Hi, On Fri, Jun 25, 2021 at 6:19 AM Joerg Roedel <joro@xxxxxxxxxx> wrote: > > Hi Douglas, > > On Thu, Jun 24, 2021 at 10:17:56AM -0700, Douglas Anderson wrote: > > The goal of this patch series is to get better SD/MMC performance on > > Qualcomm eMMC controllers and in generally nudge us forward on the > > path of allowing some devices to be in strict mode and others to be in > > non-strict mode. > > So if I understand it right, this patch-set wants a per-device decision > about setting dma-mode to strict vs. non-strict. > > I think we should get to the reason why strict mode is used by default > first. Is the default on ARM platforms to use iommu-strict mode by > default and if so, why? > > The x86 IOMMUs use non-strict mode by default (yes, it is a security > trade-off). It is certainly a good question. I will say that, as per usual, I'm fumbling around trying to solve problems in subsystems I'm not an expert at, so if something I'm saying sounds like nonsense it probably is. Please correct me. I guess I'd start out by thinking about what devices I think need to be in "strict" mode. Most of my thoughts on this are in the 3rd patch in the series. I think devices where it's important to be in strict mode fall into "Case 1" from that patch description, copied here: Case 1: IOMMUs prevent malicious code running on the peripheral (maybe a malicious peripheral or maybe someone exploited a benign peripheral) from turning into an exploit of the Linux kernel. This is particularly important if the peripheral has loadable / updatable firmware or if the peripheral has some type of general purpose processor and is processing untrusted inputs. It's also important if the device is something that can be easily plugged into the host and the device has direct DMA access itself, like a PCIe device. Using sc7180 as an example (searching for iommus in sc7180.dtsi), I'd expect these peripherals to be in strict mode: * WiFi / LTE - I'm almost certain we want this in "strict" mode. Both have loadable / updatable firmware and both do complex processing on untrusted inputs. Both have a history of being compromised over the air just by being near an attacker. Note that on sc7180 these are _not_ connected over PCI so we can't leverage any PCI mechanism for deciding strict / non-strict. * Video decode / encode - pretty sure we want this in strict. It's got loadable / updatable firmware and processing complex / untrusted inputs. * LPASS (low power audio subsystem) - I don't know a ton and I think we don't use this much on our designs, but I believe it meets the definitions for needing "strict". * The QUPs (handles UART, SPI, and i2c) - I'm not as sure here. These are much "smarter" than you'd expect. They have loadable / updatable firmware and certainly have a sort of general purpose processor in them. They also might be processing untrusted inputs, but presumably in a pretty simple way. At the moment we don't use a ton of DMA here anyway and these are pretty low speed, so I would tend to leave them as strict just to be on the safe side. I'd expect these to be non-strict: * SD/MMC - as described in this patch series. * USB - As far as I know firmware isn't updatable and has no history of being compromised. Special: * GPU - This already has a bunch of special cases, so we probably don't need to discuss here. As far as I can tell everything in sc7180.dtsi that has an "iommus" property is classified above. So, unless I'm wrong and it's totally fine to run LTE / WiFi / Video / LPASS in non-strict mode then: * We still need some way to pick strict vs. non-strict. * Since I've only identified two peripherals that I think should be non-strict, having "strict" the default seems like fewer special cases. It's also safer. In terms of thinking about x86 / AMD where the default is non-strict, I don't have any historical knowledge there. I guess the use of PCI for connecting WiFi is more common (so you can use the PCI special cases) and I'd sorta hope that WiFi is running in strict mode. For video encode / decode, perhaps x86 / AMD are just accepting the risk here because there was no kernel infrastructure for doing better? I'd also expect that x86/AMD don't have something quite as crazy as the QUPs for UART/I2C/SPI, but even if they do I wouldn't be terribly upset if they were in non-strict mode. ...so I guess maybe the super short answer to everything above is that I believe that at least WiFi ought to be in "strict" mode and it's not on PCI so we need to come up with some type of per-device solution. -Doug