Hi,
While I'm working with CXL, I have been facing migraion overhead esp.
TLB shootdown on promotion or demotion between different tiers. Yeah..
most TLB shootdowns on migration through hinting fault can be avoided
thanks to Huang Ying's work, commit 4d4b6d66db ("mm,unmap: avoid
flushing TLB in batch if PTE is inaccessible").
However, it's only for ones using hinting fault. I thought it'd be much
better if we have a general mechanism to reduce # of TLB flushes that
we can apply to any type of migration. I tried it only for tiering
migration for now tho.
I'm suggesting a mechanism to reduce TLB flushes by keeping source and
destination of folios participated in the migrations until all TLB
flushes required are done, only if those folios are not mapped with
write permission PTE entries at all.
I saw the number of TLB full flush reduced over 50% and the performance
a little bit improved but not that big with the workload I tested with,
XSBench. However, I believe that it would help more with other ones or
any real ones. It'd be appreciated to tell me if I'm missing something.
Byungchul
Byungchul Park (2):
mm/rmap: Recognize non-writable TLB entries during TLB batch flush
mm: Defer TLB flush by keeping both src and dst folios at migration
arch/x86/include/asm/tlbflush.h | 9 +
arch/x86/mm/tlb.c | 59 +++++++
include/linux/mm.h | 30 ++++
include/linux/mm_types.h | 34 ++++
include/linux/mm_types_task.h | 4 +-
include/linux/mmzone.h | 6 +
include/linux/sched.h | 5 +
init/Kconfig | 12 ++
mm/internal.h | 14 ++
mm/memory.c | 9 +-
mm/migrate.c | 287 +++++++++++++++++++++++++++++++-
mm/mm_init.c | 1 +
mm/page_alloc.c | 16 ++
mm/rmap.c | 121 +++++++++++++-
14 files changed, 595 insertions(+), 12 deletions(-)
--
2.17.1
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 have only non-writable TLB entries, prevent
TLB flush by keeping both source and destination of folios during
migration, 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. TLB flushes can be skipped if all TLB flushes required to free the
duplicated folios have been done by any reason, which doesn't have
to be done from migrations.
4. Adjust watermark check routine, __zone_watermark_ok(), with the
number of duplicated folios because those folios can be freed
and obtained right away through appropreate TLB flushes.
5. Perform TLB flushes and free the duplicated folios pending the
flushes if page allocation routine is in trouble due to memory
pressure, even more aggresively for high order allocation.
The measurement result:
Architecture - x86_64
QEMU - kvm enabled, host cpu, 2nodes((4cpus, 2GB)+(cpuless, 6GB))
Linux Kernel - v6.4, numa balancing tiering on, demotion enabled
Benchmark - XSBench with no parameter changed
run 'perf stat' using events:
(FYI, process wide result ~= system wide result(-a option))
1) itlb.itlb_flush
2) tlb_flush.dtlb_thread
3) tlb_flush.stlb_any
run 'cat /proc/vmstat' and pick up:
1) pgdemote_kswapd
2) numa_pages_migrated
3) pgmigrate_success
4) nr_tlb_remote_flush
5) nr_tlb_remote_flush_received
6) nr_tlb_local_flush_all
7) nr_tlb_local_flush_one
BEFORE - mainline v6.4
==========================================
$ perf stat -e itlb.itlb_flush,tlb_flush.dtlb_thread,tlb_flush.stlb_any ./XSBench
Performance counter stats for './XSBench':
426856 itlb.itlb_flush
6900414 tlb_flush.dtlb_thread
7303137 tlb_flush.stlb_any
33.500486566 seconds time elapsed
92.852128000 seconds user
10.526718000 seconds sys
$ cat /proc/vmstat
...
pgdemote_kswapd 1052596
numa_pages_migrated 1052359
pgmigrate_success 2161846
nr_tlb_remote_flush 72370
nr_tlb_remote_flush_received 213711
nr_tlb_local_flush_all 3385
nr_tlb_local_flush_one 198679
...
AFTER - mainline v6.4 + CONFIG_MIGRC
==========================================
$ perf stat -e itlb.itlb_flush,tlb_flush.dtlb_thread,tlb_flush.stlb_any ./XSBench
Performance counter stats for './XSBench':
179537 itlb.itlb_flush
6131135 tlb_flush.dtlb_thread
6920979 tlb_flush.stlb_any
30.396700625 seconds time elapsed
80.331252000 seconds user
10.303761000 seconds sys
$ cat /proc/vmstat
...
pgdemote_kswapd 1044602
numa_pages_migrated 1044202
pgmigrate_success 2157808
nr_tlb_remote_flush 30453
nr_tlb_remote_flush_received 88840
nr_tlb_local_flush_all 3039
nr_tlb_local_flush_one 198875
...
Signed-off-by: Byungchul Park <[email protected]>
---
arch/x86/include/asm/tlbflush.h | 7 +
arch/x86/mm/tlb.c | 52 ++++++
include/linux/mm.h | 30 ++++
include/linux/mm_types.h | 34 ++++
include/linux/mmzone.h | 6 +
include/linux/sched.h | 4 +
init/Kconfig | 12 ++
mm/internal.h | 10 ++
mm/memory.c | 9 +-
mm/migrate.c | 287 +++++++++++++++++++++++++++++++-
mm/mm_init.c | 1 +
mm/page_alloc.c | 16 ++
mm/rmap.c | 92 ++++++++++
13 files changed, 555 insertions(+), 5 deletions(-)
diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
index 63504cde364b..da987c15049e 100644
--- a/arch/x86/include/asm/tlbflush.h
+++ b/arch/x86/include/asm/tlbflush.h
@@ -279,9 +279,16 @@ static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
}
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 69d145f1fff1..54f98a50fd59 100644
--- a/arch/x86/mm/tlb.c
+++ b/arch/x86/mm/tlb.c
@@ -1210,9 +1210,40 @@ 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 int migrc_tlb_local_begin(void)
+{
+ int ret = atomic_read(&migrc_gen);
+
+ smp_mb__after_atomic();
+ return ret;
+}
+
+static inline void migrc_tlb_local_end(int gen)
+{
+ smp_mb();
+ 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);
}
/*
@@ -1237,6 +1268,22 @@ void __flush_tlb_all(void)
}
EXPORT_SYMBOL_GPL(__flush_tlb_all);
+#ifdef CONFIG_MIGRC
+static inline bool before(int a, int b)
+{
+ return a - b < 0;
+}
+
+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;
@@ -1265,6 +1312,11 @@ 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 27ce77080c79..e1f6e1fdab18 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -3816,4 +3816,34 @@ madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
}
#endif
+#ifdef CONFIG_MIGRC
+void migrc_init_page(struct page *p);
+bool migrc_pending(struct folio *f);
+void migrc_shrink(struct llist_head *h);
+void migrc_req_start(void);
+void migrc_req_end(void);
+bool migrc_req_processing(void);
+bool migrc_try_flush(void);
+void migrc_try_flush_dirty(void);
+struct migrc_req *fold_ubc_nowr_migrc_req(void);
+void free_migrc_req(struct migrc_req *req);
+int migrc_pending_nr_in_zone(struct zone *z);
+
+extern atomic_t migrc_gen;
+extern struct llist_head migrc_reqs;
+extern struct llist_head migrc_reqs_dirty;
+#else
+static inline void migrc_init_page(struct page *p) {}
+static inline bool migrc_pending(struct folio *f) { return false; }
+static inline void migrc_shrink(struct llist_head *h) {}
+static inline void migrc_req_start(void) {}
+static inline void migrc_req_end(void) {}
+static inline bool migrc_req_processing(void) { return false; }
