[v3,2/3] mm: Defer TLB flush by keeping both src and dst folios at migration

Message ID 20231030072540.38631-3-byungchul@sk.com
State New
Headers
Series Reduce TLB flushes under some specific conditions |

Commit Message

Byungchul Park Oct. 30, 2023, 7:25 a.m. UTC
  Implementation of CONFIG_MIGRC that stands for 'Migration Read Copy'.
We always face the migration overhead at either promotion or demotion,
while working with tiered memory e.g. CXL memory and found out TLB
shootdown is a quite big one that is needed to get rid of if possible.

Fortunately, TLB flush can be defered or even skipped if both source and
destination of folios during migration are kept until all TLB flushes
required will have been done, of course, only if the target PTE entries
have read only permission, more precisely speaking, don't have write
permission. Otherwise, no doubt the folio might get messed up.

To achieve that:

   1. For the folios that map only to non-writable TLB entries, prevent
      TLB flush at migration by keeping both source and destination
      folios, which will be handled later at a better time.

   2. When any non-writable TLB entry changes to writable e.g. through
      fault handler, give up CONFIG_MIGRC mechanism so as to perform
      TLB flush required right away.

   3. Temporarily stop migrc from working when the system is in very
      high memory pressure e.g. direct reclaim needed.

The measurement result:

   Architecture - x86_64
   QEMU - kvm enabled, host cpu
   Numa - 2 nodes (16 CPUs 1GB, no CPUs 8GB)
   Linux Kernel - v6.6-rc5, numa balancing tiering on, demotion enabled
   Benchmark - XSBench -p 50000000 (-p option makes the runtime longer)

   run 'perf stat' using events:
      1) itlb.itlb_flush
      2) tlb_flush.dtlb_thread
      3) tlb_flush.stlb_any
      4) dTLB-load-misses
      5) dTLB-store-misses
      6) iTLB-load-misses

   run 'cat /proc/vmstat' and pick:
      1) numa_pages_migrated
      2) pgmigrate_success
      3) nr_tlb_remote_flush
      4) nr_tlb_remote_flush_received
      5) nr_tlb_local_flush_all
      6) nr_tlb_local_flush_one

   BEFORE - mainline v6.6-rc5
   ------------------------------------------
   $ perf stat -a \
	   -e itlb.itlb_flush \
	   -e tlb_flush.dtlb_thread \
	   -e tlb_flush.stlb_any \
	   -e dTLB-load-misses \
	   -e dTLB-store-misses \
	   -e iTLB-load-misses \
	   ./XSBench -p 50000000

   Performance counter stats for 'system wide':

      20953405     itlb.itlb_flush
      114886593    tlb_flush.dtlb_thread
      88267015     tlb_flush.stlb_any
      115304095543 dTLB-load-misses
      163904743    dTLB-store-misses
      608486259	   iTLB-load-misses

   556.787113849 seconds time elapsed

   $ cat /proc/vmstat

   ...
   numa_pages_migrated 3378748
   pgmigrate_success 7720310
   nr_tlb_remote_flush 751464
   nr_tlb_remote_flush_received 10742115
   nr_tlb_local_flush_all 21899
   nr_tlb_local_flush_one 740157
   ...

   AFTER - mainline v6.6-rc5 + CONFIG_MIGRC
   ------------------------------------------
   $ perf stat -a \
	   -e itlb.itlb_flush \
	   -e tlb_flush.dtlb_thread \
	   -e tlb_flush.stlb_any \
	   -e dTLB-load-misses \
	   -e dTLB-store-misses \
	   -e iTLB-load-misses \
	   ./XSBench -p 50000000

   Performance counter stats for 'system wide':

      4353555      itlb.itlb_flush
      72482780     tlb_flush.dtlb_thread
      68226458     tlb_flush.stlb_any
      114331610808 dTLB-load-misses
      116084771    dTLB-store-misses
      377180518    iTLB-load-misses

   552.667718220 seconds time elapsed

   $ cat /proc/vmstat

   ...
   numa_pages_migrated 3339325
   pgmigrate_success 7642363
   nr_tlb_remote_flush 192913
   nr_tlb_remote_flush_received 2327426
   nr_tlb_local_flush_all 25759
   nr_tlb_local_flush_one 740454
   ...

Signed-off-by: Byungchul Park <byungchul@sk.com>
---
 arch/x86/include/asm/tlbflush.h |   7 +
 arch/x86/mm/tlb.c               |  91 +++++++++++
 include/linux/mm.h              |  25 +++
 include/linux/mm_types.h        |  49 ++++++
 include/linux/mmzone.h          |   7 +
 include/linux/page-flags.h      |   7 +
 include/linux/sched.h           |   8 +
 include/trace/events/mmflags.h  |   9 +-
 init/Kconfig                    |  14 ++
 mm/memory.c                     |  13 ++
 mm/migrate.c                    | 268 ++++++++++++++++++++++++++++++++
 mm/page_alloc.c                 |  29 +++-
 mm/rmap.c                       |  76 +++++++++
 13 files changed, 601 insertions(+), 2 deletions(-)
  

Comments

David Hildenbrand Oct. 30, 2023, 8 a.m. UTC | #1
On 30.10.23 08:25, Byungchul Park wrote:
> Implementation of CONFIG_MIGRC that stands for 'Migration Read Copy'.
> We always face the migration overhead at either promotion or demotion,
> while working with tiered memory e.g. CXL memory and found out TLB
> shootdown is a quite big one that is needed to get rid of if possible.
> 
> Fortunately, TLB flush can be defered or even skipped if both source and
> destination of folios during migration are kept until all TLB flushes
> required will have been done, of course, only if the target PTE entries
> have read only permission, more precisely speaking, don't have write
> permission. Otherwise, no doubt the folio might get messed up.
> 
> To achieve that:
> 
>     1. For the folios that map only to non-writable TLB entries, prevent
>        TLB flush at migration by keeping both source and destination
>        folios, which will be handled later at a better time.
> 
>     2. When any non-writable TLB entry changes to writable e.g. through
>        fault handler, give up CONFIG_MIGRC mechanism so as to perform
>        TLB flush required right away.
> 
>     3. Temporarily stop migrc from working when the system is in very
>        high memory pressure e.g. direct reclaim needed.
> 
> The measurement result:
> 
>     Architecture - x86_64
>     QEMU - kvm enabled, host cpu
>     Numa - 2 nodes (16 CPUs 1GB, no CPUs 8GB)
>     Linux Kernel - v6.6-rc5, numa balancing tiering on, demotion enabled
>     Benchmark - XSBench -p 50000000 (-p option makes the runtime longer)
> 
>     run 'perf stat' using events:
>        1) itlb.itlb_flush
>        2) tlb_flush.dtlb_thread
>        3) tlb_flush.stlb_any
>        4) dTLB-load-misses
>        5) dTLB-store-misses
>        6) iTLB-load-misses
> 
>     run 'cat /proc/vmstat' and pick:
>        1) numa_pages_migrated
>        2) pgmigrate_success
>        3) nr_tlb_remote_flush
>        4) nr_tlb_remote_flush_received
>        5) nr_tlb_local_flush_all
>        6) nr_tlb_local_flush_one
> 
>     BEFORE - mainline v6.6-rc5
>     ------------------------------------------
>     $ perf stat -a \
> 	   -e itlb.itlb_flush \
> 	   -e tlb_flush.dtlb_thread \
> 	   -e tlb_flush.stlb_any \
> 	   -e dTLB-load-misses \
> 	   -e dTLB-store-misses \
> 	   -e iTLB-load-misses \
> 	   ./XSBench -p 50000000
> 
>     Performance counter stats for 'system wide':
> 
>        20953405     itlb.itlb_flush
>        114886593    tlb_flush.dtlb_thread
>        88267015     tlb_flush.stlb_any
>        115304095543 dTLB-load-misses
>        163904743    dTLB-store-misses
>        608486259	   iTLB-load-misses
> 
>     556.787113849 seconds time elapsed
> 
>     $ cat /proc/vmstat
> 
>     ...
>     numa_pages_migrated 3378748
>     pgmigrate_success 7720310
>     nr_tlb_remote_flush 751464
>     nr_tlb_remote_flush_received 10742115
>     nr_tlb_local_flush_all 21899
>     nr_tlb_local_flush_one 740157
>     ...
> 
>     AFTER - mainline v6.6-rc5 + CONFIG_MIGRC
>     ------------------------------------------
>     $ perf stat -a \
> 	   -e itlb.itlb_flush \
> 	   -e tlb_flush.dtlb_thread \
> 	   -e tlb_flush.stlb_any \
> 	   -e dTLB-load-misses \
> 	   -e dTLB-store-misses \
> 	   -e iTLB-load-misses \
> 	   ./XSBench -p 50000000
> 
>     Performance counter stats for 'system wide':
> 
>        4353555      itlb.itlb_flush
>        72482780     tlb_flush.dtlb_thread
>        68226458     tlb_flush.stlb_any
>        114331610808 dTLB-load-misses
>        116084771    dTLB-store-misses
>        377180518    iTLB-load-misses
> 
>     552.667718220 seconds time elapsed
> 
>     $ cat /proc/vmstat
> 

So, an improvement of 0.74% ? How stable are the results? Serious 
question: worth the churn?

Or did I get the numbers wrong?

>   #define node_present_pages(nid)	(NODE_DATA(nid)->node_present_pages)
> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> index 5c02720c53a5..1ca2ac91aa14 100644
> --- a/include/linux/page-flags.h
> +++ b/include/linux/page-flags.h
> @@ -135,6 +135,9 @@ enum pageflags {
>   #ifdef CONFIG_ARCH_USES_PG_ARCH_X
>   	PG_arch_2,
>   	PG_arch_3,
> +#endif
> +#ifdef CONFIG_MIGRC
> +	PG_migrc,		/* Page has its copy under migrc's control */
>   #endif
>   	__NR_PAGEFLAGS,
>   
> @@ -589,6 +592,10 @@ TESTCLEARFLAG(Young, young, PF_ANY)
>   PAGEFLAG(Idle, idle, PF_ANY)
>   #endif
>   
> +#ifdef CONFIG_MIGRC
> +PAGEFLAG(Migrc, migrc, PF_ANY)
> +#endif

I assume you know this: new pageflags are frowned upon.
  
Nadav Amit Oct. 30, 2023, 8:50 a.m. UTC | #2
Hi Byungchul,

Few more comments. I think I lost concentration towards the end.

This patch should be broken into multiple patches, I think, and have more
comments.

I didn’t thoroughly review the entire “stopping” mechanism and other parts
of the code.


> On Oct 30, 2023, at 9:25 AM, Byungchul Park <byungchul@sk.com> wrote:
> 
> 
> 
> +#ifdef CONFIG_MIGRC

Do you really need a new config option?

I think you rely on the UBC stuff and that should be the only config
option you rely on.

> +extern void arch_migrc_adj(struct arch_tlbflush_unmap_batch *batch, int gen);
> +#else
> +static inline void arch_migrc_adj(struct arch_tlbflush_unmap_batch *batch, int gen) {}
> +#endif
> +
> static inline bool pte_flags_need_flush(unsigned long oldflags,
> 					unsigned long newflags,
> 					bool ignore_access)
> diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c
> index 314cd9912a88..0bdb617ffe94 100644
> --- a/arch/x86/mm/tlb.c
> +++ b/arch/x86/mm/tlb.c
> @@ -1219,9 +1219,80 @@ STATIC_NOPV void native_flush_tlb_local(void)
> 	native_write_cr3(__native_read_cr3());
> }
> 
> +#ifdef CONFIG_MIGRC
> +DEFINE_PER_CPU(int, migrc_done);

Any concern about overflow?

Following what I read later in the patch, you consider migrc_done as shared.
If that’s the case you should have used DECLARE_PER_CPU_SHARED_ALIGNED.

But I am not sure it should be shared, because that’s is used for an
optimization. See my comment below regarding this matter.

> +
> +static inline bool before(int a, int b)
> +{
> +	return a - b < 0;
> +}

I think you should be able to manage without this function. 

> +
> +static inline int migrc_tlb_local_begin(void)

Usually, it is better to leave it to the compiler to make inline decisions.
IOW, if there is no really good reason, I wouldn't use “inline" in “.c” files.

> +{
> +	int ret = atomic_read(&migrc_gen);
> +
> +	/*
> +	 * The following order should be guaranteed:
> +	 *
> +	 *    1. A migrc's request is queued at migration.
> +	 *    2. Each CPU performs __flush_tlb_local().
> +	 *    3. Marks this CPU has performed the TLB flush after 1.

(3) is not clear to me, specifically what “after 1” means.

> +	 *
> +	 * To achieve that, a kind of timestamp, migrc_gen was
> +	 * introduced to track the the order but conservatively, that
> +	 * is, it doesn't care whether TLB flush is performed more than
> +	 * it's needed but should guarantee TLB flush has been performed
> +	 * after the timestamp in migrc_done.
> +	 *
> +	 * 1 is done by migrc_unmap_done() and migrc_req_end().
> +	 * 2 is done by __flush_tlb_local().
> +	 * 3 is done by migrc_tlb_local_begin() and migrc_tlb_local_end().
> +	 *
> +	 * FYI, an example that should avoid is:

FYI?

> +	 *
> +	 *    1. Performs __flush_tlb_local().
> +	 *    2. A migrc's request is queued at migration.
> +	 *    3. Marks this CPU has performed the TLB flush after 2(?).

I think this example is not needed. But the comment can be improved.

> +	 *
> +	 * XXX: barrier() would be sufficient if the architecture
> +	 * already quarantees the order between memory access to
> +	 * migrc_gen and TLB flush.
> +	 */

Since UBC is used by two architectures IIRC, then is it needed or not?
Having an unnecessary smp_mb() is not great.

On x86 the memory write and the TLB flush are ordered. I think you also
regard the remote TLB flush and those should also be ordered (IIUC).

> +	smp_mb();
> +
> +	return ret;
> +}
> +
> +static inline void migrc_tlb_local_end(int gen)
> +{
> +	/*
> +	 * XXX: barrier() would be sufficient if the architecture
> +	 * quarantees the order between TLB flush and memory access to
> +	 * migrc_done.
> +	 */
> +	smp_mb();
> +
> +	if (before(READ_ONCE(*this_cpu_ptr(&migrc_done)), gen))

I have no idea why READ_ONCE() is needed here, and the fact that you think
it is needed is a bit concerning.

> +		WRITE_ONCE(*this_cpu_ptr(&migrc_done), gen);
> +}
> +#else
> +static inline int migrc_tlb_local_begin(void)
> +{
> +	return 0;
> +}
> +
> +static inline void migrc_tlb_local_end(int gen)
> +{
> +}
> +#endif
> +
> void flush_tlb_local(void)
> {
> +	unsigned int gen;
> +
> +	gen = migrc_tlb_local_begin();
> 	__flush_tlb_local();
> +	migrc_tlb_local_end(gen);
> }
> 
> /*
> @@ -1246,6 +1317,21 @@ void __flush_tlb_all(void)
> }
> EXPORT_SYMBOL_GPL(__flush_tlb_all);
> 
> +#ifdef CONFIG_MIGRC
> +/*
> + * Skip CPUs that have already performed TLB flushes required, which is
> + * tracked by migrc_gen and migrc_done.
> + */
> +void arch_migrc_adj(struct arch_tlbflush_unmap_batch *batch, int gen)
> +{
> +	int cpu;
> +
> +	for_each_cpu(cpu, &batch->cpumask)
> +		if (!before(READ_ONCE(*per_cpu_ptr(&migrc_done, cpu)), gen))
> +			cpumask_clear_cpu(cpu, &batch->cpumask);

This is an optimization, and I think it should have gone into a different patch
with separate performance benefit evaluation. Accessing remote CPUs caches to
decide whether to flush the remote TLB is a performance tradeoff, so this change
might not always be a win.

> +}
> +#endif
> +
> void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
> {
> 	struct flush_tlb_info *info;
> @@ -1274,6 +1360,11 @@ void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
> 	put_cpu();
> }
> 
> +void arch_tlbbatch_clean(struct arch_tlbflush_unmap_batch *batch)

arch_tlbbatch_clear() perhaps? Or maybe even better name that describes what
it does in a higher level (i.e., when it should be used?)

> +{
> +	cpumask_clear(&batch->cpumask);
> +}
> +
> void arch_tlbbatch_fold(struct arch_tlbflush_unmap_batch *bdst,
> 			struct arch_tlbflush_unmap_batch *bsrc)
> {
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index bf5d0b1b16f4..ddefe6033b66 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -4062,4 +4062,29 @@ static inline void accept_memory(phys_addr_t start, phys_addr_t end)
> 
> #endif
> 
> +#ifdef CONFIG_MIGRC
> +void migrc_init_page(struct page *p);
> +void migrc_unpend_folio(struct folio *f);
> +bool migrc_pending_folio(struct folio *f);
> +void migrc_shrink(struct llist_head *h);
> +bool migrc_try_flush_free_folios(nodemask_t *nodes);
> +void fold_ubc_nowr_to_migrc(void);
> +void free_migrc_req(struct migrc_req *req);
> +int migrc_pending_nr_in_zone(struct zone *z);
> +void migrc_stop_pending(void);
> +void migrc_resume_pending(void);

How about some comments to describe when each of those should be used?

And do they need to be in mm.h or can they be in mm/internal.h ?

> +
> +extern atomic_t migrc_gen;
> +#else
> +static inline void migrc_init_page(struct page *p) {}
> +static inline void migrc_unpend_folio(struct folio *f) {}
> +static inline bool migrc_pending_folio(struct folio *f) { return false; }
> +static inline void migrc_shrink(struct llist_head *h) {}
> +static inline bool migrc_try_flush_free_folios(nodemask_t *nodes) { return false; }
> +static inline void fold_ubc_nowr_to_migrc(void) {}
> +static inline void free_migrc_req(struct migrc_req *req) {}
> +static inline int migrc_pending_nr_in_zone(struct zone *z) { return 0; }
> +static inline void migrc_stop_pending(void) {}
> +static inline void migrc_resume_pending(void) {}
> +#endif
> #endif /* _LINUX_MM_H */
> diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> index 36c5b43999e6..55350f1d0db2 100644
> --- a/include/linux/mm_types.h
> +++ b/include/linux/mm_types.h
> @@ -101,6 +101,13 @@ struct page {
> 				/* Or, free page */
> 				struct list_head buddy_list;
> 				struct list_head pcp_list;
> +#ifdef CONFIG_MIGRC
> +				/*
> +				 * Hang onto migrc's request queues,
> +				 * waiting for TLB flushes.
> +				 */
> +				struct llist_node migrc_node;
> +#endif

Any reason for the use of llist?

