support dirtyrate at the granualrity of vcpu
cover support dirtyrate at the granualrity of vcpu
From: Hyman Huang(黄勇) <[email protected]>
Since the Dirty Ring on QEMU part has been merged recently, how to use this feature is under consideration.
In the scene of migration, it is valuable to provide a more accurante interface to track dirty memory than existing one, so that the upper layer application can make a wise decision, or whatever. More importantly, dirtyrate info at the granualrity of vcpu could provide a possibility to make migration convergent by imposing restriction on vcpu. With Dirty Ring, we can calculate dirtyrate efficiently and cheaply.
The old interface implemented by sampling pages, it consumes cpu resource, and the larger guest memory size become, the more cpu resource it consumes, namely, hard to scale. New interface has no such drawback.
Please review, thanks !
Best Regards !
Hyman Huang(黄勇) (6):
KVM: add kvm_dirty_ring_enabled function
KVM: introduce dirty_pages into CPUState migration/dirtyrate: add vcpu option for qmp calc-dirty-rate migration/dirtyrate: adjust struct DirtyRateStat migration/dirtyrate: check support of calculation for vcpu migration/dirtyrate: implement dirty-ring dirtyrate calculation
accel/kvm/kvm-all.c | 11 +++ include/hw/core/cpu.h | 2 + include/sysemu/kvm.h | 1 + migration/dirtyrate.c | 179 +++++++++++++++++++++++++++++++++++++---- migration/dirtyrate.h | 19 ++++- migration/trace-events | 1 + qapi/migration.json | 28 ++++++- 7 files changed, 222 insertions(+), 19 deletions(-)
Signature · 1 line
2.24.3
PATCH 1/6 KVM: add kvm_dirty_ring_enabled function
From: Hyman Huang(黄勇) <[email protected]>
introduce kvm_dirty_ring_enabled to show if kvm-reaper is working. dirtyrate thread could use it to check if calculation can base on dirty ring feature.
--- accel/kvm/kvm-all.c | 5 +++++ include/sysemu/kvm.h | 1 + 2 files changed, 6 insertions(+)
accel/kvm/kvm-all.c +5 -0
@@ -2293,6 +2293,11 @@ bool kvm_vcpu_id_is_valid(int vcpu_id)return vcpu_id >=0 && vcpu_id <kvm_max_vcpu_id (s); } +bool kvm_dirty_ring_enabled (void ) +{ +return kvm_state->kvm_dirty_ring_size ? true : false; +} +static int kvm_init (MachineState *ms) {MachineClass *mc =MACHINE_GET_CLASS (ms);
include/sysemu/kvm.h +1 -0
@@ -547,4 +547,5 @@ bool kvm_cpu_check_are_resettable(void);bool kvm_arch_cpu_check_are_resettable (void ); +bool kvm_dirty_ring_enabled (void );#endif
Signature · 1 line
2.24.3
PATCH 2/6 KVM: introduce dirty_pages into CPUState
In reply to [email protected]From: Hyman Huang(黄勇) <[email protected]>
dirty_pages is used to calculate dirtyrate via dirty ring, when enabled, kvm-reaper will increase the dirty pages after gfns being dirtied.
--- accel/kvm/kvm-all.c | 6 ++++++ include/hw/core/cpu.h | 2 ++ 2 files changed, 8 insertions(+)
accel/kvm/kvm-all.c +6 -0
@@ -506,6 +506,9 @@ int kvm_init_vcpu(CPUState *cpu, Error **errp) } } + cpu->dirty_pages =0 ; + cpu->stat_dirty_pages = false; + ret =kvm_arch_init_vcpu (cpu);if (ret <0 ) { error_setg_errno(errp, -ret,
@@ -739,6 +742,9 @@ static uint32_t kvm_dirty_ring_reap_one(KVMState *s, CPUState *cpu) cur->offset);dirty_gfn_set_collected (cur);trace_kvm_dirty_ring_page (cpu->cpu_index, fetch, cur->offset); +if (cpu->stat_dirty_pages) { + cpu->dirty_pages++; + } fetch++; count++; }
include/hw/core/cpu.h +2 -0
@@ -375,6 +375,8 @@ struct CPUState {struct kvm_run *kvm_run;struct kvm_dirty_gfn *kvm_dirty_gfns;uint32_t kvm_fetch_index; +uint64_t dirty_pages; +bool stat_dirty_pages;/* Used for events with 'vcpu' and *without* the 'disabled' properties */ DECLARE_BITMAP (trace_dstate_delayed,CPU_TRACE_DSTATE_MAX_EVENTS );
Signature · 1 line
2.24.3
Re: [PATCH v1 2/6] KVM: introduce dirty_pages into CPUState
include/hw/core/cpu.h struct CPUState { · line 379
Quoted 10 lines · On Tue, Jun 01, 2021 at 01:04:06AM +0800, [email protected] wrote:
include/hw/core/cpu.h +2
-0@@ -375,6 +375,8 @@ struct CPUState {struct kvm_run *kvm_run;struct kvm_dirty_gfn *kvm_dirty_gfns;uint32_t kvm_fetch_index; +uint64_t dirty_pages; +bool stat_dirty_pages;
Shall we make this bool a global one? As I don't think we'll be able to only enable it on a subset of cpus?
