| Message ID | 20220315162455.5190-2-mario.limonciello@amd.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [1/2] iommu/amd: Add support to indicate whether DMA remap support is enabled | expand |
On Tue, Mar 15, 2022 at 11:24:55AM -0500, Mario Limonciello wrote: > - * handled natively using IOMMU. It is enabled when IOMMU is > - * enabled and ACPI DMAR table has DMAR_PLATFORM_OPT_IN set. > + * handled natively using IOMMU. It is enabled when the IOMMU is > + * enabled and either: > + * ACPI DMAR table has DMAR_PLATFORM_OPT_IN set > + * or > + * ACPI IVRS table has DMA_REMAP bitset > */ > return sprintf(buf, "%d\n", > - iommu_present(&pci_bus_type) && dmar_platform_optin()); > + iommu_present(&pci_bus_type) && > + (dmar_platform_optin() || amd_ivrs_remap_support())); Yikes. No, the thunderbot code does not have any business poking into either dmar_platform_optin or amd_ivrs_remap_support. This needs a proper abstration from the IOMMU code.
[Public] > On Tue, Mar 15, 2022 at 11:24:55AM -0500, Mario Limonciello wrote: > > - * handled natively using IOMMU. It is enabled when IOMMU is > > - * enabled and ACPI DMAR table has DMAR_PLATFORM_OPT_IN set. > > + * handled natively using IOMMU. It is enabled when the IOMMU is > > + * enabled and either: > > + * ACPI DMAR table has DMAR_PLATFORM_OPT_IN set > > + * or > > + * ACPI IVRS table has DMA_REMAP bitset > > */ > > return sprintf(buf, "%d\n", > > - iommu_present(&pci_bus_type) && > dmar_platform_optin()); > > + iommu_present(&pci_bus_type) && > > + (dmar_platform_optin() || amd_ivrs_remap_support())); > > Yikes. No, the thunderbot code does not have any business poking into > either dmar_platform_optin or amd_ivrs_remap_support. This needs > a proper abstration from the IOMMU code. To make sure I follow your ask - it's to make a new generic iommu function That would check dmar/ivrs, and switch out thunderbolt domain.c to use the symbol? I'm happy to rework that if that is what you want. Do you have a preferred proposed function name for that?
On 2022-03-15 16:54, Limonciello, Mario via iommu wrote: > [Public] > > >> On Tue, Mar 15, 2022 at 11:24:55AM -0500, Mario Limonciello wrote: >>> - * handled natively using IOMMU. It is enabled when IOMMU is >>> - * enabled and ACPI DMAR table has DMAR_PLATFORM_OPT_IN set. >>> + * handled natively using IOMMU. It is enabled when the IOMMU is >>> + * enabled and either: >>> + * ACPI DMAR table has DMAR_PLATFORM_OPT_IN set >>> + * or >>> + * ACPI IVRS table has DMA_REMAP bitset >>> */ >>> return sprintf(buf, "%d\n", >>> - iommu_present(&pci_bus_type) && >> dmar_platform_optin()); >>> + iommu_present(&pci_bus_type) && >>> + (dmar_platform_optin() || amd_ivrs_remap_support())); >> >> Yikes. No, the thunderbot code does not have any business poking into >> either dmar_platform_optin or amd_ivrs_remap_support. This needs >> a proper abstration from the IOMMU code. > > To make sure I follow your ask - it's to make a new generic iommu function > That would check dmar/ivrs, and switch out thunderbolt domain.c to use the > symbol? > > I'm happy to rework that if that is what you want. > Do you have a preferred proposed function name for that? But why? Either IOMMU translation is enabled or it isn't, and if it is, what's to gain from guessing at *why* it might have been? And even if the IOMMU's firmware table did tell the IOMMU driver to enable the IOMMU, why should that be Thunderbolt's business? Furthermore, looking at patch #1 I can only conclude that this is entirely meaningless anyway. AFAICS it's literally reporting whether the firmware flag was set or not. Not whether it's actually been honoured and the IOMMU is enforcing any kind of DMA protection at all. Even on Intel where the flag does at least have some effect on the IOMMU driver, that can still be overridden. I already have a patch refactoring this to get rid of iommu_present(), but at the time I wasn't looking to closely at what it's trying to *do* with the information. If it's supposed to accurately reflect whether the Thunderbolt device is subject to IOMMU translation and not bypassed, I can fix that too (and unexport dmar_platform_optin() in the process...) Robin.