+static inline bool migrc_try_flush(void) { return false; }
+static inline void migrc_try_flush_dirty(void) {}
+static inline struct migrc_req *fold_ubc_nowr_migrc_req(void) { return NULL; }
+static inline void free_migrc_req(struct migrc_req *req) {}
+static inline int migrc_pending_nr_in_zone(struct zone *z) { return 0; }
+#endif
+
#endif /* _LINUX_MM_H */
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 306a3d1a0fa6..3be66d3eabd2 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -228,6 +228,10 @@ struct page {
#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
int _last_cpupid;
#endif
+#ifdef CONFIG_MIGRC
+ struct llist_node migrc_node;
+ unsigned int migrc_state;
+#endif
} _struct_page_alignment;
/*
@@ -1255,4 +1259,34 @@ enum {
/* See also internal only FOLL flags in mm/internal.h */
};
+#ifdef CONFIG_MIGRC
+struct migrc_req {
+ /*
+ * pages pending for TLB flush
+ */
+ struct llist_head pages;
+
+ /*
+ * llist_node of the last page in pages llist
+ */
+ struct llist_node *last;
+
+ /*
+ * for hanging onto migrc_reqs llist
+ */
+ struct llist_node llnode;
+
+ /*
+ * architecture specific batch information
+ */
+ struct arch_tlbflush_unmap_batch arch;
+
+ /*
+ * when the request hung onto migrc_reqs llist
+ */
+ int gen;
+};
+#else
+struct migrc_req {};
+#endif
#endif /* _LINUX_MM_TYPES_H */
diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index a4889c9d4055..1ec79bb63ba7 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -958,6 +958,9 @@ 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 {
@@ -1371,6 +1374,9 @@ typedef struct pglist_data {
#ifdef CONFIG_MEMORY_FAILURE
struct memory_failure_stats mf_stats;
#endif
+#ifdef CONFIG_MIGRC
+ atomic_t migrc_pending_nr;
+#endif
} pg_data_t;
#define node_present_pages(nid) (NODE_DATA(nid)->node_present_pages)
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 2232b2cdfce8..d0a46089959d 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1323,6 +1323,10 @@ struct task_struct {
struct tlbflush_unmap_batch tlb_ubc;
struct tlbflush_unmap_batch tlb_ubc_nowr;
+#ifdef CONFIG_MIGRC
+ struct migrc_req *mreq;
+ struct migrc_req *mreq_dirty;
+#endif
/* Cache last used pipe for splice(): */
struct pipe_inode_info *splice_pipe;
diff --git a/init/Kconfig b/init/Kconfig
index 32c24950c4ce..f4882c1be364 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -907,6 +907,18 @@ 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
+ 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 type of migration is heavily happening.
+
menuconfig CGROUPS
bool "Control Group support"
select KERNFS
diff --git a/mm/internal.h b/mm/internal.h
index b90d516ad41f..a8e3168614d6 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -841,6 +841,8 @@ void try_to_unmap_flush(void);
void try_to_unmap_flush_dirty(void);
void flush_tlb_batched_pending(struct mm_struct *mm);
void fold_ubc_nowr(void);
+int nr_flush_required(void);
+int nr_flush_required_nowr(void);
#else
static inline void try_to_unmap_flush(void)
{
@@ -854,6 +856,14 @@ static inline void flush_tlb_batched_pending(struct mm_struct *mm)
static inline void fold_ubc_nowr(void)
{
}
+static inline int nr_flush_required(void)
+{
+ return 0;
+}
+static inline int nr_flush_required_nowr(void)
+{
+ return 0;
+}
#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
extern const struct trace_print_flags pageflag_names[];
diff --git a/mm/memory.c b/mm/memory.c
index f69fbc251198..061f23e34d69 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3345,6 +3345,12 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
+ if (vmf->page)
+ folio = page_folio(vmf->page);
+
+ if (folio && migrc_pending(folio))
+ migrc_try_flush();
+
/*
* Shared mapping: we are guaranteed to have VM_WRITE and
* FAULT_FLAG_WRITE set at this point.
@@ -3362,9 +3368,6 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
return wp_page_shared(vmf);
}
- if (vmf->page)
- folio = page_folio(vmf->page);
-
/*
* Private mapping: create an exclusive anonymous page copy if reuse
* is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
diff --git a/mm/migrate.c b/mm/migrate.c
index 01cac26a3127..944c7e179288 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -58,6 +58,244 @@
#include "internal.h"
+#ifdef CONFIG_MIGRC
+static int sysctl_migrc_enable = 1;
+#ifdef CONFIG_SYSCTL
+static int sysctl_migrc_enable_handler(struct ctl_table *table, int write,
+ void *buffer, size_t *lenp, loff_t *ppos)
+{
+ struct ctl_table t;
+ int err;
+ int enabled = sysctl_migrc_enable;
+
+ if (write && !capable(CAP_SYS_ADMIN))
+ return -EPERM;
+
+ t = *table;
+ t.data = &enabled;
+ err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
+ if (err < 0)
+ return err;
+ if (write)
+ sysctl_migrc_enable = enabled;
+ return err;
+}
+
+static struct ctl_table migrc_sysctls[] = {
+ {
+ .procname = "migrc_enable",
+ .data = NULL, /* filled in by handler */
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = sysctl_migrc_enable_handler,
+ .extra1 = SYSCTL_ZERO,
+ .extra2 = SYSCTL_ONE,
+ },
+ {}
+};
+
+static int __init migrc_sysctl_init(void)
+{
+ register_sysctl_init("vm", migrc_sysctls);
+ return 0;
+}
+late_initcall(migrc_sysctl_init);
+#endif
+
+/*
+ * TODO: Yeah, it's a non-sense magic number. This simple value manages
+ * to work conservatively anyway. However, the value needs to be
+ * tuned and adjusted based on the internal condition of memory
+ * management subsystem later.
+ *
+ * Let's start with a simple value for now.
+ */
+static const int migrc_pending_max = 512; /* unit: page */
+
+atomic_t migrc_gen;
+LLIST_HEAD(migrc_reqs);
+LLIST_HEAD(migrc_reqs_dirty);
+
+enum {
+ MIGRC_STATE_NONE,
+ MIGRC_SRC_PENDING,
+ MIGRC_DST_PENDING,
+};
+
+#define MAX_MIGRC_REQ_NR 4096
+static struct migrc_req migrc_req_pool_static[MAX_MIGRC_REQ_NR];
+static atomic_t migrc_req_pool_idx = ATOMIC_INIT(-1);
+static LLIST_HEAD(migrc_req_pool_llist);
+static DEFINE_SPINLOCK(migrc_req_pool_lock);
+
+static struct migrc_req *alloc_migrc_req(void)
+{
+ int idx = atomic_read(&migrc_req_pool_idx);
+ struct llist_node *n;
+
+ if (idx < MAX_MIGRC_REQ_NR - 1) {
+ idx = atomic_inc_return(&migrc_req_pool_idx);
+ if (idx < MAX_MIGRC_REQ_NR)
+ return migrc_req_pool_static + idx;
+ }
+
+ spin_lock(&migrc_req_pool_lock);
+ n = llist_del_first(&migrc_req_pool_llist);
+ spin_unlock(&migrc_req_pool_lock);
+
+ return n ? llist_entry(n, struct migrc_req, llnode) : NULL;
+}
+
+void free_migrc_req(struct migrc_req *req)
+{
+ llist_add(&req->llnode, &migrc_req_pool_llist);
+}
+
+static bool migrc_full(int nid)
+{
+ struct pglist_data *node = NODE_DATA(nid);
+
+ if (migrc_pending_max == -1)
+ return false;
+
+ return atomic_read(&node->migrc_pending_nr) >= migrc_pending_max;
+}
+
+void migrc_init_page(struct page *p)
+{
+ WRITE_ONCE(p->migrc_state, MIGRC_STATE_NONE);
+}
+
+/*
+ * The list should be isolated before.
