lkml.org 
[lkml]   [2022]   [May]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH v2 00/21] KCFI support
On Tue, May 17, 2022 at 09:33:47AM +0200, Sedat Dilek wrote:
> Is your tags/kcfi-rfc-v2 the recommended LLVM toolchain for testing kcfi-rfc-v2?
>
> What I want to ask is if your commit well tested for x86 (and arm64)
> means build and boot on bare metal?

I have run kCFI (v1, I haven't had time to test v2) on x86_64 hardware,
both AMD and Intel. I have only found two failures so far: the i915
issue that I mention below and a failure in the KVM subsystem, which I
can see by just running QEMU:

[ 124.344654] CFI failure at kvm_mmu_notifier_invalidate_range_end+0xca/0x177 [kvm] (target: kvm_null_fn+0x0/0x7 [kvm]; expected type: 0x1e2c5b9c)
[ 124.344691] WARNING: CPU: 5 PID: 2767 at kvm_mmu_notifier_invalidate_range_end+0xca/0x177 [kvm]
[ 124.344708] Modules linked in: ...
[ 124.344737] CPU: 5 PID: 2767 Comm: qemu-system-x86 Tainted: P 5.18.0-rc6-debug-00033-g1d5284aff7cd #1 8c4966c7fb24f3345076747feebac5fb340a4797
[ 124.344738] Hardware name: ASUS System Product Name/PRIME Z590M-PLUS, BIOS 1203 10/27/2021
[ 124.344738] RIP: 0010:kvm_mmu_notifier_invalidate_range_end+0xca/0x177 [kvm]
[ 124.344756] Code: 48 89 da e8 88 16 86 dd 48 85 c0 89 6c 24 0c 74 30 4d 85 f6 74 36 48 c7 c5 19 c8 b4 c0 4c 8b 34 24 81 7d fa 9c 5b 2c 1e 74 02 <0f> 0b 48 89 c7 4c 89 fe 48 89 da e8 b6 16 86 dd 48 85 c0 75 e2 eb
[ 124.344757] RSP: 0018:ffffc033039d78e8 EFLAGS: 00010282
[ 124.344758] RAX: ffffa0a141ed9450 RBX: 00007fb897e9efff RCX: ffffa0a143c9d850
[ 124.344758] RDX: 00007fb897e9efff RSI: 00007fb897e9e000 RDI: ffffc03304499cf8
[ 124.344759] RBP: ffffffffc0b4c819 R08: 0000000000000000 R09: 0000000000000000
[ 124.344759] R10: 0000000000000000 R11: ffffffffc0b4c639 R12: ffffc03304499000
[ 124.344760] R13: ffffc033044a30a0 R14: ffffc033044a3120 R15: 00007fb897e9e000
[ 124.344760] FS: 00007fbca4bc3640(0000) GS:ffffa0a87f540000(0000) knlGS:0000000000000000
[ 124.344761] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 124.344761] CR2: 0000000000000000 CR3: 0000000119f74006 CR4: 0000000000772ee0
[ 124.344762] PKRU: 55555554
[ 124.344762] Call Trace:
[ 124.344762] <TASK>
[ 124.344763] __mmu_notifier_invalidate_range_end+0xa1/0xd7
[ 124.344764] wp_page_copy+0x592/0x920
[ 124.344766] __handle_mm_fault+0x820/0x8f0
[ 124.344767] handle_mm_fault+0xe0/0x227
[ 124.344768] __get_user_pages+0x17a/0x430
[ 124.344769] get_user_pages_unlocked+0xd9/0x327
[ 124.344770] hva_to_pfn+0xfa/0x3f7 [kvm 950e483805e01f967a3e331506278d766494a4cc]
[ 124.344788] kvm_faultin_pfn+0xc3/0x2f0 [kvm 950e483805e01f967a3e331506278d766494a4cc]
[ 124.344806] ? fast_page_fault+0x400/0x4c0 [kvm 950e483805e01f967a3e331506278d766494a4cc]
[ 124.344823] direct_page_fault+0x130/0x350 [kvm 950e483805e01f967a3e331506278d766494a4cc]
[ 124.344841] kvm_mmu_page_fault+0x176/0x2f7 [kvm 950e483805e01f967a3e331506278d766494a4cc]
[ 124.344858] vmx_handle_exit+0xe/0x37 [kvm_intel 647f514f18fc365045dc57db46d2bb4930b746eb]
[ 124.344863] vcpu_enter_guest+0xbff/0x1030 [kvm 950e483805e01f967a3e331506278d766494a4cc]
[ 124.344881] vcpu_run+0x65/0x2f0 [kvm 950e483805e01f967a3e331506278d766494a4cc]
[ 124.344898] kvm_arch_vcpu_ioctl_run+0x15f/0x3f7 [kvm 950e483805e01f967a3e331506278d766494a4cc]
[ 124.344916] kvm_vcpu_ioctl+0x547/0x627 [kvm 950e483805e01f967a3e331506278d766494a4cc]
[ 124.344933] ? syscall_exit_to_user_mode+0x24/0x47
[ 124.344934] ? do_syscall_64+0x7a/0x97
[ 124.344935] ? __fget_files+0xa1/0xc0
[ 124.344936] __se_sys_ioctl+0x7c/0xc0
[ 124.344937] do_syscall_64+0x6c/0x97
[ 124.344938] ? do_syscall_64+0x7a/0x97
[ 124.344939] entry_SYSCALL_64_after_hwframe+0x44/0xae
[ 124.344940] RIP: 0033:0x7fbca75f7b1f
[ 124.344940] Code: 00 48 89 44 24 18 31 c0 48 8d 44 24 60 c7 04 24 10 00 00 00 48 89 44 24 08 48 8d 44 24 20 48 89 44 24 10 b8 10 00 00 00 0f 05 <89> c2 3d 00 f0 ff ff 77 18 48 8b 44 24 18 64 48 2b 04 25 28 00 00
[ 124.344941] RSP: 002b:00007fbca4bc2550 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
[ 124.344942] RAX: ffffffffffffffda RBX: 000000000000ae80 RCX: 00007fbca75f7b1f
[ 124.344942] RDX: 0000000000000000 RSI: 000000000000ae80 RDI: 000000000000000e
[ 124.344943] RBP: 000056351ba661d0 R08: 000056351a74dc48 R09: 0000000000000004
[ 124.344943] R10: 004818cf4f129e11 R11: 0000000000000246 R12: 0000000000000000
[ 124.344943] R13: 0000000000000000 R14: 00007fbca814f004 R15: 0000000000000000
[ 124.344944] </TASK>
[ 124.344945] ---[ end trace 0000000000000000 ]---

