Re: [PATCH v5 00/16] x86: make PAT and MTRR independent from each other

From: Juergen Gross
Date: Tue Nov 08 2022 - 02:30:30 EST


On 07.11.22 20:25, Borislav Petkov wrote:
On Thu, Nov 03, 2022 at 05:15:52PM +0100, Borislav Petkov wrote:
Lemme try to find a smaller box which shows that too - that one is a
pain to bisect on.

Ok, couldn't find a smaller one (or maybe it had to be a big one to
tickle this out).

So I think it is the parallel setup thing:

x86/mtrr: Do MTRR/PAT setup on all secondary CPUs in parallel

Note that before it, it would do the configuration sequentially on each
CPU:

[ 0.759239] MTRR: prepare_set: CPU83, MSR_MTRRdefType: 0x0, read: (0xc00:0)
[ 0.759239] MTRR: set_mtrr_state: CPU83, mtrr_deftype_lo: 0xc00, mtrr_state.def_type: 0, mtrr_state.enabled: 3
[ 0.760794] MTRR: post_set: CPU83, MSR_MTRRdefType will write: (0xc00:0)
[ 0.761151] MTRR: prepare_set: CPU70, MSR_MTRRdefType: 0x0, read: (0xc00:0)
[ 0.761151] MTRR: set_mtrr_state: CPU70, mtrr_deftype_lo: 0xc00, mtrr_state.def_type: 0, mtrr_state.enabled: 3
[ 0.761151] MTRR: post_set: CPU70, MSR_MTRRdefType will write: (0xc00:0)
...

and so on.

Now, it would do it all in parallel:

[ 0.762006] MTRR: mtrr_disable: CPU70, MSR_MTRRdefType: 0x0, read: (0xc00:0)
[ 0.761916] MTRR: mtrr_disable: CPU18, MSR_MTRRdefType: 0x0, read: (0xc00:0)
[ 0.761808] MTRR: mtrr_disable: CPU82, MSR_MTRRdefType: 0x0, read: (0xc00:0)
[ 0.762593] MTRR: mtrr_disable: CPU6, MSR_MTRRdefType: 0x0, read: (0x0:0)
^^^^^^

Note that last thing. That comes from (with debug output added):

void mtrr_disable(struct cache_state *state)
{
unsigned int cpu = smp_processor_id();
u64 msrval;

/* Save MTRR state */
rdmsr(MSR_MTRRdefType, state->mtrr_deftype_lo, state->mtrr_deftype_hi);

/* Disable MTRRs, and set the default type to uncached */
mtrr_wrmsr(MSR_MTRRdefType, state->mtrr_deftype_lo & ~0xcff,
state->mtrr_deftype_hi);

rdmsrl(MSR_MTRRdefType, msrval);

pr_info("%s: CPU%d, MSR_MTRRdefType: 0x%llx, read: (0x%x:%x)\n",
__func__, cpu, msrval, state->mtrr_deftype_lo, state->mtrr_deftype_hi);
}

The "read: (0x0:0)" basically says that

state->mtrr_deftype_lo, state->mtrr_deftype_hi

are both 0 already. BUT(!), they should NOT be. The low piece is 0xc00 on most
cores except a handful and it means that MTRRs and Fixed Range are
enabled. In total, they're these cores here:

[ 0.762593] MTRR: mtrr_disable: CPU6, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762247] MTRR: mtrr_disable: CPU26, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762685] MTRR: mtrr_disable: CPU68, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762725] MTRR: mtrr_disable: CPU17, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762685] MTRR: mtrr_disable: CPU69, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762800] MTRR: mtrr_disable: CPU1, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762734] MTRR: mtrr_disable: CPU13, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762720] MTRR: mtrr_disable: CPU24, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762696] MTRR: mtrr_disable: CPU66, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762716] MTRR: mtrr_disable: CPU48, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762693] MTRR: mtrr_disable: CPU57, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762519] MTRR: mtrr_disable: CPU87, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762532] MTRR: mtrr_disable: CPU58, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762755] MTRR: mtrr_disable: CPU32, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762693] MTRR: mtrr_disable: CPU52, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762861] MTRR: mtrr_disable: CPU0, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762724] MTRR: mtrr_disable: CPU21, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762640] MTRR: mtrr_disable: CPU15, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762615] MTRR: mtrr_disable: CPU50, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762741] MTRR: mtrr_disable: CPU40, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762738] MTRR: mtrr_disable: CPU37, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762716] MTRR: mtrr_disable: CPU25, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762512] MTRR: mtrr_disable: CPU59, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762721] MTRR: mtrr_disable: CPU45, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762682] MTRR: mtrr_disable: CPU56, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762583] MTRR: mtrr_disable: CPU124, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762751] MTRR: mtrr_disable: CPU12, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762741] MTRR: mtrr_disable: CPU9, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762575] MTRR: mtrr_disable: CPU51, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762632] MTRR: mtrr_disable: CPU100, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762688] MTRR: mtrr_disable: CPU61, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762610] MTRR: mtrr_disable: CPU105, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762721] MTRR: mtrr_disable: CPU20, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.762583] MTRR: mtrr_disable: CPU47, MSR_MTRRdefType: 0x0, read: (0x0:0)