+ Christian Kellner (Bolt userspace maintainer) On 3/15/2022 13:07, Robin Murphy wrote: > On 2022-03-15 16:54, Limonciello, Mario via iommu wrote: >> [Public] >> >> >>> On Tue, Mar 15, 2022 at 11:24:55AM -0500, Mario Limonciello wrote: >>>> - * handled natively using IOMMU. It is enabled when IOMMU is >>>> - * enabled and ACPI DMAR table has DMAR_PLATFORM_OPT_IN set. >>>> + * handled natively using IOMMU. It is enabled when the IOMMU is >>>> + * enabled and either: >>>> + * ACPI DMAR table has DMAR_PLATFORM_OPT_IN set >>>> + * or >>>> + * ACPI IVRS table has DMA_REMAP bitset >>>> */ >>>> return sprintf(buf, "%d\n", >>>> - iommu_present(&pci_bus_type) && >>> dmar_platform_optin()); >>>> + iommu_present(&pci_bus_type) && >>>> + (dmar_platform_optin() || amd_ivrs_remap_support())); >>> >>> Yikes. No, the thunderbot code does not have any business poking into >>> either dmar_platform_optin or amd_ivrs_remap_support. This needs >>> a proper abstration from the IOMMU code. >> >> To make sure I follow your ask - it's to make a new generic iommu >> function >> That would check dmar/ivrs, and switch out thunderbolt domain.c to use >> the >> symbol? >> >> I'm happy to rework that if that is what you want. >> Do you have a preferred proposed function name for that? > > But why? Either IOMMU translation is enabled or it isn't, and if it is, > what's to gain from guessing at *why* it might have been? And even if > the IOMMU's firmware table did tell the IOMMU driver to enable the > IOMMU, why should that be Thunderbolt's business? A lot of this comes from baggage from early Thunderbolt 3 implementation on systems with ICM (Intel's FW CM). On those systems there was a concept of "Security Levels". This meant that downstream PCIe devices were not automatically authorized when a TBT3 device was plugged in. In those cases there was no guarantee that the IOMMU was in use and so the security was passed on to the user to make a decision. In Linux this was accomplished using the 'authorized' attribute in /sys/bus/thunderbolt/devices/$NUM/authorized. When this was set to 1 then the TBT3 device and PCIe topology behind it would be enumerated. Further documentation explaining how this works is available here: https://www.kernel.org/doc/html/latest/admin-guide/thunderbolt.html#security-levels-and-how-to-use-them (Intel based) Platforms from 2018+ w/ TBT3 started to use the IOMMU consistently at runtime but had this existing implementation of security levels to worry about. Furthermore tunnels could be created pre-boot, and so the thunderbolt driver may or may not re-create them based on policy. So a new attribute was created "iommu_dma_protection" that userspace could use as part of a policy decision to automatically authorize devices. Exporting this attribute is very similar to what Microsoft does to let the user see the security of the system. https://docs.microsoft.com/en-us/windows-hardware/design/device-experiences/oem-kernel-dma-protection In Linux today some userspace software "bolt" has a policy included by default that will automatically authorize TBT3 and USB4 (w/ PCIe) devices when iommu_dma_protection is set to 1. > > Furthermore, looking at patch #1 I can only conclude that this is > entirely meaningless anyway. AFAICS it's literally reporting whether the > firmware flag was set or not. Not whether it's actually been honoured > and the IOMMU is enforcing any kind of DMA protection at all. Even on > Intel where the flag does at least have some effect on the IOMMU driver, > that can still be overridden. Take a look at the Microsoft link I shared above. They also make policy decisions based on the information in these tables. > > I already have a patch refactoring this to get rid of iommu_present(), > but at the time I wasn't looking to closely at what it's trying to *do* > with the information. If it's supposed to accurately reflect whether the > Thunderbolt