Re: [PATCH v2 4/7] x86/xen: support nested lazy_mmu sections (again)

[David Hildenbrand <david@xxxxxxxxxx>]

On 09.09.25 11:37, Jürgen Groß wrote:
> On 09.09.25 11:13, David Hildenbrand wrote:
> > On 08.09.25 09:39, Kevin Brodsky wrote:
> > > Commit 49147beb0ccb ("x86/xen: allow nesting of same lazy mode")
> > > originally introduced support for nested lazy sections (LAZY_MMU and
> > > LAZY_CPU). It later got reverted by commit c36549ff8d84 as its
> > > implementation turned out to be intolerant to preemption.
> > >
> > > Now that the lazy_mmu API allows enter() to pass through a state to
> > > the matching leave() call, we can support nesting again for the
> > > LAZY_MMU mode in a preemption-safe manner. If xen_enter_lazy_mmu() is
> > > called inside an active lazy_mmu section, xen_lazy_mode will already
> > > be set to XEN_LAZY_MMU and we can then return LAZY_MMU_NESTED to
> > > instruct the matching xen_leave_lazy_mmu() call to leave
> > > xen_lazy_mode unchanged.
> > >
> > > The only effect of this patch is to ensure that xen_lazy_mode
> > > remains set to XEN_LAZY_MMU until the outermost lazy_mmu section
> > > ends. xen_leave_lazy_mmu() still calls xen_mc_flush()
> > > unconditionally.
> > >
> > > Signed-off-by: Kevin Brodsky <kevin.brodsky@xxxxxxx>
> > > ---
> > >    arch/x86/include/asm/paravirt.h       |  6 ++----
> > >    arch/x86/include/asm/paravirt_types.h |  4 ++--
> > >    arch/x86/xen/mmu_pv.c                 | 11 ++++++++---
> > >    3 files changed, 12 insertions(+), 9 deletions(-)
> > >
> > > diff --git a/arch/x86/include/asm/paravirt.h b/arch/x86/include/asm/paravirt.h
> > > index 65a0d394fba1..4ecd3a6b1dea 100644
> > > --- a/arch/x86/include/asm/paravirt.h
> > > +++ b/arch/x86/include/asm/paravirt.h
> > > @@ -529,14 +529,12 @@ static inline void arch_end_context_switch(struct
> > > task_struct *next)
> > >    #define  __HAVE_ARCH_ENTER_LAZY_MMU_MODE
> > >    static inline lazy_mmu_state_t arch_enter_lazy_mmu_mode(void)
> > >    {
> > > -    PVOP_VCALL0(mmu.lazy_mode.enter);
> > > -
> > > -    return LAZY_MMU_DEFAULT;
> > > +    return PVOP_CALL0(lazy_mmu_state_t, mmu.lazy_mode.enter);
> > >    }
> > >    static inline void arch_leave_lazy_mmu_mode(lazy_mmu_state_t state)
> > >    {
> > > -    PVOP_VCALL0(mmu.lazy_mode.leave);
> > > +    PVOP_VCALL1(mmu.lazy_mode.leave, state);
> > >    }
> > >    static inline void arch_flush_lazy_mmu_mode(void)
> > > diff --git a/arch/x86/include/asm/paravirt_types.h b/arch/x86/include/asm/
> > > paravirt_types.h
> > > index bc1af86868a3..b7c567ccbf32 100644
> > > --- a/arch/x86/include/asm/paravirt_types.h
> > > +++ b/arch/x86/include/asm/paravirt_types.h
> > > @@ -45,8 +45,8 @@ typedef int lazy_mmu_state_t;
> > >    struct pv_lazy_ops {
> > >        /* Set deferred update mode, used for batching operations. */
> > > -    void (*enter)(void);
> > > -    void (*leave)(void);
> > > +    lazy_mmu_state_t (*enter)(void);
> > > +    void (*leave)(lazy_mmu_state_t);
> > >        void (*flush)(void);
> > >    } __no_randomize_layout;
> > >    #endif
> > > diff --git a/arch/x86/xen/mmu_pv.c b/arch/x86/xen/mmu_pv.c
> > > index 2039d5132ca3..6e5390ff06a5 100644
> > > --- a/arch/x86/xen/mmu_pv.c
> > > +++ b/arch/x86/xen/mmu_pv.c
> > > @@ -2130,9 +2130,13 @@ static void xen_set_fixmap(unsigned idx, phys_addr_t
> > > phys, pgprot_t prot)
> > >    #endif
> > >    }
> > > -static void xen_enter_lazy_mmu(void)
> > > +static lazy_mmu_state_t xen_enter_lazy_mmu(void)
> > >    {
> > > +    if (this_cpu_read(xen_lazy_mode) == XEN_LAZY_MMU)
> > > +        return LAZY_MMU_NESTED;
> > > +
> > You mention above "preemption-safe manner" above, so I am wondering,
> > what if we get preempted immediately after doing the this_cpu_read() and get
> > scheduled on another CPU?
> This should still be correct: preemption needs a context switch to happen,
> so xen_start_context_switch() and xen_end_context_switch() are involved.
> Those are dealing with this problem by doing the right thing in the old
> and the new context.
Thanks, that makes sense. Would be valuable to add that detail to the 
patch description.
--
Cheers

David / dhildenb