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.coop...@citrix.com>
> ---
> CC: Anthony PERARD <anthony.per...@vates.tech>
> CC: Michal Orzel <michal.or...@amd.com>
> CC: Jan Beulich <jbeul...@suse.com>
> CC: Julien Grall <jul...@xen.org>
> CC: Roger Pau Monné <roger....@citrix.com>
> CC: Stefano Stabellini <sstabell...@kernel.org>
> CC: Volodymyr Babchuk <volodymyr_babc...@epam.com>
> CC: Bertrand Marquis <bertrand.marq...@arm.com>
> CC: Oleksii Kurochko <oleksii.kuroc...@gmail.com>
> CC: Shawn Anastasio <sanasta...@raptorengineering.com>
>
> 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.
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? 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.
I therefore think that while moving stuff into a separate header makes sense,
HAS_TLB_CLOCK isn't overly useful to introduce.
Jan
