmDmaLib/ArmDmaLib.c b/ArmPkg/Library/ArmDmaLib/ArmDmaLib.c index c2a44398d25a..7321388de63e 100644 --- a/ArmPkg/Library/ArmDmaLib/ArmDmaLib.c +++ b/ArmPkg/Library/ArmDmaLib/ArmDmaLib.c @@ -27,7 +27,7 @@ typedef struct { EFI_PHYSICAL_ADDRESS HostAddress; - EFI_PHYSICAL_ADDRESS DeviceAddress; + VOID *BufferAddress; UINTN NumberOfBytes; DMA_MAP_OPERATION Operation; BOOLEAN DoubleBuffer; @@ -94,7 +94,7 @@ DmaMap ( ((*NumberOfBytes & (mCpu->DmaBufferAlignment - 1)) != 0)) { // Get the cacheability of the region - Status = gDS->GetMemorySpaceDescriptor (*DeviceAddress, &GcdDescriptor); + Status = gDS->GetMemorySpaceDescriptor ((UINTN)HostAddress, &GcdDescriptor); if (EFI_ERROR(Status)) { return Status; } @@ -128,6 +128,7 @@ DmaMap ( } *DeviceAddress = ConvertToPhysicalAddress ((UINTN)Buffer); + Map->BufferAddress = Buffer; } else { Map->DoubleBuffer = FALSE; } @@ -143,7 +144,7 @@ DmaMap ( // So duplicate the check here when running in DEBUG mode, just to assert // that we are not trying to create a consistent mapping for cached memory. // - Status = gDS->GetMemorySpaceDescriptor (*DeviceAddress, &GcdDescriptor); + Status = gDS->GetMemorySpaceDescriptor ((UINTN)HostAddress, &GcdDescriptor); ASSERT_EFI_ERROR(Status); ASSERT (Operation != MapOperationBusMasterCommonBuffer || @@ -152,12 +153,11 @@ DmaMap ( DEBUG_CODE_END (); // Flush the Data Cache (should not have any effect if the memory region is uncached) - mCpu->FlushDataCache (mCpu, *DeviceAddress, *NumberOfBytes, + mCpu->FlushDataCache (mCpu, (UINTN)HostAddress, *NumberOfBytes, EfiCpuFlushTypeWriteBackInvalidate); } Map->HostAddress = (UINTN)HostAddress; - Map->DeviceAddress = *DeviceAddress; Map->NumberOfBytes = *NumberOfBytes; Map->Operation = Operation; @@ -200,10 +200,11 @@ DmaUnmap ( if (Map->Operation == MapOperationBusMasterCommonBuffer) { Status = EFI_INVALID_PARAMETER; } else if (Map->Operation == MapOperationBusMasterWrite) { - CopyMem ((VOID *)(UINTN)Map->HostAddress, (VOID *)(UINTN)Map->DeviceAddress, Map->NumberOfBytes); + CopyMem ((VOID *)(UINTN)Map->HostAddress, Map->BufferAddress, + Map->NumberOfBytes); } - DmaFreeBuffer (EFI_SIZE_TO_PAGES (Map->NumberOfBytes), (VOID *)(UINTN)Map->DeviceAddress); + DmaFreeBuffer (EFI_SIZE_TO_PAGES (Map->NumberOfBytes), Map->BufferAddress); } else { if (Map->Operation == MapOperationBusMasterWrite) {
In preparation of adding support to ArmDmalib for DMA bus masters whose view of memory is offset by a constant compared to the CPU's view, clean up some abuse of the device address. The device address is not defined in terms of the CPU's address space, and so it should not be used in CopyMem () or cache maintenance operations that require a valid mapping. This not only applies to the above use case, but also to the DebugUncachedMemoryAllocationLib that unmaps the primary, cached mapping of an allocation, and returns a host address which is an uncached alias offset by a constant. Since we should never access the device address from the CPU, there is no need to record it in the MAPINFO struct. Instead, record the buffer address in case of double buffering, since we do need to copy the contents (in case of a bus master write) and free the buffer (in all cases) when DmaUnmap() is called. Contributed-under: TianoCore Contribution Agreement 1.0 Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org> --- ArmPkg/Library/ArmDmaLib/ArmDmaLib.c | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-) -- 2.7.4 _______________________________________________ edk2-devel mailing list edk2-devel@lists.01.org https://lists.01.org/mailman/listinfo/edk2-devel