Signature · 1 line
Peter Xu
Re: [PATCH v1 2/6] KVM: introduce dirty_pages into CPUState
Quoted 14 lines · Peter Xu
include/hw/core/cpu.h struct CPUState { · line 379
Quoted 10 lines · On Tue, Jun 01, 2021 at 01:04:06AM +0800, [email protected] wrote:
include/hw/core/cpu.h +2
-0@@ -375,6 +375,8 @@ struct CPUState {struct kvm_run *kvm_run;struct kvm_dirty_gfn *kvm_dirty_gfns;uint32_t kvm_fetch_index; +uint64_t dirty_pages; +bool stat_dirty_pages;Shall we make this bool a global one? As I don't think we'll be able to only enable it on a subset of cpus?
Yes, it's a reasonable advice, i'll apply this on the next version
Quoted 1 line
Signature · 3 lines
Best regard Hyman Huang(黄勇)
Re: [PATCH v1 2/6] KVM: introduce dirty_pages into CPUState
Quoted 15 lines · Hyman Huang
Quoted 14 lines · Peter Xu
include/hw/core/cpu.h struct CPUState { · line 379
Quoted 10 lines · On Tue, Jun 01, 2021 at 01:04:06AM +0800, [email protected] wrote:
include/hw/core/cpu.h +2
-0@@ -375,6 +375,8 @@ struct CPUState {struct kvm_run *kvm_run;struct kvm_dirty_gfn *kvm_dirty_gfns;uint32_t kvm_fetch_index; +uint64_t dirty_pages; +bool stat_dirty_pages;Shall we make this bool a global one? As I don't think we'll be able to only enable it on a subset of cpus?
Quoted 1 line
Yes, it's a reasonable advice, i'll apply this on the next version
Or even drop the bool and do the accounting unconditionally? No need to do +1 for each pfn, but do it once before returning from kvm_dirty_ring_reap_one().
Signature · 1 line
Peter Xu
PATCH 3/6 migration/dirtyrate: add vcpu option for qmp calc-dirty-rate
In reply to [email protected]From: Hyman Huang(黄勇) <[email protected]>
calculate dirtyrate for each vcpu if vcpu is true, add the dirtyrate of each vcpu to the return value also.
