[PATCH 0/3] migration: Fixes to the 'background-snapshot' code

[PATCH 0/3] migration: Fixes to the 'background-snapshot' code

[Andrey Gruzdev]

This patch series contains:
 * Fix to the issue with occasionally truncated non-iterable device state
 * Solution to compatibility issues with virtio-balloon device
 * Fix to the issue when discarded or never populated pages miss UFFD
   write protection and get into migration stream in dirty state

Andrey Gruzdev (3):
  migration: Fix missing qemu_fflush() on buffer file in
    bg_migration_thread
  migration: Inhibit virtio-balloon for the duration of background
    snapshot
  migration: Pre-fault memory before starting background snasphot

 hw/virtio/virtio-balloon.c |  8 ++++--
 include/migration/misc.h   |  2 ++
 migration/migration.c      | 18 +++++++++++++-
 migration/ram.c            | 51 ++++++++++++++++++++++++++++++++++++++
 migration/ram.h            |  1 +
 5 files changed, 77 insertions(+), 3 deletions(-)

-- 
2.25.1

[PATCH 1/3] migration: Fix missing qemu_fflush() on buffer file in bg_migration_thread

[Andrey Gruzdev]

Added missing qemu_fflush() on buffer file holding precopy device state.
Increased initial QIOChannelBuffer allocation to 512KB to avoid reallocs.
Typical configurations often require >200KB for device state and VMDESC.

Signed-off-by: Andrey Gruzdev <andrey.gruzdev@virtuozzo.com>
---
 migration/migration.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/migration/migration.c b/migration/migration.c
index 36768391b6..496cf6e17b 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -3857,7 +3857,7 @@ static void *bg_migration_thread(void *opaque)
      * with vCPUs running and, finally, write stashed non-RAM part of
      * the vmstate from the buffer to the migration stream.
      */
-    s->bioc = qio_channel_buffer_new(128 * 1024);
+    s->bioc = qio_channel_buffer_new(512 * 1024);
     qio_channel_set_name(QIO_CHANNEL(s->bioc), "vmstate-buffer");
     fb = qemu_fopen_channel_output(QIO_CHANNEL(s->bioc));
     object_unref(OBJECT(s->bioc));
@@ -3911,6 +3911,8 @@ static void *bg_migration_thread(void *opaque)
     if (qemu_savevm_state_complete_precopy_non_iterable(fb, false, false)) {
         goto fail;
     }
+    qemu_fflush(fb);
+
     /* Now initialize UFFD context and start tracking RAM writes */
     if (ram_write_tracking_start()) {
         goto fail;
-- 
2.25.1

[PATCH 2/3] migration: Inhibit virtio-balloon for the duration of background snapshot

[Andrey Gruzdev]

The same thing as for incoming postcopy - we cannot deal with concurrent
RAM discards when using background snapshot feature in outgoing migration.

Signed-off-by: Andrey Gruzdev <andrey.gruzdev@virtuozzo.com>
---
 hw/virtio/virtio-balloon.c | 8 ++++++--
 include/migration/misc.h   | 2 ++
 migration/migration.c      | 8 ++++++++
 3 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/hw/virtio/virtio-balloon.c b/hw/virtio/virtio-balloon.c
index e770955176..d120bf8f43 100644
--- a/hw/virtio/virtio-balloon.c
+++ b/hw/virtio/virtio-balloon.c
@@ -66,8 +66,12 @@ static bool virtio_balloon_pbp_matches(PartiallyBalloonedPage *pbp,
 
 static bool virtio_balloon_inhibited(void)
 {
-    /* Postcopy cannot deal with concurrent discards, so it's special. */
-    return ram_block_discard_is_disabled() || migration_in_incoming_postcopy();
+    /*
+     * Postcopy cannot deal with concurrent discards,
+     * so it's special, as well as background snapshots.
+     */
+    return ram_block_discard_is_disabled() || migration_in_incoming_postcopy() ||
+            migration_in_bg_snapshot();
 }
 
 static void balloon_inflate_page(VirtIOBalloon *balloon,
diff --git a/include/migration/misc.h b/include/migration/misc.h
index bccc1b6b44..738675ef52 100644
--- a/include/migration/misc.h
+++ b/include/migration/misc.h
@@ -70,6 +70,8 @@ bool migration_in_postcopy_after_devices(MigrationState *);
 void migration_global_dump(Monitor *mon);
 /* True if incomming migration entered POSTCOPY_INCOMING_DISCARD */
 bool migration_in_incoming_postcopy(void);
+/* True if background snapshot is active */
+bool migration_in_bg_snapshot(void);
 