device is subject to IOMMU translation and not bypassed, I > can fix that too (and unexport dmar_platform_optin() in the process...) > > Robin. This patch series stems from that history. To give the best experience to end users you want hotplugged devices to be automatically authorized when software says it's safe to do so. To summarize the flow: * User plugs in device * USB4 CM will query supported tunnels * USB4 CM will create devices in /sys/bus/thunderbolt/devices for new plugged in TBT3/USB4 device * "authorized" attribute will default to "0" and PCIe tunnels are not created * Userspace gets a uevent that the device was added * Userspace (bolt) reacts by reading /sys/bus/thunderbolt/devices/domainX/iommu_dma_protection * If that is set to "1", bolt will write "1" to "authorized" and USB4 CM will create PCIe tunnels * If that is set to "0", bolt will send an event to GUI to show a popup asking to authorize the device * After user acks the authorization then it will write "1" to "authorized" and USB4 CM will create PCIe tunnels Mika, I wonder if maybe what we really want is to only use that flow for the authorized attribute when using TBT3 + ICM (or IOMMU disabled at runtime). If we're using a USB4 host, check IOMMU translation layer active like Robin suggested and then automatically authorize from the CM. Thoughts?
[Public] > -----Original Message----- > From: Limonciello, Mario > Sent: Tuesday, March 15, 2022 13:36 > To: Robin Murphy <robin.murphy@arm.com>; Christoph Hellwig > <hch@infradead.org>; christian@kellner.me; Mika Westerberg > <mika.westerberg@linux.intel.com> > Cc: Michael Jamet <michael.jamet@intel.com>; open list:THUNDERBOLT > DRIVER <linux-usb@vger.kernel.org>; open list <linux- > kernel@vger.kernel.org>; Yehezkel Bernat <YehezkelShB@gmail.com>; > open list:AMD IOMMU (AMD-VI) <iommu@lists.linux-foundation.org>; > Andreas Noever <andreas.noever@gmail.com>; Will Deacon > <will@kernel.org> > Subject: Re: [PATCH 2/2] thunderbolt: Use pre-boot DMA protection on AMD > systems > > + Christian Kellner (Bolt userspace maintainer) > > On 3/15/2022 13:07, Robin Murphy wrote: > > On 2022-03-15 16:54, Limonciello, Mario via iommu wrote: > >> [Public] > >> > >> > >>> On Tue, Mar 15, 2022 at 11:24:55AM -0500, Mario Limonciello wrote: > >>>> - * handled natively using IOMMU. It is enabled when IOMMU is > >>>> - * enabled and ACPI DMAR table has DMAR_PLATFORM_OPT_IN > set. > >>>> + * handled natively using IOMMU. It is enabled when the IOMMU is > >>>> + * enabled and either: > >>>> + * ACPI DMAR table has DMAR_PLATFORM_OPT_IN set > >>>> + * or > >>>> + * ACPI IVRS table has DMA_REMAP bitset > >>>> */ > >>>> return sprintf(buf, "%d\n", > >>>> - iommu_present(&pci_bus_type) && > >>> dmar_platform_optin()); > >>>> + iommu_present(&pci_bus_type) && > >>>> + (dmar_platform_optin() || amd_ivrs_remap_support())); > >>> > >>> Yikes. No, the thunderbot code does not have any business poking into > >>> either dmar_platform_optin or amd_ivrs_remap_support. This needs > >>> a proper abstration from the IOMMU code. > >> > >> To make sure I follow your ask - it's to make a new generic iommu > >> function > >> That would check dmar/ivrs, and switch out thunderbolt domain.c to use > >> the > >> symbol? > >> > >> I'm happy to rework that if that is what you want. > >> Do you have a preferred proposed function name for that? > > > > But why? Either IOMMU translation is enabled or it isn't, and if it is, > > what's to gain from guessing at *why* it might have been? And even if > > the IOMMU's firmware table did tell the IOMMU driver to enable the > > IOMMU, why should that be Thunderbolt's business? > A lot of this comes from baggage from early Thunderbolt 3 implementation > on systems with ICM (Intel's FW CM). On those systems there was a > concept of "Security Levels". This meant that downstream PCIe devices > were not automatically authorized when a