On Sat, Aug 29, 2026 at 09:20:39AM +0800, Geliang Tang wrote:
> From: Geliang Tang <[email protected]>
>
> The `gfp` parameter of alloc_buf() has been dead code since the
> function was introduced. The underlying allocation was always
> hardcoded to GFP_KERNEL - first via kmalloc(..., GFP_KERNEL), then
> via kmalloc_flex(*buf, sg, pages, GFP_KERNEL), and most recently
> via kmalloc_flex(*buf, sg, pages) which defaults to GFP_KERNEL.
>
> This causes a "sleeping function called from invalid context" BUG
> when the hvc console flush path calls put_chars() with IRQs
> disabled:
>
> BUG: sleeping function called from invalid context at \
> include/linux/sched/mm.h:322
> in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 416, \
> name: test_progs
> ...
> Call Trace:
> __kmalloc_noprof+0x468/0x7c0
> alloc_buf.isra.0+0x42/0x260
> put_chars+0x1e5/0x2f0
> hvc_console_print+0x29d/0x6e0
> console_emit_next_record+0x21d/0x480
> console_flush_one_record+0x431/0x6a0
> console_unlock+0xda/0x1c0
> vprintk_emit+0x300/0x350
>
> put_chars() correctly passes GFP_ATOMIC, but alloc_buf() silently
> discards it and allocates with GFP_KERNEL, which may sleep.
>
> Pass the caller-supplied gfp flag through to kmalloc_flex() so
> that atomic callers get atomic allocations.
>
> Fixes: 69050f8d6d07 ("treewide: Replace kmalloc with kmalloc_obj for
> non-scalar types")
> Signed-off-by: Geliang Tang <[email protected]>
Sorry for the delay, I took this earlier patch instead:
https://lore.kernel.org/r/[email protected]