> 			};
> 			/* See page-flags.h for PAGE_MAPPING_FLAGS */
> 			struct address_space *mapping;
> @@ -306,6 +313,13 @@ struct folio {
> 	/* private: */
> 				};
> 	/* public: */
> +#ifdef CONFIG_MIGRC
> +				/*
> +				 * Hang onto migrc's request queues,
> +				 * waiting for TLB flushes.
> +				 */
> +				struct llist_node migrc_node;
> +#endif
> 			};
> 			struct address_space *mapping;
> 			pgoff_t index;
> @@ -1372,4 +1386,39 @@ enum {
> 	/* See also internal only FOLL flags in mm/internal.h */
> };
> 
> +#ifdef CONFIG_MIGRC
> +struct migrc_req {
> +	/*
> +	 * folios pending for TLB flush
> +	 */
> +	struct llist_head folios;
> +
> +	/*
> +	 * the last folio in 'struct llist_head folios'
> +	 */
> +	struct llist_node *last;

Sounds like some unconventional synchronization scheme is going to come
later. Sounds like trouble...

> +
> +	/*
> +	 * for hanging to migrc's request queue, migrc_reqs
> +	 */
> +	struct llist_node llnode;
> +
> +	/*
> +	 * architecture specific batch data
> +	 */
> +	struct arch_tlbflush_unmap_batch arch;
> +
> +	/*
> +	 * timestamp when hanging to migrc's request queue, migrc_reqs
> +	 */
> +	int gen;
> +
> +	/*
> +	 * associated numa node
> +	 */
> +	int nid;
> +};
> +#else
> +struct migrc_req {};
> +#endif
> #endif /* _LINUX_MM_TYPES_H */
> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> index 4106fbc5b4b3..4e78d1c9ceb2 100644
> --- a/include/linux/mmzone.h
> +++ b/include/linux/mmzone.h
> @@ -980,6 +980,9 @@ struct zone {
> 	/* Zone statistics */
> 	atomic_long_t		vm_stat[NR_VM_ZONE_STAT_ITEMS];
> 	atomic_long_t		vm_numa_event[NR_VM_NUMA_EVENT_ITEMS];
> +#ifdef CONFIG_MIGRC
> +	atomic_t		migrc_pending_nr;
> +#endif
> } ____cacheline_internodealigned_in_smp;
> 
> enum pgdat_flags {
> @@ -1398,6 +1401,10 @@ typedef struct pglist_data {
> #ifdef CONFIG_MEMORY_FAILURE
> 	struct memory_failure_stats mf_stats;
> #endif
> +#ifdef CONFIG_MIGRC
> +	atomic_t migrc_pending_nr;
> +	struct llist_head migrc_reqs;
> +#endif
> } pg_data_t;
> 
> #define node_present_pages(nid)	(NODE_DATA(nid)->node_present_pages)
> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> index 5c02720c53a5..1ca2ac91aa14 100644
> --- a/include/linux/page-flags.h
> +++ b/include/linux/page-flags.h
> @@ -135,6 +135,9 @@ enum pageflags {
> #ifdef CONFIG_ARCH_USES_PG_ARCH_X
> 	PG_arch_2,
> 	PG_arch_3,
> +#endif
> +#ifdef CONFIG_MIGRC
> +	PG_migrc,		/* Page has its copy under migrc's control */

Well, you know that’s not great.

> #endif
> 	__NR_PAGEFLAGS,
> 
> @@ -589,6 +592,10 @@ TESTCLEARFLAG(Young, young, PF_ANY)
> PAGEFLAG(Idle, idle, PF_ANY)
> #endif
> 
> +#ifdef CONFIG_MIGRC
> +PAGEFLAG(Migrc, migrc, PF_ANY)
> +#endif
> +
> /*
>  * PageReported() is used to track reported free pages within the Buddy
>  * allocator. We can use the non-atomic version of the test and set
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index 63189c023357..b7c402850b2f 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -1325,6 +1325,14 @@ struct task_struct {
> 
> 	struct tlbflush_unmap_batch	tlb_ubc;
> 	struct tlbflush_unmap_batch	tlb_ubc_nowr;
> +#ifdef CONFIG_MIGRC
> +	/*
> +	 * Temporarily used while a migration is on processing before
> +	 * hanging to the associated numa node at the end of the
> +	 * migration.
> +	 */
> +	struct migrc_req		*mreq;

Why is this indirection needed?

> +#endif
> 
> 	/* Cache last used pipe for splice(): */
> 	struct pipe_inode_info		*splice_pipe;
> diff --git a/include/trace/events/mmflags.h b/include/trace/events/mmflags.h
> index 1478b9dd05fa..85b3b8859bc8 100644
> --- a/include/trace/events/mmflags.h
> +++ b/include/trace/events/mmflags.h
> @@ -95,6 +95,12 @@
> #define IF_HAVE_PG_ARCH_X(_name)
> #endif
> 
> +#ifdef CONFIG_MIGRC
> +#define IF_HAVE_PG_MIGRC(_name) ,{1UL << PG_##_name, __stringify(_name)}
> +#else
> +#define IF_HAVE_PG_MIGRC(_name)
> +#endif
> +
> #define DEF_PAGEFLAG_NAME(_name) { 1UL <<  PG_##_name, __stringify(_name) }
> 
> #define __def_pageflag_names						\
> @@ -125,7 +131,8 @@ IF_HAVE_PG_HWPOISON(hwpoison)						\
> IF_HAVE_PG_IDLE(idle)							\
> IF_HAVE_PG_IDLE(young)							\
> IF_HAVE_PG_ARCH_X(arch_2)						\
> -IF_HAVE_PG_ARCH_X(arch_3)
> +IF_HAVE_PG_ARCH_X(arch_3)						\
> +IF_HAVE_PG_MIGRC(migrc)
> 
> #define show_page_flags(flags)						\
> 	(flags) ? __print_flags(flags, "|",				\
> diff --git a/init/Kconfig b/init/Kconfig
> index 6d35728b94b2..4d88dab52059 100644
> --- a/init/Kconfig
> +++ b/init/Kconfig
> @@ -908,6 +908,20 @@ config NUMA_BALANCING_DEFAULT_ENABLED
> 	  If set, automatic NUMA balancing will be enabled if running on a NUMA
> 	  machine.
> 
> +config MIGRC
> +	bool "Deferring TLB flush by keeping read copies on migration"
> +	depends on ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> +	depends on NUMA_BALANCING
> +	depends on 64BIT
> +	default n
> +	help
> +	  TLB flush is necessary when PTE changes by migration. However,
> +	  TLB flush can be deferred if both copies of the src page and
> +	  the dst page are kept until TLB flush if they are non-writable.
> +	  System performance will be improved especially in case that
> +	  promotion and demotion types of migrations are heavily
> +	  happening.

I don’t think users should have any control over this feature.

> +
> menuconfig CGROUPS
> 	bool "Control Group support"
> 	select KERNFS
> diff --git a/mm/memory.c b/mm/memory.c
> index 6c264d2f969c..75dc48b6e15f 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -3359,6 +3359,19 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
> 	if (vmf->page)
> 		folio = page_folio(vmf->page);
> 
> +	/*
> +	 * This folio has its read copy to prevent inconsistency while
> +	 * deferring TLB flushes. However, the problem might arise if
> +	 * it's going to become writable.
> +	 *
> +	 * To prevent it, give up the deferring TLB flushes and perform
> +	 * TLB flush right away.
> +	 */
> +	if (folio && migrc_pending_folio(folio)) {
> +		migrc_unpend_folio(folio);
> +		migrc_try_flush_free_folios(NULL);

So many potential function calls… Probably they should have been combined
into one and at least migrc_pending_folio() should have been an inline
function in the header.

> +	}
> +

What about mprotect? I thought David has changed it so it can set writable
PTEs.

> 	/*
> 	 * Shared mapping: we are guaranteed to have VM_WRITE and
> 	 * FAULT_FLAG_WRITE set at this point.
> diff --git a/mm/migrate.c b/mm/migrate.c
> index 2053b54556ca..76278eca8417 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -57,6 +57,204 @@
> 
> #include "internal.h"
> 
> +#ifdef CONFIG_MIGRC
> +
> +atomic_t migrc_gen;

This would overflow.

> +
> +static struct migrc_req *alloc_migrc_req(void)
> +{
> +	return kmalloc(sizeof(struct migrc_req), GFP_KERNEL);
> +}
> +
> +void free_migrc_req(struct migrc_req *req)
> +{
> +	kfree(req);
> +}
> +
> +void migrc_init_page(struct page *p)
> +{
> +	ClearPageMigrc(p);
> +}

So many wrappers for short functions. A bit of unnecessary overheads.

> +
> +/*
> + * One of migrc's core functions that frees up all the folios in a
> + * migrc's pending queue.
> + *
> + * The list should be isolated before.
> + */
> +void migrc_shrink(struct llist_head *h)

I am not sure by the name of it that it does what I originally thought it does
seeing the word “shrink” (I thought it should be a memory shrinker to free memory
when too many pages got their TLB flush deferred).

I don’t think it should be exposed for someone to call it without actual TLB
flush.

> +{
> +	struct folio *f, *f2;
> +	struct llist_node *n;
> +
> +	n = llist_del_all(h);
> +	llist_for_each_entry_safe(f, f2, n, migrc_node) {
> +		struct pglist_data *node;
> +		struct zone *zone;
> +
> +		node = NODE_DATA(folio_nid(f));
> +		zone = page_zone(&f->page);
> +		atomic_dec(&node->migrc_pending_nr);
> +		atomic_dec(&zone->migrc_pending_nr);

Why do you need both?

> +		folio_put(f);
> +	}
> +}
> +
> +void migrc_unpend_folio(struct folio *f)
> +{
> +	folio_clear_migrc(f);
> +}
> +
> +bool migrc_pending_folio(struct folio *f)
> +{
> +	return folio_test_migrc(f);
> +}
> +
> +/*
> + * Marks the folios as being under migrc's control.
> + */
> +static void migrc_mark_folios(struct folio *fsrc, struct folio *fdst)
> +{
> +	if (!current->mreq)
> +		return;
> +
> +	folio_get(fsrc);
> +	folio_set_migrc(fsrc);
> +	folio_set_migrc(fdst);
> +	fold_ubc_nowr_to_migrc();
> +}
> +
> +static bool migrc_marked_folios(struct folio *fsrc, struct folio *fdst)
> +{
> +	/*
> +	 * XXX: The condition below is so tricky. Please suggest a
> +	 * better idea to determine if it's under migrc's control.
> +	 */
> +	return migrc_pending_folio(fsrc) && migrc_pending_folio(fdst);

Tricky == racy?

I do not follow the ordering that ensures a migrated page would not
become writable before the migration is over. Perhaps a comment on what
guarantees the order would help.

> +}
> +
> +static void migrc_undo_folios(struct folio *fsrc, struct folio *fdst)
> +{
> +	/*
> +	 * TLB flushes needed are already done at this moment so the
> +	 * flag doesn't have to be kept.
> +	 */
> +	folio_clear_migrc(fsrc);
> +	folio_clear_migrc(fdst);
> +
> +	folio_put(fsrc);
> +}
> +
> +static void migrc_expand_req(struct folio *f)
> +{
> +	struct migrc_req *req = current->mreq;
> +	struct pglist_data *node;
> +	struct zone *zone;
> +
> +	if (!req)
> +		return;
> +
> +	/*
> +	 * Focusing on numa migration, all the nids in a req are same.
> +	 */
> +	if (req->nid == -1)
> +		req->nid = folio_nid(f);
> +	else if (WARN_ON(req->nid != folio_nid(f)))
> +		return;
> +
> +	if (llist_add(&f->migrc_node, &req->folios))
> +		req->last = &f->migrc_node;
> +
> +	node = NODE_DATA(folio_nid(f));
> +	zone = page_zone(&f->page);
> +	atomic_inc(&node->migrc_pending_nr);
> +	atomic_inc(&zone->migrc_pending_nr);
> +}
> +
> +static void migrc_unmap_done(void)
> +{
> +	if (current->mreq)
> +		current->mreq->gen = atomic_inc_return(&migrc_gen);
> +}
> +
> +/*
> + * Prepares for gathering folios pending for TLB flushes, try to
> + * allocate objects needed, initialize them and make them ready.
> + */
> +static void migrc_req_start(void)
> +{
> +	struct migrc_req *req;
> +
> +	if (WARN_ON(current->mreq))
> +		return;
> +
> +	req = alloc_migrc_req();
> +	if (!req)
> +		return;
> +
> +	arch_tlbbatch_clean(&req->arch);
> +	init_llist_head(&req->folios);
> +	req->last = NULL;
> +	req->nid = -1;
> +	current->mreq = req;
> +}
> +
> +/*
> + * Hang the request with the collected folios to 'migrc_reqs' llist of
> + * the corresponding node.
> + */
> +static void migrc_req_end_if_started(void)
> +{
> +	struct migrc_req *req = current->mreq;
> +
> +	if (!req)
> +		return;
> +
> +	if (llist_empty(&req->folios))
> +		free_migrc_req(req);
> +	else
> +		llist_add(&req->llnode, &NODE_DATA(req->nid)->migrc_reqs);
> +
> +	current->mreq = NULL;
> +}
> +
> +int migrc_pending_nr_in_zone(struct zone *z)
> +{
> +	return atomic_read(&z->migrc_pending_nr);
> +}
> +
> +static atomic_t migrc_stop = ATOMIC_INIT(0);

Comments?

> +
> +void migrc_stop_pending(void)
> +{
> +	atomic_inc(&migrc_stop);
> +}
> +
> +void migrc_resume_pending(void)
> +{
> +	atomic_dec(&migrc_stop);
> +}
> +
> +static inline bool migrc_stopped_pending(void)
> +{
> +	return !!atomic_read(&migrc_stop);
> +}
> +
> +static bool can_migrc_system(void)
> +{
> +	return !migrc_stopped_pending();
> +}
> +#else
> +static inline void migrc_mark_folios(struct folio *fsrc, struct folio *fdst) {}
> +static inline bool migrc_marked_folios(struct folio *fsrc, struct folio *fdst) { return false; }
> +static inline void migrc_undo_folios(struct folio *fsrc, struct folio *fdst) {}
> +static inline void migrc_expand_req(struct folio *f) {}
> +static inline void migrc_unmap_done(void) {}
> +static inline void migrc_req_start(void) {}
> +static inline void migrc_req_end_if_started(void) {}
> +static inline bool can_migrc_system(void) { return false; }
> +#endif
> +
> bool isolate_movable_page(struct page *page, isolate_mode_t mode)
> {
> 	struct folio *folio = folio_get_nontail_page(page);
> @@ -379,6 +577,7 @@ static int folio_expected_refs(struct address_space *mapping,
> 		struct folio *folio)
> {
> 	int refs = 1;
> +
> 	if (!mapping)
> 		return refs;
> 
> @@ -406,6 +605,13 @@ int folio_migrate_mapping(struct address_space *mapping,
> 	int expected_count = folio_expected_refs(mapping, folio) + extra_count;
> 	long nr = folio_nr_pages(folio);
> 
> +	/*
> +	 * Migrc mechanism increases the reference count to defer
> +	 * freeing up folios until a better time.
> +	 */
> +	if (migrc_marked_folios(folio, newfolio))
> +		expected_count++;
> +
> 	if (!mapping) {
> 		/* Anonymous page without mapping */
> 		if (folio_ref_count(folio) != expected_count)
> @@ -1620,16 +1826,30 @@ static int migrate_pages_batch(struct list_head *from,
> 	LIST_HEAD(unmap_folios);
> 	LIST_HEAD(dst_folios);
> 	bool nosplit = (reason == MR_NUMA_MISPLACED);
> +	bool migrc_cond1;

Try to find better names.