+ */
+void migrc_shrink(struct llist_head *h)
+{
+ struct page *p;
+ struct llist_node *n;
+
+ n = llist_del_all(h);
+ llist_for_each_entry(p, n, migrc_node) {
+ if (p->migrc_state == MIGRC_SRC_PENDING) {
+ struct pglist_data *node;
+ struct zone *zone;
+
+ node = NODE_DATA(page_to_nid(p));
+ zone = page_zone(p);
+ atomic_dec(&node->migrc_pending_nr);
+ atomic_dec(&zone->migrc_pending_nr);
+ }
+ WRITE_ONCE(p->migrc_state, MIGRC_STATE_NONE);
+ folio_put(page_folio(p));
+ }
+}
+
+bool migrc_pending(struct folio *f)
+{
+ return READ_ONCE(f->page.migrc_state) != MIGRC_STATE_NONE;
+}
+
+static void migrc_expand_req(struct folio *fsrc, struct folio *fdst)
+{
+ struct migrc_req *req;
+ struct pglist_data *node;
+ struct zone *zone;
+
+ req = fold_ubc_nowr_migrc_req();
+ if (!req)
+ return;
+
+ folio_get(fsrc);
+ folio_get(fdst);
+ WRITE_ONCE(fsrc->page.migrc_state, MIGRC_SRC_PENDING);
+ WRITE_ONCE(fdst->page.migrc_state, MIGRC_DST_PENDING);
+
+ if (llist_add(&fsrc->page.migrc_node, &req->pages))
+ req->last = &fsrc->page.migrc_node;
+ llist_add(&fdst->page.migrc_node, &req->pages);
+
+ node = NODE_DATA(folio_nid(fsrc));
+ zone = page_zone(&fsrc->page);
+ atomic_inc(&node->migrc_pending_nr);
+ atomic_inc(&zone->migrc_pending_nr);
+
+ if (migrc_full(folio_nid(fsrc)))
+ migrc_try_flush();
+}
+
+void migrc_req_start(void)
+{
+ struct migrc_req *req;
+ struct migrc_req *req_dirty;
+
+ if (WARN_ON(current->mreq || current->mreq_dirty))
+ return;
+
+ req = alloc_migrc_req();
+ req_dirty = alloc_migrc_req();
+
+ if (!req || !req_dirty)
+ goto fail;
+
+ arch_tlbbatch_clean(&req->arch);
+ init_llist_head(&req->pages);
+ req->last = NULL;
+ current->mreq = req;
+
+ arch_tlbbatch_clean(&req_dirty->arch);
+ init_llist_head(&req_dirty->pages);
+ req_dirty->last = NULL;
+ current->mreq_dirty = req_dirty;
+ return;
+fail:
+ if (req_dirty)
+ free_migrc_req(req_dirty);
+ if (req)
+ free_migrc_req(req);
+}
+
+void migrc_req_end(void)
+{
+ struct migrc_req *req = current->mreq;
+ struct migrc_req *req_dirty = current->mreq_dirty;
+
+ WARN_ON((!req && req_dirty) || (req && !req_dirty));
+
+ if (!req || !req_dirty)
+ return;
+
+ if (llist_empty(&req->pages)) {
+ free_migrc_req(req);
+ } else {
+ req->gen = atomic_inc_return(&migrc_gen);
+ llist_add(&req->llnode, &migrc_reqs);
+ }
+ current->mreq = NULL;
+
+ if (llist_empty(&req_dirty->pages)) {
+ free_migrc_req(req_dirty);
+ } else {
+ req_dirty->gen = atomic_inc_return(&migrc_gen);
+ llist_add(&req_dirty->llnode, &migrc_reqs_dirty);
+ }
+ current->mreq_dirty = NULL;
+}
+
+bool migrc_req_processing(void)
+{
+ return current->mreq && current->mreq_dirty;
+}
+
+int migrc_pending_nr_in_zone(struct zone *z)
+{
+ return atomic_read(&z->migrc_pending_nr);
+}
+#else
+static const int sysctl_migrc_enable;
+static bool migrc_full(int nid) { return true; }
+static void migrc_expand_req(struct folio *fsrc, struct folio *fdst) {}
+#endif
+
bool isolate_movable_page(struct page *page, isolate_mode_t mode)
{
struct folio *folio = folio_get_nontail_page(page);
@@ -383,6 +621,9 @@ static int folio_expected_refs(struct address_space *mapping,
struct folio *folio)
{
int refs = 1;
+
+ refs += migrc_pending(folio) ? 1 : 0;
+
if (!mapping)
return refs;
@@ -1060,6 +1301,12 @@ static void migrate_folio_undo_src(struct folio *src,
bool locked,
struct list_head *ret)
{
+ /*
+ * TODO: There might be folios already pending for migrc.
+ * However, there's no way to cancel those on failure for now.
+ * Let's reflect the requirement when needed.
+ */
+
if (page_was_mapped)
remove_migration_ptes(src, src, false);
/* Drop an anon_vma reference if we took one */
@@ -1627,10 +1874,17 @@ static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
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));
+ migrc_cond1 = sysctl_migrc_enable &&
+ ((reason == MR_DEMOTION && current_is_kswapd()) ||
+ reason == MR_NUMA_MISPLACED);
+
+ if (migrc_cond1)
+ migrc_req_start();
for (pass = 0; pass < nr_pass && (retry || large_retry); pass++) {
retry = 0;
large_retry = 0;
@@ -1638,6 +1892,10 @@ static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
nr_retry_pages = 0;
list_for_each_entry_safe(folio, folio2, from, lru) {
+ int nr_required;
+ bool migrc_cond2;
+ bool migrc;
+
/*
* Large folio statistics is based on the source large
* folio. Capture required information that might get
@@ -1671,8 +1929,14 @@ static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
continue;
}
+ nr_required = nr_flush_required();
rc = migrate_folio_unmap(get_new_page, put_new_page, private,
folio, &dst, mode, reason, ret_folios);
+ migrc_cond2 = nr_required == nr_flush_required() &&
+ nr_flush_required_nowr() &&
+ !migrc_full(folio_nid(folio));
+ migrc = migrc_cond1 && migrc_cond2;
+
/*
* The rules are:
* Success: folio will be freed
@@ -1722,9 +1986,11 @@ static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
nr_large_failed += large_retry;
stats->nr_thp_failed += thp_retry;
rc_saved = rc;
- if (list_empty(&unmap_folios))
+ if (list_empty(&unmap_folios)) {
+ if (migrc_cond1)
+ migrc_req_end();
goto out;
- else
+ } else
goto move;
case -EAGAIN:
if (is_large) {
@@ -1742,6 +2008,13 @@ static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
case MIGRATEPAGE_UNMAP:
list_move_tail(&folio->lru, &unmap_folios);
list_add_tail(&dst->lru, &dst_folios);
+
+ if (migrc)
+ /*
+ * XXX: On migration failure,
+ * extra TLB flush might happen.