 /* migration/block-dirty-bitmap.c */
 void dirty_bitmap_mig_init(void);
diff --git a/migration/migration.c b/migration/migration.c
index 496cf6e17b..656d6249a6 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -1976,6 +1976,14 @@ bool migration_in_incoming_postcopy(void)
     return ps >= POSTCOPY_INCOMING_DISCARD && ps < POSTCOPY_INCOMING_END;
 }
 
+bool migration_in_bg_snapshot(void)
+{
+    MigrationState *s = migrate_get_current();
+
+    return migrate_background_snapshot() &&
+            migration_is_setup_or_active(s->state);
+}
+
 bool migration_is_idle(void)
 {
     MigrationState *s = current_migration;
-- 
2.25.1

[PATCH 3/3] migration: Pre-fault memory before starting background snasphot

[Andrey Gruzdev]

This commit solves the issue with userfault_fd WP feature that
background snapshot is based on. For any never poluated or discarded
memory page, the UFFDIO_WRITEPROTECT ioctl() would skip updating
PTE for that page, thereby loosing WP setting for it.

So we need to pre-fault pages for each RAM block to be protected
before making a userfault_fd wr-protect ioctl().

Signed-off-by: Andrey Gruzdev <andrey.gruzdev@virtuozzo.com>
---
 migration/migration.c |  6 +++++
 migration/ram.c       | 51 +++++++++++++++++++++++++++++++++++++++++++
 migration/ram.h       |  1 +
 3 files changed, 58 insertions(+)

diff --git a/migration/migration.c b/migration/migration.c
index 656d6249a6..496e88cbda 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -3872,6 +3872,12 @@ static void *bg_migration_thread(void *opaque)
 
     update_iteration_initial_status(s);
 
+    /*
+     * Prepare for tracking memory writes with UFFD-WP - populate
+     * RAM pages before protecting.
+     */
+    ram_write_tracking_prepare();
+
     qemu_savevm_state_header(s->to_dst_file);
     qemu_savevm_state_setup(s->to_dst_file);
 
diff --git a/migration/ram.c b/migration/ram.c
index 52537f14ac..825eb80030 100644
--- a/migration/ram.c
+++ b/migration/ram.c
@@ -1560,6 +1560,57 @@ out:
     return ret;
 }
 
+/*
+ * ram_block_populate_pages: populate memory in the RAM block by reading
+ *   an integer from the beginning of each page.
+ *
+ * Since it's solely used for userfault_fd WP feature, here we just
+ *   hardcode page size to TARGET_PAGE_SIZE.
+ *
+ * @bs: RAM block to populate
+ */
+volatile int ram_block_populate_pages__tmp;
+static void ram_block_populate_pages(RAMBlock *bs)
+{
+    ram_addr_t offset = 0;
+    int tmp = 0;
+
+    for (char *ptr = (char *) bs->host; offset < bs->used_length;
+            ptr += TARGET_PAGE_SIZE, offset += TARGET_PAGE_SIZE) {
+        /* Try to do it without memory writes */
+        tmp += *(volatile int *) ptr;
+    }
+    /* Create dependency on 'extern volatile int' to avoid optimizing out */
+    ram_block_populate_pages__tmp += tmp;
+}
+
+/*
+ * ram_write_tracking_prepare: prepare for UFFD-WP memory tracking
+ */
+void ram_write_tracking_prepare(void)
+{
+    RAMBlock *bs;
+
+    RCU_READ_LOCK_GUARD();
+
+    RAMBLOCK_FOREACH_NOT_IGNORED(bs) {
+        /* Nothing to do with read-only and MMIO-writable regions */
+        if (bs->mr->readonly || bs->mr->rom_device) {
+            continue;
+        }
+
+        /*
+         * Populate pages of the RAM block before enabling userfault_fd
+         * write protection.
+         *
+         * This stage is required since ioctl(UFFDIO_WRITEPROTECT) with
+         * UFFDIO_WRITEPROTECT_MODE_WP mode setting would silently skip
+         * pages with pte_none() entries in page table.
+         */
+        ram_block_populate_pages(bs);
+    }
+}
+
 /*
  * ram_write_tracking_start: start UFFD-WP memory tracking
  *
diff --git a/migration/ram.h b/migration/ram.h
index 6378bb3ebc..4833e9fd5b 100644
--- a/migration/ram.h
+++ b/migration/ram.h
@@ -82,6 +82,7 @@ void colo_incoming_start_dirty_log(void);
 /* Background snapshot */
 bool ram_write_tracking_available(void);
 bool ram_write_tracking_compatible(void);
+void ram_write_tracking_prepare(void);
 int ram_write_tracking_start(void);
 void ram_write_tracking_stop(void);
 
