On 12.11.21 08:01, Baoquan He wrote: > On 11/11/21 at 08:18pm, David Hildenbrand wrote: >> To clear a user buffer we cannot simply use memset, we have to use >> clear_user(). Using a kernel config based on rawhide Fedora and a >> virtio-mem device that registers a vmcore_cb, I can easily trigger: >> >> [ 11.327580] systemd[1]: Starting Kdump Vmcore Save Service... >> [ 11.339697] kdump[420]: Kdump is using the default log level(3). >> [ 11.370964] kdump[453]: saving to /sysroot/var/crash/127.0.0.1-2021-11-11-14:59:22/ >> [ 11.373997] kdump[458]: saving vmcore-dmesg.txt to /sysroot/var/crash/127.0.0.1-2021-11-11-14:59:22/ >> [ 11.385357] kdump[465]: saving vmcore-dmesg.txt complete >> [ 11.386722] kdump[467]: saving vmcore >> [ 16.531275] BUG: unable to handle page fault for address: 00007f2374e01000 >> [ 16.531705] #PF: supervisor write access in kernel mode >> [ 16.532037] #PF: error_code(0x0003) - permissions violation >> [ 16.532396] PGD 7a523067 P4D 7a523067 PUD 7a528067 PMD 7a525067 PTE 800000007048f867 >> [ 16.532872] Oops: 0003 [#1] PREEMPT SMP NOPTI >> [ 16.533154] CPU: 0 PID: 468 Comm: cp Not tainted 5.15.0+ #6 >> [ 16.533513] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.14.0-27-g64f37cc530f1-prebuilt.qemu.org 04/01/2014 >> [ 16.534198] RIP: 0010:read_from_oldmem.part.0.cold+0x1d/0x86 >> [ 16.534552] Code: ff ff ff e8 05 ff fe ff e9 b9 e9 7f ff 48 89 de 48 c7 c7 38 3b 60 82 e8 f1 fe fe ff 83 fd 08 72 3c 49 8d 7d 08 4c 89 e9 89 e8 <49> c7 45 00 00 00 00 00 49 c7 44 05 f8 00 00 00 00 48 83 e7 f81 >> [ 16.535670] RSP: 0018:ffffc9000073be08 EFLAGS: 00010212 >> [ 16.535998] RAX: 0000000000001000 RBX: 00000000002fd000 RCX: 00007f2374e01000 >> [ 16.536441] RDX: 0000000000000001 RSI: 00000000ffffdfff RDI: 00007f2374e01008 >> [ 16.536878] RBP: 0000000000001000 R08: 0000000000000000 R09: ffffc9000073bc50 >> [ 16.537315] R10: ffffc9000073bc48 R11: ffffffff829461a8 R12: 000000000000f000 >> [ 16.537755] R13: 00007f2374e01000 R14: 0000000000000000 R15: ffff88807bd421e8 >> [ 16.538200] FS: 00007f2374e12140(0000) GS:ffff88807f000000(0000) knlGS:0000000000000000 >> [ 16.538696] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 >> [ 16.539055] CR2: 00007f2374e01000 CR3: 000000007a4aa000 CR4: 0000000000350eb0 >> [ 16.539510] Call Trace: >> [ 16.539679] <TASK> >> [ 16.539828] read_vmcore+0x236/0x2c0 >> [ 16.540063] ? enqueue_hrtimer+0x2f/0x80 >> [ 16.540323] ? inode_security+0x22/0x60 >> [ 16.540572] proc_reg_read+0x55/0xa0 >> [ 16.540807] vfs_read+0x95/0x190 >> [ 16.541022] ksys_read+0x4f/0xc0 >> [ 16.541238] do_syscall_64+0x3b/0x90 >> [ 16.541475] entry_SYSCALL_64_after_hwframe+0x44/0xae >> >> To fix, properly use clear_user() when required. > > Looks a great fix to me, thanks for fixing this. > > Check the code, clear_user invokes access_ok to do check, then call > memset(). It's unclear to me how the bug is triggered, could you > please tell more so that I can learn? > TBH, I was testing virtio-mem+vmcore before without running into this issue, but after I retested with upstream in a different setup (different kernel config but eventually also different CPU features), I ran into this. Note that you were looking at the generic __clear_user() implementation, the x86-64 variant is different, see arch/x86/lib/usercopy_64.c I can spot that it triggers stac()/clac() (X86_SMAP): https://en.wikipedia.org/wiki/Supervisor_Mode_Access_Prevention "that allows supervisor mode programs to optionally set user-space memory mappings so that access to those mappings from supervisor mode will cause a trap. This makes it harder for malicious programs to "trick" the kernel into using instructions or data from a user-space program" Yes, that's most probably it :) -- Thanks, David / dhildenb
