Migration: Make more ram_counters atomic
v5 cover Migration: Make more ram_counters atomic
Hi
In this v5:
Not only change the type of the counters, also use the __nocheck() variants of the functions.
Please, review.
[v4]
- Change aligned_uint64_t to size_t to make (some) 32bit hosts happy.
Please review.
[v3]
- Addressed reviews
- All counters are now atomic, either Stat64 or atomic.
- Rename duplicated to zero_pages
- Rename normal to zero_pages.
Please review.
[v2]
- fix typos found by David Edmondson
- Add review-by tags.
Please review.
[v1]
On previous series we cerate ram_atomic_counters. But we basically need that all counters are atomic. So move back to only have ram_counters, just with a new type that allows the atomic counters.
Once there, move update of stats out of RAM mutex. And make multifd_bytes atomic.
Later, Juan.
Juan Quintela (2):
migration: Make dirty_pages_rate atomic
migration: Make dirty_bytes_last_sync atomic
migration/migration.c | 9 ++++++--- migration/ram.c | 7 ++++--- migration/ram.h | 4 ++-- 3 files changed, 12 insertions(+), 8 deletions(-)
Signature · 1 line
2.40.0
PATCH v5 1/2 migration: Make dirty_pages_rate atomic
In this case we use qatomic operations instead of Stat64 wrapper because there is no stat64_set(). Defining the 64 bit wrapper is trivial. The one without atomics is more interesting.
---
Don't use __nocheck() variants
--- migration/migration.c | 6 ++++-- migration/ram.c | 5 +++-- migration/ram.h | 2 +- 3 files changed, 8 insertions(+), 5 deletions(-)
migration/migration.c +4 -2
@@ -1005,7 +1005,8 @@ static void populate_ram_info(MigrationInfo *info, MigrationState *s)if (s->state !=MIGRATION_STATUS_COMPLETED ) { info->ram->remaining =ram_bytes_remaining (); - info->ram->dirty_pages_rate = ram_counters.dirty_pages_rate; + info->ram->dirty_pages_rate = +qatomic_read (&ram_counters.dirty_pages_rate); } }
@@ -2751,7 +2752,8 @@ static void migration_update_counters(MigrationState *s,* if we haven't sent anything, we don't want to * recalculate. 10000 is a small enough number for our purposes */ -if (ram_counters.dirty_pages_rate && transferred >10000 ) { +if (qatomic_read (&ram_counters.dirty_pages_rate) && + transferred >10000 ) { s->expected_downtime = ram_counters.remaining / bandwidth; }
migration/ram.c +3 -2
@@ -1129,8 +1129,9 @@ static void migration_update_rates(RAMState *rs, int64_t end_time)double compressed_size;/* calculate period counters */ - ram_counters.dirty_pages_rate = rs->num_dirty_pages_period *1000 - / (end_time - rs->time_last_bitmap_sync); +qatomic_set (&ram_counters.dirty_pages_rate, + rs->num_dirty_pages_period *1000 / + (end_time - rs->time_last_bitmap_sync));if (!page_count) {return ;
migration/ram.h +1 -1
@@ -41,7 +41,7 @@* one thread). */ typedef struct { - int64_t dirty_pages_rate; + aligned_uint64_t dirty_pages_rate; Stat64 dirty_sync_count; Stat64 dirty_sync_missed_zero_copy; Stat64 downtime_bytes;
Signature · 1 line
2.40.0
Re: [PATCH v5 1/2] migration: Make dirty_pages_rate atomic
Quoted 3 lines · Juan Quintela
In this case we use qatomic operations instead of Stat64 wrapper because there is no
stat64_set(). Defining the 64 bit wrapper is trivial. The one without atomics is more interesting.
This does not work if CONFIG_ATOMIC64 is not defined.
I actually have stat64_set in a patchset that I still have to submit, so you can reuse it:
include/qemu/stats64.h +6 -0
@@ -40,6 +40,11 @@ static inline uint64_t stat64_get(const Stat64 *s)return qatomic_read__nocheck (&s->value); } +static inline void stat64_set (Stat64 *s,uint64_t value) +{ +qatomic_set__nocheck (&s->value, value); +} +static inline void stat64_add (Stat64 *s,uint64_t value) {qatomic_add (&s->value, value);
@@ -62,6 +67,7 @@ static inline void stat64_max(Stat64 *s, }#else uint64_t stat64_get (const Stat64 *s); +void stat64_set (Stat64 *s,uint64_t value);bool stat64_min_slow (Stat64 *s,uint64_t value);bool stat64_max_slow (Stat64 *s,uint64_t value);bool stat64_add32_carry (Stat64 *s,uint32_t low,uint32_t high);
util/stats64.c +11 -0
@@ -57,6 +57,17 @@ uint64_t stat64_get(const Stat64 *s)return ((uint64_t )high <<32 ) | low; } +void stat64_set (Stat64 *s,uint64_t val) +{ +while (!stat64_wrtrylock (s)) { +cpu_relax (); + } + +qatomic_set (&s->high, val >>32 ); +qatomic_set (&s->low, val); +stat64_wrunlock (s); +} +bool stat64_add32_carry (Stat64 *s,uint32_t low,uint32_t high) {uint32_t old;
Paolo
PATCH v5 2/2 migration: Make dirty_bytes_last_sync atomic
In reply to Juan QuintelaAs we set its value, it needs to be operated with atomics. We rename it from remaining to better reflect its meaning.