> 
> 	VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
> 			!list_empty(from) && !list_is_singular(from));
> 
> +	/*
> +	 * For now, apply migrc only to numa migration.
> +	 */
> +	migrc_cond1 = can_migrc_system() &&
> +		      ((reason == MR_DEMOTION && current_is_kswapd()) ||
> +		      (reason == MR_NUMA_MISPLACED));
> +
> +	if (migrc_cond1)
> +		migrc_req_start();
> 	for (pass = 0; pass < nr_pass && retry; pass++) {
> 		retry = 0;
> 		thp_retry = 0;
> 		nr_retry_pages = 0;
> 
> 		list_for_each_entry_safe(folio, folio2, from, lru) {
> +			int nr_required;
> +			bool migrc_cond2;
> +			bool migrc;
> +
> 			is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
> 			nr_pages = folio_nr_pages(folio);
> 
> @@ -1657,9 +1877,34 @@ static int migrate_pages_batch(struct list_head *from,
> 				continue;
> 			}
> 
> +			/*
> +			 * In case that the system is in high memory
> +			 * pressure, let's stop migrc from working and
> +			 * free up all the folios in the pending queues
> +			 * not to impact the system.
> +			 */
> +			if (unlikely(migrc_cond1 && !can_migrc_system())) {
> +				migrc_cond1 = false;
> +				migrc_req_end_if_started();
> +				migrc_try_flush_free_folios(NULL);
> +			}
> +
> +			nr_required = nr_flush_required();
> 			rc = migrate_folio_unmap(get_new_folio, put_new_folio,
> 					private, folio, &dst, mode, reason,
> 					ret_folios);
> +			/*
> +			 * migrc_cond2 will be true only if:
> +			 *
> +			 *    1. There's no writable mapping at all
> +			 *    2. There's read only mapping found
> +			 *    3. Not already under migrc's control
> +			 */
> +			migrc_cond2 = nr_required == nr_flush_required() &&
> +				      nr_flush_required_nowr() &&
> +				      !migrc_pending_folio(folio);
> +			migrc = migrc_cond1 && migrc_cond2;
> +
> 			/*
> 			 * The rules are:
> 			 *	Success: folio will be freed
> @@ -1720,6 +1965,9 @@ static int migrate_pages_batch(struct list_head *from,
> 			case MIGRATEPAGE_UNMAP:
> 				list_move_tail(&folio->lru, &unmap_folios);
> 				list_add_tail(&dst->lru, &dst_folios);
> +
> +				if (migrc)
> +					migrc_mark_folios(folio, dst);
> 				break;
> 			default:
> 				/*
> @@ -1733,12 +1981,24 @@ static int migrate_pages_batch(struct list_head *from,
> 				stats->nr_failed_pages += nr_pages;
> 				break;
> 			}
> +
> +			/*
> +			 * Done with the current folio. Fold the ro
> +			 * batch data gathered, to the normal batch.
> +			 */
> +			fold_ubc_nowr();
> 		}
> 	}
> 	nr_failed += retry;
> 	stats->nr_thp_failed += thp_retry;
> 	stats->nr_failed_pages += nr_retry_pages;
> move:
> +	/*
> +	 * Generate a new timestamp that will be used to filter out
> +	 * CPUs that have already performed TLB flushes needed.
> +	 */
> +	migrc_unmap_done();
> +
> 	/* Flush TLBs for all unmapped folios */
> 	try_to_unmap_flush();
> 
> @@ -1774,6 +2034,9 @@ static int migrate_pages_batch(struct list_head *from,
> 			case MIGRATEPAGE_SUCCESS:
> 				stats->nr_succeeded += nr_pages;
> 				stats->nr_thp_succeeded += is_thp;
> +
> +				if (migrc_marked_folios(folio, dst))
> +					migrc_expand_req(folio);
> 				break;
> 			default:
> 				nr_failed++;
> @@ -1791,6 +2054,8 @@ static int migrate_pages_batch(struct list_head *from,
> 
> 	rc = rc_saved ? : nr_failed;
> out:
> +	migrc_req_end_if_started();
> +
> 	/* Cleanup remaining folios */
> 	dst = list_first_entry(&dst_folios, struct folio, lru);
> 	dst2 = list_next_entry(dst, lru);
> @@ -1798,6 +2063,9 @@ static int migrate_pages_batch(struct list_head *from,
> 		int page_was_mapped = 0;
> 		struct anon_vma *anon_vma = NULL;
> 
> +		if (migrc_marked_folios(folio, dst))
> +			migrc_undo_folios(folio, dst);
> +
> 		__migrate_folio_extract(dst, &page_was_mapped, &anon_vma);
> 		migrate_folio_undo_src(folio, page_was_mapped, anon_vma,
> 				       true, ret_folios);
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 95546f376302..2fa84c3a9681 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1535,6 +1535,9 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
> 
> 	set_page_owner(page, order, gfp_flags);
> 	page_table_check_alloc(page, order);
> +
> +	for (i = 0; i != 1 << order; ++i)
> +		migrc_init_page(page + i);

Do we really need an additional atomic operation to clear the page-flag here?

> }
> 
> static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
> @@ -2839,6 +2842,8 @@ bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
> 	long min = mark;
> 	int o;
> 
> +	free_pages += migrc_pending_nr_in_zone(z);
> +
> 	/* free_pages may go negative - that's OK */
> 	free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
> 
> @@ -2933,7 +2938,7 @@ static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
> 		long usable_free;
> 		long reserved;
> 
> -		usable_free = free_pages;
> +		usable_free = free_pages + migrc_pending_nr_in_zone(z);
> 		reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
> 
> 		/* reserved may over estimate high-atomic reserves. */
> @@ -3121,6 +3126,16 @@ get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
> 				       gfp_mask)) {
> 			int ret;
> 
> +			/*
> +			 * Free the pending folios so that the remaining
> +			 * code can use them. Check zone_watermark_fast()
> +			 * again with more accurate stats before going.
> +			 */
> +			migrc_try_flush_free_folios(ac->nodemask);
> +			if (zone_watermark_fast(zone, order, mark,
> +			    ac->highest_zoneidx, alloc_flags, gfp_mask))
> +				goto try_this_zone;
> +
> 			if (has_unaccepted_memory()) {
> 				if (try_to_accept_memory(zone, order))
> 					goto try_this_zone;
> @@ -3911,6 +3926,7 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> 	unsigned int cpuset_mems_cookie;
> 	unsigned int zonelist_iter_cookie;
> 	int reserve_flags;
> +	bool migrc_stopped = false;
> 
> restart:
> 	compaction_retries = 0;
> @@ -4042,6 +4058,15 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> 	if (page)
> 		goto got_pg;
> 
> +	/*
> +	 * The system is in very high memory pressure. Stop migrc from
> +	 * expanding its pending queue and working temporarily.
> +	 */
> +	if (!migrc_stopped) {
> +		migrc_stop_pending();
> +		migrc_stopped = true;
> +	}
> +
> 	/* Caller is not willing to reclaim, we can't balance anything */
> 	if (!can_direct_reclaim)
> 		goto nopage;
> @@ -4169,6 +4194,8 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> 	warn_alloc(gfp_mask, ac->nodemask,
> 			"page allocation failure: order:%u", order);
> got_pg:
> +	if (migrc_stopped)
> +		migrc_resume_pending();
> 	return page;
> }
> 
> diff --git a/mm/rmap.c b/mm/rmap.c
> index a045f3b57c60..d0fa7bf66e70 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -606,6 +606,82 @@ struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
> 
> #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> 
> +#ifdef CONFIG_MIGRC
> +/*
> + * Gather folios and CPUs in cpumask to handle.
> + */
> +static void migrc_gather(struct llist_head *folios,
> +			 struct arch_tlbflush_unmap_batch *arch,
> +			 struct llist_head *reqs)
> +{
> +	struct llist_node *reqs_nodes;
> +	struct migrc_req *req;
> +	struct migrc_req *req2;
> +
> +	reqs_nodes = llist_del_all(reqs);
> +	if (!reqs_nodes)
> +		return;
> +
> +	/*
> +	 * TODO: Optimize the time complexity.
> +	 */

Please mark the patch as an RFC until no more TODOs.

> +	llist_for_each_entry_safe(req, req2, reqs_nodes, llnode) {
> +		struct llist_node *n;
> +
> +		arch_migrc_adj(&req->arch, req->gen);
> +		arch_tlbbatch_fold(arch, &req->arch);
> +
> +		n = llist_del_all(&req->folios);
> +		llist_add_batch(n, req->last, folios);
> +		free_migrc_req(req);
> +	}
> +}
> +
> +bool migrc_try_flush_free_folios(nodemask_t *nodes)
> +{
> +	struct arch_tlbflush_unmap_batch arch;
> +	LLIST_HEAD(folios);

Again, I don’t understand why llist is needed.

> +	bool ret = false;
> +	int nid;
> +
> +	arch_tlbbatch_clean(&arch);
> +	nodes = nodes ?: &node_possible_map;

I would just use an if

> +	for_each_node_mask(nid, *nodes)
> +		migrc_gather(&folios, &arch, &NODE_DATA(nid)->migrc_reqs);
> +
> +	ret = !llist_empty(&folios);
> +	if (ret) {
> +		arch_tlbbatch_flush(&arch);
> +		migrc_shrink(&folios);
> +	}
> +	return ret;
> +}
> +
> +/*
> + * Fold ro batch data to the current migrc's request. In case that the
> + * request is not ready, fold it to the normal batch as a fallback.
> + */
> +void fold_ubc_nowr_to_migrc(void)
> +{
> +	struct tlbflush_unmap_batch *tlb_ubc_nowr = &current->tlb_ubc_nowr;
> +	struct migrc_req *req;
> +
> +	if (!tlb_ubc_nowr->nr_flush_required)
> +		return;
> +
> +	req = current->mreq;
> +	if (!req) {
> +		/* fallback */
> +		fold_ubc_nowr();
> +		return;
> +	}
> +
> +	arch_tlbbatch_fold(&req->arch, &tlb_ubc_nowr->arch);
> +	tlb_ubc_nowr->nr_flush_required = 0;
> +	tlb_ubc_nowr->writable = false;
> +}
> +#endif
> +
> void fold_ubc_nowr(void)
> {
> 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
> -- 
> 2.17.1
  
Byungchul Park Oct. 30, 2023, 9:58 a.m. UTC | #3
On Mon, Oct 30, 2023 at 09:00:56AM +0100, David Hildenbrand wrote:
> On 30.10.23 08:25, Byungchul Park wrote:
> > Implementation of CONFIG_MIGRC that stands for 'Migration Read Copy'.
> > We always face the migration overhead at either promotion or demotion,
> > while working with tiered memory e.g. CXL memory and found out TLB
> > shootdown is a quite big one that is needed to get rid of if possible.
> > 
> > Fortunately, TLB flush can be defered or even skipped if both source and
> > destination of folios during migration are kept until all TLB flushes
> > required will have been done, of course, only if the target PTE entries
> > have read only permission, more precisely speaking, don't have write
> > permission. Otherwise, no doubt the folio might get messed up.
> > 
> > To achieve that:
> > 
> >     1. For the folios that map only to non-writable TLB entries, prevent
> >        TLB flush at migration by keeping both source and destination
> >        folios, which will be handled later at a better time.
> > 
> >     2. When any non-writable TLB entry changes to writable e.g. through
> >        fault handler, give up CONFIG_MIGRC mechanism so as to perform
> >        TLB flush required right away.
> > 
> >     3. Temporarily stop migrc from working when the system is in very
> >        high memory pressure e.g. direct reclaim needed.
> > 
> > The measurement result:
> > 
> >     Architecture - x86_64
> >     QEMU - kvm enabled, host cpu
> >     Numa - 2 nodes (16 CPUs 1GB, no CPUs 8GB)
> >     Linux Kernel - v6.6-rc5, numa balancing tiering on, demotion enabled
> >     Benchmark - XSBench -p 50000000 (-p option makes the runtime longer)
> > 
> >     run 'perf stat' using events:
> >        1) itlb.itlb_flush
> >        2) tlb_flush.dtlb_thread
> >        3) tlb_flush.stlb_any
> >        4) dTLB-load-misses
> >        5) dTLB-store-misses
> >        6) iTLB-load-misses
> > 
> >     run 'cat /proc/vmstat' and pick:
> >        1) numa_pages_migrated
> >        2) pgmigrate_success
> >        3) nr_tlb_remote_flush
> >        4) nr_tlb_remote_flush_received
> >        5) nr_tlb_local_flush_all
> >        6) nr_tlb_local_flush_one
> > 
> >     BEFORE - mainline v6.6-rc5
> >     ------------------------------------------
> >     $ perf stat -a \
> > 	   -e itlb.itlb_flush \
> > 	   -e tlb_flush.dtlb_thread \
> > 	   -e tlb_flush.stlb_any \
> > 	   -e dTLB-load-misses \
> > 	   -e dTLB-store-misses \
> > 	   -e iTLB-load-misses \
> > 	   ./XSBench -p 50000000
> > 
> >     Performance counter stats for 'system wide':
> > 
> >        20953405     itlb.itlb_flush
> >        114886593    tlb_flush.dtlb_thread
> >        88267015     tlb_flush.stlb_any
> >        115304095543 dTLB-load-misses
> >        163904743    dTLB-store-misses
> >        608486259	   iTLB-load-misses
> > 
> >     556.787113849 seconds time elapsed
> > 
> >     $ cat /proc/vmstat
> > 
> >     ...
> >     numa_pages_migrated 3378748
> >     pgmigrate_success 7720310
> >     nr_tlb_remote_flush 751464
> >     nr_tlb_remote_flush_received 10742115
> >     nr_tlb_local_flush_all 21899
> >     nr_tlb_local_flush_one 740157
> >     ...
> > 
> >     AFTER - mainline v6.6-rc5 + CONFIG_MIGRC
> >     ------------------------------------------
> >     $ perf stat -a \
> > 	   -e itlb.itlb_flush \
> > 	   -e tlb_flush.dtlb_thread \
> > 	   -e tlb_flush.stlb_any \
> > 	   -e dTLB-load-misses \
> > 	   -e dTLB-store-misses \
> > 	   -e iTLB-load-misses \
> > 	   ./XSBench -p 50000000
> > 
> >     Performance counter stats for 'system wide':
> > 
> >        4353555      itlb.itlb_flush
> >        72482780     tlb_flush.dtlb_thread
> >        68226458     tlb_flush.stlb_any
> >        114331610808 dTLB-load-misses
> >        116084771    dTLB-store-misses
> >        377180518    iTLB-load-misses
> > 
> >     552.667718220 seconds time elapsed
> > 
> >     $ cat /proc/vmstat
> > 
> 
> So, an improvement of 0.74% ? How stable are the results? Serious question:

I'm getting very stable result.

> worth the churn?

Yes, ultimately the time wise improvement should be observed. However,
I've been focusing on the numbers of TLB flushes and TLB misses because
better result in terms of total time will be followed depending on the
test condition. We can see the result if we test with a system that:

   1. has more CPUs that would induce a crazy number of IPIs.
   2. has slow memories that makes TLB miss overhead bigger.
   3. runs workloads that is harmful at TLB miss and IPI storm.
   4. runs workloads that causes heavier numa migrations.
   5. runs workloads that has a lot of read only permission mappings.
   6. and so on.

I will share the results once I manage to meet the conditions.

By the way, I should've added IPI reduction because it also has super
big delta :)

> Or did I get the numbers wrong?
> 
> >   #define node_present_pages(nid)	(NODE_DATA(nid)->node_present_pages)
> > diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> > index 5c02720c53a5..1ca2ac91aa14 100644
> > --- a/include/linux/page-flags.h
> > +++ b/include/linux/page-flags.h
> > @@ -135,6 +135,9 @@ enum pageflags {
> >   #ifdef CONFIG_ARCH_USES_PG_ARCH_X
> >   	PG_arch_2,
> >   	PG_arch_3,
> > +#endif
> > +#ifdef CONFIG_MIGRC
> > +	PG_migrc,		/* Page has its copy under migrc's control */
> >   #endif
> >   	__NR_PAGEFLAGS,
> > @@ -589,6 +592,10 @@ TESTCLEARFLAG(Young, young, PF_ANY)
> >   PAGEFLAG(Idle, idle, PF_ANY)
> >   #endif
> > +#ifdef CONFIG_MIGRC
> > +PAGEFLAG(Migrc, migrc, PF_ANY)
> > +#endif
> 
> I assume you know this: new pageflags are frowned upon.

Sorry for that. I really didn't want to add a new headache.

	Byungchul
  
Byungchul Park Oct. 30, 2023, 12:51 p.m. UTC | #4
On Mon, Oct 30, 2023 at 08:50:20AM +0000, Nadav Amit wrote:
> Hi Byungchul,
> 
> Few more comments. I think I lost concentration towards the end.
> 
> This patch should be broken into multiple patches, I think, and have more
> comments.
> 
> I didn’t thoroughly review the entire “stopping” mechanism and other parts
> of the code.

I really spent a lot of time hesitating whether splitting it or not.
However, the test result cannot be stable without those. I'm still
confused. I think the test result should be stable at each commit,
right?

> > On Oct 30, 2023, at 9:25 AM, Byungchul Park <byungchul@sk.com> wrote:
> > 
> > 
> > 
> > +#ifdef CONFIG_MIGRC
> 
> Do you really need a new config option?

I will remove the config if more guys give the same opinion. Thank you
for the opinion anyway.

> I think you rely on the UBC stuff and that should be the only config
> option you rely on.

CONFIG_64BIT too, for a new page flag.

> > +#else
> > +static inline void arch_migrc_adj(struct arch_tlbflush_unmap_batch *batch, int gen) {}
> > +#endif
> > +
> > static inline bool pte_flags_need_flush(unsigned long oldflags,
> > 					unsigned long newflags,
> > 					bool ignore_access)
> > diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c
> > index 314cd9912a88..0bdb617ffe94 100644
> > --- a/arch/x86/mm/tlb.c
> > +++ b/arch/x86/mm/tlb.c
> > @@ -1219,9 +1219,80 @@ STATIC_NOPV void native_flush_tlb_local(void)
> > 	native_write_cr3(__native_read_cr3());
> > }
> > 
> > +#ifdef CONFIG_MIGRC
> > +DEFINE_PER_CPU(int, migrc_done);
> 
> Any concern about overflow?

No problem because I use the before() funtion below to compare two
int values.

> Following what I read later in the patch, you consider migrc_done as shared.
> If that’s the case you should have used DECLARE_PER_CPU_SHARED_ALIGNED.

Each CPU updates its local migrc_done while any CPU at any time can
happen to read all the CPUs' migrc_done at once. I will change it to
DECLARE_PER_CPU_SHARED_ALIGNED.

> But I am not sure it should be shared, because that’s is used for an
> optimization. See my comment below regarding this matter.
> 
> > +
> > +static inline bool before(int a, int b)
> > +{
> > +	return a - b < 0;
> > +}
> 
> I think you should be able to manage without this function. 

How? AFAIK, generation numbers are usually handled this way. Am I
missing something?

> > +
> > +static inline int migrc_tlb_local_begin(void)
> 
> Usually, it is better to leave it to the compiler to make inline decisions.
> IOW, if there is no really good reason, I wouldn't use “inline" in “.c” files.

Hm.. Okay I will.

> > +{
> > +	int ret = atomic_read(&migrc_gen);
> > +
> > +	/*
> > +	 * The following order should be guaranteed:
> > +	 *
> > +	 *    1. A migrc's request is queued at migration.
> > +	 *    2. Each CPU performs __flush_tlb_local().
> > +	 *    3. Marks this CPU has performed the TLB flush after 1.
> 
> (3) is not clear to me, specifically what “after 1” means.

It means 'whether TLB flush on this CPU required by a migrc's request
has been performed *since* the request was queued?'.

> > +	 *
> > +	 * To achieve that, a kind of timestamp, migrc_gen was
> > +	 * introduced to track the the order but conservatively, that
> > +	 * is, it doesn't care whether TLB flush is performed more than
> > +	 * it's needed but should guarantee TLB flush has been performed
> > +	 * after the timestamp in migrc_done.
> > +	 *
> > +	 * 1 is done by migrc_unmap_done() and migrc_req_end().
> > +	 * 2 is done by __flush_tlb_local().
> > +	 * 3 is done by migrc_tlb_local_begin() and migrc_tlb_local_end().
> > +	 *
> > +	 * FYI, an example that should avoid is:
> 
> FYI?
> 
> > +	 *
> > +	 *    1. Performs __flush_tlb_local().
> > +	 *    2. A migrc's request is queued at migration.
> > +	 *    3. Marks this CPU has performed the TLB flush after 2(?).
> 
> I think this example is not needed. But the comment can be improved.