-- 
2.25.1

Re: [PATCH 2/3] migration: Inhibit virtio-balloon for the duration of background snapshot

[David Hildenbrand]

On 18.03.21 18:46, Andrey Gruzdev wrote:
> The same thing as for incoming postcopy - we cannot deal with concurrent
> RAM discards when using background snapshot feature in outgoing migration.
> 
> Signed-off-by: Andrey Gruzdev <andrey.gruzdev@virtuozzo.com>
> ---
>   hw/virtio/virtio-balloon.c | 8 ++++++--
>   include/migration/misc.h   | 2 ++
>   migration/migration.c      | 8 ++++++++
>   3 files changed, 16 insertions(+), 2 deletions(-)
> 
> diff --git a/hw/virtio/virtio-balloon.c b/hw/virtio/virtio-balloon.c
> index e770955176..d120bf8f43 100644
> --- a/hw/virtio/virtio-balloon.c
> +++ b/hw/virtio/virtio-balloon.c
> @@ -66,8 +66,12 @@ static bool virtio_balloon_pbp_matches(PartiallyBalloonedPage *pbp,
>   
>   static bool virtio_balloon_inhibited(void)
>   {
> -    /* Postcopy cannot deal with concurrent discards, so it's special. */
> -    return ram_block_discard_is_disabled() || migration_in_incoming_postcopy();
> +    /*
> +     * Postcopy cannot deal with concurrent discards,
> +     * so it's special, as well as background snapshots.
> +     */
> +    return ram_block_discard_is_disabled() || migration_in_incoming_postcopy() ||
> +            migration_in_bg_snapshot();
>   }
>   
>   static void balloon_inflate_page(VirtIOBalloon *balloon,
> diff --git a/include/migration/misc.h b/include/migration/misc.h
> index bccc1b6b44..738675ef52 100644
> --- a/include/migration/misc.h
> +++ b/include/migration/misc.h
> @@ -70,6 +70,8 @@ bool migration_in_postcopy_after_devices(MigrationState *);
>   void migration_global_dump(Monitor *mon);
>   /* True if incomming migration entered POSTCOPY_INCOMING_DISCARD */
>   bool migration_in_incoming_postcopy(void);
> +/* True if background snapshot is active */
> +bool migration_in_bg_snapshot(void);
>   
>   /* migration/block-dirty-bitmap.c */
>   void dirty_bitmap_mig_init(void);
> diff --git a/migration/migration.c b/migration/migration.c
> index 496cf6e17b..656d6249a6 100644
> --- a/migration/migration.c
> +++ b/migration/migration.c
> @@ -1976,6 +1976,14 @@ bool migration_in_incoming_postcopy(void)
>       return ps >= POSTCOPY_INCOMING_DISCARD && ps < POSTCOPY_INCOMING_END;
>   }
>   
> +bool migration_in_bg_snapshot(void)
> +{
> +    MigrationState *s = migrate_get_current();
> +
> +    return migrate_background_snapshot() &&
> +            migration_is_setup_or_active(s->state);
> +}
> +
>   bool migration_is_idle(void)
>   {
>       MigrationState *s = current_migration;
> 

