Message ID | 20180801200418.1325826-1-jeremy.linton@arm.com (mailing list archive) |
---|---|
Headers | show |
Series | harden alloc_pages against bogus nid | expand |
On Wed, 1 Aug 2018 15:04:16 -0500 Jeremy Linton <jeremy.linton@arm.com> wrote: > The thread "avoid alloc memory on offline node" > > https://lkml.org/lkml/2018/6/7/251 > > Asked at one point why the kzalloc_node was crashing rather than > returning memory from a valid node. The thread ended up fixing > the immediate causes of the crash but left open the case of bad > proximity values being in DSDT tables without corrisponding > SRAT/SLIT entries as is happening on another machine. > > Its also easy to fix that, but we should also harden the allocator > sufficiently that it doesn't crash when passed an invalid node id. > There are a couple possible ways to do this, and i've attached two > separate patches which individually fix that problem. > > The first detects the offline node before calling > the new_slab code path when it becomes apparent that the allocation isn't > going to succeed. The second actually hardens node_zonelist() and > prepare_alloc_pages() in the face of NODE_DATA(nid) returning a NULL > zonelist. This latter case happens if the node has never been initialized > or is possibly out of range. There are other places (NODE_DATA & > online_node) which should be checking if the node id's are > MAX_NUMNODES. > What is it that leads to a caller requesting memory from an invalid node? A race against offlining? If so then that's a lack of appropriate locking, isn't it? I don't see a problem with emitting a warning and then selecting a different node so we can keep running. But we do want that warning, so we can understand the root cause and fix it?
Hi, On 08/01/2018 04:50 PM, Andrew Morton wrote: > On Wed, 1 Aug 2018 15:04:16 -0500 Jeremy Linton <jeremy.linton@arm.com> wrote: > >> The thread "avoid alloc memory on offline node" >> >> https://lkml.org/lkml/2018/6/7/251 >> >> Asked at one point why the kzalloc_node was crashing rather than >> returning memory from a valid node. The thread ended up fixing >> the immediate causes of the crash but left open the case of bad >> proximity values being in DSDT tables without corrisponding >> SRAT/SLIT entries as is happening on another machine. >> >> Its also easy to fix that, but we should also harden the allocator >> sufficiently that it doesn't crash when passed an invalid node id. >> There are a couple possible ways to do this, and i've attached two >> separate patches which individually fix that problem. >> >> The first detects the offline node before calling >> the new_slab code path when it becomes apparent that the allocation isn't >> going to succeed. The second actually hardens node_zonelist() and >> prepare_alloc_pages() in the face of NODE_DATA(nid) returning a NULL >> zonelist. This latter case happens if the node has never been initialized >> or is possibly out of range. There are other places (NODE_DATA & >> online_node) which should be checking if the node id's are > MAX_NUMNODES. >> > > What is it that leads to a caller requesting memory from an invalid > node? A race against offlining? If so then that's a lack of > appropriate locking, isn't it? There were a couple unrelated cases, both having to do with the PXN associated with a PCI port. The first case AFAIK, the domain wasn't really invalid if the entire SRAT was parsed and nodes created even when there weren't associated CPUs. The second case (a different machine) is simply a PXN value that is completely invalid (no associated SLIT/SRAT/etc entries) due to firmware making a mistake when a socket isn't populated. There have been a few other suggested or merged patches for the individual problems above, this set is just an attempt at avoiding a full crash if/when another similar problem happens. > > I don't see a problem with emitting a warning and then selecting a > different node so we can keep running. But we do want that warning, so > we can understand the root cause and fix it? Yes, we do want to know when an invalid id is passed, i will add the VM_WARN in the first one. The second one I wasn't sure about as failing prepare_alloc_pages() generates a couple of error messages, but the system then continues operation. I guess my question though is which method (or both/something else?) is the preferred way to harden this up? Thanks for looking at this.
