RE: [PATCH 3/3] COLO: Optimize memory back-up process

[Zhanghailiang <zhang.zhanghailiang@huawei.com>]

Hi Dave,

> -----Original Message-----
> From: Dr. David Alan Gilbert [mailto:dgilbert@redhat.com]
> Sent: Friday, February 21, 2020 2:25 AM
> To: Zhanghailiang <zhang.zhanghailiang@huawei.com>
> Cc: qemu-devel@nongnu.org; quintela@redhat.com; chen.zhang@intel.com;
> danielcho@qnap.com
> Subject: Re: [PATCH 3/3] COLO: Optimize memory back-up process
> 
> * Hailiang Zhang (zhang.zhanghailiang@huawei.com) wrote:
> > This patch will reduce the downtime of VM for the initial process,
> > Privously, we copied all these memory in preparing stage of COLO
> > while we need to stop VM, which is a time-consuming process.
> > Here we optimize it by a trick, back-up every page while in migration
> > process while COLO is enabled, though it affects the speed of the
> > migration, but it obviously reduce the downtime of back-up all SVM'S
> > memory in COLO preparing stage.
> >
> > Signed-off-by: Hailiang Zhang <zhang.zhanghailiang@huawei.com>
> 
> OK, I think this is right, but it took me quite a while to understand,
> I think one of the comments below might not be right:
> 

> > ---
> >  migration/colo.c |  3 +++
> >  migration/ram.c  | 35 +++++++++++++++++++++++++++--------
> >  migration/ram.h  |  1 +
> >  3 files changed, 31 insertions(+), 8 deletions(-)
> >
> > diff --git a/migration/colo.c b/migration/colo.c
> > index d30c6bc4ad..febf010571 100644
> > --- a/migration/colo.c
> > +++ b/migration/colo.c
> > @@ -26,6 +26,7 @@
> >  #include "qemu/main-loop.h"
> >  #include "qemu/rcu.h"
> >  #include "migration/failover.h"
> > +#include "migration/ram.h"
> >  #ifdef CONFIG_REPLICATION
> >  #include "replication.h"
> >  #endif
> > @@ -906,6 +907,8 @@ void *colo_process_incoming_thread(void
> *opaque)
> >       */
> >      qemu_file_set_blocking(mis->from_src_file, true);
> >
> > +    colo_incoming_start_dirty_log();
> > +
> >      bioc = qio_channel_buffer_new(COLO_BUFFER_BASE_SIZE);
> >      fb = qemu_fopen_channel_input(QIO_CHANNEL(bioc));
> >      object_unref(OBJECT(bioc));
> > diff --git a/migration/ram.c b/migration/ram.c
> > index ed23ed1c7c..24a8aa3527 100644
> > --- a/migration/ram.c
> > +++ b/migration/ram.c
> > @@ -2986,7 +2986,6 @@ int colo_init_ram_cache(void)
> >                  }
> >                  return -errno;
> >              }
> > -            memcpy(block->colo_cache, block->host,
> block->used_length);
> >          }
> >      }
> >
> > @@ -3005,12 +3004,16 @@ int colo_init_ram_cache(void)
> >              bitmap_set(block->bmap, 0, pages);
> >          }
> >      }
> > +
> > +    return 0;
> > +}
> > +
> > +void colo_incoming_start_dirty_log(void)
> > +{
> >      ram_state = g_new0(RAMState, 1);
> >      ram_state->migration_dirty_pages = 0;
> >      qemu_mutex_init(&ram_state->bitmap_mutex);
> >      memory_global_dirty_log_start();
> > -
> > -    return 0;
> >  }
> >
> >  /* It is need to hold the global lock to call this helper */
> > @@ -3348,7 +3351,7 @@ static int ram_load_precopy(QEMUFile *f)
> >
> >      while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) {
> >          ram_addr_t addr, total_ram_bytes;
> > -        void *host = NULL;
> > +        void *host = NULL, *host_bak = NULL;
> >          uint8_t ch;
> >
> >          /*
> > @@ -3378,13 +3381,26 @@ static int ram_load_precopy(QEMUFile *f)
> >          if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE |
> >                       RAM_SAVE_FLAG_COMPRESS_PAGE |
> RAM_SAVE_FLAG_XBZRLE)) {
> >              RAMBlock *block = ram_block_from_stream(f, flags);
> > -
> >              /*
> > -             * After going into COLO, we should load the Page into
> colo_cache.
> > +             * After going into COLO, we should load the Page into
> colo_cache
> > +             * NOTE: We need to keep a copy of SVM's ram in
> colo_cache.
> > +             * Privously, we copied all these memory in preparing stage
> of COLO
> > +             * while we need to stop VM, which is a time-consuming
> process.
> > +             * Here we optimize it by a trick, back-up every page while
> in
> > +             * migration process while COLO is enabled, though it
> affects the
> > +             * speed of the migration, but it obviously reduce the
> downtime of
> > +             * back-up all SVM'S memory in COLO preparing stage.
> >               */
> > -            if (migration_incoming_in_colo_state()) {
> > +            if (migration_incoming_colo_enabled()) {
> >                  host = colo_cache_from_block_offset(block, addr);
> > -            } else {
> > +                /*
> > +                 * After going into COLO, load the Page into
> colo_cache.
> > +                 */
> > +                if (!migration_incoming_in_colo_state()) {
> > +                    host_bak = host;
> > +                }
> > +            }
> > +            if (!migration_incoming_in_colo_state()) {
> >                  host = host_from_ram_block_offset(block, addr);
> 
> So this works out as quite complicated:
>    a) In normal migration we do the last one and just set:
>          host = host_from_ram_block_offset(block, addr);
>          host_bak = NULL
> 
>    b) At the start, when colo_enabled, but !in_colo_state
>          host = colo_cache
>          host_bak = host
>          host = host_from_ram_block_offset
> 
>    c) in_colo_state
>          host = colo_cache
>          host_bak = NULL
> 
> 
> (b) is pretty confusing, setting host twice; can't we tidy that up?
> 
> Also, that last comment 'After going into COLO' I think is really
>   'Before COLO state, copy from ram into cache'
> 