Reviewed-by: David Hildenbrand <david@redhat.com>

-- 
Thanks,

David / dhildenb

Re: [PATCH 2/3] migration: Inhibit virtio-balloon for the duration of background snapshot

[Andrey Gruzdev]

On 18.03.2021 21:16, David Hildenbrand wrote:
> On 18.03.21 18:46, Andrey Gruzdev wrote:
>> The same thing as for incoming postcopy - we cannot deal with concurrent
>> RAM discards when using background snapshot feature in outgoing 
>> migration.
>>
>> Signed-off-by: Andrey Gruzdev <andrey.gruzdev@virtuozzo.com>
>> ---
>>   hw/virtio/virtio-balloon.c | 8 ++++++--
>>   include/migration/misc.h   | 2 ++
>>   migration/migration.c      | 8 ++++++++
>>   3 files changed, 16 insertions(+), 2 deletions(-)
>>
>> diff --git a/hw/virtio/virtio-balloon.c b/hw/virtio/virtio-balloon.c
>> index e770955176..d120bf8f43 100644
>> --- a/hw/virtio/virtio-balloon.c
>> +++ b/hw/virtio/virtio-balloon.c
>> @@ -66,8 +66,12 @@ static bool 
>> virtio_balloon_pbp_matches(PartiallyBalloonedPage *pbp,
>>     static bool virtio_balloon_inhibited(void)
>>   {
>> -    /* Postcopy cannot deal with concurrent discards, so it's 
>> special. */
>> -    return ram_block_discard_is_disabled() || 
>> migration_in_incoming_postcopy();
>> +    /*
>> +     * Postcopy cannot deal with concurrent discards,
>> +     * so it's special, as well as background snapshots.
>> +     */
>> +    return ram_block_discard_is_disabled() || 
>> migration_in_incoming_postcopy() ||
>> +            migration_in_bg_snapshot();
>>   }
>>     static void balloon_inflate_page(VirtIOBalloon *balloon,
>> diff --git a/include/migration/misc.h b/include/migration/misc.h
>> index bccc1b6b44..738675ef52 100644
>> --- a/include/migration/misc.h
>> +++ b/include/migration/misc.h
>> @@ -70,6 +70,8 @@ bool 
>> migration_in_postcopy_after_devices(MigrationState *);
>>   void migration_global_dump(Monitor *mon);
>>   /* True if incomming migration entered POSTCOPY_INCOMING_DISCARD */
>>   bool migration_in_incoming_postcopy(void);
>> +/* True if background snapshot is active */
>> +bool migration_in_bg_snapshot(void);
>>     /* migration/block-dirty-bitmap.c */
>>   void dirty_bitmap_mig_init(void);
>> diff --git a/migration/migration.c b/migration/migration.c
>> index 496cf6e17b..656d6249a6 100644
>> --- a/migration/migration.c
>> +++ b/migration/migration.c
>> @@ -1976,6 +1976,14 @@ bool migration_in_incoming_postcopy(void)
>>       return ps >= POSTCOPY_INCOMING_DISCARD && ps < 
>> POSTCOPY_INCOMING_END;
>>   }
>>   +bool migration_in_bg_snapshot(void)
>> +{
>> +    MigrationState *s = migrate_get_current();
>> +
>> +    return migrate_background_snapshot() &&
>> +            migration_is_setup_or_active(s->state);
>> +}
>> +
>>   bool migration_is_idle(void)
>>   {
>>       MigrationState *s = current_migration;
>>
>
> Reviewed-by: David Hildenbrand <david@redhat.com>
>
Thanks!

