On Mon, Nov 9, 2020 at 5:13 PM Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote: > > Peter Collingbourne <pcc@xxxxxxxxxx> writes: > > > The kernel currently clears the tag bits (i.e. bits 56-63) in the fault > > address exposed via siginfo.si_addr and sigcontext.fault_address. However, > > the tag bits may be needed by tools in order to accurately diagnose > > memory errors, such as HWASan [1] or future tools based on the Memory > > Tagging Extension (MTE). > > > > We should not stop clearing these bits in the existing fault address > > fields, because there may be existing userspace applications that are > > expecting the tag bits to be cleared. Instead, create a new pair of > > fields in siginfo._sigfault, and store the tag bits of FAR_EL1 there, > > together with a mask specifying which bits are valid. > > > > A flag is added to si_faultflags to allow userspace to determine whether > > the values in the fields are valid. > > I think I am missing some things: > > Today it is documented that the tag bits are cleared, and so we can't > use the highbits to hold the tag bits by default. > > Why do you need to deliver which tag bits are valid? That feels like an > implementation detail that is needed to setup the tag bits. It feels > like it would be constant per process. So I don't understand why the > siginfo needs to report information the process should already have. It isn't constant as it may vary depending on the specific type of fault being delivered. For example on arm64 the architecture only provides us with bits 56-59 of the tag for tag check faults, while all other data aborts also provide bits 60-63. Now although the user program may distinguish the two cases by checking the si_code, we would also like to allow future architecture extensions to provide bits 60-63 for tag check faults as well and allow distinguishing between "bits 60-63 were zero" and "bits 60-63 were unknown" (which is important when providing error reports). I suppose that you could imagine that, if the "bits 60-63 available" extension ever comes to pass, it could be exposed via a bit in getauxval(AT_HWCAP2) (or maybe something like AT_HWCAP3 depending on how long it takes to arrive) and that would provide a way for userspace to know which bits are valid. But it seems like allowing this to vary per signal is relatively cheap, allows the information to be made available trivially architecture independently and keeps our options open for the future (because we don't know if some future architecture will actually make this a per-signal attribute). That being said, maybe we could solve this problem at the point at which we do encounter such an architecture though. > Want prevents adding a sigaction sa_flag SA_EXPOSE_TABITS that when set > causes the high bits to be set, and when clear (the default) will have > the signal delivery code clear those bits. > > That should be enough for code that wants the tag bits to ask for them. > As userspace would need to be updated to get the new bits > > Even if you have chained handlers. The chaining mechanism would need to > be updated and it could call the aware handlers first then clear the tag > bits and call the rest of the handlers. > > It feels like always passing the tag bits in the address and then > clearing them in the copy to userspace if the signal handler is > not ready for them would be easier to maintain. I think that approach might work. Although it may make life harder for callers of ptrace(PTRACE_SETSIGINFO) since they may need to know the value of the bit in order to prepare a correct siginfo structure, if we can reasonably expect them to always be delivering an exact copy of a signal that was received before then maybe that is okay. Assuming that this is an alternative to introducing si_addr_tag_bits_mask, the userspace code would need to use the flag bit support detection protocol for SA_EXPOSE_TAGBITS in order to be able to distinguish between "no bits valid" and "some bits valid", and then use an architecture-specific mechanism to determine exactly which bits are valid. Is that okay for a generic feature? Peter
