[PATCH v2 3/6] powerpc/32s: Only leave NX unset on segments used for modules

[Christophe Leroy]

On 08/21/2020 05:11 AM, Christophe Leroy wrote:
> 
> 
> Le 21/08/2020 à 00:00, Andreas Schwab a écrit :
>> On Jun 29 2020, Christophe Leroy wrote:
>>
>>> Instead of leaving NX unset on all segments above the start
>>> of vmalloc space, only leave NX unset on segments used for
>>> modules.
>>
>> I'm getting this crash:
>>
>> kernel tried to execute exec-protected page (f294b000) - exploit 
>> attempt (uid: 0)
>> BUG: Unable to handle kernel instruction fetch
>> Faulting instruction address: 0xf294b000
>> Oops: Kernel access of bad area, sig: 11 [#1]
>> BE PAGE_SIZE=4K MMU=Hash PowerMac
>> Modules linked in: pata_macio(+)
>> CPU: 0 PID: 87 Comm: udevd Not tainted 5.8.0-rc2-test #49
>> NIP:  f294b000 LR: 0005c60 CTR: f294b000
>> REGS: f18d9cc0 TRAP: 0400  Not tainted  (5.8.0-rc2-test)
>> MSR:  10009032 <E,ME,IR,DR,RI>  CR: 84222422  XER: 20000000
>> GPR00: c0005c14 f18d9d78 ef30ca20 00000000 ef0000e0 c00993d0 ef6da038 
>> 0000005e
>> GPR08: c09050b8 c08b0000 00000000 f18d9d78 44222422 10072070 00000000 
>> 0fefaca4
>> GPR16: 1006a00c f294d50b 00000120 00000124 c0096ea8 0000000e ef2776c0 
>> ef2776e4
>> GPR24: f18fd6e8 00000001 c086fe64 c086fe04 00000000 c08b0000 f294b000 
>> ffffffff
>> NIP [f294b000] pata_macio_init+0x0/0xc0 [pata_macio]
>> LR [c0005c60] do_one_initcall+0x6c/0x160
>> Call Trace:
>> [f18d9d78] [c0005c14] do_one_initcall+0x20/0x160 (unreliable)
>> [f18d9dd8] [c009a22c] do_init_module+0x60/0x1c0
>> [f18d9df8] [c00993d8] load_module+0x16a8/0x1c14
>> [f18d9ea8] [c0099aa4] sys_finit_module+0x8c/0x94
>> [f18d9f38] [c0012174] ret_from_syscall+0x0/0x34
>> --- interrupt: c01 at 0xfdb4318
>>     LR = 0xfeee9c0
>> Instruction dump:
>> XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
>> XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX <3d20c08b> 3d40c086 9421ffe0 8129106c
>> ---[ end trace 85a98cc836109871 ]---
>>
> 
> Please try the patch at 
> https://patchwork.ozlabs.org/project/linuxppc-dev/patch/07884ed033c31e074747b7eb8eaa329d15db07ec.1596641219.git.christophe.leroy@csgroup.eu/ 
> 
> 
> And if you are using KAsan, also take 
> https://patchwork.ozlabs.org/project/linuxppc-dev/patch/6eddca2d5611fd57312a88eae31278c87a8fc99d.1596641224.git.christophe.leroy@csgroup.eu/ 
> 
> 
> Allthough I have some doubt that it will fix it, because the faulting 
> instruction address is at 0xf294b000 which is within the vmalloc area. 
> In the likely case the patch doesn't fix the issue, can you provide your 
> .config and a dump of /sys/kernel/debug/powerpc/segment_registers (You 
> have to have CONFIG_PPC_PTDUMP enabled for that) and also the below part 
> from boot log.
> 
> [    0.000000] Memory: 509556K/524288K available (7088K kernel code, 
> 592K rwdata, 1304K rodata, 356K init, 803K bss, 14732K reserved, 0K 
> cma-reserved)
> [    0.000000] Kernel virtual memory layout:
> [    0.000000]   * 0xff7ff000..0xfffff000  : fixmap
> [    0.000000]   * 0xff7fd000..0xff7ff000  : early ioremap
> [    0.000000]   * 0xe1000000..0xff7fd000  : vmalloc & ioremap
> 


I found the issue, when VMALLOC_END is above 0xf0000000, 
ALIGN(VMALLOC_END, SZ_256M) is 0 so the test is always false.

The below change should fix it.

diff --git a/arch/powerpc/mm/book3s32/mmu.c b/arch/powerpc/mm/book3s32/mmu.c
index 82ae9e06a773..d426eaf76bb0 100644
--- a/arch/powerpc/mm/book3s32/mmu.c
+++ b/arch/powerpc/mm/book3s32/mmu.c
@@ -194,12 +194,12 @@ static bool is_module_segment(unsigned long addr)
  #ifdef MODULES_VADDR
  	if (addr < ALIGN_DOWN(MODULES_VADDR, SZ_256M))
  		return false;
-	if (addr >= ALIGN(MODULES_END, SZ_256M))
+	if (addr > ALIGN(MODULES_END, SZ_256M) - 1)
  		return false;
  #else
  	if (addr < ALIGN_DOWN(VMALLOC_START, SZ_256M))
  		return false;
-	if (addr >= ALIGN(VMALLOC_END, SZ_256M))
+	if (addr > ALIGN(VMALLOC_END, SZ_256M) - 1)
  		return false;
  #endif
  	return true;


Christophe