Data corruption on i.MX6 IPU in arm_copy_from_user()

[Krzysztof Hałasa]

Hello,

I've encountered an interesting case of data corruption while accessing
IPU (Image Processing Unit) on i.MX6 (rev1.2, Cortex A9). What I'm doing
here is basically:

openat(AT_FDCWD, "/dev/mem", O_RDWR|O_SYNC) = 3
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_SHARED, 3, 0x2630000) = ptr
write(1, ptr, 32)                = 32

Normally, the write() should end up with:
 04008A00 02FF03FF 02FF03FF 00000000 00000000 00000000 00000000 00000000

However, with current kernels, the first 32 bits (the first IPU
 register) are dropped:
 02FF03FF 02FF03FF 00000000 00000000 00000000 00000000 00000000 00000000

0x2630000 is IPU1 CSI0 address (i.e., a register block). The same
happens with other IPU regions. Writes shorter than 8 * 32 bits are not
affected.

write() uses arm_copy_from_user() and since commit f441882a5229:
    ARM: 8812/1: Optimise copy_{from/to}_user for !CPU_USE_DOMAINS

    ARMv6+ processors do not use CONFIG_CPU_USE_DOMAINS and use privileged
    ldr/str instructions in copy_{from/to}_user.  They are currently
    unnecessarily using single ldr/str instructions and can use ldm/stm
    instructions instead like memcpy does (but with appropriate fixup
    tables).

apparently uses 8 * 32-bit ldmia instruction to copy data:
    .macro ldr8w ptr reg1 reg2 reg3 reg4 reg5 reg6 reg7 reg8 abort
    USERL(\abort, ldmia \ptr!, {\reg1, \reg2, \reg3, \reg4, \reg5, \reg6, \reg7, \reg8})
    .endm

Before this commit it used ldr instruction (single 32-bit value) and the
problem didn't show up (reverting f441882a5229 on v5.11 fixes it as
well). The i.MX6 errata doesn't seem to list this problem.

I wonder what the theory says about this case. Is it at all valid to
read 8 IPU registers at a time using LDM instruction? If so, should
something be done with this problem, or should it be left as is?
-- 
Krzysztof Hałasa

Sieć Badawcza Łukasiewicz
Przemysłowy Instytut Automatyki i Pomiarów PIAP
Al. Jerozolimskie 202, 02-486 Warszawa