Re: [PATCH v4] x86/bugs: Add a separate config for each mitigation

[Josh Poimboeuf]

On Mon, Oct 16, 2023 at 11:45:38AM +0200, Ingo Molnar wrote:
> * Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> > On Fri, Oct 13, 2023 at 12:50:59PM +0200, Ingo Molnar wrote:
> > > * Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> > > > On Thu, Oct 12, 2023 at 08:51:24PM +0200, Ingo Molnar wrote:
> > > > > > Another way to avoid ifdeffery:
> > > > > > 
> > > > > > static enum retbleed_mitigation_cmd retbleed_cmd __ro_after_init =
> > > > > > 	IS_ENABLED(CONFIG_MITIGATION_RETBLEED) ? RETBLEED_CMD_AUTO : RETBLEED_CMD_OFF;
> > > > > 
> > > > > I think we could make it a simple:
> > > > > 
> > > > > 	static enum retbleed_mitigation_cmd retbleed_cmd __ro_after_init = IS_ENABLED(CONFIG_MITIGATION_RETBLEED);
> > > > > 
> > > > > Because RETBLEED_CMD_AUTO && RETBLEED_CMD_OFF maps naturally to 1 and 0. 
> > > > > Maybe add a comment to the enum to maintain this property in the future 
> > > > > too.
> > > > 
> > > > Hm, that both obfuscates the default and makes it fragile.  The fact
> > > > that it would need a comment to try to prevent breaking it in the future
> > > > is a clue that maybe we shouldn't do it ;-)
> > > 
> > > Can be enforced with BUILD_BUG_ON().
> > 
> > That replaces fragility with brittleness.  If we change a default then
> > we have to go rearrange the corresponding enum, and update the
> > BUILD_BUG_ONs.
> 
> How realistic is that? A world in which an enum named '*_OFF' isn't zero 
> and the most obvious second enum isn't 'auto' would be unconditionally sad 
> IMO...

I'm glad you asked ;-)

Not only is it realistic, it already seems to be the case for the
majority of the enums which are used for defaults:

enum l1tf_mitigations {
	L1TF_MITIGATION_OFF,
	L1TF_MITIGATION_FLUSH_NOWARN,
	L1TF_MITIGATION_FLUSH, /* default */
	L1TF_MITIGATION_FLUSH_NOSMT,
	L1TF_MITIGATION_FULL,
	L1TF_MITIGATION_FULL_FORCE
};

enum taa_mitigations {
	TAA_MITIGATION_OFF,
	TAA_MITIGATION_UCODE_NEEDED,
	TAA_MITIGATION_VERW, /* default */
	TAA_MITIGATION_TSX_DISABLED,
};

enum mmio_mitigations {
	MMIO_MITIGATION_OFF,
	MMIO_MITIGATION_UCODE_NEEDED,
	MMIO_MITIGATION_VERW, /* default */
};

enum srbds_mitigations {
	SRBDS_MITIGATION_OFF,
	SRBDS_MITIGATION_UCODE_NEEDED,
	SRBDS_MITIGATION_FULL, /* default */
	SRBDS_MITIGATION_TSX_OFF,
	SRBDS_MITIGATION_HYPERVISOR,
};

enum l1d_flush_mitigations {
	L1D_FLUSH_OFF = 0, / * default */
	L1D_FLUSH_ON,
};

enum gds_mitigations {
	GDS_MITIGATION_OFF,
	GDS_MITIGATION_UCODE_NEEDED,
	GDS_MITIGATION_FORCE,
	GDS_MITIGATION_FULL, /* default */
	GDS_MITIGATION_FULL_LOCKED,
	GDS_MITIGATION_HYPERVISOR,
};

enum srso_mitigation_cmd {
	SRSO_CMD_OFF,
	SRSO_CMD_MICROCODE,
	SRSO_CMD_SAFE_RET, /* default */
	SRSO_CMD_IBPB,
	SRSO_CMD_IBPB_ON_VMEXIT,
};

It's dangerous (and in fact antithetical to the concept of enums!) to
make assumptions about enum values, for both existing and future code.

> > More importantly, it's still less readable because the reader now has to 
> > go read the enum values to cross-reference the hard-coded values of 0 and 
> > 1 with the enums which are used everywhere else.
> 
> They'd have to do that anyway, to make sense of the enum jungle.

Why?  The enum values are completely opaque to the code as far as I can
tell.  Changing that will just make the code even more obtuse.

I don't see any benefit in hard-coding them.  What's the point?

-- 
Josh