+ */
+ migrc_expand_req(folio, dst);
break;
default:
/*
@@ -1760,6 +2033,7 @@ static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
stats->nr_failed_pages += nr_pages;
break;
}
+ fold_ubc_nowr();
}
}
nr_failed += retry;
@@ -1767,6 +2041,15 @@ static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
stats->nr_thp_failed += thp_retry;
stats->nr_failed_pages += nr_retry_pages;
move:
+ /*
+ * Should be prior to try_to_unmap_flush() so that
+ * migrc_try_flush() that will be performed later based on the
+ * gen # assigned in migrc_req_end(), can take benefit of the
+ * TLB flushes in try_to_unmap_flush().
+ */
+ if (migrc_cond1)
+ migrc_req_end();
+
/* Flush TLBs for all unmapped folios */
try_to_unmap_flush();
diff --git a/mm/mm_init.c b/mm/mm_init.c
index 7f7f9c677854..87cbddc7d780 100644
--- a/mm/mm_init.c
+++ b/mm/mm_init.c
@@ -558,6 +558,7 @@ static void __meminit __init_single_page(struct page *page, unsigned long pfn,
page_mapcount_reset(page);
page_cpupid_reset_last(page);
page_kasan_tag_reset(page);
+ migrc_init_page(page);
INIT_LIST_HEAD(&page->lru);
#ifdef WANT_PAGE_VIRTUAL
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 47421bedc12b..167dadb0d817 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -3176,6 +3176,11 @@ bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
long min = mark;
int o;
+ /*
+ * There are pages that can be freed by migrc_try_flush().
+ */
+ 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);
@@ -4254,6 +4259,7 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
unsigned int zonelist_iter_cookie;
int reserve_flags;
+ migrc_try_flush();
restart:
compaction_retries = 0;
no_progress_loops = 0;
@@ -4769,6 +4775,16 @@ struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
if (likely(page))
goto out;
+ if (order && migrc_try_flush()) {
+ /*
+ * Try again after freeing migrc's pending pages in case
+ * of high order allocation.
+ */
+ page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
+ if (likely(page))
+ goto out;
+ }
+
alloc_gfp = gfp;
ac.spread_dirty_pages = false;
diff --git a/mm/rmap.c b/mm/rmap.c
index d18460a48485..5b251eb01cd4 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -606,6 +606,86 @@ struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
+#ifdef CONFIG_MIGRC
+static bool __migrc_try_flush(struct llist_head *h)
+{
+ struct arch_tlbflush_unmap_batch arch;
+ struct llist_node *reqs;
+ struct migrc_req *req;
+ struct migrc_req *req2;
+ LLIST_HEAD(pages);
+
+ reqs = llist_del_all(h);
+ if (!reqs)
+ return false;
+
+ arch_tlbbatch_clean(&arch);
+
+ /*
+ * TODO: Optimize the time complexity.
+ */
+ llist_for_each_entry_safe(req, req2, reqs, llnode) {
+ struct llist_node *n;
+
+ arch_migrc_adj(&req->arch, req->gen);
+ arch_tlbbatch_fold(&arch, &req->arch);
+
+ n = llist_del_all(&req->pages);
+ llist_add_batch(n, req->last, &pages);
+ free_migrc_req(req);
+ }
+
+ arch_tlbbatch_flush(&arch);
+ migrc_shrink(&pages);
+ return true;
+}
+
+bool migrc_try_flush(void)
+{
+ bool ret;
+
+ if (migrc_req_processing()) {
+ migrc_req_end();
+ migrc_req_start();
+ }
+ ret = __migrc_try_flush(&migrc_reqs);
+ ret = ret || __migrc_try_flush(&migrc_reqs_dirty);
+
+ return ret;
+}
+
+void migrc_try_flush_dirty(void)
+{
+ if (migrc_req_processing()) {
+ migrc_req_end();
+ migrc_req_start();
+ }
+ __migrc_try_flush(&migrc_reqs_dirty);
+}
+
+struct migrc_req *fold_ubc_nowr_migrc_req(void)
+{
+ struct tlbflush_unmap_batch *tlb_ubc_nowr = ¤t->tlb_ubc_nowr;
+ struct migrc_req *req;
+ bool dirty;
+
+ if (!tlb_ubc_nowr->nr_flush_required)
+ return NULL;
+
+ dirty = tlb_ubc_nowr->writable;
+ req = dirty ? current->mreq_dirty : current->mreq;
+ if (!req) {
+ fold_ubc_nowr();
+ return NULL;
+ }
+
+ arch_tlbbatch_fold(&req->arch, &tlb_ubc_nowr->arch);
+ tlb_ubc_nowr->nr_flush_required = 0;
+ tlb_ubc_nowr->writable = false;
+ return req;
+}
+#endif
+
void fold_ubc_nowr(void)
{
struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
@@ -621,6 +701,16 @@ void fold_ubc_nowr(void)
tlb_ubc_nowr->writable = false;
}
+int nr_flush_required(void)
+{
+ return current->tlb_ubc.nr_flush_required;
+}
+
+int nr_flush_required_nowr(void)
+{
+ return current->tlb_ubc_nowr.nr_flush_required;
+}
+
/*
* Flush TLB entries for recently unmapped pages from remote CPUs. It is
* important if a PTE was dirty when it was unmapped that it's flushed
@@ -648,6 +738,8 @@ void try_to_unmap_flush_dirty(void)
if (tlb_ubc->writable || tlb_ubc_nowr->writable)
try_to_unmap_flush();
+
+ migrc_try_flush_dirty();
}
/*
--
2.17.1
Functionally, no change. This is a preparation for CONFIG_MIGRC that
requires to recognize non-writable TLB entries and makes use of them to
batch more aggressively or even skip TLB flushes.
While at it, changed struct tlbflush_unmap's ->flush_required(boolean)
to ->nr_flush_required(int) in order to take account of not only whether
it has been requested or not, but also the exact number of the requests.
That will be used in CONFIG_MIGRC implementation too.