The code logic here is not good, I have fixed this in V2 like this :)

+            host = host_from_ram_block_offset(block, addr);
             /*
-             * After going into COLO, we should load the Page into colo_cache.
+             * After going into COLO stage, we should not load the page
+             * into SVM's memory diretly, we put them into colo_cache firstly.
+             * NOTE: We need to keep a copy of SVM's ram in colo_cache.
+             * Privously, we copied all these memory in preparing stage of COLO
+             * while we need to stop VM, which is a time-consuming process.
+             * Here we optimize it by a trick, back-up every page while in
+             * migration process while COLO is enabled, though it affects the
+             * speed of the migration, but it obviously reduce the downtime of
+             * back-up all SVM'S memory in COLO preparing stage.
              */
-            if (migration_incoming_in_colo_state()) {
-                host = colo_cache_from_block_offset(block, addr);
-            } else {
-                host = host_from_ram_block_offset(block, addr);
+            if (migration_incoming_colo_enabled()) {
+                if (migration_incoming_in_colo_state()) {
+                    /* In COLO stage, put all pages into cache temporarily */
+                    host = colo_cache_from_block_offset(block, addr);
+                } else {
+                   /*
+                    * In migration stage but before COLO stage,
+                    * Put all pages into both cache and SVM's memory.
+                    */
+                    host_bak = colo_cache_from_block_offset(block, addr);
+                }


Thanks,
Hailiang

> Dave
> 
> >              }
> >              if (!host) {
> > @@ -3506,6 +3522,9 @@ static int ram_load_precopy(QEMUFile *f)
> >          if (!ret) {
> >              ret = qemu_file_get_error(f);
> >          }
> > +        if (!ret && host_bak && host) {
> > +            memcpy(host_bak, host, TARGET_PAGE_SIZE);
> > +        }
> >      }
> >
> >      ret |= wait_for_decompress_done();
> > diff --git a/migration/ram.h b/migration/ram.h
> > index a553d40751..5ceaff7cb4 100644
> > --- a/migration/ram.h
> > +++ b/migration/ram.h
> > @@ -66,5 +66,6 @@ int ram_dirty_bitmap_reload(MigrationState *s,
> RAMBlock *rb);
> >  /* ram cache */
> >  int colo_init_ram_cache(void);
> >  void colo_release_ram_cache(void);
> > +void colo_incoming_start_dirty_log(void);
> >
> >  #endif
> > --
> > 2.21.0
> >
> >
> --
> Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK