Re: [PATCH 4/8] xen/common: Split tlb-clock.h out of mm.h

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 13/03/2025 12:59 pm, Jan Beulich wrote:
> On 12.03.2025 18:45, Andrew Cooper wrote:
>> xen/mm.h includes asm/tlbflush.h almost at the end, which creates a horrible
>> tangle.  This is in order to provide two common files with an abstraction 
>> over
>> the x86-specific TLB clock logic.
>>
>> First, introduce CONFIG_HAS_TLB_CLOCK, selected by x86 only.  Next, introduce
>> xen/tlb-clock.h, providing empty stubs, and include this into memory.c and
>> page_alloc.c
>>
>> No functional change.
>>
>> Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
>> ---
>> CC: Anthony PERARD <anthony.perard@xxxxxxxxxx>
>> CC: Michal Orzel <michal.orzel@xxxxxxx>
>> CC: Jan Beulich <jbeulich@xxxxxxxx>
>> CC: Julien Grall <julien@xxxxxxx>
>> CC: Roger Pau Monné <roger.pau@xxxxxxxxxx>
>> CC: Stefano Stabellini <sstabellini@xxxxxxxxxx>
>> CC: Volodymyr Babchuk <Volodymyr_Babchuk@xxxxxxxx>
>> CC: Bertrand Marquis <bertrand.marquis@xxxxxxx>
>> CC: Oleksii Kurochko <oleksii.kurochko@xxxxxxxxx>
>> CC: Shawn Anastasio <sanastasio@xxxxxxxxxxxxxxxxxxxxx>
>>
>> There is still a mess here with the common vs x86 split, but it's better
>> contained than before.
>> ---
>>  xen/arch/x86/Kconfig        |  1 +
>>  xen/common/Kconfig          |  3 +++
>>  xen/common/memory.c         |  1 +
>>  xen/common/page_alloc.c     |  1 +
>>  xen/include/xen/mm.h        | 27 --------------------
>>  xen/include/xen/tlb-clock.h | 49 +++++++++++++++++++++++++++++++++++++
>>  6 files changed, 55 insertions(+), 27 deletions(-)
>>  create mode 100644 xen/include/xen/tlb-clock.h
>>
>> diff --git a/xen/arch/x86/Kconfig b/xen/arch/x86/Kconfig
>> index 6e41bc0fb435..e9a166ee3dd0 100644
>> --- a/xen/arch/x86/Kconfig
>> +++ b/xen/arch/x86/Kconfig
>> @@ -28,6 +28,7 @@ config X86
>>      select HAS_PCI_MSI
>>      select HAS_PIRQ
>>      select HAS_SCHED_GRANULARITY
>> +    select HAS_TLB_CLOCK
>>      select HAS_UBSAN
>>      select HAS_VMAP
>>      select HAS_VPCI if HVM
>> diff --git a/xen/common/Kconfig b/xen/common/Kconfig
>> index 6166327f4d14..dcf7d9d00d0a 100644
>> --- a/xen/common/Kconfig
>> +++ b/xen/common/Kconfig
>> @@ -83,6 +83,9 @@ config HAS_PMAP
>>  config HAS_SCHED_GRANULARITY
>>      bool
>>  
>> +config HAS_TLB_CLOCK
>> +    bool
>> +
>>  config HAS_UBSAN
>>      bool
>>  
>> diff --git a/xen/common/memory.c b/xen/common/memory.c
>> index 61a94b23abae..9138fd096696 100644
>> --- a/xen/common/memory.c
>> +++ b/xen/common/memory.c
>> @@ -24,6 +24,7 @@
>>  #include <xen/perfc.h>
>>  #include <xen/sched.h>
>>  #include <xen/sections.h>
>> +#include <xen/tlb-clock.h>
>>  #include <xen/trace.h>
>>  
>>  #include <asm/current.h>
>> diff --git a/xen/common/page_alloc.c b/xen/common/page_alloc.c
>> index bc029ea797a2..b90c3d7988b4 100644
>> --- a/xen/common/page_alloc.c
>> +++ b/xen/common/page_alloc.c
>> @@ -137,6 +137,7 @@
>>  #include <xen/sections.h>
>>  #include <xen/softirq.h>
>>  #include <xen/spinlock.h>
>> +#include <xen/tlb-clock.h>
>>  #include <xen/vm_event.h>
>>  #include <xen/xvmalloc.h>
>>  
>> diff --git a/xen/include/xen/mm.h b/xen/include/xen/mm.h
>> index 45000cc3f64b..fff36ff903d6 100644
>> --- a/xen/include/xen/mm.h
>> +++ b/xen/include/xen/mm.h
>> @@ -588,33 +588,6 @@ unsigned long get_upper_mfn_bound(void);
>>  
>>  #include <asm/flushtlb.h>
>>  
>> -static inline void accumulate_tlbflush(bool *need_tlbflush,
>> -                                       const struct page_info *page,
>> -                                       uint32_t *tlbflush_timestamp)
>> -{
>> -    if ( page->u.free.need_tlbflush &&
>> -         page->tlbflush_timestamp <= tlbflush_current_time() &&
>> -         (!*need_tlbflush ||
>> -          page->tlbflush_timestamp > *tlbflush_timestamp) )
>> -    {
>> -        *need_tlbflush = true;
>> -        *tlbflush_timestamp = page->tlbflush_timestamp;
>> -    }
>> -}
>> -
>> -static inline void filtered_flush_tlb_mask(uint32_t tlbflush_timestamp)
>> -{
>> -    cpumask_t mask;
>> -
>> -    cpumask_copy(&mask, &cpu_online_map);
>> -    tlbflush_filter(&mask, tlbflush_timestamp);
>> -    if ( !cpumask_empty(&mask) )
>> -    {
>> -        perfc_incr(need_flush_tlb_flush);
>> -        arch_flush_tlb_mask(&mask);
>> -    }
>> -}
>> -
>>  enum XENSHARE_flags {
>>      SHARE_rw,
>>      SHARE_ro,
>> diff --git a/xen/include/xen/tlb-clock.h b/xen/include/xen/tlb-clock.h
>> new file mode 100644
>> index 000000000000..796c0be7fbef
>> --- /dev/null
>> +++ b/xen/include/xen/tlb-clock.h
>> @@ -0,0 +1,49 @@
>> +/* SPDX-License-Identifier: GPL-2.0-only */
>> +#ifndef XEN_TLB_CLOCK_H
>> +#define XEN_TLB_CLOCK_H
>> +
>> +#include <xen/types.h>
>> +
>> +#ifdef CONFIG_HAS_TLB_CLOCK
>> +
>> +#include <xen/mm.h>
>> +
>> +#include <asm/flushtlb.h>
>> +
>> +static inline void accumulate_tlbflush(
>> +    bool *need_tlbflush, const struct page_info *page,
>> +    uint32_t *tlbflush_timestamp)
>> +{
>> +    if ( page->u.free.need_tlbflush &&
>> +         page->tlbflush_timestamp <= tlbflush_current_time() &&
>> +         (!*need_tlbflush ||
>> +          page->tlbflush_timestamp > *tlbflush_timestamp) )
>> +    {
>> +        *need_tlbflush = true;
>> +        *tlbflush_timestamp = page->tlbflush_timestamp;
>> +    }
>> +}
>> +
>> +static inline void filtered_flush_tlb_mask(uint32_t tlbflush_timestamp)
>> +{
>> +    cpumask_t mask;
>> +
>> +    cpumask_copy(&mask, &cpu_online_map);
>> +    tlbflush_filter(&mask, tlbflush_timestamp);
>> +    if ( !cpumask_empty(&mask) )
>> +    {
>> +        perfc_incr(need_flush_tlb_flush);
> Would this perf counter then perhaps also want to become dependent upon
> HAS_TLB_CLOCK=y (or become an arch-specific one)? It's used elsewhere in x86,
> but not for any of the other arch-es.