Signed-off-by: Byungchul Park <[email protected]>
---
arch/x86/include/asm/tlbflush.h | 2 ++
arch/x86/mm/tlb.c | 7 +++++++
include/linux/mm_types_task.h | 4 ++--
include/linux/sched.h | 1 +
mm/internal.h | 4 ++++
mm/rmap.c | 29 ++++++++++++++++++++++++-----
6 files changed, 40 insertions(+), 7 deletions(-)
diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
index 75bfaa421030..63504cde364b 100644
--- a/arch/x86/include/asm/tlbflush.h
+++ b/arch/x86/include/asm/tlbflush.h
@@ -279,6 +279,8 @@ static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
}
extern void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch);
+extern void arch_tlbbatch_fold(struct arch_tlbflush_unmap_batch *bdst,
+ struct arch_tlbflush_unmap_batch *bsrc);
static inline bool pte_flags_need_flush(unsigned long oldflags,
unsigned long newflags,
diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c
index 267acf27480a..69d145f1fff1 100644
--- a/arch/x86/mm/tlb.c
+++ b/arch/x86/mm/tlb.c
@@ -1265,6 +1265,13 @@ void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
put_cpu();
}
+void arch_tlbbatch_fold(struct arch_tlbflush_unmap_batch *bdst,
+ struct arch_tlbflush_unmap_batch *bsrc)
+{
+ cpumask_or(&bdst->cpumask, &bdst->cpumask, &bsrc->cpumask);
+ cpumask_clear(&bsrc->cpumask);
+}
+
/*
* Blindly accessing user memory from NMI context can be dangerous
* if we're in the middle of switching the current user task or
diff --git a/include/linux/mm_types_task.h b/include/linux/mm_types_task.h
index 5414b5c6a103..6f3bb757eb46 100644
--- a/include/linux/mm_types_task.h
+++ b/include/linux/mm_types_task.h
@@ -59,8 +59,8 @@ struct tlbflush_unmap_batch {
*/
struct arch_tlbflush_unmap_batch arch;
- /* True if a flush is needed. */
- bool flush_required;
+ /* The number of flush requested. */
+ int nr_flush_required;
/*
* If true then the PTE was dirty when unmapped. The entry must be
diff --git a/include/linux/sched.h b/include/linux/sched.h
index eed5d65b8d1f..2232b2cdfce8 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1322,6 +1322,7 @@ struct task_struct {
#endif
struct tlbflush_unmap_batch tlb_ubc;
+ struct tlbflush_unmap_batch tlb_ubc_nowr;
/* Cache last used pipe for splice(): */
struct pipe_inode_info *splice_pipe;
diff --git a/mm/internal.h b/mm/internal.h
index 68410c6d97ac..b90d516ad41f 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -840,6 +840,7 @@ extern struct workqueue_struct *mm_percpu_wq;
void try_to_unmap_flush(void);
void try_to_unmap_flush_dirty(void);
void flush_tlb_batched_pending(struct mm_struct *mm);
+void fold_ubc_nowr(void);
#else
static inline void try_to_unmap_flush(void)
{
@@ -850,6 +851,9 @@ static inline void try_to_unmap_flush_dirty(void)
static inline void flush_tlb_batched_pending(struct mm_struct *mm)
{
}
+static inline void fold_ubc_nowr(void)
+{
+}
#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
extern const struct trace_print_flags pageflag_names[];
diff --git a/mm/rmap.c b/mm/rmap.c
index 19392e090bec..d18460a48485 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -605,6 +605,22 @@ struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
}
#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
+
+void fold_ubc_nowr(void)
+{
+ struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
+ struct tlbflush_unmap_batch *tlb_ubc_nowr = ¤t->tlb_ubc_nowr;
+
+ if (!tlb_ubc_nowr->nr_flush_required)
+ return;
+
+ arch_tlbbatch_fold(&tlb_ubc->arch, &tlb_ubc_nowr->arch);
+ tlb_ubc->writable = tlb_ubc->writable || tlb_ubc_nowr->writable;
+ tlb_ubc->nr_flush_required += tlb_ubc_nowr->nr_flush_required;
+ tlb_ubc_nowr->nr_flush_required = 0;
+ tlb_ubc_nowr->writable = false;
+}
+
/*
* Flush TLB entries for recently unmapped pages from remote CPUs. It is
* important if a PTE was dirty when it was unmapped that it's flushed
@@ -615,11 +631,12 @@ void try_to_unmap_flush(void)
{
struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
- if (!tlb_ubc->flush_required)
+ fold_ubc_nowr();
+ if (!tlb_ubc->nr_flush_required)
return;
arch_tlbbatch_flush(&tlb_ubc->arch);
- tlb_ubc->flush_required = false;
+ tlb_ubc->nr_flush_required = 0;
tlb_ubc->writable = false;
}
@@ -627,8 +644,9 @@ void try_to_unmap_flush(void)
void try_to_unmap_flush_dirty(void)
{
struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
+ struct tlbflush_unmap_batch *tlb_ubc_nowr = ¤t->tlb_ubc_nowr;
- if (tlb_ubc->writable)
+ if (tlb_ubc->writable || tlb_ubc_nowr->writable)
try_to_unmap_flush();
}
@@ -644,15 +662,16 @@ void try_to_unmap_flush_dirty(void)
static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval)
{
- struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
+ struct tlbflush_unmap_batch *tlb_ubc;
int batch;
bool writable = pte_dirty(pteval);
if (!pte_accessible(mm, pteval))
return;
+ tlb_ubc = pte_write(pteval) ? ¤t->tlb_ubc : ¤t->tlb_ubc_nowr;
arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
- tlb_ubc->flush_required = true;
+ tlb_ubc->nr_flush_required += 1;
/*
* Ensure compiler does not re-order the setting of tlb_flush_batched
--
2.17.1
On 4 Aug 2023, at 2:18, 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.
So this would only reduce or eliminate TLB flushes? The same goal should
be achievable with batched TLB flush, right? You probably can group
to-be-migrated pages into a read-only group and a writable group and
migrate the read-only group first then the writable group. It would
reduce or eliminate the TLB flushes for the read-only group of pages, right?
>
> To achieve that:
>
> 1. For the folios that have only non-writable TLB entries, prevent
> TLB flush by keeping both source and destination of folios during
> migration, which will be handled later at a better time.
In this case, the page table points to the destination folio, but TLB
would cache the old translation pointing to the source folio. I wonder
if there would be any correctness issue.
>
> 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. TLB flushes can be skipped if all TLB flushes required to free the
> duplicated folios have been done by any reason, which doesn't have
> to be done from migrations.
>
> 4. Adjust watermark check routine, __zone_watermark_ok(), with the
> number of duplicated folios because those folios can be freed
> and obtained right away through appropreate TLB flushes.
>
> 5. Perform TLB flushes and free the duplicated folios pending the
> flushes if page allocation routine is in trouble due to memory
> pressure, even more aggresively for high order allocation.
>
> The measurement result:
>
> Architecture - x86_64
> QEMU - kvm enabled, host cpu, 2nodes((4cpus, 2GB)+(cpuless, 6GB))
> Linux Kernel - v6.4, numa balancing tiering on, demotion enabled
> Benchmark - XSBench with no parameter changed
>
> run 'perf stat' using events:
> (FYI, process wide result ~= system wide result(-a option))
> 1) itlb.itlb_flush
> 2) tlb_flush.dtlb_thread
> 3) tlb_flush.stlb_any
>
> run 'cat /proc/vmstat' and pick up:
> 1) pgdemote_kswapd
> 2) numa_pages_migrated
> 3) pgmigrate_success
> 4) nr_tlb_remote_flush
> 5) nr_tlb_remote_flush_received
> 6) nr_tlb_local_flush_all
> 7) nr_tlb_local_flush_one
>
> BEFORE - mainline v6.4
> ==========================================
>
> $ perf stat -e itlb.itlb_flush,tlb_flush.dtlb_thread,tlb_flush.stlb_any ./XSBench
>
> Performance counter stats for './XSBench':
>
> 426856 itlb.itlb_flush
> 6900414 tlb_flush.dtlb_thread
> 7303137 tlb_flush.stlb_any
>
> 33.500486566 seconds time elapsed
> 92.852128000 seconds user
> 10.526718000 seconds sys
>
> $ cat /proc/vmstat
>
> ...
