Re: [PATCH -v2] ext4: enable scoped NOFS when starting a handle in nojournal mode

On Jul 16, 2026, at 08:48, Theodore Ts'o <[email protected]> wrote:
> 
> The jbd2 layer enables NOFS mode using memalloc_nofs_{save,restore}()
> while a handle is active.  We need to do the same in nojournal mode so
> that it is safe to remove GFP_NOFS flags while a jbd2 handle is
> active.
> 
> This will require that we actually allocate a real handle, but with an
> h_invalid flag set, so there is a place to put the saved memalloc
> context.

If you wanted to avoid an allocation for each nojournal handle on 64-bit
systems, you could put a magic number in the high 32 bits of the handle
pointer (which is unlikely to conflict with real memory addresses),
and the PF_* flags in the low 32 bits.

While the memory savings would be minimal (32 bytes for minimum kmalloc),
it would avoid a potential new allocation failure path for that mode.

> diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h
> index b68561187e90..7348fdadc810 100644
> --- a/include/linux/jbd2.h
> +++ b/include/linux/jbd2.h
> @@ -513,6 +513,7 @@ struct jbd2_journal_handle
>  	unsigned int h_sync: 1;
>  	unsigned int h_reserved: 1;
>  	unsigned int h_aborted: 1;
> +	unsigned int h_invalid: 1;
>  	unsigned int h_type: 8;
>  	unsigned int h_line_no: 16;

Not sure if it makes much difference, but should h_line_no be aligned on a
16-bit offset and h_type aligned on an 8-bit offset by adding an
"unsigned int h_unused:4" field before h_type, or reorder them so the 1-bit
fields are at the end?  That would allow the integer fields to be accessed
directly rather than also having to shift them.

Cheers, Andreas