I also think this is not necessary but I added it for better
understanding. So you are asking me to improve the comment right? Lemme
try it.

> > +	 *
> > +	 * XXX: barrier() would be sufficient if the architecture
> > +	 * already quarantees the order between memory access to
> > +	 * migrc_gen and TLB flush.
> > +	 */
> 
> Since UBC is used by two architectures IIRC, then is it needed or not?
> Having an unnecessary smp_mb() is not great.

Totally right.

> On x86 the memory write and the TLB flush are ordered. I think you also

What about a memory read and TLB flush? If yes too, then I can use
barrier() instead of smp_mb(). Thanks a lot!

> regard the remote TLB flush and those should also be ordered (IIUC).
> 
> > +	smp_mb();
> > +
> > +	return ret;
> > +}
> > +
> > +static inline void migrc_tlb_local_end(int gen)
> > +{
> > +	/*
> > +	 * XXX: barrier() would be sufficient if the architecture
> > +	 * quarantees the order between TLB flush and memory access to
> > +	 * migrc_done.
> > +	 */
> > +	smp_mb();
> > +
> > +	if (before(READ_ONCE(*this_cpu_ptr(&migrc_done)), gen))
> 
> I have no idea why READ_ONCE() is needed here, and the fact that you think
> it is needed is a bit concerning.

I was worried if we can quarantee the read *actually* accesses the
memory every time when repeatedly getting here. However, yeah.. I don't
have to.. I think the compilers would work well for that. I will remove
it. Thank you.

> > +		WRITE_ONCE(*this_cpu_ptr(&migrc_done), gen);
> > +}
> > +#else
> > +static inline int migrc_tlb_local_begin(void)
> > +{
> > +	return 0;
> > +}
> > +
> > +static inline void migrc_tlb_local_end(int gen)
> > +{
> > +}
> > +#endif
> > +
> > void flush_tlb_local(void)
> > {
> > +	unsigned int gen;
> > +
> > +	gen = migrc_tlb_local_begin();
> > 	__flush_tlb_local();
> > +	migrc_tlb_local_end(gen);
> > }
> > 
> > /*
> > @@ -1246,6 +1317,21 @@ void __flush_tlb_all(void)
> > }
> > EXPORT_SYMBOL_GPL(__flush_tlb_all);
> > 
> > +#ifdef CONFIG_MIGRC
> > +/*
> > + * Skip CPUs that have already performed TLB flushes required, which is
> > + * tracked by migrc_gen and migrc_done.
> > + */
> > +void arch_migrc_adj(struct arch_tlbflush_unmap_batch *batch, int gen)
> > +{
> > +	int cpu;
> > +
> > +	for_each_cpu(cpu, &batch->cpumask)
> > +		if (!before(READ_ONCE(*per_cpu_ptr(&migrc_done, cpu)), gen))
> > +			cpumask_clear_cpu(cpu, &batch->cpumask);
> 
> This is an optimization, and I think it should have gone into a different patch
> with separate performance benefit evaluation. Accessing remote CPUs caches to
> decide whether to flush the remote TLB is a performance tradeoff, so this change
> might not always be a win.

Right. I will.

> > +}
> > +#endif
> > +
> > void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
> > {
> > 	struct flush_tlb_info *info;
> > @@ -1274,6 +1360,11 @@ void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
> > 	put_cpu();
> > }
> > 
> > +void arch_tlbbatch_clean(struct arch_tlbflush_unmap_batch *batch)
> 
> arch_tlbbatch_clear() perhaps? Or maybe even better name that describes what
> it does in a higher level (i.e., when it should be used?)

I use this at each preparation phase when I want to gather batched
information e.g. to gather which CPUs need TLB flush.

arch_tlbbatch_init()
arch_tlbbatch_clear()
arch_tlbbatch_reset()
arch_tlbbatch_?()

> > +{
> > +	cpumask_clear(&batch->cpumask);
> > +}
> > +
> > void arch_tlbbatch_fold(struct arch_tlbflush_unmap_batch *bdst,
> > 			struct arch_tlbflush_unmap_batch *bsrc)
> > {
> > diff --git a/include/linux/mm.h b/include/linux/mm.h
> > index bf5d0b1b16f4..ddefe6033b66 100644
> > --- a/include/linux/mm.h
> > +++ b/include/linux/mm.h
> > @@ -4062,4 +4062,29 @@ static inline void accept_memory(phys_addr_t start, phys_addr_t end)
> > 
> > #endif
> > 
> > +#ifdef CONFIG_MIGRC
> > +void migrc_init_page(struct page *p);
> > +void migrc_unpend_folio(struct folio *f);
> > +bool migrc_pending_folio(struct folio *f);
> > +void migrc_shrink(struct llist_head *h);
> > +bool migrc_try_flush_free_folios(nodemask_t *nodes);
> > +void fold_ubc_nowr_to_migrc(void);
> > +void free_migrc_req(struct migrc_req *req);
> > +int migrc_pending_nr_in_zone(struct zone *z);
> > +void migrc_stop_pending(void);
> > +void migrc_resume_pending(void);
> 
> How about some comments to describe when each of those should be used?

Sounds great. I will add comments.

> And do they need to be in mm.h or can they be in mm/internal.h ?

Otherwise, architecture code like x86 cannot use these functions.

> > +extern atomic_t migrc_gen;
> > +#else
> > +static inline void migrc_init_page(struct page *p) {}
> > +static inline void migrc_unpend_folio(struct folio *f) {}
> > +static inline bool migrc_pending_folio(struct folio *f) { return false; }
> > +static inline void migrc_shrink(struct llist_head *h) {}
> > +static inline bool migrc_try_flush_free_folios(nodemask_t *nodes) { return false; }
> > +static inline void fold_ubc_nowr_to_migrc(void) {}
> > +static inline void free_migrc_req(struct migrc_req *req) {}
> > +static inline int migrc_pending_nr_in_zone(struct zone *z) { return 0; }
> > +static inline void migrc_stop_pending(void) {}
> > +static inline void migrc_resume_pending(void) {}
> > +#endif
> > #endif /* _LINUX_MM_H */
> > diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> > index 36c5b43999e6..55350f1d0db2 100644
> > --- a/include/linux/mm_types.h
> > +++ b/include/linux/mm_types.h
> > @@ -101,6 +101,13 @@ struct page {
> > 				/* Or, free page */
> > 				struct list_head buddy_list;
> > 				struct list_head pcp_list;
> > +#ifdef CONFIG_MIGRC
> > +				/*
> > +				 * Hang onto migrc's request queues,
> > +				 * waiting for TLB flushes.
> > +				 */
> > +				struct llist_node migrc_node;
> > +#endif
> 
> Any reason for the use of llist?

For lockless design at the beginning.. but for now.. you are right. I
don't have to keep it as llist. I will change it so as to use lru for
that purpose instead.

I'm suprised a lot how much you understand the code. Respect.

> > 			};
> > 			/* See page-flags.h for PAGE_MAPPING_FLAGS */
> > 			struct address_space *mapping;
> > @@ -306,6 +313,13 @@ struct folio {
> > 	/* private: */
> > 				};
> > 	/* public: */
> > +#ifdef CONFIG_MIGRC
> > +				/*
> > +				 * Hang onto migrc's request queues,
> > +				 * waiting for TLB flushes.
> > +				 */
> > +				struct llist_node migrc_node;
> > +#endif
> > 			};
> > 			struct address_space *mapping;
> > 			pgoff_t index;
> > @@ -1372,4 +1386,39 @@ enum {
> > 	/* See also internal only FOLL flags in mm/internal.h */
> > };
> > 
> > +#ifdef CONFIG_MIGRC
> > +struct migrc_req {
> > +	/*
> > +	 * folios pending for TLB flush
> > +	 */
> > +	struct llist_head folios;
> > +
> > +	/*
> > +	 * the last folio in 'struct llist_head folios'
> > +	 */
> > +	struct llist_node *last;
> 
> Sounds like some unconventional synchronization scheme is going to come
> later. Sounds like trouble...

I don't think so. What would make trouble with this? Anyway, I wouldn't
need this anymore if I use list_head instead of llist.

> > +
> > +	/*
> > +	 * for hanging to migrc's request queue, migrc_reqs
> > +	 */
> > +	struct llist_node llnode;
> > +
> > +	/*
> > +	 * architecture specific batch data
> > +	 */
> > +	struct arch_tlbflush_unmap_batch arch;
> > +
> > +	/*
> > +	 * timestamp when hanging to migrc's request queue, migrc_reqs
> > +	 */
> > +	int gen;
> > +
> > +	/*
> > +	 * associated numa node
> > +	 */
> > +	int nid;
> > +};
> > +#else
> > +struct migrc_req {};
> > +#endif
> > #endif /* _LINUX_MM_TYPES_H */
> > diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> > index 4106fbc5b4b3..4e78d1c9ceb2 100644
> > --- a/include/linux/mmzone.h
> > +++ b/include/linux/mmzone.h
> > @@ -980,6 +980,9 @@ struct zone {
> > 	/* Zone statistics */
> > 	atomic_long_t		vm_stat[NR_VM_ZONE_STAT_ITEMS];
> > 	atomic_long_t		vm_numa_event[NR_VM_NUMA_EVENT_ITEMS];
> > +#ifdef CONFIG_MIGRC
> > +	atomic_t		migrc_pending_nr;
> > +#endif
> > } ____cacheline_internodealigned_in_smp;
> > 
> > enum pgdat_flags {
> > @@ -1398,6 +1401,10 @@ typedef struct pglist_data {
> > #ifdef CONFIG_MEMORY_FAILURE
> > 	struct memory_failure_stats mf_stats;
> > #endif
> > +#ifdef CONFIG_MIGRC
> > +	atomic_t migrc_pending_nr;
> > +	struct llist_head migrc_reqs;
> > +#endif
> > } pg_data_t;
> > 
> > #define node_present_pages(nid)	(NODE_DATA(nid)->node_present_pages)
> > diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> > index 5c02720c53a5..1ca2ac91aa14 100644
> > --- a/include/linux/page-flags.h
> > +++ b/include/linux/page-flags.h
> > @@ -135,6 +135,9 @@ enum pageflags {
> > #ifdef CONFIG_ARCH_USES_PG_ARCH_X
> > 	PG_arch_2,
> > 	PG_arch_3,
> > +#endif
> > +#ifdef CONFIG_MIGRC
> > +	PG_migrc,		/* Page has its copy under migrc's control */
> 
> Well, you know that’s not great.

Yeah. I really didn't want to add this. Any idea to distinguish folios
that have their copies in migrc's pending list from others?

Migrc needs to initiate TLB flush needed when facing do_wp_page() fault
handler that chages any mapping associated to those folios under migrc's
control, from wp permission to writable.

> > #endif
> > 	__NR_PAGEFLAGS,
> > 
> > @@ -589,6 +592,10 @@ TESTCLEARFLAG(Young, young, PF_ANY)
> > PAGEFLAG(Idle, idle, PF_ANY)
> > #endif
> > 
> > +#ifdef CONFIG_MIGRC
> > +PAGEFLAG(Migrc, migrc, PF_ANY)
> > +#endif
> > +
> > /*
> >  * PageReported() is used to track reported free pages within the Buddy
> >  * allocator. We can use the non-atomic version of the test and set
> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index 63189c023357..b7c402850b2f 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -1325,6 +1325,14 @@ struct task_struct {
> > 
> > 	struct tlbflush_unmap_batch	tlb_ubc;
> > 	struct tlbflush_unmap_batch	tlb_ubc_nowr;
> > +#ifdef CONFIG_MIGRC
> > +	/*
> > +	 * Temporarily used while a migration is on processing before
> > +	 * hanging to the associated numa node at the end of the
> > +	 * migration.
> > +	 */
> > +	struct migrc_req		*mreq;
> 
> Why is this indirection needed?

To gather folios in the local context without any synchronization like
tlb_ubc. Or I need to pass the pointer to a ton of functions as an
additional parameter but it looks worse. I guess that's why tlb_ubc is
used that way too.

> > +#endif
> > 
> > 	/* Cache last used pipe for splice(): */
> > 	struct pipe_inode_info		*splice_pipe;
> > diff --git a/include/trace/events/mmflags.h b/include/trace/events/mmflags.h
> > index 1478b9dd05fa..85b3b8859bc8 100644
> > --- a/include/trace/events/mmflags.h
> > +++ b/include/trace/events/mmflags.h
> > @@ -95,6 +95,12 @@
> > #define IF_HAVE_PG_ARCH_X(_name)
> > #endif
> > 
> > +#ifdef CONFIG_MIGRC
> > +#define IF_HAVE_PG_MIGRC(_name) ,{1UL << PG_##_name, __stringify(_name)}
> > +#else
> > +#define IF_HAVE_PG_MIGRC(_name)
> > +#endif
> > +
> > #define DEF_PAGEFLAG_NAME(_name) { 1UL <<  PG_##_name, __stringify(_name) }
> > 
> > #define __def_pageflag_names						\
> > @@ -125,7 +131,8 @@ IF_HAVE_PG_HWPOISON(hwpoison)						\
> > IF_HAVE_PG_IDLE(idle)							\
> > IF_HAVE_PG_IDLE(young)							\
> > IF_HAVE_PG_ARCH_X(arch_2)						\
> > -IF_HAVE_PG_ARCH_X(arch_3)
> > +IF_HAVE_PG_ARCH_X(arch_3)						\
> > +IF_HAVE_PG_MIGRC(migrc)
> > 
> > #define show_page_flags(flags)						\
> > 	(flags) ? __print_flags(flags, "|",				\
> > diff --git a/init/Kconfig b/init/Kconfig
> > index 6d35728b94b2..4d88dab52059 100644
> > --- a/init/Kconfig
> > +++ b/init/Kconfig
> > @@ -908,6 +908,20 @@ config NUMA_BALANCING_DEFAULT_ENABLED
> > 	  If set, automatic NUMA balancing will be enabled if running on a NUMA
> > 	  machine.
> > 
> > +config MIGRC
> > +	bool "Deferring TLB flush by keeping read copies on migration"
> > +	depends on ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> > +	depends on NUMA_BALANCING
> > +	depends on 64BIT
> > +	default n
> > +	help
> > +	  TLB flush is necessary when PTE changes by migration. However,
> > +	  TLB flush can be deferred if both copies of the src page and
> > +	  the dst page are kept until TLB flush if they are non-writable.
> > +	  System performance will be improved especially in case that
> > +	  promotion and demotion types of migrations are heavily
> > +	  happening.
> 
> I don’t think users should have any control over this feature.
> 
> > +
> > menuconfig CGROUPS
> > 	bool "Control Group support"
> > 	select KERNFS
> > diff --git a/mm/memory.c b/mm/memory.c
> > index 6c264d2f969c..75dc48b6e15f 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -3359,6 +3359,19 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
> > 	if (vmf->page)
> > 		folio = page_folio(vmf->page);
> > 
> > +	/*
> > +	 * This folio has its read copy to prevent inconsistency while
> > +	 * deferring TLB flushes. However, the problem might arise if
> > +	 * it's going to become writable.
> > +	 *
> > +	 * To prevent it, give up the deferring TLB flushes and perform
> > +	 * TLB flush right away.
> > +	 */
> > +	if (folio && migrc_pending_folio(folio)) {
> > +		migrc_unpend_folio(folio);
> > +		migrc_try_flush_free_folios(NULL);
> 
> So many potential function calls… Probably they should have been combined
> into one and at least migrc_pending_folio() should have been an inline
> function in the header.

I will try to change it as you mention.

> > +	}
> > +
> 
> What about mprotect? I thought David has changed it so it can set writable
> PTEs.

I will check it out.

> > 	/*
> > 	 * Shared mapping: we are guaranteed to have VM_WRITE and
> > 	 * FAULT_FLAG_WRITE set at this point.
> > diff --git a/mm/migrate.c b/mm/migrate.c
> > index 2053b54556ca..76278eca8417 100644
> > --- a/mm/migrate.c
> > +++ b/mm/migrate.c
> > @@ -57,6 +57,204 @@
> > 
> > #include "internal.h"
> > 
> > +#ifdef CONFIG_MIGRC
> > +
> > +atomic_t migrc_gen;
> 
> This would overflow.

It's safe. As I mentioned, I use before() function to compare two
generation numbers.

> > +
> > +static struct migrc_req *alloc_migrc_req(void)
> > +{
> > +	return kmalloc(sizeof(struct migrc_req), GFP_KERNEL);
> > +}
> > +
> > +void free_migrc_req(struct migrc_req *req)
> > +{
> > +	kfree(req);
> > +}
> > +
> > +void migrc_init_page(struct page *p)
> > +{
> > +	ClearPageMigrc(p);
> > +}
> 
> So many wrappers for short functions. A bit of unnecessary overheads.

Hmm.. Okay. Lemme try to get rid of unnecessary wrappers.

> > +
> > +/*
> > + * One of migrc's core functions that frees up all the folios in a
> > + * migrc's pending queue.
> > + *
> > + * The list should be isolated before.
> > + */
> > +void migrc_shrink(struct llist_head *h)
> 
> I am not sure by the name of it that it does what I originally thought it does
> seeing the word “shrink” (I thought it should be a memory shrinker to free memory
> when too many pages got their TLB flush deferred).
> 
> I don’t think it should be exposed for someone to call it without actual TLB
> flush.

I might need a better name. I will rename it.

> > +{
> > +	struct folio *f, *f2;
> > +	struct llist_node *n;
> > +
> > +	n = llist_del_all(h);
> > +	llist_for_each_entry_safe(f, f2, n, migrc_node) {
> > +		struct pglist_data *node;
> > +		struct zone *zone;
> > +
> > +		node = NODE_DATA(folio_nid(f));
> > +		zone = page_zone(&f->page);
> > +		atomic_dec(&node->migrc_pending_nr);
> > +		atomic_dec(&zone->migrc_pending_nr);
> 
> Why do you need both?

node->migrc_pending_nr is not used anymore in the current version. I
will remove it.

