On Wed, Oct 07, 2020 at 08:20:58PM +0300, Jarkko Sakkinen wrote: > On Wed, Oct 07, 2020 at 06:13:49PM +0200, Jethro Beekman wrote: > > On 2020-10-07 17:49, Jarkko Sakkinen wrote: > > > On Tue, Oct 06, 2020 at 06:13:28PM +0300, Jarkko Sakkinen wrote: > > >> On Mon, Oct 05, 2020 at 03:56:52PM -0700, Sean Christopherson wrote: > > >>> On Wed, Sep 30, 2020 at 02:45:54PM +0300, Jarkko Sakkinen wrote: > > >>>> On Wed, Sep 30, 2020 at 09:12:06AM +0200, Jethro Beekman wrote: > > >>>>> On 2020-09-30 03:16, Jarkko Sakkinen wrote: > > >>>>>> On Tue, Sep 29, 2020 at 05:52:48PM +0200, Jethro Beekman wrote: > > >>>>>>> Since the latest API changes, I'm unable to load a large enclave. The > > >>>>>>> test program at > > >>>>>>> https://github.com/fortanix/rust-sgx/blob/sgx-load-large-enclave-test/src/main.rs > > >>>>>>> always fails with ENOMEM after loading 0xffd6 pages. > > >>>>>>> > > >>>>>>> I've tested this with v36, if there's reason to believe it has been > > >>>>>>> fixed I'd be happy to try it out on a newer patch set. > > >>>>>> > > >>>>>> I recommend using v39-rc1 tag that I created for testing because API is > > >>>>>> reverted back to be compatible with v36. > > >>>>> > > >>>>> Not sure what you're saying. I tested with v36. You're saying v39-rc1 > > >>>>> will be the same? Or did you fix the issue since v36? > > >>>> > > >>>> v37 and v38 has an API change that is reverted in v39: > > >>>> > > >>>> https://lore.kernel.org/linux-sgx/20200921195822.GA58176@xxxxxxxxxxxxxxx/ > > >>>> > > >>>> I'm not sure of the root cause yet but you asked to try to out a newer > > >>>> patch set and v39-rc1 is the best option. > > >>>> > > >>>> There was off-by-one error in enclave maximum size calculation fixed in > > >>>> v37 (it was actually a bug in SDM inherited to the code) but that should > > >>>> not result the situation you just described. > > >>> > > >>> My money is on the XArray changes, that's the most notable change in v36 and > > >>> IIRC the only thing that touched EPC/memory management. > > >> > > >> Yeah, that's what we've been speculating for some days now. That's > > >> somewhat deprecated email. It all started to enroll when I asked > > >> Haitao to turn CONFIG_PROVE_LOCKING on, and we got the information > > >> required to root cause the bug. > > > > > > I run the failing test and filtered SGX mmap's and ioctl's with this > > > eBPF script: > > > > > > kretprobe:sgx_ioctl /retval != 0/ > > > { > > > printf("sgx_ioctl: %d\n", retval) > > > } > > > > > > kretprobe:sgx_mmap /retval != 0/ > > > { > > > printf("sgx_mmap: %d\n", retval) > > > } > > > > > > This results zero positives, i.e. empty output, when run with bpftrace. > > > > > > I'd go instead after RLIMIT_AS [*]. > > > > > > With these conclusions, I'm done with this bug. > > > > > > > How can it be RLIMIT_AS? With the current flow, you mmap the whole range before mmaping the individual pages over it? > > > > Also, I can easily load a 1GB enclave with the old driver. > > > > Also: > > > > $ ulimit -v > > unlimited > > ➜ ~ (master) ✔ sudo bpftrace sgx_ret.bt > Attaching 3 probes... > ksys_mmap_pgoff: -12 > ^C > > ~ (master) ✔ cat sgx_ret.bt > kretprobe:sgx_ioctl /retval != 0/ > { > printf("sgx_ioctl: %d\n", retval) > } > > kretprobe:sgx_mmap /retval != 0/ > { > printf("sgx_mmap: %d\n", retval) > } > > kretprobe:ksys_mmap_pgoff /retval == (uint64)-12/ > { > printf("ksys_mmap_pgoff: %d\n", retval) > } > > This shows that it fails before reaching sgx_mmap(). ➜ ~ (master) ✔ sudo bpftrace -e 'kr:ksys_mmap_pgoff { @[comm] = count(); }' Attaching 1 probe... ^C @[zsh]: 44 @[git]: 47 @[date]: 48 @[network.sh]: 48 @[battery.sh]: 56 @[which]: 84 @[cargo]: 94 @[head]: 96 @[iw]: 126 @[uname]: 144 @[cat]: 168 @[sh]: 175 @[sed]: 198 @[bash]: 216 @[ping]: 222 @[ls]: 324 @[sgx-load-large-]: 65510 65510 is the default value for /proc/sys/vm/max_map_count [*]. [*] https://www.kernel.org/doc/Documentation/sysctl/vm.txt /Jarkko
