Hi, On Tue, Jun 22, 2021 at 3:10 PM Robin Murphy <robin.murphy@xxxxxxx> wrote: > > On 2021-06-22 17:06, Doug Anderson wrote: > > Hi, > > > > On Tue, Jun 22, 2021 at 4:35 AM Robin Murphy <robin.murphy@xxxxxxx> wrote: > >> > >> Hi Doug, > >> > >> On 2021-06-22 00:52, Douglas Anderson wrote: > >>> > >>> This patch attempts to put forward a proposal for enabling non-strict > >>> DMA on a device-by-device basis. The patch series requests non-strict > >>> DMA for the Qualcomm SDHCI controller as a first device to enable, > >>> getting a nice bump in performance with what's believed to be a very > >>> small drop in security / safety (see the patch for the full argument). > >>> > >>> As part of this patch series I am end up slightly cleaning up some of > >>> the interactions between the PCI subsystem and the IOMMU subsystem but > >>> I don't go all the way to fully remove all the tentacles. Specifically > >>> this patch series only concerns itself with a single aspect: strict > >>> vs. non-strict mode for the IOMMU. I'm hoping that this will be easier > >>> to talk about / reason about for more subsystems compared to overall > >>> deciding what it means for a device to be "external" or "untrusted". > >>> > >>> If something like this patch series ends up being landable, it will > >>> undoubtedly need coordination between many maintainers to land. I > >>> believe it's fully bisectable but later patches in the series > >>> definitely depend on earlier ones. Sorry for the long CC list. :( > >> > >> Unfortunately, this doesn't work. In normal operation, the default > >> domains should be established long before individual drivers are even > >> loaded (if they are modules), let alone anywhere near probing. The fact > >> that iommu_probe_device() sometimes gets called far too late off the > >> back of driver probe is an unfortunate artefact of the original > >> probe-deferral scheme, and causes other problems like potentially > >> malformed groups - I've been forming a plan to fix that for a while now, > >> so I for one really can't condone anything trying to rely on it. > >> Non-deterministic behaviour based on driver probe order for multi-device > >> groups is part of the existing problem, and your proposal seems equally > >> vulnerable to that too. > > > > Doh! :( I definitely can't say I understand the iommu subsystem > > amazingly well. It was working for me, but I could believe that I was > > somehow violating a rule somewhere. > > > > I'm having a bit of a hard time understanding where the problem is > > though. Is there any chance that you missed the part of my series > > where I introduced a "pre_probe" step? Specifically, I see this: > > > > * really_probe() is called w/ a driver and a device. > > * -> calls dev->bus->dma_configure() w/ a "struct device *" > > * -> eventually calls iommu_probe_device() w/ the device. > > This... > > > * -> calls iommu_alloc_default_domain() w/ the device > > * -> calls iommu_group_alloc_default_domain() > > * -> always allocates a new domain > > > > ...so we always have a "struct device" when a domain is allocated if > > that domain is going to be associated with a device. > > > > I will agree that iommu_probe_device() is called before the driver > > probe, but unless I missed something it's after the device driver is > > loaded. ...and assuming something like patch #1 in this series looks > > OK then iommu_probe_device() will be called after "pre_probe". > > > > So assuming I'm not missing something, I'm not actually relying the > > IOMMU getting init off the back of driver probe. > > ...is implicitly that. Sorry that it's not obvious. > > The "proper" flow is that iommu_probe_device() is called for everything > which already exists during the IOMMU driver's own probe when it calls > bus_set_iommu(), then at BUS_NOTIFY_ADD_DEVICE time for everything which > appears subsequently. The only trouble is, to observe it in action on a > DT-based system you'd currently have to go back at least 4 years, before > 09515ef5ddad... > > Basically there were two issues: firstly we need the of_xlate step > before add_device (now probe_device) for a DT-based IOMMU driver to know > whether it should care about the given device or not. When -EPROBE_DEFER > was the only tool we had to ensure probe ordering, and resolving the > "iommus" DT property the only place to decide that, delaying it all > until driver probe time was the only reasonable option, however ugly. > The iommu_probe_device() "replay" in {of,acpi}_iommu_configure() is > merely doing its best to fake up the previous behaviour. Try binding a > dummy driver to your device first, then unbind it and bind the real one, > and you'll see that iommu_probe_device() doesn't run the second or > subsequent times. Now that we have fw_devlink to take care of ordering, I probably haven't been following fw_devlink as closely as I should, but does it completely replace -EPROBE_DEFER? From what I can tell right now it helps optimize the boot ordering but it's not mandatory (AKA Linux should boot/work fine with fw_devlink=off)? > the main reason for this weirdness is largely gone, so I'm keen to start > getting rid of the divergence again as far as possible. Fundamentally, > IOMMU drivers are supposed to be aware of all devices which the kernel > knows about, regardless of whether they have a driver available or not. > > The second issue is that when we have multiple IOMMU instances, the > initial bus_set_iommu() "replay" is only useful for the first instance, > so devices managed by other instances which aren't up and running yet > will be glossed over. Currently this ends up being papered over by the > solution to the first point on DT systems, while the x86 drivers hide > their individual IOMMU units behind a single IOMMU API "instance", so > it's been having little impact in practice. However, improving the core > API's multi-instance support is an increasingly pressing issue now that > new more varied systems are showing up, and it's that which is really > going to be