TBT3 device was plugged in. In > those cases there was no guarantee that the IOMMU was in use and so the > security was passed on to the user to make a decision. > > In Linux this was accomplished using the 'authorized' attribute in > /sys/bus/thunderbolt/devices/$NUM/authorized. When this was set to 1 > then the TBT3 device and PCIe topology behind it would be enumerated. > > Further documentation explaining how this works is available here: > https://www.kernel.org/doc/html/latest/admin- > guide/thunderbolt.html#security-levels-and-how-to-use-them > > (Intel based) Platforms from 2018+ w/ TBT3 started to use the IOMMU > consistently at runtime but had this existing implementation of security > levels to worry about. Furthermore tunnels could be created pre-boot, > and so the thunderbolt driver may or may not re-create them based on > policy. > > So a new attribute was created "iommu_dma_protection" that userspace > could use as part of a policy decision to automatically authorize > devices. Exporting this attribute is very similar to what Microsoft > does to let the user see the security of the system. > > https://docs.microsoft.com/en-us/windows-hardware/design/device- > experiences/oem-kernel-dma-protection > > In Linux today some userspace software "bolt" has a policy included by > default that will automatically authorize TBT3 and USB4 (w/ PCIe) > devices when iommu_dma_protection is set to 1. > > > > > Furthermore, looking at patch #1 I can only conclude that this is > > entirely meaningless anyway. AFAICS it's literally reporting whether the > > firmware flag was set or not. Not whether it's actually been honoured > > and the IOMMU is enforcing any kind of DMA protection at all. Even on > > Intel where the flag does at least have some effect on the IOMMU driver, > > that can still be overridden. > > Take a look at the Microsoft link I shared above. They also make policy > decisions based on the information in these tables. > > > > > I already have a patch refactoring this to get rid of iommu_present(), > > but at the time I wasn't looking to closely at what it's trying to *do* > > with the information. If it's supposed to accurately reflect whether the > > Thunderbolt device is subject to IOMMU translation and not bypassed, I > > can fix that too (and unexport dmar_platform_optin() in the process...) > > > > Robin. > > This patch series stems from that history. To give the best experience > to end users you want hotplugged devices to be automatically authorized > when software says it's safe to do so. > > To summarize the flow: > * User plugs in device > * USB4 CM will query supported tunnels > * USB4 CM will create devices in /sys/bus/thunderbolt/devices for new > plugged in TBT3/USB4 device > * "authorized" attribute will default to "0" and PCIe tunnels are not > created > * Userspace gets a uevent that the device was added > * Userspace (bolt) reacts by reading > /sys/bus/thunderbolt/devices/domainX/iommu_dma_protection > * If that is set to "1", bolt will write "1" to "authorized" and USB4 > CM will create PCIe tunnels > * If that is set to "0", bolt will send an event to GUI to show a popup > asking to authorize the device > * After user acks the authorization then it will write "1" to > "authorized" and USB4 CM will create PCIe tunnels > > > Mika, > > I wonder if maybe what we really want is to only use that flow for the > authorized attribute when using TBT3 + ICM (or IOMMU disabled at > runtime). If we're using a USB4 host, check IOMMU translation layer > active like Robin suggested and then automatically authorize from the CM. > > Thoughts? > > I put an RFC together with what this idea looks like, comments welcome. https://lore.kernel.org/linux-usb/20220315213008.5357-1-mario.limonciello@amd.com/T/#u
