On Mon, 28 Jan 2019 17:13:22 -0600 Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote: > On Mon, Jan 28, 2019 at 11:31:08AM +0000, Jonathan Cameron wrote: > > On Thu, 20 Dec 2018 13:57:14 -0600 > > Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote: > > > On Thu, Dec 20, 2018 at 09:13:12AM -0800, Dave Hansen wrote: > > > > On 12/20/18 7:12 AM, Bjorn Helgaas wrote: > > > > >> Other than the error we might be able to use acpi_map_pxm_to_online_node > > > > >> for this, or call both acpi_map_pxm_to_node and acpi_map_pxm_to_online_node > > > > >> and compare the answers to verify we are getting the node we want? > > > > > Where are we at with this? It'd be nice to resolve it for v4.21, but > > > > > it's a little out of my comfort zone, so I don't want to apply it > > > > > unless there's clear consensus that this is the right fix. > > > > > > > > I still think the fix in this patch sweeps the problem under the rug too > > > > much. But, it just might be the best single fix for backports, for > > > > instance. > > > > > > Sounds like we should first find the best fix, then worry about how to > > > backport it. So I think we have a little more noodling to do, and > > > I'll defer this for now. > > > > > > Bjorn > > > > Hi All, > > > > I'd definitely appreciate some guidance on what the 'right' fix is. > > We are starting to get real performance issues reported as a result of not > > being able to use this patch on mainline. > > > > 5-10% performance drop on some networking benchmarks. > > I guess the performance drop must be from calling kmalloc_node() with > the wrong node number because we currently ignore _PXM for the NIC? > And to get that performance back, you need both the previous patch to > pay attention to _PXM (https://lore.kernel.org/linux-pci/1527768879-88161-2-git-send-email-xiexiuqi@xxxxxxxxxx) > and this patch (to set "numa_off=1" to avoid the regression the _PXM > patch by itself would cause)? Exactly. > > > As a brief summary (having added linux-mm / linux-acpi) the issue is: > > > > 1) ACPI allows _PXM to be applied to pci devices (including root ports for > > example, but any device is fine). > > 2) Due to the ordering of when the fw node was set for PCI devices this wasn't > > taking effect. Easy to solve by just adding the numa node if provided in > > pci_acpi_setup (which is late enough) > > 3) A patch to fix that was applied to the PCIe tree > > https://patchwork.kernel.org/patch/10597777/ > > but we got non booting regressions on some threadripper platforms. > > That turned out to be because they don't have SRAT, but do have PXM entries. > > (i.e. broken firmware). Naturally Bjorn reverted this very quickly! > > Here's the beginning of the current thread, for anybody coming in > late: https://lore.kernel.org/linux-pci/20181211094737.71554-1-Jonathan.Cameron@xxxxxxxxxx). > > The current patch proposes setting "numa_off=1" in the x86 version of > dummy_numa_init(), on the assumption (from the changelog) that: > > It is invalid under the ACPI spec to specify new NUMA nodes using > _PXM if they have no presence in SRAT. > > Do you have a reference for this? I looked and couldn't find a clear > statement in the spec to that effect. The _PXM description (ACPI > v6.2, sec 6.1.14) says that two devices with the same _PXM value are > in the same proximity domain, but it doesn't seem to require an SRAT. No comment (feel free to guess why). *sigh* > > But I guess it doesn't really matter whether it's invalid; that > situation exists in the field, so we have to handle it gracefully. > > Martin reported the regression from 3) above and attached useful logs, > which unfortunately aren't in the archives because the mailing list rejects > attachments. To preserve them, I opened https://bugzilla.kernel.org/show_bug.cgi?id=202443 > and attached the logs there. Cool. Thanks for doing that. > > > I proposed this fix which was to do the same as on Arm and clearly > > mark numa as off when SRAT isn't present on an ACPI system. > > https://elixir.bootlin.com/linux/latest/source/arch/arm64/mm/numa.c#L460 > > https://elixir.bootlin.com/linux/latest/source/arch/x86/mm/numa.c#L688 > > There are several threads we could pull on while untangling this. > > We use dummy_numa_init() when we don't have static NUMA info from ACPI > SRAT or DT. On arm64 (but not x86), it sets numa_off=1 when we don't > have that static info. I think neither should set numa_off=1 because > we should allow for future information, e.g., from _PXM. > > I think acpi_numa_init() is being a little