--- migration/dirtyrate.c | 5 ++++- migration/dirtyrate.h | 1 + qapi/migration.json | 28 ++++++++++++++++++++++++++-- 3 files changed, 31 insertions(+), 3 deletions(-)
migration/dirtyrate.c +4 -1
@@ -383,7 +383,10 @@ void *get_dirtyrate_thread(void *arg)return NULL ; } -void qmp_calc_dirty_rate (int64_t calc_time,Error **errp) +void qmp_calc_dirty_rate (int64_t calc_time, +bool has_vcpu, +bool vcpu, +Error **errp) {static struct DirtyRateConfig config;QemuThread thread;
migration/dirtyrate.h +1 -0
@@ -38,6 +38,7 @@struct DirtyRateConfig {uint64_t sample_pages_per_gigabytes;/* sample pages per GB */ int64_t sample_period_seconds;/* time duration between two sampling */ +bool vcpu;/* calculate dirtyrate for each vcpu using dirty ring */ };/*
qapi/migration.json +26 -2
@@ -1708,6 +1708,21 @@ { 'event': 'UNPLUG_PRIMARY', 'data': { 'device-id': 'str' } } +## +# @DirtyRateVcpu: +# +# Dirty rate of vcpu. +# +# @id: vcpu index. +# +# @dirty-rate: dirty rate. +# +# Since: 6.1 +# +## +{ 'struct': 'DirtyRateVcpu', + 'data': { 'id': 'int', 'dirty-rate': 'int64' } } + ## # @DirtyRateStatus: #
@@ -1740,6 +1755,10 @@ # # @calc-time: time in units of second for sample dirty pages # +# @vcpu: calculate dirtyrate for each vcpu (Since 6.1) +# +# @vcpu-dirty-rate: dirtyrate for each vcpu (Since 6.1) +# # Since: 5.2 # ##
@@ -1747,7 +1766,9 @@ 'data': {'*dirty-rate': 'int64', 'status': 'DirtyRateStatus', 'start-time': 'int64', - 'calc-time': 'int64'} } + 'calc-time': 'int64', + '*vcpu': 'bool', + '*vcpu-dirty-rate': [ 'DirtyRateVcpu' ] } } ## # @calc-dirty-rate:
@@ -1756,13 +1777,16 @@ # # @calc-time: time in units of second for sample dirty pages # +# @vcpu: calculate vcpu dirty rate if true, the default value is +# false (since 6.1) +# # Since: 5.2 # # Example: # {"command": "calc-dirty-rate", "data": {"calc-time": 1} } # ## -{ 'command': 'calc-dirty-rate', 'data': {'calc-time': 'int64'} } +{ 'command': 'calc-dirty-rate', 'data': {'calc-time': 'int64', '*vcpu': 'bool'} } ## # @query-dirty-rate:
Signature · 1 line
2.24.3
PATCH 4/6 migration/dirtyrate: adjust struct DirtyRateStat
In reply to [email protected]From: Hyman Huang(黄勇) <[email protected]>
use union to store stat data of two mutual exclusive methods.
--- migration/dirtyrate.c | 32 ++++++++++++++++++++------------ migration/dirtyrate.h | 18 +++++++++++++++--- 2 files changed, 35 insertions(+), 15 deletions(-)
migration/dirtyrate.c +20 -12
@@ -78,31 +78,39 @@ static struct DirtyRateInfo *query_dirty_rate_info(void)return info; } -static void init_dirtyrate_stat (int64_t start_time,int64_t calc_time) +static void init_dirtyrate_stat (int64_t start_time, +int64_t calc_time, +struct DirtyRateConfig config) { - DirtyStat.total_dirty_samples =0 ; - DirtyStat.total_sample_count =0 ; - DirtyStat.total_block_mem_MB =0 ; DirtyStat.dirty_rate =-1 ; DirtyStat.start_time = start_time; DirtyStat.calc_time = calc_time; + +if (config.vcpu) { + DirtyStat.method.vcpu.nvcpu =-1 ; + DirtyStat.method.vcpu.rates =NULL ; + }else { + DirtyStat.method.vm.total_dirty_samples =0 ; + DirtyStat.method.vm.total_sample_count =0 ; + DirtyStat.method.vm.total_block_mem_MB =0 ; + } }static void update_dirtyrate_stat (struct RamblockDirtyInfo *info) { - DirtyStat.total_dirty_samples += info->sample_dirty_count; - DirtyStat.total_sample_count += info->sample_pages_count; + DirtyStat.method.vm.total_dirty_samples += info->sample_dirty_count; + DirtyStat.method.vm.total_sample_count += info->sample_pages_count;/* size of total pages in MB */ - DirtyStat.total_block_mem_MB += (info->ramblock_pages * + DirtyStat.method.vm.total_block_mem_MB += (info->ramblock_pages *TARGET_PAGE_SIZE ) >>20 ; }static void update_dirtyrate (uint64_t msec) {uint64_t dirtyrate; -uint64_t total_dirty_samples = DirtyStat.total_dirty_samples; -uint64_t total_sample_count = DirtyStat.total_sample_count; -uint64_t total_block_mem_MB = DirtyStat.total_block_mem_MB; +uint64_t total_dirty_samples = DirtyStat.method.vm.total_dirty_samples; +uint64_t total_sample_count = DirtyStat.method.vm.total_sample_count; +uint64_t total_block_mem_MB = DirtyStat.method.vm.total_block_mem_MB; dirtyrate = total_dirty_samples * total_block_mem_MB *1000 / (total_sample_count * msec);
@@ -315,7 +323,7 @@ static bool compare_page_hash_info(struct RamblockDirtyInfo *info,update_dirtyrate_stat (block_dinfo); } -if (DirtyStat.total_sample_count ==0 ) { +if (DirtyStat.method.vm.total_sample_count ==0 ) {return false; }
@@ -371,7 +379,7 @@ void *get_dirtyrate_thread(void *arg) start_time =qemu_clock_get_ms (QEMU_CLOCK_REALTIME ) /1000 ; calc_time = config.sample_period_seconds; -init_dirtyrate_stat (start_time, calc_time); +init_dirtyrate_stat (start_time, calc_time, config);calculate_dirtyrate (config);
migration/dirtyrate.h +15 -3
@@ -54,16 +54,28 @@ struct RamblockDirtyInfo {uint32_t *hash_result;/* array of hash result for sampled pages */ }; +typedef struct SampleVMStat { +uint64_t total_dirty_samples;/* total dirty sampled page */ +uint64_t total_sample_count;/* total sampled pages */ +uint64_t total_block_mem_MB;/* size of total sampled pages in MB */ +}SampleVMStat ; + +typedef struct VcpuStat { +int nvcpu;/* number of vcpu */ +DirtyRateVcpu *rates;/* array of dirty rate for each vcpu */ +}VcpuStat ; +/* * Store calculation statistics for each measure. */ struct DirtyRateStat { -uint64_t total_dirty_samples;/* total dirty sampled page */ -uint64_t total_sample_count;/* total sampled pages */ -uint64_t total_block_mem_MB;/* size of total sampled pages in MB */ int64_t dirty_rate;/* dirty rate in MB/s */ int64_t start_time;/* calculation start time in units of second */ int64_t calc_time;/* time duration of two sampling in units of second */ +union { +SampleVMStat vm; +VcpuStat vcpu; + } method; };void *get_dirtyrate_thread (void *arg);
Signature · 1 line
2.24.3
PATCH 5/6 migration/dirtyrate: check support of calculation for vcpu
In reply to [email protected]From: Hyman Huang(黄勇) <[email protected]>
vcpu method only works when kvm dirty ring is enabled, use kvm_dirty_ring_enabled to probe if dirty ring is enabled.
--- migration/dirtyrate.c | 10 ++++++++++ 1 file changed, 10 insertions(+)
migration/dirtyrate.c +10 -0
@@ -16,6 +16,7 @@#include "cpu.h" #include "exec/ramblock.h" #include "qemu/rcu_queue.h" +#include "sysemu/kvm.h" #include "qapi/qapi-commands-migration.h" #include "ram.h" #include "trace.h"
@@ -415,6 +416,14 @@ void qmp_calc_dirty_rate(int64_t calc_time,return ; } +/* +* Vcpu method only works when kvm dirty ring is enabled. +*/ +if (has_vcpu && vcpu && !kvm_dirty_ring_enabled ()) { +error_setg (errp,"kvm dirty ring is disabled, use sample method." ); +return ; + } +/* * Init calculation state as unstarted. */
@@ -427,6 +436,7 @@ void qmp_calc_dirty_rate(int64_t calc_time, config.sample_period_seconds = calc_time; config.sample_pages_per_gigabytes =DIRTYRATE_DEFAULT_SAMPLE_PAGES ; + config.vcpu = has_vcpu ? vcpu : false;qemu_thread_create (&thread,"get_dirtyrate" , get_dirtyrate_thread, (void *)&config,QEMU_THREAD_DETACHED ); }
Signature · 1 line
2.24.3
PATCH 6/6 migration/dirtyrate: implement dirty-ring dirtyrate calculation
In reply to [email protected]From: Hyman Huang(黄勇) <[email protected]>
use dirty ring feature to implement dirtyrate calculation. to enable it, set vcpu option as true in qmp calc-dirty-rate.