> > +		folio_put(f);
> > +	}
> > +}
> > +
> > +void migrc_unpend_folio(struct folio *f)
> > +{
> > +	folio_clear_migrc(f);
> > +}
> > +
> > +bool migrc_pending_folio(struct folio *f)
> > +{
> > +	return folio_test_migrc(f);
> > +}
> > +
> > +/*
> > + * Marks the folios as being under migrc's control.
> > + */
> > +static void migrc_mark_folios(struct folio *fsrc, struct folio *fdst)
> > +{
> > +	if (!current->mreq)
> > +		return;
> > +
> > +	folio_get(fsrc);
> > +	folio_set_migrc(fsrc);
> > +	folio_set_migrc(fdst);
> > +	fold_ubc_nowr_to_migrc();
> > +}
> > +
> > +static bool migrc_marked_folios(struct folio *fsrc, struct folio *fdst)
> > +{
> > +	/*
> > +	 * XXX: The condition below is so tricky. Please suggest a
> > +	 * better idea to determine if it's under migrc's control.
> > +	 */
> > +	return migrc_pending_folio(fsrc) && migrc_pending_folio(fdst);
> 
> Tricky == racy?

No. I meant the way to implement is tricky. 

> I do not follow the ordering that ensures a migrated page would not
> become writable before the migration is over. Perhaps a comment on what
> guarantees the order would help.

I will check it and add comments explaining it.

> > +}
> > +
> > +static void migrc_undo_folios(struct folio *fsrc, struct folio *fdst)
> > +{
> > +	/*
> > +	 * TLB flushes needed are already done at this moment so the
> > +	 * flag doesn't have to be kept.
> > +	 */
> > +	folio_clear_migrc(fsrc);
> > +	folio_clear_migrc(fdst);
> > +
> > +	folio_put(fsrc);
> > +}
> > +
> > +static void migrc_expand_req(struct folio *f)
> > +{
> > +	struct migrc_req *req = current->mreq;
> > +	struct pglist_data *node;
> > +	struct zone *zone;
> > +
> > +	if (!req)
> > +		return;
> > +
> > +	/*
> > +	 * Focusing on numa migration, all the nids in a req are same.
> > +	 */
> > +	if (req->nid == -1)
> > +		req->nid = folio_nid(f);
> > +	else if (WARN_ON(req->nid != folio_nid(f)))
> > +		return;
> > +
> > +	if (llist_add(&f->migrc_node, &req->folios))
> > +		req->last = &f->migrc_node;
> > +
> > +	node = NODE_DATA(folio_nid(f));
> > +	zone = page_zone(&f->page);
> > +	atomic_inc(&node->migrc_pending_nr);
> > +	atomic_inc(&zone->migrc_pending_nr);
> > +}
> > +
> > +static void migrc_unmap_done(void)
> > +{
> > +	if (current->mreq)
> > +		current->mreq->gen = atomic_inc_return(&migrc_gen);
> > +}
> > +
> > +/*
> > + * Prepares for gathering folios pending for TLB flushes, try to
> > + * allocate objects needed, initialize them and make them ready.
> > + */
> > +static void migrc_req_start(void)
> > +{
> > +	struct migrc_req *req;
> > +
> > +	if (WARN_ON(current->mreq))
> > +		return;
> > +
> > +	req = alloc_migrc_req();
> > +	if (!req)
> > +		return;
> > +
> > +	arch_tlbbatch_clean(&req->arch);
> > +	init_llist_head(&req->folios);
> > +	req->last = NULL;
> > +	req->nid = -1;
> > +	current->mreq = req;
> > +}
> > +
> > +/*
> > + * Hang the request with the collected folios to 'migrc_reqs' llist of
> > + * the corresponding node.
> > + */
> > +static void migrc_req_end_if_started(void)
> > +{
> > +	struct migrc_req *req = current->mreq;
> > +
> > +	if (!req)
> > +		return;
> > +
> > +	if (llist_empty(&req->folios))
> > +		free_migrc_req(req);
> > +	else
> > +		llist_add(&req->llnode, &NODE_DATA(req->nid)->migrc_reqs);
> > +
> > +	current->mreq = NULL;
> > +}
> > +
> > +int migrc_pending_nr_in_zone(struct zone *z)
> > +{
> > +	return atomic_read(&z->migrc_pending_nr);
> > +}
> > +
> > +static atomic_t migrc_stop = ATOMIC_INIT(0);
> 
> Comments?

I will add comments on it.

> > +
> > +void migrc_stop_pending(void)
> > +{
> > +	atomic_inc(&migrc_stop);
> > +}
> > +
> > +void migrc_resume_pending(void)
> > +{
> > +	atomic_dec(&migrc_stop);
> > +}
> > +
> > +static inline bool migrc_stopped_pending(void)
> > +{
> > +	return !!atomic_read(&migrc_stop);
> > +}
> > +
> > +static bool can_migrc_system(void)
> > +{
> > +	return !migrc_stopped_pending();
> > +}
> > +#else
> > +static inline void migrc_mark_folios(struct folio *fsrc, struct folio *fdst) {}
> > +static inline bool migrc_marked_folios(struct folio *fsrc, struct folio *fdst) { return false; }
> > +static inline void migrc_undo_folios(struct folio *fsrc, struct folio *fdst) {}
> > +static inline void migrc_expand_req(struct folio *f) {}
> > +static inline void migrc_unmap_done(void) {}
> > +static inline void migrc_req_start(void) {}
> > +static inline void migrc_req_end_if_started(void) {}
> > +static inline bool can_migrc_system(void) { return false; }
> > +#endif
> > +
> > bool isolate_movable_page(struct page *page, isolate_mode_t mode)
> > {
> > 	struct folio *folio = folio_get_nontail_page(page);
> > @@ -379,6 +577,7 @@ static int folio_expected_refs(struct address_space *mapping,
> > 		struct folio *folio)
> > {
> > 	int refs = 1;
> > +
> > 	if (!mapping)
> > 		return refs;
> > 
> > @@ -406,6 +605,13 @@ int folio_migrate_mapping(struct address_space *mapping,
> > 	int expected_count = folio_expected_refs(mapping, folio) + extra_count;
> > 	long nr = folio_nr_pages(folio);
> > 
> > +	/*
> > +	 * Migrc mechanism increases the reference count to defer
> > +	 * freeing up folios until a better time.
> > +	 */
> > +	if (migrc_marked_folios(folio, newfolio))
> > +		expected_count++;
> > +
> > 	if (!mapping) {
> > 		/* Anonymous page without mapping */
> > 		if (folio_ref_count(folio) != expected_count)
> > @@ -1620,16 +1826,30 @@ static int migrate_pages_batch(struct list_head *from,
> > 	LIST_HEAD(unmap_folios);
> > 	LIST_HEAD(dst_folios);
> > 	bool nosplit = (reason == MR_NUMA_MISPLACED);
> > +	bool migrc_cond1;
> 
> Try to find better names.

Okay.

> > 	VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
> > 			!list_empty(from) && !list_is_singular(from));
> > 
> > +	/*
> > +	 * For now, apply migrc only to numa migration.
> > +	 */
> > +	migrc_cond1 = can_migrc_system() &&
> > +		      ((reason == MR_DEMOTION && current_is_kswapd()) ||
> > +		      (reason == MR_NUMA_MISPLACED));
> > +
> > +	if (migrc_cond1)
> > +		migrc_req_start();
> > 	for (pass = 0; pass < nr_pass && retry; pass++) {
> > 		retry = 0;
> > 		thp_retry = 0;
> > 		nr_retry_pages = 0;
> > 
> > 		list_for_each_entry_safe(folio, folio2, from, lru) {
> > +			int nr_required;
> > +			bool migrc_cond2;
> > +			bool migrc;
> > +
> > 			is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
> > 			nr_pages = folio_nr_pages(folio);
> > 
> > @@ -1657,9 +1877,34 @@ static int migrate_pages_batch(struct list_head *from,
> > 				continue;
> > 			}
> > 
> > +			/*
> > +			 * In case that the system is in high memory
> > +			 * pressure, let's stop migrc from working and
> > +			 * free up all the folios in the pending queues
> > +			 * not to impact the system.
> > +			 */
> > +			if (unlikely(migrc_cond1 && !can_migrc_system())) {
> > +				migrc_cond1 = false;
> > +				migrc_req_end_if_started();
> > +				migrc_try_flush_free_folios(NULL);
> > +			}
> > +
> > +			nr_required = nr_flush_required();
> > 			rc = migrate_folio_unmap(get_new_folio, put_new_folio,
> > 					private, folio, &dst, mode, reason,
> > 					ret_folios);
> > +			/*
> > +			 * migrc_cond2 will be true only if:
> > +			 *
> > +			 *    1. There's no writable mapping at all
> > +			 *    2. There's read only mapping found
> > +			 *    3. Not already under migrc's control
> > +			 */
> > +			migrc_cond2 = nr_required == nr_flush_required() &&
> > +				      nr_flush_required_nowr() &&
> > +				      !migrc_pending_folio(folio);
> > +			migrc = migrc_cond1 && migrc_cond2;
> > +
> > 			/*
> > 			 * The rules are:
> > 			 *	Success: folio will be freed
> > @@ -1720,6 +1965,9 @@ static int migrate_pages_batch(struct list_head *from,
> > 			case MIGRATEPAGE_UNMAP:
> > 				list_move_tail(&folio->lru, &unmap_folios);
> > 				list_add_tail(&dst->lru, &dst_folios);
> > +
> > +				if (migrc)
> > +					migrc_mark_folios(folio, dst);
> > 				break;
> > 			default:
> > 				/*
> > @@ -1733,12 +1981,24 @@ static int migrate_pages_batch(struct list_head *from,
> > 				stats->nr_failed_pages += nr_pages;
> > 				break;
> > 			}
> > +
> > +			/*
> > +			 * Done with the current folio. Fold the ro
> > +			 * batch data gathered, to the normal batch.
> > +			 */
> > +			fold_ubc_nowr();
> > 		}
> > 	}
> > 	nr_failed += retry;
> > 	stats->nr_thp_failed += thp_retry;
> > 	stats->nr_failed_pages += nr_retry_pages;
> > move:
> > +	/*
> > +	 * Generate a new timestamp that will be used to filter out
> > +	 * CPUs that have already performed TLB flushes needed.
> > +	 */
> > +	migrc_unmap_done();
> > +
> > 	/* Flush TLBs for all unmapped folios */
> > 	try_to_unmap_flush();
> > 
> > @@ -1774,6 +2034,9 @@ static int migrate_pages_batch(struct list_head *from,
> > 			case MIGRATEPAGE_SUCCESS:
> > 				stats->nr_succeeded += nr_pages;
> > 				stats->nr_thp_succeeded += is_thp;
> > +
> > +				if (migrc_marked_folios(folio, dst))
> > +					migrc_expand_req(folio);
> > 				break;
> > 			default:
> > 				nr_failed++;
> > @@ -1791,6 +2054,8 @@ static int migrate_pages_batch(struct list_head *from,
> > 
> > 	rc = rc_saved ? : nr_failed;
> > out:
> > +	migrc_req_end_if_started();
> > +
> > 	/* Cleanup remaining folios */
> > 	dst = list_first_entry(&dst_folios, struct folio, lru);
> > 	dst2 = list_next_entry(dst, lru);
> > @@ -1798,6 +2063,9 @@ static int migrate_pages_batch(struct list_head *from,
> > 		int page_was_mapped = 0;
> > 		struct anon_vma *anon_vma = NULL;
> > 
> > +		if (migrc_marked_folios(folio, dst))
> > +			migrc_undo_folios(folio, dst);
> > +
> > 		__migrate_folio_extract(dst, &page_was_mapped, &anon_vma);
> > 		migrate_folio_undo_src(folio, page_was_mapped, anon_vma,
> > 				       true, ret_folios);
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 95546f376302..2fa84c3a9681 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -1535,6 +1535,9 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
> > 
> > 	set_page_owner(page, order, gfp_flags);
> > 	page_table_check_alloc(page, order);
> > +
> > +	for (i = 0; i != 1 << order; ++i)
> > +		migrc_init_page(page + i);
> 
> Do we really need an additional atomic operation to clear the page-flag here?

Of course not. I will change it.

> > }
> > 
> > static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
> > @@ -2839,6 +2842,8 @@ bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
> > 	long min = mark;
> > 	int o;
> > 
> > +	free_pages += migrc_pending_nr_in_zone(z);
> > +
> > 	/* free_pages may go negative - that's OK */
> > 	free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
> > 
> > @@ -2933,7 +2938,7 @@ static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
> > 		long usable_free;
> > 		long reserved;
> > 
> > -		usable_free = free_pages;
> > +		usable_free = free_pages + migrc_pending_nr_in_zone(z);
> > 		reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
> > 
> > 		/* reserved may over estimate high-atomic reserves. */
> > @@ -3121,6 +3126,16 @@ get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
> > 				       gfp_mask)) {
> > 			int ret;
> > 
> > +			/*
> > +			 * Free the pending folios so that the remaining
> > +			 * code can use them. Check zone_watermark_fast()
> > +			 * again with more accurate stats before going.
> > +			 */
> > +			migrc_try_flush_free_folios(ac->nodemask);
> > +			if (zone_watermark_fast(zone, order, mark,
> > +			    ac->highest_zoneidx, alloc_flags, gfp_mask))
> > +				goto try_this_zone;
> > +
> > 			if (has_unaccepted_memory()) {
> > 				if (try_to_accept_memory(zone, order))
> > 					goto try_this_zone;
> > @@ -3911,6 +3926,7 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> > 	unsigned int cpuset_mems_cookie;
> > 	unsigned int zonelist_iter_cookie;
> > 	int reserve_flags;
> > +	bool migrc_stopped = false;
> > 
> > restart:
> > 	compaction_retries = 0;
> > @@ -4042,6 +4058,15 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> > 	if (page)
> > 		goto got_pg;
> > 
> > +	/*
> > +	 * The system is in very high memory pressure. Stop migrc from
> > +	 * expanding its pending queue and working temporarily.
> > +	 */
> > +	if (!migrc_stopped) {
> > +		migrc_stop_pending();
> > +		migrc_stopped = true;
> > +	}
> > +
> > 	/* Caller is not willing to reclaim, we can't balance anything */
> > 	if (!can_direct_reclaim)
> > 		goto nopage;
> > @@ -4169,6 +4194,8 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> > 	warn_alloc(gfp_mask, ac->nodemask,
> > 			"page allocation failure: order:%u", order);
> > got_pg:
> > +	if (migrc_stopped)
> > +		migrc_resume_pending();
> > 	return page;
> > }
> > 
> > diff --git a/mm/rmap.c b/mm/rmap.c
> > index a045f3b57c60..d0fa7bf66e70 100644
> > --- a/mm/rmap.c
> > +++ b/mm/rmap.c
> > @@ -606,6 +606,82 @@ struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
> > 
> > #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> > 
> > +#ifdef CONFIG_MIGRC
> > +/*
> > + * Gather folios and CPUs in cpumask to handle.
> > + */
> > +static void migrc_gather(struct llist_head *folios,
> > +			 struct arch_tlbflush_unmap_batch *arch,
> > +			 struct llist_head *reqs)
> > +{
> > +	struct llist_node *reqs_nodes;
> > +	struct migrc_req *req;
> > +	struct migrc_req *req2;
> > +
> > +	reqs_nodes = llist_del_all(reqs);
> > +	if (!reqs_nodes)
> > +		return;
> > +
> > +	/*
> > +	 * TODO: Optimize the time complexity.
> > +	 */
> 
> Please mark the patch as an RFC until no more TODOs.

I removed RFC because those are longterm TODOs that are not essential.

> > +	llist_for_each_entry_safe(req, req2, reqs_nodes, llnode) {
> > +		struct llist_node *n;
> > +
> > +		arch_migrc_adj(&req->arch, req->gen);
> > +		arch_tlbbatch_fold(arch, &req->arch);
> > +
> > +		n = llist_del_all(&req->folios);
> > +		llist_add_batch(n, req->last, folios);
> > +		free_migrc_req(req);
> > +	}
> > +}
> > +
> > +bool migrc_try_flush_free_folios(nodemask_t *nodes)
> > +{
> > +	struct arch_tlbflush_unmap_batch arch;
> > +	LLIST_HEAD(folios);
> 
> Again, I don’t understand why llist is needed.

Yep. I will change it to list_head. Thanks!

Appreciate all your great comments.

	Byungchul
  
Nadav Amit Oct. 30, 2023, 3:58 p.m. UTC | #5
> On Oct 30, 2023, at 2:51 PM, Byungchul Park <byungchul@sk.com> wrote:
> 
> I really spent a lot of time hesitating whether splitting it or not.
> However, the test result cannot be stable without those. I'm still
> confused. I think the test result should be stable at each commit,
> right?

Of course. You can extract the optimization we mentioned, and perhaps
have more preparatory patches.

Just a couple of comments that may also help breaking the patches:

1. The “stopping” logic is a bit not great. Try to see if you can
somehow use shrinker or OOM infrastructure instead.

2. Regarding “overflows”, it’s not always a question of whether an
overflow would happen naturally, but whether a malicious process can
trigger it.

Regards,
Nadav
  
Byungchul Park Oct. 30, 2023, 10:40 p.m. UTC | #6
On Mon, Oct 30, 2023 at 03:58:44PM +0000, Nadav Amit wrote:
> 
> 
> > On Oct 30, 2023, at 2:51 PM, Byungchul Park <byungchul@sk.com> wrote:
> > 
> > I really spent a lot of time hesitating whether splitting it or not.
> > However, the test result cannot be stable without those. I'm still
> > confused. I think the test result should be stable at each commit,
> > right?
> 
> Of course. You can extract the optimization we mentioned, and perhaps
> have more preparatory patches.
> 
> Just a couple of comments that may also help breaking the patches:
> 
> 1. The “stopping” logic is a bit not great. Try to see if you can
> somehow use shrinker or OOM infrastructure instead.

The stopping means "temporarily pausing" expanding migrc's pending
queue, not shrinking folios.. Yeah my fault. I will rename it another
not to make guys get it wrong.