-- 
Andrey Gruzdev, Principal Engineer
Virtuozzo GmbH  +7-903-247-6397
                 virtuzzo.com

Re: [PATCH 3/3] migration: Pre-fault memory before starting background snasphot

[David Hildenbrand]

> +/*
> + * ram_block_populate_pages: populate memory in the RAM block by reading
> + *   an integer from the beginning of each page.
> + *
> + * Since it's solely used for userfault_fd WP feature, here we just
> + *   hardcode page size to TARGET_PAGE_SIZE.
> + *
> + * @bs: RAM block to populate
> + */
> +volatile int ram_block_populate_pages__tmp;
> +static void ram_block_populate_pages(RAMBlock *bs)
> +{
> +    ram_addr_t offset = 0;
> +    int tmp = 0;
> +
> +    for (char *ptr = (char *) bs->host; offset < bs->used_length;
> +            ptr += TARGET_PAGE_SIZE, offset += TARGET_PAGE_SIZE) {

You'll want qemu_real_host_page_size instead of TARGET_PAGE_SIZE

> +        /* Try to do it without memory writes */
> +        tmp += *(volatile int *) ptr;
> +    }


The following is slightly simpler and doesn't rely on volatile semantics [1].
Should work on any arch I guess.

static void ram_block_populate_pages(RAMBlock *bs)
{
     char *ptr = (char *) bs->host;
     ram_addr_t offset;

     for (offset = 0; offset < bs->used_length;
          offset += qemu_real_host_page_size) {
	char tmp = *(volatile char *)(ptr + offset)

	/* Don't optimize the read out. */
	asm volatile ("" : "+r" (tmp));
}

Compiles to

     for (offset = 0; offset < bs->used_length;
     316d:       48 8b 4b 30             mov    0x30(%rbx),%rcx
     char *ptr = (char *) bs->host;
     3171:       48 8b 73 18             mov    0x18(%rbx),%rsi
     for (offset = 0; offset < bs->used_length;
     3175:       48 85 c9                test   %rcx,%rcx
     3178:       74 ce                   je     3148 <ram_write_tracking_prepare+0x58>
          offset += qemu_real_host_page_size) {
     317a:       48 8b 05 00 00 00 00    mov    0x0(%rip),%rax        # 3181 <ram_write_tracking_prepare+0x91>
     3181:       48 8b 38                mov    (%rax),%rdi
     3184:       0f 1f 40 00             nopl   0x0(%rax)
         char tmp = *(volatile char *)(ptr + offset);
     3188:       48 8d 04 16             lea    (%rsi,%rdx,1),%rax
     318c:       0f b6 00                movzbl (%rax),%eax
          offset += qemu_real_host_page_size) {
     318f:       48 01 fa                add    %rdi,%rdx
     for (offset = 0; offset < bs->used_length;
     3192:       48 39 ca                cmp    %rcx,%rdx
     3195:       72 f1                   jb     3188 <ram_write_tracking_prepare+0x98>


[1] https://programfan.github.io/blog/2015/04/27/prevent-gcc-optimize-away-code/


I'll send patches soon to take care of virtio-mem via RamDiscardManager -
to skip populating the parts that are supposed to remain discarded and not migrated.
Unfortunately, the RamDiscardManager patches are still stuck waiting for
acks ... and now we're in soft-freeze.

-- 
Thanks,

David / dhildenb

Re: [PATCH 3/3] migration: Pre-fault memory before starting background snasphot

[David Hildenbrand]

On 19.03.21 10:28, David Hildenbrand wrote:
>> +/*
>> + * ram_block_populate_pages: populate memory in the RAM block by reading
>> + *   an integer from the beginning of each page.
>> + *
>> + * Since it's solely used for userfault_fd WP feature, here we just
>> + *   hardcode page size to TARGET_PAGE_SIZE.
>> + *
>> + * @bs: RAM block to populate
>> + */
>> +volatile int ram_block_populate_pages__tmp;
>> +static void ram_block_populate_pages(RAMBlock *bs)
>> +{
>> +    ram_addr_t offset = 0;
>> +    int tmp = 0;
>> +
>> +    for (char *ptr = (char *) bs->host; offset < bs->used_length;
>> +            ptr += TARGET_PAGE_SIZE, offset += TARGET_PAGE_SIZE) {
> 
> You'll want qemu_real_host_page_size instead of TARGET_PAGE_SIZE
> 
>> +        /* Try to do it without memory writes */
>> +        tmp += *(volatile int *) ptr;
>> +    }
> 
> 
> The following is slightly simpler and doesn't rely on volatile semantics [1].
> Should work on any arch I guess.
> 
> static void ram_block_populate_pages(RAMBlock *bs)
> {
>       char *ptr = (char *) bs->host;
>       ram_addr_t offset;
> 
>       for (offset = 0; offset < bs->used_length;
>            offset += qemu_real_host_page_size) {
> 	char tmp = *(volatile char *)(ptr + offset)

I wanted to do a "= *(ptr + offset)" here.

