Re: [PATCH 4/4] block: add configurable error injection

[Christoph Hellwig <hch@lst.de>]

On Mon, Jun 08, 2026 at 08:53:22AM -0600, Jens Axboe wrote:
> > +	if (!test_bit(GD_ERROR_INJECT, &bio->bi_bdev->bd_disk->state))
> > +		return false;
> > +	return __blk_error_inject(bio);
> > +}
> 
> I really hate this part, that's a pretty deep set of pointer chasings to
> figure out if injection is enabled or not,

It's to the bdev we use everywhere, and then to the disk which we use
in a lot of places in the submission path.

The only easy way to reduce it would be to move the state to the
block_device.  We currently don't do partitions in debugfs, but maybe
we should?

> when in practice error
> injection is only ever enabled for specific test cases and distros
> invariably will set CONFIG_BLK_ERROR_INJECTION because they turn on
> every damn thing under the sun.
> 
> IOW, that won't fly for the hot path. Maybe a static key would be useful
> here?

a static_key makes sense here, probably including the legacy error
injection.