> 2. Regarding “overflows”, it’s not always a question of whether an
> overflow would happen naturally, but whether a malicious process can
> trigger it.

I understand what you are worried about. However, it's intended that
the variable is going to overflow. And the overflow doesn't matter if we
are aware of it and careful in handling it. See time_after() in
include/linux/jiffies.h. That would help you understand what I mean.

	Byungchul
> 
> Regards,
> Nadav
>
  
Huang, Ying Nov. 1, 2023, 3:06 a.m. UTC | #7
Byungchul Park <byungchul@sk.com> writes:

> On Mon, Oct 30, 2023 at 09:00:56AM +0100, David Hildenbrand wrote:
>> On 30.10.23 08:25, Byungchul Park wrote:
>> > Implementation of CONFIG_MIGRC that stands for 'Migration Read Copy'.
>> > We always face the migration overhead at either promotion or demotion,
>> > while working with tiered memory e.g. CXL memory and found out TLB
>> > shootdown is a quite big one that is needed to get rid of if possible.
>> > 
>> > Fortunately, TLB flush can be defered or even skipped if both source and
>> > destination of folios during migration are kept until all TLB flushes
>> > required will have been done, of course, only if the target PTE entries
>> > have read only permission, more precisely speaking, don't have write
>> > permission. Otherwise, no doubt the folio might get messed up.
>> > 
>> > To achieve that:
>> > 
>> >     1. For the folios that map only to non-writable TLB entries, prevent
>> >        TLB flush at migration by keeping both source and destination
>> >        folios, which will be handled later at a better time.
>> > 
>> >     2. When any non-writable TLB entry changes to writable e.g. through
>> >        fault handler, give up CONFIG_MIGRC mechanism so as to perform
>> >        TLB flush required right away.
>> > 
>> >     3. Temporarily stop migrc from working when the system is in very
>> >        high memory pressure e.g. direct reclaim needed.
>> > 
>> > The measurement result:
>> > 
>> >     Architecture - x86_64
>> >     QEMU - kvm enabled, host cpu
>> >     Numa - 2 nodes (16 CPUs 1GB, no CPUs 8GB)
>> >     Linux Kernel - v6.6-rc5, numa balancing tiering on, demotion enabled
>> >     Benchmark - XSBench -p 50000000 (-p option makes the runtime longer)
>> > 
>> >     run 'perf stat' using events:
>> >        1) itlb.itlb_flush
>> >        2) tlb_flush.dtlb_thread
>> >        3) tlb_flush.stlb_any
>> >        4) dTLB-load-misses
>> >        5) dTLB-store-misses
>> >        6) iTLB-load-misses
>> > 
>> >     run 'cat /proc/vmstat' and pick:
>> >        1) numa_pages_migrated
>> >        2) pgmigrate_success
>> >        3) nr_tlb_remote_flush
>> >        4) nr_tlb_remote_flush_received
>> >        5) nr_tlb_local_flush_all
>> >        6) nr_tlb_local_flush_one
>> > 
>> >     BEFORE - mainline v6.6-rc5
>> >     ------------------------------------------
>> >     $ perf stat -a \
>> > 	   -e itlb.itlb_flush \
>> > 	   -e tlb_flush.dtlb_thread \
>> > 	   -e tlb_flush.stlb_any \
>> > 	   -e dTLB-load-misses \
>> > 	   -e dTLB-store-misses \
>> > 	   -e iTLB-load-misses \
>> > 	   ./XSBench -p 50000000
>> > 
>> >     Performance counter stats for 'system wide':
>> > 
>> >        20953405     itlb.itlb_flush
>> >        114886593    tlb_flush.dtlb_thread
>> >        88267015     tlb_flush.stlb_any
>> >        115304095543 dTLB-load-misses
>> >        163904743    dTLB-store-misses
>> >        608486259	   iTLB-load-misses
>> > 
>> >     556.787113849 seconds time elapsed
>> > 
>> >     $ cat /proc/vmstat
>> > 
>> >     ...
>> >     numa_pages_migrated 3378748
>> >     pgmigrate_success 7720310
>> >     nr_tlb_remote_flush 751464
>> >     nr_tlb_remote_flush_received 10742115
>> >     nr_tlb_local_flush_all 21899
>> >     nr_tlb_local_flush_one 740157
>> >     ...
>> > 
>> >     AFTER - mainline v6.6-rc5 + CONFIG_MIGRC
>> >     ------------------------------------------
>> >     $ perf stat -a \
>> > 	   -e itlb.itlb_flush \
>> > 	   -e tlb_flush.dtlb_thread \
>> > 	   -e tlb_flush.stlb_any \
>> > 	   -e dTLB-load-misses \
>> > 	   -e dTLB-store-misses \
>> > 	   -e iTLB-load-misses \
>> > 	   ./XSBench -p 50000000
>> > 
>> >     Performance counter stats for 'system wide':
>> > 
>> >        4353555      itlb.itlb_flush
>> >        72482780     tlb_flush.dtlb_thread
>> >        68226458     tlb_flush.stlb_any
>> >        114331610808 dTLB-load-misses
>> >        116084771    dTLB-store-misses
>> >        377180518    iTLB-load-misses
>> > 
>> >     552.667718220 seconds time elapsed
>> > 
>> >     $ cat /proc/vmstat
>> > 
>> 
>> So, an improvement of 0.74% ? How stable are the results? Serious question:
>
> I'm getting very stable result.
>
>> worth the churn?
>
> Yes, ultimately the time wise improvement should be observed. However,
> I've been focusing on the numbers of TLB flushes and TLB misses because
> better result in terms of total time will be followed depending on the
> test condition. We can see the result if we test with a system that:
>
>    1. has more CPUs that would induce a crazy number of IPIs.

FYI, the TLB flushing IPI number reduces much with commit 7e12beb8ca2a
("migrate_pages: batch flushing TLB") if multiple pages are migrated
together.

--
Best Regards,
Huang, Ying

>    2. has slow memories that makes TLB miss overhead bigger.
>    3. runs workloads that is harmful at TLB miss and IPI storm.
>    4. runs workloads that causes heavier numa migrations.
>    5. runs workloads that has a lot of read only permission mappings.
>    6. and so on.
>
> I will share the results once I manage to meet the conditions.
>
> By the way, I should've added IPI reduction because it also has super
> big delta :)
>
>> Or did I get the numbers wrong?
>> 
>> >   #define node_present_pages(nid)	(NODE_DATA(nid)->node_present_pages)
>> > diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
>> > index 5c02720c53a5..1ca2ac91aa14 100644
>> > --- a/include/linux/page-flags.h
>> > +++ b/include/linux/page-flags.h
>> > @@ -135,6 +135,9 @@ enum pageflags {
>> >   #ifdef CONFIG_ARCH_USES_PG_ARCH_X
>> >   	PG_arch_2,
>> >   	PG_arch_3,
>> > +#endif
>> > +#ifdef CONFIG_MIGRC
>> > +	PG_migrc,		/* Page has its copy under migrc's control */
>> >   #endif
>> >   	__NR_PAGEFLAGS,
>> > @@ -589,6 +592,10 @@ TESTCLEARFLAG(Young, young, PF_ANY)
>> >   PAGEFLAG(Idle, idle, PF_ANY)
>> >   #endif
>> > +#ifdef CONFIG_MIGRC
>> > +PAGEFLAG(Migrc, migrc, PF_ANY)
>> > +#endif
>> 
>> I assume you know this: new pageflags are frowned upon.
>
> Sorry for that. I really didn't want to add a new headache.
>
> 	Byungchul
  
Byungchul Park Nov. 8, 2023, 4:12 a.m. UTC | #8
On Mon, Oct 30, 2023 at 09:51:30PM +0900, Byungchul Park wrote:
> > > diff --git a/mm/memory.c b/mm/memory.c
> > > index 6c264d2f969c..75dc48b6e15f 100644
> > > --- a/mm/memory.c
> > > +++ b/mm/memory.c
> > > @@ -3359,6 +3359,19 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
> > > 	if (vmf->page)
> > > 		folio = page_folio(vmf->page);
> > > 
> > > +	/*
> > > +	 * This folio has its read copy to prevent inconsistency while
> > > +	 * deferring TLB flushes. However, the problem might arise if
> > > +	 * it's going to become writable.
> > > +	 *
> > > +	 * To prevent it, give up the deferring TLB flushes and perform
> > > +	 * TLB flush right away.
> > > +	 */
> > > +	if (folio && migrc_pending_folio(folio)) {
> > > +		migrc_unpend_folio(folio);
> > > +		migrc_try_flush_free_folios(NULL);
> > 
> > So many potential function calls… Probably they should have been combined
> > into one and at least migrc_pending_folio() should have been an inline
> > function in the header.
> 
> I will try to change it as you mention.
> 
> > > +	}
> > > +
> > 
> > What about mprotect? I thought David has changed it so it can set writable
> > PTEs.
> 
> I will check it out.

I found mprotect stuff is already performing TLB flushes needed for it.
So some redundant TLB flushes might happen by migrc but it's not that
harmful I think. Thanks.

	Byungchul
  
Byungchul Park Nov. 9, 2023, 5:35 a.m. UTC | #9
On Mon, Oct 30, 2023 at 09:51:30PM +0900, Byungchul Park wrote:
> On Mon, Oct 30, 2023 at 08:50:20AM +0000, Nadav Amit wrote:
> > I do not follow the ordering that ensures a migrated page would not
> > become writable before the migration is over. Perhaps a comment on what
> > guarantees the order would help.

IIUC, it'd be safe because unmap is performed first at the beginning of
each migration so as to be safe with page fault handlers. So I don't
think it should be a concern. Or am I missing something?

	Byungchul
  
Nadav Amit Nov. 9, 2023, 10:16 a.m. UTC | #10
> On Nov 8, 2023, at 6:12 AM, Byungchul Park <byungchul@sk.com> wrote:
> 
> !! External Email
> 
> On Mon, Oct 30, 2023 at 09:51:30PM +0900, Byungchul Park wrote:
>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>> index 6c264d2f969c..75dc48b6e15f 100644
>>>> --- a/mm/memory.c
>>>> +++ b/mm/memory.c
>>>> @@ -3359,6 +3359,19 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
>>>>  if (vmf->page)
>>>>          folio = page_folio(vmf->page);
>>>> 
>>>> + /*
>>>> +  * This folio has its read copy to prevent inconsistency while
>>>> +  * deferring TLB flushes. However, the problem might arise if
>>>> +  * it's going to become writable.
>>>> +  *
>>>> +  * To prevent it, give up the deferring TLB flushes and perform
>>>> +  * TLB flush right away.
>>>> +  */
>>>> + if (folio && migrc_pending_folio(folio)) {
>>>> +         migrc_unpend_folio(folio);
>>>> +         migrc_try_flush_free_folios(NULL);
>>> 
>>> So many potential function calls… Probably they should have been combined
>>> into one and at least migrc_pending_folio() should have been an inline
>>> function in the header.
>> 
>> I will try to change it as you mention.
>> 
>>>> + }
>>>> +
>>> 
>>> What about mprotect? I thought David has changed it so it can set writable
>>> PTEs.
>> 
>> I will check it out.
> 
> I found mprotect stuff is already performing TLB flushes needed for it.
> So some redundant TLB flushes might happen by migrc but it's not that
> harmful I think. Thanks.

Let me explain the scenario I am concerned with. Assume page P is RO, and
moves from Psrc to Pdst. Pointer “p” points to P. Initially (*p == 0).

Let’s also assume we also have an atomic variable “a”. Initially (a == 0).

I hope I got the migration function names right, but I hope the problem
itself can be clear regardless. 

CPU0			CPU1			CPU2		CPU3
----			----			----		----
			(user-mode)		(user-mode)		

			Access *p
			[Psrc cached in TLB]
 
migrate_pages_batch()
-> migrate_folio_unmap()

[ PTE updated, 
  still no flush ]

								mprotect(p,
									RW)

								[ Psrc is
								  RW ]

								[ flush
								  deferred]


						*p = 1  # Pdst
						
						xchg(&a, 1)
			mfence
			if (a == 1)
			  assert(*p == 1);


				
Now at this point the assertion might fail. CPU2 wrote into Pdst, whereas
CPU1 reads from Psrc. But based on x86 memory model, userspace might not
expect this scenario to be possible, hence leading to bugs.
  
Byungchul Park Nov. 10, 2023, 1:02 a.m. UTC | #11
On Thu, Nov 09, 2023 at 10:16:57AM +0000, Nadav Amit wrote:
> 
> 
> > On Nov 8, 2023, at 6:12 AM, Byungchul Park <byungchul@sk.com> wrote:
> > 
> > !! External Email
> > 
> > On Mon, Oct 30, 2023 at 09:51:30PM +0900, Byungchul Park wrote:
> >>>> diff --git a/mm/memory.c b/mm/memory.c
> >>>> index 6c264d2f969c..75dc48b6e15f 100644
> >>>> --- a/mm/memory.c
> >>>> +++ b/mm/memory.c
> >>>> @@ -3359,6 +3359,19 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
> >>>>  if (vmf->page)
> >>>>          folio = page_folio(vmf->page);
> >>>> 
> >>>> + /*
> >>>> +  * This folio has its read copy to prevent inconsistency while
> >>>> +  * deferring TLB flushes. However, the problem might arise if
> >>>> +  * it's going to become writable.
> >>>> +  *
> >>>> +  * To prevent it, give up the deferring TLB flushes and perform
> >>>> +  * TLB flush right away.
> >>>> +  */
> >>>> + if (folio && migrc_pending_folio(folio)) {
> >>>> +         migrc_unpend_folio(folio);
> >>>> +         migrc_try_flush_free_folios(NULL);
> >>> 
> >>> So many potential function calls… Probably they should have been combined
> >>> into one and at least migrc_pending_folio() should have been an inline
> >>> function in the header.
> >> 
> >> I will try to change it as you mention.
> >> 
> >>>> + }
> >>>> +
> >>> 
> >>> What about mprotect? I thought David has changed it so it can set writable
> >>> PTEs.
> >> 
> >> I will check it out.
> > 
> > I found mprotect stuff is already performing TLB flushes needed for it.
> > So some redundant TLB flushes might happen by migrc but it's not that
> > harmful I think. Thanks.
> 
> Let me explain the scenario I am concerned with. Assume page P is RO, and
> moves from Psrc to Pdst. Pointer “p” points to P. Initially (*p == 0).
> 
> Let’s also assume we also have an atomic variable “a”. Initially (a == 0).
> 
> I hope I got the migration function names right, but I hope the problem
> itself can be clear regardless. 
> 
> CPU0			CPU1			CPU2		CPU3
> ----			----			----		----
> 			(user-mode)		(user-mode)		
> 
> 			Access *p
> 			[Psrc cached in TLB]
>  
> migrate_pages_batch()
> -> migrate_folio_unmap()
> 
> [ PTE updated, 
>   still no flush ]
> 
> 								mprotect(p,
> 									RW)

Here,

mprotect()
   do_mprotect_pkey()
      tlb_finish_mmu()
         tlb_flush_mmu()

I thought TLB flush for mprotect() is performed by tlb_flush_mmu() so
any cached TLB entries on other CPUs can have chance to update. Could
you correct me if I get it wrong? Thanks.

	Byungchul

> 
> 								[ Psrc is
> 								  RW ]
> 
> 								[ flush
> 								  deferred]
> 
> 
> 						*p = 1  # Pdst
> 						
> 						xchg(&a, 1)
> 			mfence
> 			if (a == 1)
> 			  assert(*p == 1);
> 
> 
> 				
> Now at this point the assertion might fail. CPU2 wrote into Pdst, whereas
> CPU1 reads from Psrc. But based on x86 memory model, userspace might not
> expect this scenario to be possible, hence leading to bugs.
  
Byungchul Park Nov. 10, 2023, 3:13 a.m. UTC | #12
On Fri, Nov 10, 2023 at 10:02:01AM +0900, Byungchul Park wrote:
> On Thu, Nov 09, 2023 at 10:16:57AM +0000, Nadav Amit wrote:
> > 
> > 
> > > On Nov 8, 2023, at 6:12 AM, Byungchul Park <byungchul@sk.com> wrote:
> > > 
> > > !! External Email
> > > 
> > > On Mon, Oct 30, 2023 at 09:51:30PM +0900, Byungchul Park wrote:
> > >>>> diff --git a/mm/memory.c b/mm/memory.c
> > >>>> index 6c264d2f969c..75dc48b6e15f 100644
> > >>>> --- a/mm/memory.c
> > >>>> +++ b/mm/memory.c
> > >>>> @@ -3359,6 +3359,19 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
> > >>>>  if (vmf->page)
> > >>>>          folio = page_folio(vmf->page);
> > >>>> 
> > >>>> + /*
> > >>>> +  * This folio has its read copy to prevent inconsistency while
> > >>>> +  * deferring TLB flushes. However, the problem might arise if
> > >>>> +  * it's going to become writable.
> > >>>> +  *
> > >>>> +  * To prevent it, give up the deferring TLB flushes and perform
> > >>>> +  * TLB flush right away.
> > >>>> +  */
> > >>>> + if (folio && migrc_pending_folio(folio)) {
> > >>>> +         migrc_unpend_folio(folio);
> > >>>> +         migrc_try_flush_free_folios(NULL);
> > >>> 
> > >>> So many potential function calls… Probably they should have been combined
> > >>> into one and at least migrc_pending_folio() should have been an inline
> > >>> function in the header.
> > >> 
> > >> I will try to change it as you mention.
> > >> 
> > >>>> + }
> > >>>> +
> > >>> 
> > >>> What about mprotect? I thought David has changed it so it can set writable
> > >>> PTEs.
> > >> 
> > >> I will check it out.
> > > 
> > > I found mprotect stuff is already performing TLB flushes needed for it.
> > > So some redundant TLB flushes might happen by migrc but it's not that
> > > harmful I think. Thanks.
> > 
> > Let me explain the scenario I am concerned with. Assume page P is RO, and
> > moves from Psrc to Pdst. Pointer “p” points to P. Initially (*p == 0).
> > 
> > Let’s also assume we also have an atomic variable “a”. Initially (a == 0).
> > 
> > I hope I got the migration function names right, but I hope the problem
> > itself can be clear regardless. 
> > 
> > CPU0			CPU1			CPU2		CPU3
> > ----			----			----		----
> > 			(user-mode)		(user-mode)		
> > 
> > 			Access *p
> > 			[Psrc cached in TLB]
> >  
> > migrate_pages_batch()
> > -> migrate_folio_unmap()
> > 
> > [ PTE updated, 
> >   still no flush ]
> > 
> > 								mprotect(p,
> > 									RW)
> 
> Here,
> 
> mprotect()
>    do_mprotect_pkey()
>       tlb_finish_mmu()
>          tlb_flush_mmu()
> 
> I thought TLB flush for mprotect() is performed by tlb_flush_mmu() so
> any cached TLB entries on other CPUs can have chance to update. Could
> you correct me if I get it wrong? Thanks.