> 
> 	/* Don't optimize the read out. */
> 	asm volatile ("" : "+r" (tmp));

So this is the only volatile thing that the compiler must guarantee to 
not optimize away.


-- 
Thanks,

David / dhildenb

Re: [PATCH 3/3] migration: Pre-fault memory before starting background snasphot

[Andrey Gruzdev]

On 19.03.2021 12:28, David Hildenbrand wrote:
>> +/*
>> + * ram_block_populate_pages: populate memory in the RAM block by 
>> reading
>> + *   an integer from the beginning of each page.
>> + *
>> + * Since it's solely used for userfault_fd WP feature, here we just
>> + *   hardcode page size to TARGET_PAGE_SIZE.
>> + *
>> + * @bs: RAM block to populate
>> + */
>> +volatile int ram_block_populate_pages__tmp;
>> +static void ram_block_populate_pages(RAMBlock *bs)
>> +{
>> +    ram_addr_t offset = 0;
>> +    int tmp = 0;
>> +
>> +    for (char *ptr = (char *) bs->host; offset < bs->used_length;
>> +            ptr += TARGET_PAGE_SIZE, offset += TARGET_PAGE_SIZE) {
>
> You'll want qemu_real_host_page_size instead of TARGET_PAGE_SIZE
>
Ok.
>> +        /* Try to do it without memory writes */
>> +        tmp += *(volatile int *) ptr;
>> +    }
>
>
> The following is slightly simpler and doesn't rely on volatile 
> semantics [1].
> Should work on any arch I guess.
>
> static void ram_block_populate_pages(RAMBlock *bs)
> {
>     char *ptr = (char *) bs->host;
>     ram_addr_t offset;
>
>     for (offset = 0; offset < bs->used_length;
>          offset += qemu_real_host_page_size) {
>     char tmp = *(volatile char *)(ptr + offset)
>
>     /* Don't optimize the read out. */
>     asm volatile ("" : "+r" (tmp));
> }
>
Thanks, good option, I'll change the code.

> Compiles to
>
>     for (offset = 0; offset < bs->used_length;
>     316d:       48 8b 4b 30             mov    0x30(%rbx),%rcx
>     char *ptr = (char *) bs->host;
>     3171:       48 8b 73 18             mov    0x18(%rbx),%rsi
>     for (offset = 0; offset < bs->used_length;
>     3175:       48 85 c9                test   %rcx,%rcx
>     3178:       74 ce                   je     3148 
> <ram_write_tracking_prepare+0x58>
>          offset += qemu_real_host_page_size) {
>     317a:       48 8b 05 00 00 00 00    mov 0x0(%rip),%rax        # 
> 3181 <ram_write_tracking_prepare+0x91>
>     3181:       48 8b 38                mov    (%rax),%rdi
>     3184:       0f 1f 40 00             nopl   0x0(%rax)
>         char tmp = *(volatile char *)(ptr + offset);
>     3188:       48 8d 04 16             lea    (%rsi,%rdx,1),%rax
>     318c:       0f b6 00                movzbl (%rax),%eax
>          offset += qemu_real_host_page_size) {
>     318f:       48 01 fa                add    %rdi,%rdx
>     for (offset = 0; offset < bs->used_length;
>     3192:       48 39 ca                cmp    %rcx,%rdx
>     3195:       72 f1                   jb     3188 
> <ram_write_tracking_prepare+0x98>
>
>
> [1] 
> https://programfan.github.io/blog/2015/04/27/prevent-gcc-optimize-away-code/
>
>
> I'll send patches soon to take care of virtio-mem via RamDiscardManager -
> to skip populating the parts that are supposed to remain discarded and 
> not migrated.
> Unfortunately, the RamDiscardManager patches are still stuck waiting for
> acks ... and now we're in soft-freeze.
>
RamDiscardManager patches - do they also modify migration code?
I mean which part is responsible of not migrating discarded ranges.