> pgdemote_kswapd 1052596
> numa_pages_migrated 1052359
> pgmigrate_success 2161846
> nr_tlb_remote_flush 72370
> nr_tlb_remote_flush_received 213711
> nr_tlb_local_flush_all 3385
> nr_tlb_local_flush_one 198679
> ...
>
> AFTER - mainline v6.4 + CONFIG_MIGRC
> ==========================================
>
> $ perf stat -e itlb.itlb_flush,tlb_flush.dtlb_thread,tlb_flush.stlb_any ./XSBench
>
> Performance counter stats for './XSBench':
>
> 179537 itlb.itlb_flush
> 6131135 tlb_flush.dtlb_thread
> 6920979 tlb_flush.stlb_any
>
> 30.396700625 seconds time elapsed
> 80.331252000 seconds user
> 10.303761000 seconds sys
>
> $ cat /proc/vmstat
>
> ...
> pgdemote_kswapd 1044602
> numa_pages_migrated 1044202
> pgmigrate_success 2157808
> nr_tlb_remote_flush 30453
> nr_tlb_remote_flush_received 88840
> nr_tlb_local_flush_all 3039
> nr_tlb_local_flush_one 198875
> ...
>
> Signed-off-by: Byungchul Park <[email protected]>
> ---
--
Best Regards,
Yan, Zi
> On Aug 3, 2023, at 11:18 PM, Byungchul Park <[email protected]> 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 have only non-writable TLB entries, prevent
> TLB flush by keeping both source and destination of folios during
> migration, 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. TLB flushes can be skipped if all TLB flushes required to free the
> duplicated folios have been done by any reason, which doesn't have
> to be done from migrations.
>
> 4. Adjust watermark check routine, __zone_watermark_ok(), with the
> number of duplicated folios because those folios can be freed
> and obtained right away through appropreate TLB flushes.
>
> 5. Perform TLB flushes and free the duplicated folios pending the
> flushes if page allocation routine is in trouble due to memory
> pressure, even more aggresively for high order allocation.
So I think that what you want to do may be possible, but I think it worth
checking once an RFC that can be reviewed is posted. The complexity and
overheads would then need to be evaluated.
The patch in its current form, I am afraid, is very very hard to review.
It is way too big and is missing comments. Having CONFIG_MIGRC makes no
sense (I guess it is intended to be a “chicken-bit”). Variable and
function names are not informative. The memory barriers are handle
improperly (please check again the smp_mb__after_atomic() rules).
Actually, when it comes to concurrency, there are many things I did not
understand from a glance at the code when it comes to concurrency: the
use of llist_add when (I think?) the llist is not shared (I think?); the
use of WRITE_ONCE() for synchronization; migrc_gen scheme (and BTW, since
such a counter might overflow it should be atomic64).
But much more importantly, going up one level, there are several issues
that should be addressed/considered/discussed:
a. It seems to me that when a new PTE is established (e.g., following
an mmap()), and there are pending deferred flushes, a full TLB flush
would also be required. So your point (2) would need to be extended.
b. When a reference to the page is taken in other means (get_user_pages()),
a TLB flush might also be needed.
c. If we start deferring TLB flushes for a long time, and throughout that
time many events (TLB flush, page-faults, etc.) might require a *full*
TLB flush, that might have negative impact.
d. The interactions with other mechanisms that inspect the PTE to make
decisions and might not take into account the fact a TLB flush was not
done need to be considered. The interaction with mmu_gather has been
taken for, but there is a question of whether something here might
break it.
Now there are many things in the patch that need to be addressed and are
unacceptable in their current form (e.g., migrc_try_flush() flushing
potentially twice the same cores), but reviewing this patch in its
current form is too tedious for me.
[ BTW: for future versions, consider cc'ing Peter Zijlstra, Andy
Lutomirski and Dave Hansen. ]
On Fri, Aug 04, 2023 at 12:08:07PM -0400, Zi Yan wrote:
> On 4 Aug 2023, at 2:18, 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.
>
> So this would only reduce or eliminate TLB flushes? The same goal should
> be achievable with batched TLB flush, right? You probably can group
CONFIG_MIGRC achieves the reduction with:
1) Batch TLB flush,
2) defer TLB flush aggressively for non-writable folios.
The 2nd benifit is one that this patch set focuses on more.
> to-be-migrated pages into a read-only group and a writable group and
> migrate the read-only group first then the writable group. It would
> reduce or eliminate the TLB flushes for the read-only group of pages, right?
I'm not sure we are in the same page.
> > To achieve that:
> >
> > 1. For the folios that have only non-writable TLB entries, prevent
> > TLB flush by keeping both source and destination of folios during
> > migration, which will be handled later at a better time.
>
> In this case, the page table points to the destination folio, but TLB
> would cache the old translation pointing to the source folio. I wonder
> if there would be any correctness issue.
Kinda yes. That's why I keep the old source folio until guaranteed all
TLB flushes required will be done or special cases e.g. memory pressure
or the permission changes to writable through fault handler.
Worth noting that of course this works only with non-writable folios to
keep the consistency and correctness.
It'd be apprecisated to point out if I miss something.
Byungchul
> > 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. TLB flushes can be skipped if all TLB flushes required to free the
> > duplicated folios have been done by any reason, which doesn't have
> > to be done from migrations.
> >
> > 4. Adjust watermark check routine, __zone_watermark_ok(), with the
> > number of duplicated folios because those folios can be freed
> > and obtained right away through appropreate TLB flushes.
> >
> > 5. Perform TLB flushes and free the duplicated folios pending the
> > flushes if page allocation routine is in trouble due to memory
> > pressure, even more aggresively for high order allocation.
> >
> > The measurement result:
> >
> > Architecture - x86_64
> > QEMU - kvm enabled, host cpu, 2nodes((4cpus, 2GB)+(cpuless, 6GB))
> > Linux Kernel - v6.4, numa balancing tiering on, demotion enabled
> > Benchmark - XSBench with no parameter changed
> >
> > run 'perf stat' using events:
> > (FYI, process wide result ~= system wide result(-a option))
> > 1) itlb.itlb_flush
> > 2) tlb_flush.dtlb_thread
> > 3) tlb_flush.stlb_any
> >
> > run 'cat /proc/vmstat' and pick up:
> > 1) pgdemote_kswapd
> > 2) numa_pages_migrated
> > 3) pgmigrate_success
> > 4) nr_tlb_remote_flush
> > 5) nr_tlb_remote_flush_received
> > 6) nr_tlb_local_flush_all
> > 7) nr_tlb_local_flush_one
> >
> > BEFORE - mainline v6.4
> > ==========================================
> >
> > $ perf stat -e itlb.itlb_flush,tlb_flush.dtlb_thread,tlb_flush.stlb_any ./XSBench
> >
> > Performance counter stats for './XSBench':
> >
> > 426856 itlb.itlb_flush
> > 6900414 tlb_flush.dtlb_thread
> > 7303137 tlb_flush.stlb_any
> >
> > 33.500486566 seconds time elapsed
> > 92.852128000 seconds user
> > 10.526718000 seconds sys
> >
> > $ cat /proc/vmstat
> >
> > ...