driving the aforementioned changes. FWIW the plan I > currently have is to hang things off iommu_device_register() instead. I tried to follow all the above the best I can and hopefully understood some of it. ;-) OK, so what do you think about these points? * I think you are arguing that IOMMU strictness requirements needs to be known regardless of whether we have a driver for a given device or not. So the whole pre_probe stuff is just not right. * One thing that's been proposed is putting strictness info in the device tree. In theory on device-tree systems you could still put strictness info in the per-device nodes assuming that all devices are actually described in the device tree. That sounds nice, but I can believe it would be pretty hard to parse backward like this. AKA when the IOMMU probes it would need to find all the device tree nodes that would end up grouped together parse their details to find out if they should be non-strict. * Instead of putting the details in per-device nodes, maybe it makes sense to accept that we should be specifying these things at the IOMMU level? If specifying things at the device tree level then the device-tree node of the IOMMU itself would just have a list of things that should be strict/non-strict. ...this could potentially be merged with a hardcoded list of things in the IOMMU driver based on the IOMMU compatible string. Do those sound right? I still haven't totally grokked the ideal way to identify devices. I guess on Qualcomm systems each device is in its own group and so could have its own strictness levels? ...or would it be better to list by "stream ID" or something like that? If we do something like this then maybe that's a solution that could land short-ish term? We would know right at init time whether a given domain should be strict or non-strict and there'd be no requirements to transition it. > >> FWIW we already have a go-faster knob for people who want to tweak the > >> security/performance compromise for specific devices, namely the sysfs > >> interface for changing a group's domain type before binding the relevant > >> driver(s). Is that something you could use in your application, say from > >> an initramfs script? > > > > We've never had an initramfs script in Chrome OS. I don't know all the > > history of why (I'm trying to check), but I'm nearly certain it was a > > conscious decision. Probably it has to do with the fact that we're not > > trying to build a generic distribution where a single boot source can > > boot a huge variety of hardware. We generally have one kernel for a > > class of devices. I believe avoiding the initramfs just keeps things > > simpler. > > > > I think trying to revamp Chrome OS to switch to an initramfs type > > system would be a pretty big undertaking since (as I understand it) > > you can't just run a little command and then return to the normal boot > > flow. Once you switch to initramfs you're committing to finding / > > setting up the rootfs yourself and on Chrome OS I believe that means a > > whole bunch of dm-verity work. > > > > > > ...so probably the initramfs is a no-go for me, but I'm still crossing > > my fingers that the pre_probe() might be legit if you take a second > > look at it? > > That's fair enough - TBH the current sysfs interface is a pretty > specialist sport primarily for datacentre folks who can afford to take > down their 40GBE NIC or whatever momentarily for a longer-term payoff, > but it was worth exploring - I'm assuming the SDHCI holds your root > filesystem so you wouldn't be able to do the unbinding dance from real > userspace. As I said, the idea of embedding any sort of data in > individual client drivers is a non-starter in general since it only has > any hope of working on DT platforms (maybe arm64 ACPI too?), and only > for very much the wrong reasons. > > If this is something primarily demanded by QCom platforms in the short > term, At the moment I'm mostly focused on QCom platforms, but that's mostly because they're the first board that I've dealt with that has iommus on pretty much every peripheral under the sun. That's a _good_ thing, but it also means it's where we're hitting all these performance issues. I believe that the usage of "iommus everywhere" is about to become a lot more widespread across the SoC ecosystem, especially as they are important for virtualization and that seems to be a hot-topic these days. Basically: I could land a short-term hack a solution for me (and probably this is the right thing for me to do?), but I'm very much interested in finding a better solution, ideally something that we could achieve in something less than a year? Am I dreaming? > I'm tempted to say just try it with more device-matching magic in > arm-smmu-qcom. Yeah, I had this working (at least as far as my testing went) downstream in the chromeos-5.4 tree. I basically just implemented "init_context" for the arm-smmu-qcom driver and then if the "struct device *" passed in matched the SDMMC compatible string I just ORed in "IO_PGTABLE_QUIRK_NON_STRICT". That actually seemed to work fine in the downstream kernel tree but not upstream. I believe it broke in commit a250c23f15c2 ("iommu: remove DOMAIN_ATTR_DMA_USE_FLUSH_QUEUE") because the flush queue stopped getting initted. I think I could revamp my patch series to continue to have the strictness attributes in the "struct iommu_domain" but just get rid of the bits in the "struct device" and obviously get rid of the "pre_probe" patch. Would something like that be acceptable? > Otherwise, the idea of growing the sysfs interface to > allow switching a DMA domain from default-strict to non-strict is > certainly an interesting prospect. Going a step beyond that to bring up > a flush queue 'live' without rebuilding the whole group and domain could > get ugly when it comes to drivers' interaction with io-pgtable, but I > think it might be *technically* feasible... It seems like it would be nice, but maybe with the above we could get by without having to do this? -Doug