Statistics always return the real reamaining bytes. This was used to store how much pages where dirty on the previous generation, so we can calculate the expected downtime as: dirty_bytes_last_sync / current_bandwith.
If we use the actual remaining bytes, we would see a very small value at the end of the iteration.
---
I am open to use ram_bytes_remaining() in its only use and be more "optimistic" about the downtime.
Don't use __nocheck() functions.
--- migration/migration.c | 3 ++- migration/ram.c | 2 +- migration/ram.h | 2 +- 3 files changed, 4 insertions(+), 3 deletions(-)
migration/migration.c +2 -1
@@ -2754,7 +2754,8 @@ static void migration_update_counters(MigrationState *s,*/ if (qatomic_read (&ram_counters.dirty_pages_rate) && transferred >10000 ) { - s->expected_downtime = ram_counters.remaining / bandwidth; + s->expected_downtime = +qatomic_read (&ram_counters.dirty_bytes_last_sync) / bandwidth; }qemu_file_reset_rate_limit (s->to_dst_file);
migration/ram.c +1 -1
@@ -1223,7 +1223,7 @@ static void migration_bitmap_sync(RAMState *rs)RAMBLOCK_FOREACH_NOT_IGNORED (block) {ramblock_sync_dirty_bitmap (rs, block); } - ram_counters.remaining =ram_bytes_remaining (); +qatomic_set (&ram_counters.dirty_bytes_last_sync,ram_bytes_remaining ()); }qemu_mutex_unlock (&rs->bitmap_mutex);
migration/ram.h +1 -1
@@ -41,6 +41,7 @@* one thread). */ typedef struct { + aligned_uint64_t dirty_bytes_last_sync; aligned_uint64_t dirty_pages_rate; Stat64 dirty_sync_count; Stat64 dirty_sync_missed_zero_copy;
@@ -51,7 +52,6 @@ typedef struct {Stat64 postcopy_bytes;Stat64 postcopy_requests;Stat64 precopy_bytes; -int64_t remaining;Stat64 transferred; } RAMStats;
Signature · 1 line
2.40.0
Re: [PATCH v5 2/2] migration: Make dirty_bytes_last_sync atomic
Quoted 1 line · Juan Quintela
Don't use
__nocheck()functions.
Doesn't this break on 32-bit platforms?
#if defined(__x86_64__) || defined(__sparc__) || defined(__mips64)
# define ATOMIC_REG_SIZE 8 #else # define ATOMIC_REG_SIZE sizeof(void *) #endif
#define qatomic_set(ptr, i) do { \
qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \
qatomic_set__nocheck(ptr, i); \
} while(0)So if sizeof(void*) == 4 it would trigger a compile-time assertion.
Paolo
Re: [PATCH v5 2/2] migration: Make dirty_bytes_last_sync atomic
Quoted 11 lines · Paolo Bonzini
Quoted 1 line · Juan Quintela
Don't use
__nocheck()functions.Doesn't this break on 32-bit platforms?
#if
defined(__x86_64__)||defined(__sparc__)||defined(__mips64)# define ATOMIC_REG_SIZE 8 #else # define ATOMIC_REG_SIZE sizeof(void *) #endif
Quoted 8 lines
#define qatomic_set(ptr, i) do { \ qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \ qatomic_set__nocheck(ptr, i); \ } while(0)So if
sizeof(void*)== 4 it would trigger a compile-time assertion.Paolo
Yeap.
Really I was waiting for stat64_set() that you have just sent, will do it on top of that.
And yes, I already broke the build with that patch.
Later, Juan.
Juan Quintela <[email protected]> wrote:
self-NACK
Working on top of paolo stat64_set() function.
Quoted 46 lines
Hi
In this v5:
Not only change the type of the counters, also use the
__nocheck()variants of the functions.Please, review.
[v4]
- Changealigned_uint64_ttosize_tto make (some) 32bit hosts happy.Please review.
[v3]
- Addressed reviews
- All counters are now atomic, either Stat64 or atomic.
- Rename duplicated tozero_pages
- Rename normal tozero_pages.Please review.
[v2]
- fix typos found by David Edmondson
- Add review-by tags.Please review.
[v1]
On previous series we cerateram_atomic_counters. But we basically need that all counters are atomic. So move back to only haveram_counters, just with a new type that allows the atomic counters.Once there, move update of stats out of RAM mutex. And make
multifd_bytesatomic.Later, Juan.
Juan Quintela (2):
migration: Makedirty_pages_rateatomic
migration: Makedirty_bytes_last_syncatomicmigration/migration.c | 9 ++++++--- migration/ram.c | 7 ++++--- migration/ram.h | 4 ++-- 3 files changed, 12 insertions(+), 8 deletions(-)