too aggressive when it > returns failure if it finds no SRAT or if it finds an SRAT with no > ACPI_SRAT_TYPE_MEMORY_AFFINITY entries. > > Also from your changelog: > > When the PCI code later comes along and calls acpi_get_node() for > any PCI card below the root port, it navigates up the ACPI tree > until it finds the _PXM value in the root port. This value is then > passed to acpi_map_pxm_to_node(). > > As numa_off has not been set on x86 it tries to allocate a NUMA > node, from the unused set, without setting up all the infrastructure > that would normally accompany such a call. We have not identified > exactly which driver is causing the subsequent hang for Martin. > > So the problem seems to be that when we get the _PXM value (in the > acpi_get_node() path), there's some infrastructure we don't set up? > I'm not sure what exactly this is -- I see that when we have an SRAT, > acpi_numa_memory_affinity() does a little more, but nothing that > would account for a problem if we call acpi_map_pxm_to_node() without > an SRAT. > > Maybe it results in an issue when we call kmalloc_node() using this > _PXM value that SRAT didn't tell us about? If so, that's reminiscent > of these earlier discussions about kmalloc_node() returning something > useless if the requested node is not online: > > https://lkml.kernel.org/r/1527768879-88161-2-git-send-email-xiexiuqi@xxxxxxxxxx > https://lore.kernel.org/linux-arm-kernel/20180801173132.19739-1-punit.agrawal@xxxxxxx/ > > As far as I know, that was never really resolved. The immediate > problem of using passing an invalid node number to kmalloc_node() was > avoided by using kmalloc() instead. Yes, that's definitely still a problem (or was last time I checked) > > > Dave's response was that we needed to fix the underlying issue of > > trying to allocate from non existent NUMA nodes. > > Oops, sorry for telling you what you obviously already know! I guess > I didn't internalize this sentence before writing the above. Not to worry, your description was a lot better than mine! Thanks. > > Bottom line, I totally agree that it would be better to fix the > underlying issue without trying to avoid it by disabling NUMA. I don't agree on this point. I think two layers make sense. If there is no NUMA description in DT or ACPI, why not just stop anything from using it at all? The firmware has basically declared there is no point, why not save a bit of complexity (and use an existing tested code path) but setting numa_off? However, if there is NUMA description, but with bugs then we should protect in depth. A simple example being that we declare 2 nodes, but then use _PXM for a third. I've done that by accident and blows up in a nasty fashion (not done it for a while, but probably still true). Given DSDT is only parsed long after SRAT we can just check on _PXM queries. Or I suppose we could do a verification parse for all _PXM entries and put out some warnings if they don't match SRAT entries? > > > Whilst I agree with that in principle (having managed to provide > > tables doing exactly that during development a few times!), I'm not > > sure the path to doing so is clear and so this has been stalled for > > a few months. There is to my mind still a strong argument, even > > with such protection in place, that we should still be short cutting > > it so that you get the same paths if you deliberately disable numa, > > and if you have no SRAT and hence can't have NUMA. > > I guess we need to resolve the question of whether NUMA without SRAT > is possible. It's certainly unclear of whether it has any meaning. If we allow for the fact that the intent of ACPI was never to allow this (and a bit of history checking verified this as best as anyone can remember), then what do we do with the few platforms that do use _PXM to nodes that haven't been defined? Note we have never actually supported them as we weren't using the values provided, so there is no regression if we simply rule them as not valid. It's also unclear that it was ever intentional for these platforms, rather than something that got through compliance tests because no one was using it. Thanks for your detailed insight and help! Jonathan > > > So given I have some 'mild for now' screaming going on, I'd > > definitely appreciate input on how to move forward! > > > > There are lots of places this could be worked around, e.g. we could > > sanity check in the acpi_get_pxm call. I'm not sure what side > > effects that would have and also what cases it wouldn't cover. > > > > Thanks, > > > > Jonathan