-- 
Andrey Gruzdev, Principal Engineer
Virtuozzo GmbH  +7-903-247-6397
                 virtuzzo.com

Re: [PATCH 3/3] migration: Pre-fault memory before starting background snasphot

[Andrey Gruzdev]

On 19.03.2021 12:32, David Hildenbrand wrote:
> On 19.03.21 10:28, David Hildenbrand wrote:
>>> +/*
>>> + * ram_block_populate_pages: populate memory in the RAM block by 
>>> reading
>>> + *   an integer from the beginning of each page.
>>> + *
>>> + * Since it's solely used for userfault_fd WP feature, here we just
>>> + *   hardcode page size to TARGET_PAGE_SIZE.
>>> + *
>>> + * @bs: RAM block to populate
>>> + */
>>> +volatile int ram_block_populate_pages__tmp;
>>> +static void ram_block_populate_pages(RAMBlock *bs)
>>> +{
>>> +    ram_addr_t offset = 0;
>>> +    int tmp = 0;
>>> +
>>> +    for (char *ptr = (char *) bs->host; offset < bs->used_length;
>>> +            ptr += TARGET_PAGE_SIZE, offset += TARGET_PAGE_SIZE) {
>>
>> You'll want qemu_real_host_page_size instead of TARGET_PAGE_SIZE
>>
>>> +        /* Try to do it without memory writes */
>>> +        tmp += *(volatile int *) ptr;
>>> +    }
>>
>>
>> The following is slightly simpler and doesn't rely on volatile 
>> semantics [1].
>> Should work on any arch I guess.
>>
>> static void ram_block_populate_pages(RAMBlock *bs)
>> {
>>       char *ptr = (char *) bs->host;
>>       ram_addr_t offset;
>>
>>       for (offset = 0; offset < bs->used_length;
>>            offset += qemu_real_host_page_size) {
>>     char tmp = *(volatile char *)(ptr + offset)
>
> I wanted to do a "= *(ptr + offset)" here.
>
Yep
>>
>>     /* Don't optimize the read out. */
>>     asm volatile ("" : "+r" (tmp));
>
> So this is the only volatile thing that the compiler must guarantee to 
> not optimize away.
>
>

-- 
Andrey Gruzdev, Principal Engineer
Virtuozzo GmbH  +7-903-247-6397
                 virtuzzo.com

Re: [PATCH 3/3] migration: Pre-fault memory before starting background snasphot

[David Hildenbrand]