Now, if I add MFENCEs around those RDMSRs:

void mtrr_disable(struct cache_state *state)
{
unsigned int cpu = smp_processor_id();
u64 msrval;

/* Save MTRR state */
rdmsr(MSR_MTRRdefType, state->mtrr_deftype_lo, state->mtrr_deftype_hi);

__mb();

/* Disable MTRRs, and set the default type to uncached */
mtrr_wrmsr(MSR_MTRRdefType, state->mtrr_deftype_lo & ~0xcff,
state->mtrr_deftype_hi);

__mb();

rdmsrl(MSR_MTRRdefType, msrval);

pr_info("%s: CPU%d, MSR_MTRRdefType: 0x%llx, read: (0x%x:%x)\n",
__func__, cpu, msrval, state->mtrr_deftype_lo, state->mtrr_deftype_hi);

__mb();
}

the amount of cores becomes less:

Probably not because of the fencing, but because of the different timing.


[ 0.765260] MTRR: mtrr_disable: CPU6, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765462] MTRR: mtrr_disable: CPU5, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765242] MTRR: mtrr_disable: CPU22, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765522] MTRR: mtrr_disable: CPU0, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765474] MTRR: mtrr_disable: CPU1, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765207] MTRR: mtrr_disable: CPU54, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765225] MTRR: mtrr_disable: CPU8, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765282] MTRR: mtrr_disable: CPU88, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765150] MTRR: mtrr_disable: CPU119, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765370] MTRR: mtrr_disable: CPU49, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765395] MTRR: mtrr_disable: CPU16, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765348] MTRR: mtrr_disable: CPU52, MSR_MTRRdefType: 0x0, read: (0x0:0)
[ 0.765270] MTRR: mtrr_disable: CPU58, MSR_MTRRdefType: 0x0, read: (0x0:0)

which basically hints at some speculative fun where we end up reading
the MSR *after* the write to it has already happened. After this thing:

/* Disable MTRRs, and set the default type to uncached */
mtrr_wrmsr(MSR_MTRRdefType, state->mtrr_deftype_lo & ~0xcff,
state->mtrr_deftype_hi);

and thus when we read it, we already read the disabled state. But this
is only a conjecture because I still have no clear idea how TF would
that even happen?!?

Yeah, and why doesn't it happen when we handle only one cpu at a time?

There might be some interaction between the cpus ...


Needless to say, this fixes it, ofc:

diff --git a/arch/x86/kernel/cpu/cacheinfo.c b/arch/x86/kernel/cpu/cacheinfo.c
index 3805a6d32d37..4a685898caf3 100644
--- a/arch/x86/kernel/cpu/cacheinfo.c
+++ b/arch/x86/kernel/cpu/cacheinfo.c
@@ -1116,12 +1116,14 @@ void cache_enable(struct cache_state *state)
__write_cr4(state->cr4);
}
+static DEFINE_RAW_SPINLOCK(set_atomicity_lock);
+
static void cache_cpu_init(void)
{
unsigned long flags;
struct cache_state state = { };
- local_irq_save(flags);
+ raw_spin_lock_irqsave(&set_atomicity_lock, flags);
cache_disable(&state);
if (memory_caching_control & CACHE_MTRR)
@@ -1131,7 +1133,7 @@ static void cache_cpu_init(void)
pat_cpu_init();
cache_enable(&state);
- local_irq_restore(flags);
+ raw_spin_unlock_irqrestore(&set_atomicity_lock, flags);
}
static bool cache_aps_delayed_init = true;

---

and frankly, considering how we have bigger fish to fry, I'd say we do
it the old way and leave that can'o'worms half-opened.

I agree to keep this patch out of the series for now.


Unless you wanna continue poking at it. I can give you access to that
box at work...

Yes, please. I suspect there are some additional requirements for updating
MTRR in parallel, or this is "just" a cpu bug.


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key

Attachment: OpenPGP_signature
Description: OpenPGP digital signature