On MT7620-class platforms (CONFIG_NET_RALINK_MT7620) we observe sporadic
wrong-jump-targets, kernel oopses, hanging, corrupted backtraces or even
"half-written" instructions when the compiler emits a direct 'jal imm26'
call.
This is triggered in:
- the small random helpers inside get_random_u32_below(), and
- the blkcg_maybe_throttle_current() call in resume_user_mode_work().
This patch forces those two call sites to use an indirect call via
a volatile function pointer (load into register + jalr) when building
for MT7620, avoiding embedding a 26-bit immediate jump target.
Additionally, on MT7620 builds the exec path in fs/exec.c is modified:
- skip arch_align_stack() + PAGE_ALIGN() in setup_arg_pages()
because the micro-randomization (< PAGE_SIZE) implemented by many
ports (including MT7620) is negated immediately by PAGE_ALIGN().
Skipping the redundant PAGE_ALIGN() reduces exposure to the
problematic code pattern.
These changes are targeted workarounds for MT7620; behavioral logic is unchanged.
Signed-off-by: Mieczyslaw Nalewaj <namiltd@yahoo.com>
Link: https://github.com/openwrt/openwrt/pull/20553
(cherry picked from commit 22ccb445e611018ede702e6c0798640353632cce)
Link: https://github.com/openwrt/openwrt/pull/22216
Signed-off-by: Robert Marko <robimarko@gmail.com>
(cherry picked from commit f7005097ae35508ad619251ce7ed127580218e61)
ae1a602e44