I guess you tried to inform me that x86 mmu automatically keeps the
consistancy based on cached TLB entries. Right? If yes, I should do
something on that path. If not, it's not problematic. Thoughts?

	Byungchul

> 	Byungchul
> 
> > 
> > 								[ Psrc is
> > 								  RW ]
> > 
> > 								[ flush
> > 								  deferred]
> > 
> > 
> > 						*p = 1  # Pdst
> > 						
> > 						xchg(&a, 1)
> > 			mfence
> > 			if (a == 1)
> > 			  assert(*p == 1);
> > 
> > 
> > 				
> > Now at this point the assertion might fail. CPU2 wrote into Pdst, whereas
> > CPU1 reads from Psrc. But based on x86 memory model, userspace might not
> > expect this scenario to be possible, hence leading to bugs.
  
Nadav Amit Nov. 10, 2023, 10:18 p.m. UTC | #13
> On Nov 10, 2023, at 5:13 AM, Byungchul Park <byungchul@sk.com> wrote:
> 
> 
>> Here,
>> 
>> mprotect()
>>   do_mprotect_pkey()
>>      tlb_finish_mmu()
>>         tlb_flush_mmu()
>> 
>> I thought TLB flush for mprotect() is performed by tlb_flush_mmu() so
>> any cached TLB entries on other CPUs can have chance to update. Could
>> you correct me if I get it wrong? Thanks.
> 
> I guess you tried to inform me that x86 mmu automatically keeps the
> consistancy based on cached TLB entries. Right? If yes, I should do
> something on that path. If not, it's not problematic. Thoughts?

Perhaps I lost something in this whole scheme. Overall, I find it overly
complicated and somewhat hard to follow.

Ideally, a solution should reduce the number of TLB flushing mechanisms
and not introduce yet another one that would further increase the already
high complexity.

Anyhow, I a bit convoluted 2 scenarios, and I’m not sure whether they
are a potential problem.

(1) Assume there is a RO page migration and you skipped a TLB flush.
Then you set a RO PTE entry for that page. Afterwards, you have mprotect()
that updates the PTE for that page to be RW.

Now, tlb_finish_mmu() will do a TLB flush eventually in the mprotect()
flow, but until it is done, you might have one CPU have RO pointing to
the source page (no TLB flush, right?) and another having RW access
that were loaded from the updated PTE. Did I miss a TLB flush that
should take place beforehand?


(2) Indeed we encountered many problems when TLB flushing decisions
are based on PTEs that are read from the page-tables and those do
not reflect the values that are held in the TLBs.


Overall, I think that a solution is to consolidate the TLB batching
mechanisms and hold per a group of PTEs (e.g., 512 PTEs held in one
page-table) the deferred TLB flush generation. Track the “done” TLB
flush generation, if the deferred is greater than the “done”, perform
a TLB flush.

This scheme is a bit similar to what you were doing, but more general,
and easier to follow. I think that its value might be more in simplifying
the code and reasoning than just performance.

IIRC, it kind of reminds the FreeBSD deferred TLB flushing scheme.
  
Byungchul Park Nov. 15, 2023, 5:48 a.m. UTC | #14
On Fri, Nov 10, 2023 at 10:18:06PM +0000, Nadav Amit wrote:
> > On Nov 10, 2023, at 5:13 AM, Byungchul Park <byungchul@sk.com> wrote:
> > 
> > 
> >> Here,
> >> 
> >> mprotect()
> >>   do_mprotect_pkey()
> >>      tlb_finish_mmu()
> >>         tlb_flush_mmu()
> >> 
> >> I thought TLB flush for mprotect() is performed by tlb_flush_mmu() so
> >> any cached TLB entries on other CPUs can have chance to update. Could
> >> you correct me if I get it wrong? Thanks.
> > 
> > I guess you tried to inform me that x86 mmu automatically keeps the
> > consistancy based on cached TLB entries. Right? If yes, I should do
> > something on that path. If not, it's not problematic. Thoughts?
> 
> Perhaps I lost something in this whole scheme. Overall, I find it overly
> complicated and somewhat hard to follow.
> 
> Ideally, a solution should reduce the number of TLB flushing mechanisms
> and not introduce yet another one that would further increase the already
> high complexity.
> 
> Anyhow, I a bit convoluted 2 scenarios, and I’m not sure whether they
> are a potential problem.
>
> (1) Assume there is a RO page migration and you skipped a TLB flush.
> Then you set a RO PTE entry for that page. Afterwards, you have mprotect()
> that updates the PTE for that page to be RW.
> 
> Now, tlb_finish_mmu() will do a TLB flush eventually in the mprotect()
> flow, but until it is done, you might have one CPU have RO pointing to
> the source page (no TLB flush, right?) and another having RW access

Is it possible for there to be another having RW access even before
mprotect() is done?

> that were loaded from the updated PTE. Did I miss a TLB flush that
> should take place beforehand?
> 
> 
> (2) Indeed we encountered many problems when TLB flushing decisions
> are based on PTEs that are read from the page-tables and those do
> not reflect the values that are held in the TLBs.

Do you have any example to help me understand what you mean?

> Overall, I think that a solution is to consolidate the TLB batching
> mechanisms and hold per a group of PTEs (e.g., 512 PTEs held in one
> page-table) the deferred TLB flush generation. Track the “done” TLB
> flush generation, if the deferred is greater than the “done”, perform
> a TLB flush.
>
> This scheme is a bit similar to what you were doing, but more general,
> and easier to follow. I think that its value might be more in simplifying
> the code and reasoning than just performance.

Worth noting that, at the beginning, I tried to reduce TLB flushes by:

	1. deferring and performing TLB flushes needed in a batch,
	2. skipping some TLB flushes that have already been done.

However, I gave up the 2nd optimization to keep only architecture
independent code in the patch set. So the current migrc version couldn't
skip TLB flushes but does only deferring and batching TLB flushes until
it's inevitable.

To cut a long story short, migrc let any other TLB flushes go performed
as is, *except* a special case e.i. RO mapped folios, during migration.

	Byungchul
  

Patch

diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
index 9d1361d2119c..cc52d83d17ee 100644
--- a/arch/x86/include/asm/tlbflush.h
+++ b/arch/x86/include/asm/tlbflush.h
@@ -292,9 +292,16 @@  static inline void arch_flush_tlb_batched_pending(struct mm_struct *mm)
 }
 
 extern void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch);
+extern void arch_tlbbatch_clean(struct arch_tlbflush_unmap_batch *batch);
 extern void arch_tlbbatch_fold(struct arch_tlbflush_unmap_batch *bdst,
 			       struct arch_tlbflush_unmap_batch *bsrc);
 
+#ifdef CONFIG_MIGRC
+extern void arch_migrc_adj(struct arch_tlbflush_unmap_batch *batch, int gen);
+#else
+static inline void arch_migrc_adj(struct arch_tlbflush_unmap_batch *batch, int gen) {}
+#endif
+
 static inline bool pte_flags_need_flush(unsigned long oldflags,
 					unsigned long newflags,
 					bool ignore_access)
diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c
index 314cd9912a88..0bdb617ffe94 100644
--- a/arch/x86/mm/tlb.c
+++ b/arch/x86/mm/tlb.c
@@ -1219,9 +1219,80 @@  STATIC_NOPV void native_flush_tlb_local(void)
 	native_write_cr3(__native_read_cr3());
 }
 
+#ifdef CONFIG_MIGRC
+DEFINE_PER_CPU(int, migrc_done);
+
+static inline bool before(int a, int b)
+{
+	return a - b < 0;
+}
+
+static inline int migrc_tlb_local_begin(void)
+{
+	int ret = atomic_read(&migrc_gen);
+
+	/*
+	 * The following order should be guaranteed:
+	 *
+	 *    1. A migrc's request is queued at migration.
+	 *    2. Each CPU performs __flush_tlb_local().
+	 *    3. Marks this CPU has performed the TLB flush after 1.
+	 *
+	 * To achieve that, a kind of timestamp, migrc_gen was
+	 * introduced to track the the order but conservatively, that
+	 * is, it doesn't care whether TLB flush is performed more than
+	 * it's needed but should guarantee TLB flush has been performed
+	 * after the timestamp in migrc_done.
+	 *
+	 * 1 is done by migrc_unmap_done() and migrc_req_end().
+	 * 2 is done by __flush_tlb_local().
+	 * 3 is done by migrc_tlb_local_begin() and migrc_tlb_local_end().
+	 *
+	 * FYI, an example that should avoid is:
+	 *
+	 *    1. Performs __flush_tlb_local().
+	 *    2. A migrc's request is queued at migration.
+	 *    3. Marks this CPU has performed the TLB flush after 2(?).
+	 *
+	 * XXX: barrier() would be sufficient if the architecture
+	 * already quarantees the order between memory access to
+	 * migrc_gen and TLB flush.
+	 */
+	smp_mb();
+
+	return ret;
+}
+
+static inline void migrc_tlb_local_end(int gen)
+{
+	/*
+	 * XXX: barrier() would be sufficient if the architecture
+	 * quarantees the order between TLB flush and memory access to
+	 * migrc_done.
+	 */
+	smp_mb();
+
+	if (before(READ_ONCE(*this_cpu_ptr(&migrc_done)), gen))
+		WRITE_ONCE(*this_cpu_ptr(&migrc_done), gen);
+}
+#else
+static inline int migrc_tlb_local_begin(void)
+{
+	return 0;
+}
+
+static inline void migrc_tlb_local_end(int gen)
+{
+}
+#endif
+
 void flush_tlb_local(void)
 {
+	unsigned int gen;
+
+	gen = migrc_tlb_local_begin();
 	__flush_tlb_local();
+	migrc_tlb_local_end(gen);
 }
 
 /*
@@ -1246,6 +1317,21 @@  void __flush_tlb_all(void)
 }
 EXPORT_SYMBOL_GPL(__flush_tlb_all);
 
+#ifdef CONFIG_MIGRC
+/*
+ * Skip CPUs that have already performed TLB flushes required, which is
+ * tracked by migrc_gen and migrc_done.
+ */
+void arch_migrc_adj(struct arch_tlbflush_unmap_batch *batch, int gen)
+{
+	int cpu;
+
+	for_each_cpu(cpu, &batch->cpumask)
+		if (!before(READ_ONCE(*per_cpu_ptr(&migrc_done, cpu)), gen))
+			cpumask_clear_cpu(cpu, &batch->cpumask);
+}
+#endif
+
 void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
 {
 	struct flush_tlb_info *info;
@@ -1274,6 +1360,11 @@  void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
 	put_cpu();
 }
 
+void arch_tlbbatch_clean(struct arch_tlbflush_unmap_batch *batch)
+{
+	cpumask_clear(&batch->cpumask);
+}
+
 void arch_tlbbatch_fold(struct arch_tlbflush_unmap_batch *bdst,
 			struct arch_tlbflush_unmap_batch *bsrc)
 {
diff --git a/include/linux/mm.h b/include/linux/mm.h
index bf5d0b1b16f4..ddefe6033b66 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -4062,4 +4062,29 @@  static inline void accept_memory(phys_addr_t start, phys_addr_t end)
 
 #endif
 
+#ifdef CONFIG_MIGRC
+void migrc_init_page(struct page *p);
+void migrc_unpend_folio(struct folio *f);
+bool migrc_pending_folio(struct folio *f);
+void migrc_shrink(struct llist_head *h);
+bool migrc_try_flush_free_folios(nodemask_t *nodes);
+void fold_ubc_nowr_to_migrc(void);
+void free_migrc_req(struct migrc_req *req);
+int migrc_pending_nr_in_zone(struct zone *z);
+void migrc_stop_pending(void);
+void migrc_resume_pending(void);
+
+extern atomic_t migrc_gen;
+#else
+static inline void migrc_init_page(struct page *p) {}
+static inline void migrc_unpend_folio(struct folio *f) {}
+static inline bool migrc_pending_folio(struct folio *f) { return false; }
+static inline void migrc_shrink(struct llist_head *h) {}
+static inline bool migrc_try_flush_free_folios(nodemask_t *nodes) { return false; }
+static inline void fold_ubc_nowr_to_migrc(void) {}
+static inline void free_migrc_req(struct migrc_req *req) {}
+static inline int migrc_pending_nr_in_zone(struct zone *z) { return 0; }
+static inline void migrc_stop_pending(void) {}
+static inline void migrc_resume_pending(void) {}
+#endif
 #endif /* _LINUX_MM_H */
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 36c5b43999e6..55350f1d0db2 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -101,6 +101,13 @@  struct page {
 				/* Or, free page */
 				struct list_head buddy_list;
 				struct list_head pcp_list;
+#ifdef CONFIG_MIGRC
+				/*
+				 * Hang onto migrc's request queues,
+				 * waiting for TLB flushes.
+				 */
+				struct llist_node migrc_node;
+#endif
 			};
 			/* See page-flags.h for PAGE_MAPPING_FLAGS */
 			struct address_space *mapping;
@@ -306,6 +313,13 @@  struct folio {
 	/* private: */
 				};
 	/* public: */
+#ifdef CONFIG_MIGRC
+				/*
+				 * Hang onto migrc's request queues,
+				 * waiting for TLB flushes.
+				 */
+				struct llist_node migrc_node;
+#endif
 			};
 			struct address_space *mapping;
 			pgoff_t index;
@@ -1372,4 +1386,39 @@  enum {
 	/* See also internal only FOLL flags in mm/internal.h */
 };
 
+#ifdef CONFIG_MIGRC
+struct migrc_req {
+	/*
+	 * folios pending for TLB flush
+	 */
+	struct llist_head folios;
+
+	/*
+	 * the last folio in 'struct llist_head folios'
+	 */
+	struct llist_node *last;
+
+	/*
+	 * for hanging to migrc's request queue, migrc_reqs
+	 */
+	struct llist_node llnode;
+
+	/*
+	 * architecture specific batch data
+	 */
+	struct arch_tlbflush_unmap_batch arch;
+
+	/*
+	 * timestamp when hanging to migrc's request queue, migrc_reqs
+	 */
+	int gen;
+
+	/*
+	 * associated numa node
+	 */
+	int nid;
+};
+#else
+struct migrc_req {};
+#endif
 #endif /* _LINUX_MM_TYPES_H */
diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 4106fbc5b4b3..4e78d1c9ceb2 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -980,6 +980,9 @@  struct zone {
 	/* Zone statistics */
 	atomic_long_t		vm_stat[NR_VM_ZONE_STAT_ITEMS];
 	atomic_long_t		vm_numa_event[NR_VM_NUMA_EVENT_ITEMS];
+#ifdef CONFIG_MIGRC
+	atomic_t		migrc_pending_nr;
+#endif
 } ____cacheline_internodealigned_in_smp;
 
 enum pgdat_flags {
@@ -1398,6 +1401,10 @@  typedef struct pglist_data {
 #ifdef CONFIG_MEMORY_FAILURE
 	struct memory_failure_stats mf_stats;
 #endif
+#ifdef CONFIG_MIGRC
+	atomic_t migrc_pending_nr;
+	struct llist_head migrc_reqs;
+#endif
 } pg_data_t;
 
 #define node_present_pages(nid)	(NODE_DATA(nid)->node_present_pages)
diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 5c02720c53a5..1ca2ac91aa14 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -135,6 +135,9 @@  enum pageflags {
 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
 	PG_arch_2,
 	PG_arch_3,
+#endif
+#ifdef CONFIG_MIGRC
+	PG_migrc,		/* Page has its copy under migrc's control */
 #endif
 	__NR_PAGEFLAGS,
 
@@ -589,6 +592,10 @@  TESTCLEARFLAG(Young, young, PF_ANY)
 PAGEFLAG(Idle, idle, PF_ANY)
 #endif
 
+#ifdef CONFIG_MIGRC
+PAGEFLAG(Migrc, migrc, PF_ANY)
+#endif
+
 /*
  * PageReported() is used to track reported free pages within the Buddy
  * allocator. We can use the non-atomic version of the test and set
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 63189c023357..b7c402850b2f 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1325,6 +1325,14 @@  struct task_struct {
 
 	struct tlbflush_unmap_batch	tlb_ubc;
 	struct tlbflush_unmap_batch	tlb_ubc_nowr;
+#ifdef CONFIG_MIGRC
+	/*
+	 * Temporarily used while a migration is on processing before
+	 * hanging to the associated numa node at the end of the
+	 * migration.
+	 */
+	struct migrc_req		*mreq;
+#endif
 
 	/* Cache last used pipe for splice(): */
 	struct pipe_inode_info		*splice_pipe;
diff --git a/include/trace/events/mmflags.h b/include/trace/events/mmflags.h
index 1478b9dd05fa..85b3b8859bc8 100644
--- a/include/trace/events/mmflags.h
+++ b/include/trace/events/mmflags.h
@@ -95,6 +95,12 @@ 
 #define IF_HAVE_PG_ARCH_X(_name)
 #endif
 
+#ifdef CONFIG_MIGRC
+#define IF_HAVE_PG_MIGRC(_name) ,{1UL << PG_##_name, __stringify(_name)}
+#else
+#define IF_HAVE_PG_MIGRC(_name)
+#endif
+
 #define DEF_PAGEFLAG_NAME(_name) { 1UL <<  PG_##_name, __stringify(_name) }
 
 #define __def_pageflag_names						\
@@ -125,7 +131,8 @@  IF_HAVE_PG_HWPOISON(hwpoison)						\
 IF_HAVE_PG_IDLE(idle)							\
 IF_HAVE_PG_IDLE(young)							\
 IF_HAVE_PG_ARCH_X(arch_2)						\
-IF_HAVE_PG_ARCH_X(arch_3)
+IF_HAVE_PG_ARCH_X(arch_3)						\
+IF_HAVE_PG_MIGRC(migrc)
 
 #define show_page_flags(flags)						\
 	(flags) ? __print_flags(flags, "|",				\
diff --git a/init/Kconfig b/init/Kconfig
index 6d35728b94b2..4d88dab52059 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -908,6 +908,20 @@  config NUMA_BALANCING_DEFAULT_ENABLED
 	  If set, automatic NUMA balancing will be enabled if running on a NUMA
 	  machine.
 