I don't do any bare metal testing for the tc-build known good revision
bumps, just QEMU tests with boot-utils.

> Just for the records:
> You definitely need a pre-LLVM-15 toolchain + KCFI sanitizer patch?
> LLVM-14?

The feature won't land in LLVM 14, so there is little point to
backporting and testing it on LLVM 14. This is a work in progress
feature so it has to target the main branch. I have fairly good coverage
of tip of tree locally and we have solid coverage through CI so we'll
know if something major happens, it should be pretty safe to test Sami's
series.

> > > Nathan has a i915 cfi patch in His personal kernel.org Git.
> > > Is this relevant to kcfi?
> >
> > It fixes a type mismatch, so in that sense it's relevant.
> >
>
> Here the link to patch "drm/i915: Fix CFI violation with show_dynamic_id()":
>
> https://git.kernel.org/pub/scm/linux/kernel/git/nathan/linux.git/commit/?h=submitted/i915-cfi-fix&id=53735be6dc53453fcfbac658e847b54360e73871

This is now in the i915 tree so that branch is going to disappear:

https://cgit.freedesktop.org/drm/drm-intel/commit/?id=18fb42db05a0b93ab5dd5eab5315e50eaa3ca620

> You say no need to build your kernel with LTO...
> That sounds good.
> Currently, I build my kernels with Clang-14 and CONFIG_LTO_CLANG_THIN=y.
> Does something speak against using CONFIG_LTO_CLANG_THIN=y with KCFI support?
> Build-time?
> Disc-usage?

I wouldn't expect the build time or disk usage to significantly increase
with kCFI. I ran a quick benchmark with Arch Linux's configuration with
ThinLTO then ThinLTO + kCFI on an AMD EPYC 7513:

Benchmark 1: ThinLTO
Time (abs ≡): 166.036 s [User: 4687.951 s, System: 1636.767 s]

Benchmark 2: ThinLTO + kCFI
Time (abs ≡): 168.739 s [User: 4682.020 s, System: 1638.109 s]

Summary
'ThinLTO' ran
1.02 times faster than 'ThinLTO + kCFI'

Cheers,
Nathan

\
 
 \ /
  Last update: 2022-05-17 20:50    [W:0.089 / U:2.260 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site