--- migration/dirtyrate.c | 146 ++++++++++++++++++++++++++++++++++++++--- migration/trace-events | 1 + 2 files changed, 139 insertions(+), 8 deletions(-)
migration/dirtyrate.c +138 -8
@@ -16,14 +16,22 @@#include "cpu.h" #include "exec/ramblock.h" #include "qemu/rcu_queue.h" +#include "qemu/main-loop.h" #include "sysemu/kvm.h" #include "qapi/qapi-commands-migration.h" #include "ram.h" #include "trace.h" #include "dirtyrate.h" +typedef enum { +CALC_NONE =0 , +CALC_DIRTY_RING , +CALC_SAMPLE_PAGES , +}CalcMethod ; +static int CalculatingState =DIRTY_RATE_STATUS_UNSTARTED ;static struct DirtyRateStat DirtyStat; +static CalcMethod last_method =CALC_NONE ;static int64_t set_sample_page_period (int64_t msec,int64_t initial_time) {
@@ -64,6 +72,7 @@ static struct DirtyRateInfo *query_dirty_rate_info(void) {int64_t dirty_rate = DirtyStat.dirty_rate;struct DirtyRateInfo *info =g_malloc0 (sizeof (DirtyRateInfo)); +DirtyRateVcpuList *head =NULL , **tail = &head;if (qatomic_read (&CalculatingState) ==DIRTY_RATE_STATUS_MEASURED ) { info->has_dirty_rate = true;
@@ -73,6 +82,22 @@ static struct DirtyRateInfo *query_dirty_rate_info(void) info->status = CalculatingState; info->start_time = DirtyStat.start_time; info->calc_time = DirtyStat.calc_time; + info->has_vcpu = true; + +if (last_method ==CALC_DIRTY_RING ) { +int i =0 ; + info->vcpu = true; + info->has_vcpu_dirty_rate = true; +for (i =0 ; i < DirtyStat.method.vcpu.nvcpu; i++) { +DirtyRateVcpu *rate =g_malloc0 (sizeof (DirtyRateVcpu)); + rate->id = DirtyStat.method.vcpu.rates[i].id; + rate->dirty_rate = DirtyStat.method.vcpu.rates[i].dirty_rate; +QAPI_LIST_APPEND (tail, rate); + } + info->vcpu_dirty_rate = head; + }else { + info->vcpu = false; + }trace_query_dirty_rate_info (DirtyRateStatus_str (CalculatingState));
@@ -87,13 +112,29 @@ static void init_dirtyrate_stat(int64_t start_time, DirtyStat.start_time = start_time; DirtyStat.calc_time = calc_time; -if (config.vcpu) { - DirtyStat.method.vcpu.nvcpu =-1 ; - DirtyStat.method.vcpu.rates =NULL ; - }else { - DirtyStat.method.vm.total_dirty_samples =0 ; - DirtyStat.method.vm.total_sample_count =0 ; - DirtyStat.method.vm.total_block_mem_MB =0 ; +switch (last_method) { +case CALC_NONE : +case CALC_SAMPLE_PAGES : +if (config.vcpu) { + DirtyStat.method.vcpu.nvcpu =-1 ; + DirtyStat.method.vcpu.rates =NULL ; + }else { + DirtyStat.method.vm.total_dirty_samples =0 ; + DirtyStat.method.vm.total_sample_count =0 ; + DirtyStat.method.vm.total_block_mem_MB =0 ; + } +break ; +case CALC_DIRTY_RING : +if (!config.vcpu) { +g_free (DirtyStat.method.vcpu.rates); + DirtyStat.method.vcpu.rates =NULL ; + DirtyStat.method.vm.total_dirty_samples =0 ; + DirtyStat.method.vm.total_sample_count =0 ; + DirtyStat.method.vm.total_block_mem_MB =0 ; + } +break ; +default : +break ; } }
@@ -331,7 +372,84 @@ static bool compare_page_hash_info(struct RamblockDirtyInfo *info,return true; } -static void calculate_dirtyrate (struct DirtyRateConfig config) +static void stat_dirtypages (CPUState *cpu,bool start) +{ + cpu->stat_dirty_pages = start; +} + +static void start_kvm_dirty_log (void ) +{ +qemu_mutex_lock_iothread (); +memory_global_dirty_log_start (); +qemu_mutex_unlock_iothread (); +} + +static void stop_kvm_dirty_log (void ) +{ +qemu_mutex_lock_iothread (); +memory_global_dirty_log_stop (); +qemu_mutex_unlock_iothread (); +} + +static int64_t do_calculate_dirtyrate_vcpu (CPUState *cpu) +{ +uint64_t memory_size_MB; +int64_t time_s; + + memory_size_MB = (cpu->dirty_pages *TARGET_PAGE_SIZE ) >>20 ; + time_s = DirtyStat.calc_time; + +return memory_size_MB / time_s; +} + +static void calculate_dirtyrate_vcpu (struct DirtyRateConfig config) +{ +CPUState *cpu; +int64_t msec =0 ; +int64_t start_time; +uint64_t dirtyrate =0 ; +uint64_t dirtyrate_sum =0 ; +int nvcpu, i =0 ; + +CPU_FOREACH (cpu) { +stat_dirtypages (cpu, true); + nvcpu++; + } + + DirtyStat.method.vcpu.nvcpu = nvcpu; + +if (last_method !