There's nothing inherently x86-specific about our tlb-clk implementation.

I don't have time to do it, but if another arch wants to add support,
then it's probably just a case of shuffling some bits out of
asm/flushtlb.h into tlb-clock.h.

>
> However, see below.
>
>> +        arch_flush_tlb_mask(&mask);
>> +    }
>> +}
>> +
>> +#else /* !CONFIG_HAS_TLB_CLOCK */
>> +
>> +struct page_info;
>> +static inline void accumulate_tlbflush(
>> +    bool *need_tlbflush, const struct page_info *page,
>> +    uint32_t *tlbflush_timestamp) {}
>> +static inline void filtered_flush_tlb_mask(uint32_t tlbflush_timestamp) {}
> Is doing nothing here correct?

Yeah, it's not, but this only occurred to me after sending the series.

Interestingly, CI is green across the board for ARM, which suggests to
me that this logic isn't getting a workout.

>  mark_page_free() can set a page's
> ->u.free.need_tlbflush. And with that flag set the full
>
> static inline void accumulate_tlbflush(
>     bool *need_tlbflush, const struct page_info *page,
>     uint32_t *tlbflush_timestamp)
> {
>     if ( page->u.free.need_tlbflush &&
>          page->tlbflush_timestamp <= tlbflush_current_time() &&
>          (!*need_tlbflush ||
>           page->tlbflush_timestamp > *tlbflush_timestamp) )
>     {
>         *need_tlbflush = true;
>         *tlbflush_timestamp = page->tlbflush_timestamp;
>     }
> }
>
> reduces to (considering that tlbflush_current_time() resolves to constant 0,
> which also implies every page's ->tlbflush_timestamp is only ever 0)
>
> static inline void accumulate_tlbflush(
>     bool *need_tlbflush, const struct page_info *page,
>     uint32_t *tlbflush_timestamp)
> {
>     if ( !*need_tlbflush )
>         *need_tlbflush = true;
> }
>
> which means a not-stubbed-out filtered_flush_tlb_mask(), with 
> tlbflush_filter()
> doing nothing, would actually invoke arch_flush_tlb_mask() (with all online 
> CPUs
> set in the mask) when called. And arch_flush_tlb_mask() isn't a no-op on Arm.

Yes.  Sadly, fixing this (without Eclair complaining in the middle of
the series) isn't as easy as I'd hoped.

> I therefore think that while moving stuff into a separate header makes sense,
> HAS_TLB_CLOCK isn't overly useful to introduce.

It takes a cpumask_t off the stack, because we can pass cpu_online_mask
straight into arch_flush_tlb_mask(), and it removes a bitmap_copy that
the compiler can't optimise out.

~Andrew