On Wed, 29 Oct 2025 at 19:08, Al Viro [off-list ref] wrote:

On Tue, Oct 28, 2025 at 10:34:51PM +0100, Ard Biesheuvel wrote:

quoted

I'll let James respond to the specifics of your suggestion, but I'll
just note that this code has a rather convoluted history, as we used
to have two separate pseudo-filesystem drivers, up until a few years
ago: the sysfs based 'efivars' and this efivarfs driver. Given that
modifications in one needed to be visible in the other, they shared a
linked list that shadowed the state of the underlying variable store.
'efivars' was removed years ago, but it was only recently that James
replaced the linked list in this driver with the dentry cache as the
shadow mechanism.

Hmm...  Another question about that code: is efivar_get_variable()
safe outside of efivar_lock()?

Not according to its kerneldoc

/*
 * efivar_get_variable() - retrieve a variable identified by name/vendor
 *
 * Must be called with efivars_lock held.
 */

But actually, I'm not convinced this is accurate. The reason for
locking at this level is mainly to ensure that a SetVariable() call
does not interfere with an ongoing enumeration calling
GetNextVariable() in a loop. The individual EFI runtime calls are
still serialized at a lower level, given that the firmware is not
reentrant, and so holding efivars_lock does not provide anything
meaningful for a GetVariable() call.