=CALC_DIRTY_RING ) { + DirtyStat.method.vcpu.rates = +g_malloc0 (sizeof (DirtyRateVcpu) * nvcpu); + } + + start_time =qemu_clock_get_ms (QEMU_CLOCK_REALTIME ); + DirtyStat.start_time = start_time /1000 ; + +start_kvm_dirty_log (); + + msec = config.sample_period_seconds *1000 ; + msec =set_sample_page_period (msec, start_time); + DirtyStat.calc_time = msec /1000 ; + +CPU_FOREACH (cpu) { +stat_dirtypages (cpu, false); + } + +stop_kvm_dirty_log (); + +CPU_FOREACH (cpu) { + dirtyrate =do_calculate_dirtyrate_vcpu (cpu); + DirtyStat.method.vcpu.rates[i].id = cpu->cpu_index; + DirtyStat.method.vcpu.rates[i].dirty_rate = dirtyrate; + dirtyrate_sum += dirtyrate; + i++; + } + + DirtyStat.dirty_rate = dirtyrate_sum / nvcpu; +} + +static void calculate_dirtyrate_sample_vm (struct DirtyRateConfig config) {struct RamblockDirtyInfo *block_dinfo =NULL ;int block_count =0 ;
@@ -364,6 +482,18 @@ out:rcu_unregister_thread (); } +static void calculate_dirtyrate (struct DirtyRateConfig config) +{ +if (config.vcpu) { +calculate_dirtyrate_vcpu (config); + last_method =CALC_DIRTY_RING ; + }else { +calculate_dirtyrate_sample_vm (config); + last_method =CALC_SAMPLE_PAGES ; + } +trace_calculate_dirtyrate (DirtyStat.dirty_rate); +} +void *get_dirtyrate_thread (void *arg) {struct DirtyRateConfig config = *(struct DirtyRateConfig *)arg;
migration/trace-events +1 -0
@@ -330,6 +330,7 @@ get_ramblock_vfn_hash(const char *idstr, uint64_t vfn, uint32_t crc) "ramblock n calc_page_dirty_rate(const char *idstr, uint32_t new_crc, uint32_t old_crc) "ramblock name: %s, new crc: %" PRIu32 ", old crc: %" PRIu32 skip_sample_ramblock(const char *idstr, uint64_t ramblock_size) "ramblock name: %s, ramblock size: %" PRIu64 find_page_matched(const char *idstr) "ramblock %s addr or size changed" +calculate_dirtyrate(int64_t dirtyrate) "dirty rate: %" PRIi64 # block.c migration_block_init_shared(const char *blk_device_name) "Start migration for %s with shared base image"
Signature · 1 line
2.24.3
Quoted 12 lines · On Tue, Jun 01, 2021 at 01:02:45AM +0800, [email protected] wrote:
From: Hyman Huang(黄勇) <[email protected]>
Since the Dirty Ring on QEMU part has been merged recently, how to use this feature is under consideration.
In the scene of migration, it is valuable to provide a more accurante interface to track dirty memory than existing one, so that the upper layer application can make a wise decision, or whatever. More importantly, dirtyrate info at the granualrity of vcpu could provide a possibility to make migration convergent by imposing restriction on vcpu. With Dirty Ring, we can calculate dirtyrate efficiently and cheaply.
Quoted 3 lines
The old interface implemented by sampling pages, it consumes cpu resource, and the larger guest memory size become, the more cpu resource it consumes, namely, hard to scale. New interface has no such drawback.
Yong,
Thanks for working on this!
Some high-level comments:
- The layout of the patch looks a bit odd. E.g., you introduced the new "vcpu" qmp parameter in patch 3, however it's not yet implemented, meanwhile I feel like you squashed mostly all the rest into patch 6. It's okay to use a single big patch, but IMHO better to not declare that flag in QMP before it's working, so ideally that should be the last patch to do that.
From that POV: patch 1/2/4 look ok to be separated; perhaps squash patch 3/5/6 into one single patch to enable the new method as the last one?
