On Thu, Jan 26, 2023 at 8:08 AM Christophe Leroy <christophe.leroy@xxxxxxxxxx> wrote: > > On powerpc64, you can build a kernel with KASAN as soon as you build it > with RADIX MMU support. However if the CPU doesn't have RADIX MMU, > KASAN isn't enabled at init and the following Oops is encountered. > > [ 0.000000][ T0] KASAN not enabled as it requires radix! > > [ 4.484295][ T26] BUG: Unable to handle kernel data access at 0xc00e000000804a04 > [ 4.485270][ T26] Faulting instruction address: 0xc00000000062ec6c > [ 4.485748][ T26] Oops: Kernel access of bad area, sig: 11 [#1] > [ 4.485920][ T26] BE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=2048 NUMA pSeries > [ 4.486259][ T26] Modules linked in: > [ 4.486637][ T26] CPU: 0 PID: 26 Comm: kworker/u2:2 Not tainted 6.2.0-rc3-02590-gf8a023b0a805 #249 > [ 4.486907][ T26] Hardware name: IBM pSeries (emulated by qemu) POWER9 (raw) 0x4e1200 0xf000005 of:SLOF,HEAD pSeries > [ 4.487445][ T26] Workqueue: eval_map_wq .tracer_init_tracefs_work_func > [ 4.488744][ T26] NIP: c00000000062ec6c LR: c00000000062bb84 CTR: c0000000002ebcd0 > [ 4.488867][ T26] REGS: c0000000049175c0 TRAP: 0380 Not tainted (6.2.0-rc3-02590-gf8a023b0a805) > [ 4.489028][ T26] MSR: 8000000002009032 <SF,VEC,EE,ME,IR,DR,RI> CR: 44002808 XER: 00000000 > [ 4.489584][ T26] CFAR: c00000000062bb80 IRQMASK: 0 > [ 4.489584][ T26] GPR00: c0000000005624d4 c000000004917860 c000000001cfc000 1800000000804a04 > [ 4.489584][ T26] GPR04: c0000000003a2650 0000000000000cc0 c00000000000d3d8 c00000000000d3d8 > [ 4.489584][ T26] GPR08: c0000000049175b0 a80e000000000000 0000000000000000 0000000017d78400 > [ 4.489584][ T26] GPR12: 0000000044002204 c000000003790000 c00000000435003c c0000000043f1c40 > [ 4.489584][ T26] GPR16: c0000000043f1c68 c0000000043501a0 c000000002106138 c0000000043f1c08 > [ 4.489584][ T26] GPR20: c0000000043f1c10 c0000000043f1c20 c000000004146c40 c000000002fdb7f8 > [ 4.489584][ T26] GPR24: c000000002fdb834 c000000003685e00 c000000004025030 c000000003522e90 > [ 4.489584][ T26] GPR28: 0000000000000cc0 c0000000003a2650 c000000004025020 c000000004025020 > [ 4.491201][ T26] NIP [c00000000062ec6c] .kasan_byte_accessible+0xc/0x20 > [ 4.491430][ T26] LR [c00000000062bb84] .__kasan_check_byte+0x24/0x90 > [ 4.491767][ T26] Call Trace: > [ 4.491941][ T26] [c000000004917860] [c00000000062ae70] .__kasan_kmalloc+0xc0/0x110 (unreliable) > [ 4.492270][ T26] [c0000000049178f0] [c0000000005624d4] .krealloc+0x54/0x1c0 > [ 4.492453][ T26] [c000000004917990] [c0000000003a2650] .create_trace_option_files+0x280/0x530 > [ 4.492613][ T26] [c000000004917a90] [c000000002050d90] .tracer_init_tracefs_work_func+0x274/0x2c0 > [ 4.492771][ T26] [c000000004917b40] [c0000000001f9948] .process_one_work+0x578/0x9f0 > [ 4.492927][ T26] [c000000004917c30] [c0000000001f9ebc] .worker_thread+0xfc/0x950 > [ 4.493084][ T26] [c000000004917d60] [c00000000020be84] .kthread+0x1a4/0x1b0 > [ 4.493232][ T26] [c000000004917e10] [c00000000000d3d8] .ret_from_kernel_thread+0x58/0x60 > [ 4.495642][ T26] Code: 60000000 7cc802a6 38a00000 4bfffc78 60000000 7cc802a6 38a00001 4bfffc68 60000000 3d20a80e 7863e8c2 792907c6 <7c6348ae> 20630007 78630fe0 68630001 > [ 4.496704][ T26] ---[ end trace 0000000000000000 ]--- > > The Oops is due to kasan_byte_accessible() not checking the readiness > of KASAN. Add missing call to kasan_arch_is_ready() and bail out when > not ready. The same problem is observed with ____kasan_kfree_large() > so fix it the same. > > Also, as KASAN is not available and no shadow area is allocated for > linear memory mapping, there is no point in allocating shadow mem for > vmalloc memory as shown below in /sys/kernel/debug/kernel_page_tables > > ---[ kasan shadow mem start ]--- > 0xc00f000000000000-0xc00f00000006ffff 0x00000000040f0000 448K r w pte valid present dirty accessed > 0xc00f000000860000-0xc00f00000086ffff 0x000000000ac10000 64K r w pte valid present dirty accessed > 0xc00f3ffffffe0000-0xc00f3fffffffffff 0x0000000004d10000 128K r w pte valid present dirty accessed > ---[ kasan shadow mem end ]--- > > So, also verify KASAN readiness before allocating and poisoning > shadow mem for VMAs. Hi Cristophe, Would it possible to unify kasan_arch_is_ready with the already existing kasan_enabled check? Both functions seem to be serving a similar purpose: for example this patch adds kasan_arch_is_ready into __kasan_poison_vmalloc, which is called by kasan_poison_vmalloc when kasan_enabled returns true. The kasan_enabled is only implemented for HW_TAGS right now, but it should be easy enough to make it work other cases by kasan_flag_enabled into common.c and adding __wrappers for shadow-related functions into include/linux/kasan.h. This way architectures won't need to define their own static key and duplicate the functionality. I don't mind having this patch applied as is, considering that it's a fix. However, if the unification that I mentioned is possible, that would be a nice improvement. Thanks!
