Re: [PATCH v2] x86/livepatch: Fix livepatch application when CET is active

[Jan Beulich <jbeulich@xxxxxxxx>]

On 17.04.2023 15:52, Andrew Cooper wrote:
> Right now, trying to apply a livepatch on any system with CET shstk (AMD Zen3
> or later, Intel Tiger Lake or Sapphire Rapids and later) fails as follows:
> 
>   (XEN) livepatch: lp: Verifying enabled expectations for all functions
>   (XEN) common/livepatch.c:1591: livepatch: lp: timeout is 30000000ns
>   (XEN) common/livepatch.c:1703: livepatch: lp: CPU28 - IPIing the other 127 
> CPUs
>   (XEN) livepatch: lp: Applying 1 functions
>   (XEN) hi_func: Hi! (called 1 times)
>   (XEN) Hook executing.
>   (XEN) Assertion 'local_irq_is_enabled() || cpumask_subset(mask, 
> cpumask_of(cpu))' failed at arch/x86/smp.c:265
>   (XEN) *** DOUBLE FAULT ***
>   <many double faults>
> 
> The assertion failure is from a global (system wide) TLB flush initiated by
> modify_xen_mappings().  I'm not entirely sure when this broke, and I'm not
> sure exactly what causes the #DF's, but it doesn't really matter either
> because they highlight a latent bug that I'd overlooked with the CET-SS vs
> patching work the first place.
> 
> While we're careful to arrange for the patching CPU to avoid encountering
> non-shstk memory with transient shstk perms, other CPUs can pick these
> mappings up too if they need to re-walk for uarch reasons.
> 
> Another bug is that for livepatching, we only disable CET if shadow stacks are
> in use.  Running on Intel CET systems when Xen is only using CET-IBT will
> crash in arch_livepatch_quiesce() when trying to clear CR0.WP with CR4.CET
> still active.
> 
> Also, we never went and cleared the dirty bits on .rodata.  This would
> matter (for the same reason it matters on .text - it becomes a valid target
> for WRSS), but we never actually patch .rodata anyway.
> 
> Therefore rework how we do patching for both alternatives and livepatches.
> 
> Introduce modify_xen_mappings_lite() with a purpose similar to
> modify_xen_mappings(), but stripped down to the bare minimum as it's used in
> weird contexts.  Leave all complexity to the caller to handle.
> 
> Instead of patching by clearing CR0.WP (and having to jump through some
> fragile hoops to disable CET in order to do this), just transiently relax the
> permissions on .text via l2_identmap[].
> 
> Note that neither alternatives nor livepatching edit .rodata, so we don't need
> to relax those permissions at this juncture.
> 
> The perms are relaxed globally, but is safe enough.  Alternatives run before
> we boot APs, and Livepatching runs in a quiesced state where the other CPUs
> are not doing anything interesting.
> 
> This approach is far more robust.
> 
> Fixes: 48cdc15a424f ("x86/alternatives: Clear CR4.CET when clearing CR0.WP")
> Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>

Reviewed-by: Jan Beulich <jbeulich@xxxxxxxx>

One further remark, though:

> @@ -5879,6 +5880,73 @@ int destroy_xen_mappings(unsigned long s, unsigned 
> long e)
>      return modify_xen_mappings(s, e, _PAGE_NONE);
>  }
>  
> +/*
> + * Similar to modify_xen_mappings(), but used by the alternatives and
> + * livepatch in weird contexts.  All synchronization, TLB flushing, etc is 
> the
> + * responsibility of the caller, and *MUST* not be introduced here.
> + *
> + * Must be limited to XEN_VIRT_{START,END}, i.e. over l2_xenmap[].
> + * Must be called with present flags, and over present mappings.
> + * Must be called on leaf page boundaries.

This last sentence, while wording-wise correct, could do with making more
explicit that it is the caller's responsibility to know whether large page
mappings are in use, due to ...

> + */
> +void init_or_livepatch modify_xen_mappings_lite(
> +    unsigned long s, unsigned long e, unsigned int _nf)
> +{
> +    unsigned long v = s, fm, nf;
> +
> +    /* Set of valid PTE bits which may be altered. */
> +#define FLAGS_MASK 
> (_PAGE_NX|_PAGE_DIRTY|_PAGE_ACCESSED|_PAGE_RW|_PAGE_PRESENT)
> +    fm = put_pte_flags(FLAGS_MASK);
> +    nf = put_pte_flags(_nf & FLAGS_MASK);
> +#undef FLAGS_MASK
> +
> +    ASSERT(nf & _PAGE_PRESENT);
> +    ASSERT(IS_ALIGNED(s, PAGE_SIZE) && s >= XEN_VIRT_START);
> +    ASSERT(IS_ALIGNED(e, PAGE_SIZE) && e <= XEN_VIRT_END);
> +
> +    while ( v < e )
> +    {
> +        l2_pgentry_t *pl2e = &l2_xenmap[l2_table_offset(v)];
> +        l2_pgentry_t l2e = l2e_read_atomic(pl2e);
> +        unsigned int l2f = l2e_get_flags(l2e);
> +
> +        ASSERT(l2f & _PAGE_PRESENT);
> +
> +        if ( l2e_get_flags(l2e) & _PAGE_PSE )
> +        {
> +            ASSERT(l1_table_offset(v) == 0);

... this.

Jan