+config MIGRC
+	bool "Deferring TLB flush by keeping read copies on migration"
+	depends on ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
+	depends on NUMA_BALANCING
+	depends on 64BIT
+	default n
+	help
+	  TLB flush is necessary when PTE changes by migration. However,
+	  TLB flush can be deferred if both copies of the src page and
+	  the dst page are kept until TLB flush if they are non-writable.
+	  System performance will be improved especially in case that
+	  promotion and demotion types of migrations are heavily
+	  happening.
+
 menuconfig CGROUPS
 	bool "Control Group support"
 	select KERNFS
diff --git a/mm/memory.c b/mm/memory.c
index 6c264d2f969c..75dc48b6e15f 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3359,6 +3359,19 @@  static vm_fault_t do_wp_page(struct vm_fault *vmf)
 	if (vmf->page)
 		folio = page_folio(vmf->page);
 
+	/*
+	 * This folio has its read copy to prevent inconsistency while
+	 * deferring TLB flushes. However, the problem might arise if
+	 * it's going to become writable.
+	 *
+	 * To prevent it, give up the deferring TLB flushes and perform
+	 * TLB flush right away.
+	 */
+	if (folio && migrc_pending_folio(folio)) {
+		migrc_unpend_folio(folio);
+		migrc_try_flush_free_folios(NULL);
+	}
+
 	/*
 	 * Shared mapping: we are guaranteed to have VM_WRITE and
 	 * FAULT_FLAG_WRITE set at this point.
diff --git a/mm/migrate.c b/mm/migrate.c
index 2053b54556ca..76278eca8417 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -57,6 +57,204 @@ 
 
 #include "internal.h"
 
+#ifdef CONFIG_MIGRC
+
+atomic_t migrc_gen;
+
+static struct migrc_req *alloc_migrc_req(void)
+{
+	return kmalloc(sizeof(struct migrc_req), GFP_KERNEL);
+}
+
+void free_migrc_req(struct migrc_req *req)
+{
+	kfree(req);
+}
+
+void migrc_init_page(struct page *p)
+{
+	ClearPageMigrc(p);
+}
+
+/*
+ * One of migrc's core functions that frees up all the folios in a
+ * migrc's pending queue.
+ *
+ * The list should be isolated before.
+ */
+void migrc_shrink(struct llist_head *h)
+{
+	struct folio *f, *f2;
+	struct llist_node *n;
+
+	n = llist_del_all(h);
+	llist_for_each_entry_safe(f, f2, n, migrc_node) {
+		struct pglist_data *node;
+		struct zone *zone;
+
+		node = NODE_DATA(folio_nid(f));
+		zone = page_zone(&f->page);
+		atomic_dec(&node->migrc_pending_nr);
+		atomic_dec(&zone->migrc_pending_nr);
+		folio_put(f);
+	}
+}
+
+void migrc_unpend_folio(struct folio *f)
+{
+	folio_clear_migrc(f);
+}
+
+bool migrc_pending_folio(struct folio *f)
+{
+	return folio_test_migrc(f);
+}
+
+/*
+ * Marks the folios as being under migrc's control.
+ */
+static void migrc_mark_folios(struct folio *fsrc, struct folio *fdst)
+{
+	if (!current->mreq)
+		return;
+
+	folio_get(fsrc);
+	folio_set_migrc(fsrc);
+	folio_set_migrc(fdst);
+	fold_ubc_nowr_to_migrc();
+}
+
+static bool migrc_marked_folios(struct folio *fsrc, struct folio *fdst)
+{
+	/*
+	 * XXX: The condition below is so tricky. Please suggest a
+	 * better idea to determine if it's under migrc's control.
+	 */
+	return migrc_pending_folio(fsrc) && migrc_pending_folio(fdst);
+}
+
+static void migrc_undo_folios(struct folio *fsrc, struct folio *fdst)
+{
+	/*
+	 * TLB flushes needed are already done at this moment so the
+	 * flag doesn't have to be kept.
+	 */
+	folio_clear_migrc(fsrc);
+	folio_clear_migrc(fdst);
+
+	folio_put(fsrc);
+}
+
+static void migrc_expand_req(struct folio *f)
+{
+	struct migrc_req *req = current->mreq;
+	struct pglist_data *node;
+	struct zone *zone;
+
+	if (!req)
+		return;
+
+	/*
+	 * Focusing on numa migration, all the nids in a req are same.
+	 */
+	if (req->nid == -1)
+		req->nid = folio_nid(f);
+	else if (WARN_ON(req->nid != folio_nid(f)))
+		return;
+
+	if (llist_add(&f->migrc_node, &req->folios))
+		req->last = &f->migrc_node;
+
+	node = NODE_DATA(folio_nid(f));
+	zone = page_zone(&f->page);
+	atomic_inc(&node->migrc_pending_nr);
+	atomic_inc(&zone->migrc_pending_nr);
+}
+
+static void migrc_unmap_done(void)
+{
+	if (current->mreq)
+		current->mreq->gen = atomic_inc_return(&migrc_gen);
+}
+
+/*
+ * Prepares for gathering folios pending for TLB flushes, try to
+ * allocate objects needed, initialize them and make them ready.
+ */
+static void migrc_req_start(void)
+{
+	struct migrc_req *req;
+
+	if (WARN_ON(current->mreq))
+		return;
+
+	req = alloc_migrc_req();
+	if (!req)
+		return;
+
+	arch_tlbbatch_clean(&req->arch);
+	init_llist_head(&req->folios);
+	req->last = NULL;
+	req->nid = -1;
+	current->mreq = req;
+}
+
+/*
+ * Hang the request with the collected folios to 'migrc_reqs' llist of
+ * the corresponding node.
+ */
+static void migrc_req_end_if_started(void)
+{
+	struct migrc_req *req = current->mreq;
+
+	if (!req)
+		return;
+
+	if (llist_empty(&req->folios))
+		free_migrc_req(req);
+	else
+		llist_add(&req->llnode, &NODE_DATA(req->nid)->migrc_reqs);
+
+	current->mreq = NULL;
+}
+
+int migrc_pending_nr_in_zone(struct zone *z)
+{
+	return atomic_read(&z->migrc_pending_nr);
+}
+
+static atomic_t migrc_stop = ATOMIC_INIT(0);
+
+void migrc_stop_pending(void)
+{
+	atomic_inc(&migrc_stop);
+}
+
+void migrc_resume_pending(void)
+{
+	atomic_dec(&migrc_stop);
+}
+
+static inline bool migrc_stopped_pending(void)
+{
+	return !!atomic_read(&migrc_stop);
+}
+
+static bool can_migrc_system(void)
+{
+	return !migrc_stopped_pending();
+}
+#else
+static inline void migrc_mark_folios(struct folio *fsrc, struct folio *fdst) {}
+static inline bool migrc_marked_folios(struct folio *fsrc, struct folio *fdst) { return false; }
+static inline void migrc_undo_folios(struct folio *fsrc, struct folio *fdst) {}
+static inline void migrc_expand_req(struct folio *f) {}
+static inline void migrc_unmap_done(void) {}
+static inline void migrc_req_start(void) {}
+static inline void migrc_req_end_if_started(void) {}
+static inline bool can_migrc_system(void) { return false; }
+#endif
+
 bool isolate_movable_page(struct page *page, isolate_mode_t mode)
 {
 	struct folio *folio = folio_get_nontail_page(page);
@@ -379,6 +577,7 @@  static int folio_expected_refs(struct address_space *mapping,
 		struct folio *folio)
 {
 	int refs = 1;
+
 	if (!mapping)
 		return refs;
 
@@ -406,6 +605,13 @@  int folio_migrate_mapping(struct address_space *mapping,
 	int expected_count = folio_expected_refs(mapping, folio) + extra_count;
 	long nr = folio_nr_pages(folio);
 
+	/*
+	 * Migrc mechanism increases the reference count to defer
+	 * freeing up folios until a better time.
+	 */
+	if (migrc_marked_folios(folio, newfolio))
+		expected_count++;
+
 	if (!mapping) {
 		/* Anonymous page without mapping */
 		if (folio_ref_count(folio) != expected_count)
@@ -1620,16 +1826,30 @@  static int migrate_pages_batch(struct list_head *from,
 	LIST_HEAD(unmap_folios);
 	LIST_HEAD(dst_folios);
 	bool nosplit = (reason == MR_NUMA_MISPLACED);
+	bool migrc_cond1;
 
 	VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
 			!list_empty(from) && !list_is_singular(from));
 
+	/*
+	 * For now, apply migrc only to numa migration.
+	 */
+	migrc_cond1 = can_migrc_system() &&
+		      ((reason == MR_DEMOTION && current_is_kswapd()) ||
+		      (reason == MR_NUMA_MISPLACED));
+
+	if (migrc_cond1)
+		migrc_req_start();
 	for (pass = 0; pass < nr_pass && retry; pass++) {
 		retry = 0;
 		thp_retry = 0;
 		nr_retry_pages = 0;
 
 		list_for_each_entry_safe(folio, folio2, from, lru) {
+			int nr_required;
+			bool migrc_cond2;
+			bool migrc;
+
 			is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
 			nr_pages = folio_nr_pages(folio);
 
@@ -1657,9 +1877,34 @@  static int migrate_pages_batch(struct list_head *from,
 				continue;
 			}
 
+			/*
+			 * In case that the system is in high memory
+			 * pressure, let's stop migrc from working and
+			 * free up all the folios in the pending queues
+			 * not to impact the system.
+			 */
+			if (unlikely(migrc_cond1 && !can_migrc_system())) {
+				migrc_cond1 = false;
+				migrc_req_end_if_started();
+				migrc_try_flush_free_folios(NULL);
+			}
+
+			nr_required = nr_flush_required();
 			rc = migrate_folio_unmap(get_new_folio, put_new_folio,
 					private, folio, &dst, mode, reason,
 					ret_folios);
+			/*
+			 * migrc_cond2 will be true only if:
+			 *
+			 *    1. There's no writable mapping at all
+			 *    2. There's read only mapping found
+			 *    3. Not already under migrc's control
+			 */
+			migrc_cond2 = nr_required == nr_flush_required() &&
+				      nr_flush_required_nowr() &&
+				      !migrc_pending_folio(folio);
+			migrc = migrc_cond1 && migrc_cond2;
+
 			/*
 			 * The rules are:
 			 *	Success: folio will be freed
@@ -1720,6 +1965,9 @@  static int migrate_pages_batch(struct list_head *from,
 			case MIGRATEPAGE_UNMAP:
 				list_move_tail(&folio->lru, &unmap_folios);
 				list_add_tail(&dst->lru, &dst_folios);
+
+				if (migrc)
+					migrc_mark_folios(folio, dst);
 				break;
 			default:
 				/*
@@ -1733,12 +1981,24 @@  static int migrate_pages_batch(struct list_head *from,
 				stats->nr_failed_pages += nr_pages;
 				break;
 			}
+
+			/*
+			 * Done with the current folio. Fold the ro
+			 * batch data gathered, to the normal batch.
+			 */
+			fold_ubc_nowr();
 		}
 	}
 	nr_failed += retry;
 	stats->nr_thp_failed += thp_retry;
 	stats->nr_failed_pages += nr_retry_pages;
 move:
+	/*
+	 * Generate a new timestamp that will be used to filter out
+	 * CPUs that have already performed TLB flushes needed.
+	 */
+	migrc_unmap_done();
+
 	/* Flush TLBs for all unmapped folios */
 	try_to_unmap_flush();
 
@@ -1774,6 +2034,9 @@  static int migrate_pages_batch(struct list_head *from,
 			case MIGRATEPAGE_SUCCESS:
 				stats->nr_succeeded += nr_pages;
 				stats->nr_thp_succeeded += is_thp;
+
+				if (migrc_marked_folios(folio, dst))
+					migrc_expand_req(folio);
 				break;
 			default:
 				nr_failed++;
@@ -1791,6 +2054,8 @@  static int migrate_pages_batch(struct list_head *from,
 
 	rc = rc_saved ? : nr_failed;
 out:
+	migrc_req_end_if_started();
+
 	/* Cleanup remaining folios */
 	dst = list_first_entry(&dst_folios, struct folio, lru);
 	dst2 = list_next_entry(dst, lru);
@@ -1798,6 +2063,9 @@  static int migrate_pages_batch(struct list_head *from,
 		int page_was_mapped = 0;
 		struct anon_vma *anon_vma = NULL;
 
+		if (migrc_marked_folios(folio, dst))
+			migrc_undo_folios(folio, dst);
+
 		__migrate_folio_extract(dst, &page_was_mapped, &anon_vma);
 		migrate_folio_undo_src(folio, page_was_mapped, anon_vma,
 				       true, ret_folios);
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 95546f376302..2fa84c3a9681 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1535,6 +1535,9 @@  inline void post_alloc_hook(struct page *page, unsigned int order,
 
 	set_page_owner(page, order, gfp_flags);
 	page_table_check_alloc(page, order);
+
+	for (i = 0; i != 1 << order; ++i)
+		migrc_init_page(page + i);
 }
 
 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
@@ -2839,6 +2842,8 @@  bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
 	long min = mark;
 	int o;
 
+	free_pages += migrc_pending_nr_in_zone(z);
+
 	/* free_pages may go negative - that's OK */
 	free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
 
@@ -2933,7 +2938,7 @@  static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
 		long usable_free;
 		long reserved;
 
-		usable_free = free_pages;
+		usable_free = free_pages + migrc_pending_nr_in_zone(z);
 		reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
 
 		/* reserved may over estimate high-atomic reserves. */
@@ -3121,6 +3126,16 @@  get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
 				       gfp_mask)) {
 			int ret;
 
+			/*
+			 * Free the pending folios so that the remaining
+			 * code can use them. Check zone_watermark_fast()
+			 * again with more accurate stats before going.
+			 */
+			migrc_try_flush_free_folios(ac->nodemask);
+			if (zone_watermark_fast(zone, order, mark,
+			    ac->highest_zoneidx, alloc_flags, gfp_mask))
+				goto try_this_zone;
+
 			if (has_unaccepted_memory()) {
 				if (try_to_accept_memory(zone, order))
 					goto try_this_zone;
@@ -3911,6 +3926,7 @@  __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
 	unsigned int cpuset_mems_cookie;
 	unsigned int zonelist_iter_cookie;
 	int reserve_flags;
+	bool migrc_stopped = false;
 
 restart:
 	compaction_retries = 0;
@@ -4042,6 +4058,15 @@  __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
 	if (page)
 		goto got_pg;
 
+	/*
+	 * The system is in very high memory pressure. Stop migrc from
+	 * expanding its pending queue and working temporarily.
+	 */
+	if (!migrc_stopped) {
+		migrc_stop_pending();
+		migrc_stopped = true;
+	}
+
 	/* Caller is not willing to reclaim, we can't balance anything */
 	if (!can_direct_reclaim)
 		goto nopage;
@@ -4169,6 +4194,8 @@  __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
 	warn_alloc(gfp_mask, ac->nodemask,
 			"page allocation failure: order:%u", order);
 got_pg:
+	if (migrc_stopped)
+		migrc_resume_pending();
 	return page;
 }
 
diff --git a/mm/rmap.c b/mm/rmap.c
index a045f3b57c60..d0fa7bf66e70 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -606,6 +606,82 @@  struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
 
 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 
+#ifdef CONFIG_MIGRC
+/*
+ * Gather folios and CPUs in cpumask to handle.
+ */
+static void migrc_gather(struct llist_head *folios,
+			 struct arch_tlbflush_unmap_batch *arch,
+			 struct llist_head *reqs)
+{
+	struct llist_node *reqs_nodes;
+	struct migrc_req *req;
+	struct migrc_req *req2;
+
+	reqs_nodes = llist_del_all(reqs);
+	if (!reqs_nodes)
+		return;
+
+	/*
+	 * TODO: Optimize the time complexity.
+	 */
+	llist_for_each_entry_safe(req, req2, reqs_nodes, llnode) {
+		struct llist_node *n;
+
+		arch_migrc_adj(&req->arch, req->gen);
+		arch_tlbbatch_fold(arch, &req->arch);
+
+		n = llist_del_all(&req->folios);
+		llist_add_batch(n, req->last, folios);
+		free_migrc_req(req);
+	}
+}
+
+bool migrc_try_flush_free_folios(nodemask_t *nodes)
+{
+	struct arch_tlbflush_unmap_batch arch;
+	LLIST_HEAD(folios);
+	bool ret = false;
+	int nid;
+
+	arch_tlbbatch_clean(&arch);
+	nodes = nodes ?: &node_possible_map;
+	for_each_node_mask(nid, *nodes)
+		migrc_gather(&folios, &arch, &NODE_DATA(nid)->migrc_reqs);
+
+	ret = !llist_empty(&folios);
+	if (ret) {
+		arch_tlbbatch_flush(&arch);
+		migrc_shrink(&folios);
+	}
+	return ret;
+}
+
+/*
+ * Fold ro batch data to the current migrc's request. In case that the
+ * request is not ready, fold it to the normal batch as a fallback.
+ */
+void fold_ubc_nowr_to_migrc(void)
+{
+	struct tlbflush_unmap_batch *tlb_ubc_nowr = &current->tlb_ubc_nowr;
+	struct migrc_req *req;
+
+	if (!tlb_ubc_nowr->nr_flush_required)
+		return;
+
+	req = current->mreq;
+	if (!req) {
+		/* fallback */
+		fold_ubc_nowr();
+		return;
+	}
+
+	arch_tlbbatch_fold(&req->arch, &tlb_ubc_nowr->arch);
+	tlb_ubc_nowr->nr_flush_required = 0;
+	tlb_ubc_nowr->writable = false;
+}
+#endif
+
 void fold_ubc_nowr(void)
 {
 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;