On 2022-03-15 18:36, Limonciello, Mario wrote: > + Christian Kellner (Bolt userspace maintainer) > > On 3/15/2022 13:07, Robin Murphy wrote: >> On 2022-03-15 16:54, Limonciello, Mario via iommu wrote: >>> [Public] >>> >>> >>>> On Tue, Mar 15, 2022 at 11:24:55AM -0500, Mario Limonciello wrote: >>>>> - * handled natively using IOMMU. It is enabled when IOMMU is >>>>> - * enabled and ACPI DMAR table has DMAR_PLATFORM_OPT_IN set. >>>>> + * handled natively using IOMMU. It is enabled when the IOMMU is >>>>> + * enabled and either: >>>>> + * ACPI DMAR table has DMAR_PLATFORM_OPT_IN set >>>>> + * or >>>>> + * ACPI IVRS table has DMA_REMAP bitset >>>>> */ >>>>> return sprintf(buf, "%d\n", >>>>> - iommu_present(&pci_bus_type) && >>>> dmar_platform_optin()); >>>>> + iommu_present(&pci_bus_type) && >>>>> + (dmar_platform_optin() || amd_ivrs_remap_support())); >>>> >>>> Yikes. No, the thunderbot code does not have any business poking into >>>> either dmar_platform_optin or amd_ivrs_remap_support. This needs >>>> a proper abstration from the IOMMU code. >>> >>> To make sure I follow your ask - it's to make a new generic iommu >>> function >>> That would check dmar/ivrs, and switch out thunderbolt domain.c to >>> use the >>> symbol? >>> >>> I'm happy to rework that if that is what you want. >>> Do you have a preferred proposed function name for that? >> >> But why? Either IOMMU translation is enabled or it isn't, and if it >> is, what's to gain from guessing at *why* it might have been? And even >> if the IOMMU's firmware table did tell the IOMMU driver to enable the >> IOMMU, why should that be Thunderbolt's business? > A lot of this comes from baggage from early Thunderbolt 3 implementation > on systems with ICM (Intel's FW CM). On those systems there was a > concept of "Security Levels". This meant that downstream PCIe devices > were not automatically authorized when a TBT3 device was plugged in. In > those cases there was no guarantee that the IOMMU was in use and so the > security was passed on to the user to make a decision. > > In Linux this was accomplished using the 'authorized' attribute in > /sys/bus/thunderbolt/devices/$NUM/authorized. When this was set to 1 > then the TBT3 device and PCIe topology behind it would be enumerated. > > Further documentation explaining how this works is available here: > https://www.kernel.org/doc/html/latest/admin-guide/thunderbolt.html#security-levels-and-how-to-use-them > > > (Intel based) Platforms from 2018+ w/ TBT3 started to use the IOMMU > consistently at runtime but had this existing implementation of security > levels to worry about. Furthermore tunnels could be created pre-boot, > and so the thunderbolt driver may or may not re-create them based on > policy. > > So a new attribute was created "iommu_dma_protection" that userspace > could use as part of a policy decision to automatically authorize > devices. Exporting this attribute is very similar to what Microsoft > does to let the user see the security of the system. > > https://docs.microsoft.com/en-us/windows-hardware/design/device-experiences/oem-kernel-dma-protection > > > In Linux today some userspace software "bolt" has a policy included by > default that will automatically authorize TBT3 and USB4 (w/ PCIe) > devices when iommu_dma_protection is set to 1. > >> >> Furthermore, looking at patch #1 I can only conclude that this is >> entirely meaningless anyway. AFAICS it's literally reporting whether >> the firmware flag was set or not. Not whether it's actually been >> honoured and the IOMMU is enforcing any kind of DMA protection at all. >> Even on Intel where the flag does at least have some effect on the >> IOMMU driver, that can still be overridden. > > Take a look at the Microsoft link I shared above. They also make policy > decisions based on the information in these tables. > >> >> I already have a patch refactoring this to get rid of iommu_present(), >> but at the time I wasn't looking to closely at what it's trying to >> *do* with the information. If it's supposed to accurately reflect >> whether the Thunderbolt device is subject to IOMMU translation and not >> bypassed, I can fix that too (and unexport dmar_platform_optin() in >> the process...) >> >> Robin. > > This patch series stems from that history. To give the best experience > to end users you want hotplugged devices to be automatically authorized > when software says it's safe to do so. > > To summarize the flow: > * User plugs in device > * USB4 CM will query supported tunnels > * USB4 CM will create devices in /sys/bus/thunderbolt/devices for new > plugged in TBT3/USB4 device > * "authorized" attribute will default to "0" and PCIe tunnels are not > created > * Userspace gets a uevent that the device was added > * Userspace (bolt) reacts by reading > /sys/bus/thunderbolt/devices/domainX/iommu_dma_protection > * If that is set to "1", bolt will write "1" to "authorized" and USB4 > CM will create PCIe tunnels > * If that is set to "0", bolt will send an event to GUI to show a popup > asking to authorize the device > * After user acks the authorization then it will write "1" to > "authorized" and USB4 CM will create PCIe tunnels > > > Mika, > > I wonder if maybe what we really want is to only use that flow for the > authorized attribute when using TBT3 + ICM (or IOMMU disabled at > runtime). If we're using a USB4 host, check IOMMU translation layer > active like Robin suggested and then automatically authorize from the CM. Thanks for the explanation. I don't think there's anything wrong with that flow per se - fundamentally, whether it's relayed through userspace or done automagically inside the kernel doesn't change the end result - but it does seem to confirm my suspicion that even now it's not actually working as intended and may end up letting devices be authorised in circumstances that they probably shouldn't be. It's absolutely fine for Thunderbolt to care about whether a device currently has IOMMU translation enabled (and to expose that to userspace in its own way if it wants to), but that's generic IOMMU API stuff, no firmware-poking required :) Tomorrow I'll rework the patch out of my iommu_present() cleanup stack to do the right thing, and share it. Cheers, Robin.
Hi Mario,
Thank you for the patch! Yet something to improve:
[auto build test ERROR on joro-iommu/next]
[also build test ERROR on arm-perf/for-next/perf linus/master v5.17-rc8 next-20220315]
[If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch]
url: https://github.com/0day-ci/linux/commits/Mario-Limonciello/iommu-amd-Add-support-to-indicate-whether-DMA-remap-support-is-enabled/20220316-002821
base: https://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu.git next
config: x86_64-randconfig-a012 (https://download.01.org/0day-ci/archive/20220316/202203160904.VB4alCdg-lkp@intel.com/config)
compiler: clang version 15.0.0 (https://github.com/llvm/llvm-project a6b2f50fb47da3baeee10b1906da6e30ac5d26ec)
reproduce (this is a W=1 build):
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
chmod +x ~/bin/make.cross
# https://github.com/0day-ci/linux/commit/9b0b7079d348c607cba7af4c87eaae1a79e52d91
git remote add linux-review https://github.com/0day-ci/linux
git fetch --no-tags linux-review Mario-Limonciello/iommu-amd-Add-support-to-indicate-whether-DMA-remap-support-is-enabled/20220316-002821
git checkout 9b0b7079d348c607cba7af4c87eaae1a79e52d91
# save the config file to linux build tree
mkdir build_dir
COMPILER_INSTALL_PATH=$HOME/0day COMPILER=clang make.cross W=1 O=build_dir ARCH=x86_64 SHELL=/bin/bash drivers/
If you fix the issue, kindly add following tag as appropriate
Reported-by: kernel test robot <lkp@intel.com>
All errors (new ones prefixed by >>):
In file included from drivers/thunderbolt/domain.c:9:
>> include/linux/amd-iommu.h:159:52: error: use of undeclared identifier 'ENODEV'
static inline int amd_iommu_detect(void) { return -ENODEV; }
^
1 error generated.
vim +/ENODEV +159 include/linux/amd-iommu.h
6a9401a7ac13e6 arch/x86/include/asm/amd_iommu.h Joerg Roedel 2009-11-20 158
480125ba49ba62 arch/x86/include/asm/amd_iommu.h Konrad Rzeszutek Wilk 2010-08-26 @159 static inline int amd_iommu_detect(void) { return -ENODEV; }
6a9401a7ac13e6 arch/x86/include/asm/amd_iommu.h Joerg Roedel 2009-11-20 160
---
0-DAY CI Kernel Test Service
https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org
Hi, On Tue, Mar 15, 2022 at 01:36:11PM -0500, Limonciello, Mario wrote: > + Christian Kellner (Bolt userspace maintainer) > > On 3/15/2022 13:07, Robin Murphy wrote: > > On 2022-03-15 16:54, Limonciello, Mario via iommu wrote: > > > [Public] > > > > > > > > > > On Tue, Mar 15, 2022 at 11:24:55AM -0500, Mario Limonciello wrote: > > > > > - * handled natively using IOMMU. It is enabled when IOMMU is > > > > > - * enabled and ACPI DMAR table has DMAR_PLATFORM_OPT_IN set. > > > > > + * handled natively using IOMMU. It is enabled when the IOMMU is > > > > > + * enabled and either: > > > > > + * ACPI DMAR table has DMAR_PLATFORM_OPT_IN set > > > > > + * or > > > > > + * ACPI IVRS table has DMA_REMAP bitset > > > > > */ > > > > > return sprintf(buf, "%d\n", > > > > > - iommu_present(&pci_bus_type) && > > > > dmar_platform_optin()); > > > > > + iommu_present(&pci_bus_type) && > > > > > + (dmar_platform_optin() || amd_ivrs_remap_support())); > > > > > > > > Yikes. No, the thunderbot code does not have any business poking into > > > > either dmar_platform_optin or amd_ivrs_remap_support. This needs > > > > a proper abstration from the IOMMU code. I agree. When it was originally added it was only the DMAR (Intel) based platforms that provided this hint so adding an abstraction for that did not make much sense. Now, since we are seeing more and more USB4 host controllers and many of them support PCIe tunneling (and IOMMU) adding an API makes more sense. > > > > > > To make sure I follow your ask - it's to make a new generic iommu > > > function > > > That would check dmar/ivrs, and switch out thunderbolt domain.c to > > > use the > > > symbol? > > > > > > I'm happy to rework that if that is what you want. > > > Do you have a preferred proposed function name for that? > > > > But why? Either IOMMU translation is enabled or it isn't, and if it is, > > what's to gain from guessing at *why* it might have been? And even if > > the IOMMU's firmware table did tell the IOMMU driver to enable the > > IOMMU, why should that be Thunderbolt's business? > A lot of this comes from baggage from early Thunderbolt 3 implementation on > systems with ICM (Intel's FW CM). On those systems there was a concept of > "Security Levels". This meant that downstream PCIe devices were not > automatically authorized when a TBT3 device was plugged in. In those cases > there was no guarantee that the IOMMU was in use and so the security was > passed on to the user to make a decision. > > In Linux this was accomplished using the 'authorized' attribute in > /sys/bus/thunderbolt/devices/$NUM/authorized. When this was set to 1 then > the TBT3 device and PCIe topology behind it would be enumerated. > > Further documentation explaining how this works is available here: > https://www.kernel.org/doc/html/latest/admin-guide/thunderbolt.html#security-levels-and-how-to-use-them > > (Intel based) Platforms from 2018+ w/ TBT3 started to use the IOMMU > consistently at runtime but had this existing implementation of security > levels to worry about. Furthermore tunnels could be created pre-boot, and > so the thunderbolt driver may or may not re-create them based on policy. > > So a new attribute was created "iommu_dma_protection" that userspace could > use as part of a policy decision to automatically authorize devices. > Exporting this attribute is very similar to what Microsoft does to let the > user see the security of the system. > > https://docs.microsoft.com/en-us/windows-hardware/design/device-experiences/oem-kernel-dma-protection > > In Linux today some userspace software "bolt" has a policy included by > default that will automatically authorize TBT3 and USB4 (w/ PCIe) devices > when iommu_dma_protection is set to 1. > > > > > Furthermore, looking at patch #1 I can only conclude that this is > > entirely meaningless anyway. AFAICS it's literally reporting whether the > > firmware flag was set or not. Not whether it's actually been honoured > > and the IOMMU is enforcing any kind of DMA protection at all. Even on > > Intel where the flag does at least have some effect on the IOMMU driver, > > that can still be overridden. > > Take a look at the Microsoft link I shared above. They also make policy > decisions based on the information in these tables. > > > > > I already have a patch refactoring this to get rid of iommu_present(), > > but at the time I wasn't looking to closely at what it's trying to *do* > > with the information. If it's supposed to accurately reflect whether the > > Thunderbolt device is subject to IOMMU translation and not bypassed, I > > can fix that too (and unexport dmar_platform_optin() in the process...) > > > > Robin. > > This patch series stems from that history. To give the best experience to > end users you want hotplugged devices to be automatically authorized when > software says it's safe to do so. > > To summarize the flow: > * User plugs in device > * USB4 CM will query supported tunnels > * USB4 CM will create devices in /sys/bus/thunderbolt/devices for new > plugged in TBT3/USB4 device > * "authorized" attribute will default to "0" and PCIe tunnels are not > created > * Userspace gets a uevent that the device was added > * Userspace (bolt) reacts by reading > /sys/bus/thunderbolt/devices/domainX/iommu_dma_protection > * If that is set to "1", bolt will write "1" to "authorized" and USB4 CM > will create PCIe tunnels > * If that is set to "0", bolt will send an event to GUI to show a popup > asking to authorize the device > * After user acks the authorization then it will write "1" to "authorized" > and USB4 CM will create PCIe tunnels > > > Mika, > > I wonder if maybe what we really want is to only use that flow for the > authorized attribute when using TBT3 + ICM (or IOMMU disabled at runtime). > If we're using a USB4 host, check IOMMU translation layer active like Robin > suggested and then automatically authorize from the CM. I would still leave that policy to userspace to decide.
diff --git a/drivers/thunderbolt/domain.c b/drivers/thunderbolt/domain.c index 7018d959f775..e03790735c12 100644 --- a/drivers/thunderbolt/domain.c +++ b/drivers/thunderbolt/domain.c @@ -6,6 +6,7 @@ * Author: Mika Westerberg <mika.westerberg@linux.intel.com> */ +#include <linux/amd-iommu.h> #include <linux/device.h> #include <linux/dmar.h> #include <linux/idr.h> @@ -259,11 +260,15 @@ static ssize_t iommu_dma_protection_show(struct device *dev, { /* * Kernel DMA protection is a feature where Thunderbolt security is - * handled natively using IOMMU. It is enabled when IOMMU is - * enabled and ACPI DMAR table has DMAR_PLATFORM_OPT_IN set. + * handled natively using IOMMU. It is enabled when the IOMMU is + * enabled and either: + * ACPI DMAR table has DMAR_PLATFORM_OPT_IN set + * or + * ACPI IVRS table has DMA_REMAP bitset */ return sprintf(buf, "%d\n", - iommu_present(&pci_bus_type) && dmar_platform_optin()); + iommu_present(&pci_bus_type) && + (dmar_platform_optin() || amd_ivrs_remap_support())); } static DEVICE_ATTR_RO(iommu_dma_protection);
The information is exported from the IOMMU driver whether or not pre-boot DMA protection has been enabled on AMD systems. Use this information to properly set iomma_dma_protection. Link: https://docs.microsoft.com/en-us/windows-hardware/design/device-experiences/oem-kernel-dma-protection Link: https://www.amd.com/system/files/TechDocs/48882_IOMMU.pdf Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> --- drivers/thunderbolt/domain.c | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-)