- You used "vcpu" across the patchset to show the per-vcpu new method. Shall we rename it globally to "per_vcpu" or "vcpu_based"? A raw "vcpu" looks more like a struct pointer not a boolean.
- Using memory_global_dirty_log_start|stop() may not be wise too IMHO, at least we need to make sure it's not during migration, otherwise we could call the stop() before migration ends then that'll be a problem..
Maybe we can start to make global_dirty_log a bitmask? Then we define:
GLOBAL_DIRTY_MIGRATION
GLOBAL_DIRTY_DIRTY_RATE
All references to global_dirty_log should mostly be untouched because any bit set there should justify that global dirty logging is enabled (either for migration or for dirty rate measurement).
Migration starting half-way of dirty rate measurement seems okay too even taking things like init-all-set into account, afaict.. as long as dirty rate code never touches the qemu dirty bitmap, but only do the accounting when collecting the pages...
Feel free to think more about it on any other potential conflict with migration, but in general seems working to me.
- Would you consider picking up my HMP patch and let HMP work from the 1st day?
- Please Cc the author of dirty rate too (Chuan Zheng <[email protected]>), while I already started to do so in this email.
Thanks,
Signature · 1 line
Peter Xu
Quoted 23 lines · Peter Xu
Quoted 12 lines · On Tue, Jun 01, 2021 at 01:02:45AM +0800, [email protected] wrote:
From: Hyman Huang(黄勇) <[email protected]>
Since the Dirty Ring on QEMU part has been merged recently, how to use this feature is under consideration.
In the scene of migration, it is valuable to provide a more accurante interface to track dirty memory than existing one, so that the upper layer application can make a wise decision, or whatever. More importantly, dirtyrate info at the granualrity of vcpu could provide a possibility to make migration convergent by imposing restriction on vcpu. With Dirty Ring, we can calculate dirtyrate efficiently and cheaply.
Quoted 3 lines
The old interface implemented by sampling pages, it consumes cpu resource, and the larger guest memory size become, the more cpu resource it consumes, namely, hard to scale. New interface has no such drawback.
Yong,
Thanks for working on this!
Some high-level comments:
Quoted 9 lines
- The layout of the patch looks a bit odd. E.g., you introduced the new "vcpu" qmp parameter in patch 3, however it's not yet implemented, meanwhile I feel like you squashed mostly all the rest into patch 6. It's okay to use a single big patch, but IMHO better to not declare that flag in QMP before it's working, so ideally that should be the last patch to do that.
From that POV: patch 1/2/4 look ok to be separated; perhaps squash patch 3/5/6 into one single patch to enable the new method as the last one?
Yeah previously the concern is make the patch clear and small, however with the comment of each commit, it seems ok. As you said, it's okay to use a single big patch, i'll adjust the patch set style base on your advice.
Quoted 4 lines
- You used "vcpu" across the patchset to show the per-vcpu new method. Shall we rename it globally to "
per_vcpu" or "vcpu_based"? A raw "vcpu" looks more like a struct pointer not a boolean.
Indeed, actually the initial name of the option is "per_vcpu". : ). i'll fix this.
Quoted 3 lines
- Using memory_global_dirty_log_start|stop() may not be wise too IMHO, at least we need to make sure it's not during migration, otherwise we could call the
stop()before migration ends then that'll be a problem..
Yeah, this may be a serious problem, thanks for your timely advice.
Quoted 10 lines
Maybe we can start to make
global_dirty_loga bitmask? Then we define:
GLOBAL_DIRTY_MIGRATION
GLOBAL_DIRTY_DIRTY_RATEAll references to
global_dirty_logshould mostly be untouched because any bit set there should justify that global dirty logging is enabled (either for migration or for dirty rate measurement).
Quoted 8 lines
Migration starting half-way of dirty rate measurement seems okay too even taking things like init-all-set into account, afaict.. as long as dirty rate code never touches the qemu dirty bitmap, but only do the accounting when collecting the pages...
Feel free to think more about it on any other potential conflict with migration, but in general seems working to me.
I'll apply this on the next version.
Quoted 5 lines
- Would you consider picking up my HMP patch and let HMP work from the 1st day?
- Please Cc the author of dirty rate too (Chuan Zheng <[email protected]>), while I already started to do so in this email.
I'd be glad to do this above two.
Quoted 2 lines
Thanks,
Thanks Peter!
Signature · 3 lines
Best regard Hyman Huang(黄勇)