On Wed, 1 Aug 2018 17:56:46 -0500 Jeremy Linton <jeremy.linton@arm.com> wrote: > Hi, > > On 08/01/2018 04:50 PM, Andrew Morton wrote: > > On Wed, 1 Aug 2018 15:04:16 -0500 Jeremy Linton <jeremy.linton@arm.com> wrote: > > > >> The thread "avoid alloc memory on offline node" > >> > >> https://lkml.org/lkml/2018/6/7/251 > >> > >> Asked at one point why the kzalloc_node was crashing rather than > >> returning memory from a valid node. The thread ended up fixing > >> the immediate causes of the crash but left open the case of bad > >> proximity values being in DSDT tables without corrisponding > >> SRAT/SLIT entries as is happening on another machine. > >> > >> Its also easy to fix that, but we should also harden the allocator > >> sufficiently that it doesn't crash when passed an invalid node id. > >> There are a couple possible ways to do this, and i've attached two > >> separate patches which individually fix that problem. > >> > >> The first detects the offline node before calling > >> the new_slab code path when it becomes apparent that the allocation isn't > >> going to succeed. The second actually hardens node_zonelist() and > >> prepare_alloc_pages() in the face of NODE_DATA(nid) returning a NULL > >> zonelist. This latter case happens if the node has never been initialized > >> or is possibly out of range. There are other places (NODE_DATA & > >> online_node) which should be checking if the node id's are > MAX_NUMNODES. > >> > > > > What is it that leads to a caller requesting memory from an invalid > > node? A race against offlining? If so then that's a lack of > > appropriate locking, isn't it? > > There were a couple unrelated cases, both having to do with the PXN > associated with a PCI port. The first case AFAIK, the domain wasn't > really invalid if the entire SRAT was parsed and nodes created even when > there weren't associated CPUs. The second case (a different machine) is > simply a PXN value that is completely invalid (no associated > SLIT/SRAT/etc entries) due to firmware making a mistake when a socket > isn't populated. > > There have been a few other suggested or merged patches for the > individual problems above, this set is just an attempt at avoiding a > full crash if/when another similar problem happens. Please add the above info to the changelog. > > > > > I don't see a problem with emitting a warning and then selecting a > > different node so we can keep running. But we do want that warning, so > > we can understand the root cause and fix it? > > Yes, we do want to know when an invalid id is passed, i will add the > VM_WARN in the first one. > > The second one I wasn't sure about as failing prepare_alloc_pages() > generates a couple of error messages, but the system then continues > operation. > > I guess my question though is which method (or both/something else?) is > the preferred way to harden this up? The first patch looked neater. Can we get a WARN_ON in there as well?
Hi, On 08/01/2018 07:14 PM, Andrew Morton wrote: > On Wed, 1 Aug 2018 17:56:46 -0500 Jeremy Linton <jeremy.linton@arm.com> wrote: > >> Hi, >> >> On 08/01/2018 04:50 PM, Andrew Morton wrote: >>> On Wed, 1 Aug 2018 15:04:16 -0500 Jeremy Linton <jeremy.linton@arm.com> wrote: >>> >>>> The thread "avoid alloc memory on offline node" >>>> >>>> https://lkml.org/lkml/2018/6/7/251 >>>> >>>> Asked at one point why the kzalloc_node was crashing rather than >>>> returning memory from a valid node. The thread ended up fixing >>>> the immediate causes of the crash but left open the case of bad >>>> proximity values being in DSDT tables without corrisponding >>>> SRAT/SLIT entries as is happening on another machine. >>>> >>>> Its also easy to fix that, but we should also harden the allocator >>>> sufficiently that it doesn't crash when passed an invalid node id. >>>> There are a couple possible ways to do this, and i've attached two >>>> separate patches which individually fix that problem. >>>> >>>> The first detects the offline node before calling >>>> the new_slab code path when it becomes apparent that the allocation isn't >>>> going to succeed. The second actually hardens node_zonelist() and >>>> prepare_alloc_pages() in the face of NODE_DATA(nid) returning a NULL >>>> zonelist. This latter case happens if the node has never been initialized >>>> or is possibly out of range. There are other places (NODE_DATA & >>>> online_node) which should be checking if the node id's are > MAX_NUMNODES. >>>> >>> >>> What is it that leads to a caller requesting memory from an invalid >>> node? A race against offlining? If so then that's a lack of >>> appropriate locking, isn't it? >> >> There were a couple unrelated cases, both having to do with the PXN >> associated with a PCI port. The first case AFAIK, the domain wasn't >> really invalid if the entire SRAT was parsed and nodes created even when >> there weren't associated CPUs. The second case (a different machine) is >> simply a PXN value that is completely invalid (no associated >> SLIT/SRAT/etc entries) due to firmware making a mistake when a socket >> isn't populated. >> >> There have been a few other suggested or merged patches for the >> individual problems above, this set is just an attempt at avoiding a >> full crash if/when another similar problem happens. > > Please add the above info to the changelog. Sure. > >> >>> >>> I don't see a problem with emitting a warning and then selecting a >>> different node so we can keep running. But we do want that warning, so >>> we can understand the root cause and fix it? >> >> Yes, we do want to know when an invalid id is passed, i will add the >> VM_WARN in the first one. >> >> The second one I wasn't sure about as failing prepare_alloc_pages() >> generates a couple of error messages, but the system then continues >> operation. >> >> I guess my question though is which method (or both/something else?) is >> the preferred way to harden this up? > > The first patch looked neater. Can we get a WARN_ON in there as well? > Yes, Thanks,