> > pgdemote_kswapd 1052596
> > numa_pages_migrated 1052359
> > pgmigrate_success 2161846
> > nr_tlb_remote_flush 72370
> > nr_tlb_remote_flush_received 213711
> > nr_tlb_local_flush_all 3385
> > nr_tlb_local_flush_one 198679
> > ...
> >
> > AFTER - mainline v6.4 + CONFIG_MIGRC
> > ==========================================
> >
> > $ perf stat -e itlb.itlb_flush,tlb_flush.dtlb_thread,tlb_flush.stlb_any ./XSBench
> >
> > Performance counter stats for './XSBench':
> >
> > 179537 itlb.itlb_flush
> > 6131135 tlb_flush.dtlb_thread
> > 6920979 tlb_flush.stlb_any
> >
> > 30.396700625 seconds time elapsed
> > 80.331252000 seconds user
> > 10.303761000 seconds sys
> >
> > $ cat /proc/vmstat
> >
> > ...
> > pgdemote_kswapd 1044602
> > numa_pages_migrated 1044202
> > pgmigrate_success 2157808
> > nr_tlb_remote_flush 30453
> > nr_tlb_remote_flush_received 88840
> > nr_tlb_local_flush_all 3039
> > nr_tlb_local_flush_one 198875
> > ...
> >
> > Signed-off-by: Byungchul Park <[email protected]>
> > ---
>
> --
> Best Regards,
> Yan, Zi
On Fri, Aug 04, 2023 at 05:32:30PM +0000, Nadav Amit wrote:
> > On Aug 3, 2023, at 11:18 PM, Byungchul Park <[email protected]> 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 have only non-writable TLB entries, prevent
> > TLB flush by keeping both source and destination of folios during
> > migration, 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. TLB flushes can be skipped if all TLB flushes required to free the
> > duplicated folios have been done by any reason, which doesn't have
> > to be done from migrations.
> >
> > 4. Adjust watermark check routine, __zone_watermark_ok(), with the
> > number of duplicated folios because those folios can be freed
> > and obtained right away through appropreate TLB flushes.
> >
> > 5. Perform TLB flushes and free the duplicated folios pending the
> > flushes if page allocation routine is in trouble due to memory
> > pressure, even more aggresively for high order allocation.
>
> So I think that what you want to do may be possible, but I think it worth
> checking once an RFC that can be reviewed is posted. The complexity and
> overheads would then need to be evaluated.
>
> The patch in its current form, I am afraid, is very very hard to review.
> It is way too big and is missing comments. Having CONFIG_MIGRC makes no
Sorry for that. I will split this patch set into more pieces and try to
add sufficient comments, and then will repost it in the next spin.
> sense (I guess it is intended to be a “chicken-bit”). Variable and
Exactly.
> function names are not informative. The memory barriers are handle
Yeah, 'naming' is the hardest one to do. Lemme try to rename those.
> improperly (please check again the smp_mb__after_atomic() rules).
Thank you. I meant to put a smp_mb() along with atomic_read() after
that, between reading migrc_gen and TLB flush. I will check it more.
> Actually, when it comes to concurrency, there are many things I did not
> understand from a glance at the code when it comes to concurrency: the
> use of llist_add when (I think?) the llist is not shared (I think?); the
A llist isolated for handling TLB flush and freeing folios is a stack
variable so it's not shared. However, there is another type of llist
that is a global llist so as to be shared, that is for collecting
all the requests in the system.
> use of WRITE_ONCE() for synchronization; migrc_gen scheme (and BTW, since
> such a counter might overflow it should be atomic64).
Sure. It would overflow but it's not a big problem because we can
compare between the generation numbers with '(int)(a - b) < 0' trick
unless more than the MAX_INT/2 numbers of requests happen at the same
time. I assumed it's barely gonna happen.
> But much more importantly, going up one level, there are several issues
> that should be addressed/considered/discussed:
>
> a. It seems to me that when a new PTE is established (e.g., following
> an mmap()), and there are pending deferred flushes, a full TLB flush
> would also be required. So your point (2) would need to be extended.
It has nothing to do with tiering migration. So I think It would work as
it was with the original code, say, it would perform TLB flush needed
for the PTE change. I don't think CONFIG_MIGRC needs to consider the
case. Could you explain what would be a probelm in this case in more
detail?
> b. When a reference to the page is taken in other means (get_user_pages()),
> a TLB flush might also be needed.
All TLB flush would be performed as it was, except TLB flushes at
tiering migration where CONFIG_MIGRC works. I might miss things.. Please
explain in more detail what you think is a problem.
> c. If we start deferring TLB flushes for a long time, and throughout that
> time many events (TLB flush, page-faults, etc.) might require a *full*
> TLB flush, that might have negative impact.
I let it work as it was, except tiering migration. It'd help me get you
if you describe in more detail. I might need to make this patch set more
readable first tho. However, I'd like to understand what you are
concerning exactly so that I can answer like either 'you don't have to
worry about that because blur blur' or 'Oh my I will fix it thank you'.
> d. The interactions with other mechanisms that inspect the PTE to make
> decisions and might not take into account the fact a TLB flush was not
> done need to be considered. The interaction with mmu_gather has been
> taken for, but there is a question of whether something here might
> break it.
I also think I need to be more careful when it comes to mmu_gather
things. My opinion on this is, it'd be okay becasue CONFIG_MIGRC only
works with tiering migration and let other pathes where TLB flush is
required go as it was. Howerver, it definitely should be considered if
those routines that require TLB flush during tiering migration. Am I
missing something?
> Now there are many things in the patch that need to be addressed and are
> unacceptable in their current form (e.g., migrc_try_flush() flushing
> potentially twice the same cores), but reviewing this patch in its
No. It doens't flush twice cuz it keeps whether each CPU has been TLB
flushed that was requested, or not ;)
> current form is too tedious for me.
I will reform it.
> [ BTW: for future versions, consider cc'ing Peter Zijlstra, Andy
> Lutomirski and Dave Hansen. ]
I will.
Appreciate all your comments!
Byungchul
On Fri, Aug 04, 2023 at 05:32:30PM +0000, Nadav Amit wrote:
> The patch in its current form, I am afraid, is very very hard to review.
> It is way too big and is missing comments. Having CONFIG_MIGRC makes no
> sense (I guess it is intended to be a “chicken-bit”). Variable and
> function names are not informative. The memory barriers are handle
> improperly (please check again the smp_mb__after_atomic() rules).
I checked it. I found what I was wrong at. Thank you for pointing it out.
I will fix it from the next spin.
Byungchul
On Tue, Aug 15, 2023 at 09:27:26AM +0800, Huang, Ying wrote:
> Byungchul Park <[email protected]> writes:
>
> > 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 have only non-writable TLB entries, prevent
> > TLB flush by keeping both source and destination of folios during
> > migration, 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. TLB flushes can be skipped if all TLB flushes required to free the
> > duplicated folios have been done by any reason, which doesn't have
> > to be done from migrations.
> >
> > 4. Adjust watermark check routine, __zone_watermark_ok(), with the
> > number of duplicated folios because those folios can be freed
> > and obtained right away through appropreate TLB flushes.
> >
> > 5. Perform TLB flushes and free the duplicated folios pending the
> > flushes if page allocation routine is in trouble due to memory
> > pressure, even more aggresively for high order allocation.
>
> Is the optimization restricted for page migration only? Can it be used
> for other places? Like page reclaiming?
Just to make sure, are you talking about the (5) description? For now,
it's performed at the beginning of __alloc_pages_slowpath(), say, before
page recaiming. Do you think it'd be meaningful to perform it during page
reclaiming? Or do you mean something else?
> > The measurement result:
> >
> > Architecture - x86_64
> > QEMU - kvm enabled, host cpu, 2nodes((4cpus, 2GB)+(cpuless, 6GB))
> > Linux Kernel - v6.4, numa balancing tiering on, demotion enabled
> > Benchmark - XSBench with no parameter changed
> >
> > run 'perf stat' using events:
> > (FYI, process wide result ~= system wide result(-a option))
> > 1) itlb.itlb_flush
> > 2) tlb_flush.dtlb_thread
> > 3) tlb_flush.stlb_any
> >
> > run 'cat /proc/vmstat' and pick up:
> > 1) pgdemote_kswapd
> > 2) numa_pages_migrated
> > 3) pgmigrate_success
> > 4) nr_tlb_remote_flush
> > 5) nr_tlb_remote_flush_received
> > 6) nr_tlb_local_flush_all
> > 7) nr_tlb_local_flush_one
> >
> > BEFORE - mainline v6.4
> > ==========================================
> >
> > $ perf stat -e itlb.itlb_flush,tlb_flush.dtlb_thread,tlb_flush.stlb_any ./XSBench
> >
> > Performance counter stats for './XSBench':
> >
> > 426856 itlb.itlb_flush
> > 6900414 tlb_flush.dtlb_thread
> > 7303137 tlb_flush.stlb_any
> >
> > 33.500486566 seconds time elapsed
> > 92.852128000 seconds user
> > 10.526718000 seconds sys
> >
> > $ cat /proc/vmstat
> >
> > ...
> > pgdemote_kswapd 1052596
> > numa_pages_migrated 1052359
> > pgmigrate_success 2161846
> > nr_tlb_remote_flush 72370
> > nr_tlb_remote_flush_received 213711
> > nr_tlb_local_flush_all 3385
> > nr_tlb_local_flush_one 198679
> > ...
> >
> > AFTER - mainline v6.4 + CONFIG_MIGRC
> > ==========================================
> >
> > $ perf stat -e itlb.itlb_flush,tlb_flush.dtlb_thread,tlb_flush.stlb_any ./XSBench
> >
> > Performance counter stats for './XSBench':
> >
> > 179537 itlb.itlb_flush
> > 6131135 tlb_flush.dtlb_thread
> > 6920979 tlb_flush.stlb_any
>
> It appears that the number of "itlb.itlb_flush" changes much, but not
> for other 2 events. Because the text segment of the executable file is
> mapped as read-only? And most other pages are mapped read-write?
Yes, for this benchmarch, XSBench. I didn't noticed that until checking
it using perf event either.
> > 30.396700625 seconds time elapsed
> > 80.331252000 seconds user
> > 10.303761000 seconds sys
> >
> > $ cat /proc/vmstat
> >
> > ...
> > pgdemote_kswapd 1044602
> > numa_pages_migrated 1044202
> > pgmigrate_success 2157808
> > nr_tlb_remote_flush 30453
> > nr_tlb_remote_flush_received 88840
> > nr_tlb_local_flush_all 3039
> > nr_tlb_local_flush_one 198875
> > ...
> >
> > Signed-off-by: Byungchul Park <[email protected]>
[...]
> > diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> > index 306a3d1a0fa6..3be66d3eabd2 100644
> > --- a/include/linux/mm_types.h
> > +++ b/include/linux/mm_types.h
> > @@ -228,6 +228,10 @@ struct page {
> > #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
> > int _last_cpupid;
> > #endif
> > +#ifdef CONFIG_MIGRC
> > + struct llist_node migrc_node;
> > + unsigned int migrc_state;
> > +#endif
>
> We cannot enlarge "struct page".
This is what I worried about. Do you have a better idea? I don't think
they fit onto page_ext or something.
Byungchul
On Wed, Aug 16, 2023 at 09:01:12AM +0800, Huang, Ying wrote:
> Byungchul Park <[email protected]> writes:
>
> > On Tue, Aug 15, 2023 at 09:27:26AM +0800, Huang, Ying wrote:
> >> Byungchul Park <[email protected]> writes:
> >>
> >> > 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 have only non-writable TLB entries, prevent
> >> > TLB flush by keeping both source and destination of folios during
> >> > migration, 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. TLB flushes can be skipped if all TLB flushes required to free the
> >> > duplicated folios have been done by any reason, which doesn't have
> >> > to be done from migrations.
> >> >
> >> > 4. Adjust watermark check routine, __zone_watermark_ok(), with the
> >> > number of duplicated folios because those folios can be freed
> >> > and obtained right away through appropreate TLB flushes.
> >> >
> >> > 5. Perform TLB flushes and free the duplicated folios pending the
> >> > flushes if page allocation routine is in trouble due to memory
> >> > pressure, even more aggresively for high order allocation.
> >>
> >> Is the optimization restricted for page migration only? Can it be used
> >> for other places? Like page reclaiming?
> >
> > Just to make sure, are you talking about the (5) description? For now,
> > it's performed at the beginning of __alloc_pages_slowpath(), say, before
> > page recaiming. Do you think it'd be meaningful to perform it during page
> > reclaiming? Or do you mean something else?
>
> Not for (5). TLB needs to be flushed during page reclaiming too. Can
> similar method be used to reduce TLB flushing there too?
Hm.. The mechanism can be used in any places where page mapping is
changing but it requires not to have write permission that might mess up
consistancy with more than one copy of page.
JFYI, one of future works is to detect read mostly pages and turn them
to read only to make use of them iff it gives a better performance.
Byungchul
On Wed, Aug 16, 2023 at 09:01:12AM +0800, Huang, Ying wrote:
> Byungchul Park <[email protected]> writes:
>
> > On Tue, Aug 15, 2023 at 09:27:26AM +0800, Huang, Ying wrote:
> >> Byungchul Park <[email protected]> writes:
> >>
> >> > 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 have only non-writable TLB entries, prevent
> >> > TLB flush by keeping both source and destination of folios during
> >> > migration, 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. TLB flushes can be skipped if all TLB flushes required to free the
> >> > duplicated folios have been done by any reason, which doesn't have
> >> > to be done from migrations.
> >> >
> >> > 4. Adjust watermark check routine, __zone_watermark_ok(), with the
> >> > number of duplicated folios because those folios can be freed
> >> > and obtained right away through appropreate TLB flushes.
> >> >
> >> > 5. Perform TLB flushes and free the duplicated folios pending the
> >> > flushes if page allocation routine is in trouble due to memory
> >> > pressure, even more aggresively for high order allocation.
> >>
> >> Is the optimization restricted for page migration only? Can it be used
> >> for other places? Like page reclaiming?
> >
> > Just to make sure, are you talking about the (5) description? For now,
> > it's performed at the beginning of __alloc_pages_slowpath(), say, before
> > page recaiming. Do you think it'd be meaningful to perform it during page
> > reclaiming? Or do you mean something else?
>
> Not for (5). TLB needs to be flushed during page reclaiming too. Can
> similar method be used to reduce TLB flushing there too?
If you were talking about unmapping for swap while reclaiming, then I
think yes. The case can also take benefit from CONFIG_MIGRC.
Byungchul