On 19.03.21 12:05, Andrey Gruzdev wrote:
> On 19.03.2021 12:28, David Hildenbrand wrote:
>>> +/*
>>> + * ram_block_populate_pages: populate memory in the RAM block by
>>> reading
>>> + *   an integer from the beginning of each page.
>>> + *
>>> + * Since it's solely used for userfault_fd WP feature, here we just
>>> + *   hardcode page size to TARGET_PAGE_SIZE.
>>> + *
>>> + * @bs: RAM block to populate
>>> + */
>>> +volatile int ram_block_populate_pages__tmp;
>>> +static void ram_block_populate_pages(RAMBlock *bs)
>>> +{
>>> +    ram_addr_t offset = 0;
>>> +    int tmp = 0;
>>> +
>>> +    for (char *ptr = (char *) bs->host; offset < bs->used_length;
>>> +            ptr += TARGET_PAGE_SIZE, offset += TARGET_PAGE_SIZE) {
>>
>> You'll want qemu_real_host_page_size instead of TARGET_PAGE_SIZE
>>
> Ok.
>>> +        /* Try to do it without memory writes */
>>> +        tmp += *(volatile int *) ptr;
>>> +    }
>>
>>
>> The following is slightly simpler and doesn't rely on volatile
>> semantics [1].
>> Should work on any arch I guess.
>>
>> static void ram_block_populate_pages(RAMBlock *bs)
>> {
>>      char *ptr = (char *) bs->host;
>>      ram_addr_t offset;
>>
>>      for (offset = 0; offset < bs->used_length;
>>           offset += qemu_real_host_page_size) {
>>      char tmp = *(volatile char *)(ptr + offset)
>>
>>      /* Don't optimize the read out. */
>>      asm volatile ("" : "+r" (tmp));
>> }
>>
> Thanks, good option, I'll change the code.
> 
>> Compiles to
>>
>>      for (offset = 0; offset < bs->used_length;
>>      316d:       48 8b 4b 30             mov    0x30(%rbx),%rcx
>>      char *ptr = (char *) bs->host;
>>      3171:       48 8b 73 18             mov    0x18(%rbx),%rsi
>>      for (offset = 0; offset < bs->used_length;
>>      3175:       48 85 c9                test   %rcx,%rcx
>>      3178:       74 ce                   je     3148
>> <ram_write_tracking_prepare+0x58>
>>           offset += qemu_real_host_page_size) {
>>      317a:       48 8b 05 00 00 00 00    mov 0x0(%rip),%rax        #
>> 3181 <ram_write_tracking_prepare+0x91>
>>      3181:       48 8b 38                mov    (%rax),%rdi
>>      3184:       0f 1f 40 00             nopl   0x0(%rax)
>>          char tmp = *(volatile char *)(ptr + offset);
>>      3188:       48 8d 04 16             lea    (%rsi,%rdx,1),%rax
>>      318c:       0f b6 00                movzbl (%rax),%eax
>>           offset += qemu_real_host_page_size) {
>>      318f:       48 01 fa                add    %rdi,%rdx
>>      for (offset = 0; offset < bs->used_length;
>>      3192:       48 39 ca                cmp    %rcx,%rdx
>>      3195:       72 f1                   jb     3188
>> <ram_write_tracking_prepare+0x98>
>>
>>
>> [1]
>> https://programfan.github.io/blog/2015/04/27/prevent-gcc-optimize-away-code/
>>
>>
>> I'll send patches soon to take care of virtio-mem via RamDiscardManager -
>> to skip populating the parts that are supposed to remain discarded and
>> not migrated.
>> Unfortunately, the RamDiscardManager patches are still stuck waiting for
>> acks ... and now we're in soft-freeze.
>>
> RamDiscardManager patches - do they also modify migration code?
> I mean which part is responsible of not migrating discarded ranges.

I haven't shared relevant patches yet that touch migration code. I'm 
planning on doing that once the generic RamDiscardManager has all 
relevant acks. I'll put you on cc.

-- 
Thanks,

David / dhildenb

Re: [PATCH 3/3] migration: Pre-fault memory before starting background snasphot

[Andrey Gruzdev]

On 19.03.2021 14:27, David Hildenbrand wrote:
> On 19.03.21 12:05, Andrey Gruzdev wrote:
>> On 19.03.2021 12:28, David Hildenbrand wrote:
>>>> +/*
>>>> + * ram_block_populate_pages: populate memory in the RAM block by
>>>> reading
>>>> + *   an integer from the beginning of each page.
>>>> + *
>>>> + * Since it's solely used for userfault_fd WP feature, here we just
>>>> + *   hardcode page size to TARGET_PAGE_SIZE.
>>>> + *
>>>> + * @bs: RAM block to populate
>>>> + */
>>>> +volatile int ram_block_populate_pages__tmp;
>>>> +static void ram_block_populate_pages(RAMBlock *bs)
>>>> +{
>>>> +    ram_addr_t offset = 0;
>>>> +    int tmp = 0;
>>>> +
>>>> +    for (char *ptr = (char *) bs->host; offset < bs->used_length;
>>>> +            ptr += TARGET_PAGE_SIZE, offset += TARGET_PAGE_SIZE) {
>>>
>>> You'll want qemu_real_host_page_size instead of TARGET_PAGE_SIZE
>>>
>> Ok.
>>>> +        /* Try to do it without memory writes */
>>>> +        tmp += *(volatile int *) ptr;
>>>> +    }
>>>
>>>
>>> The following is slightly simpler and doesn't rely on volatile
>>> semantics [1].
>>> Should work on any arch I guess.
>>>
>>> static void ram_block_populate_pages(RAMBlock *bs)
>>> {
>>>      char *ptr = (char *) bs->host;
>>>      ram_addr_t offset;
>>>
>>>      for (offset = 0; offset < bs->used_length;
>>>           offset += qemu_real_host_page_size) {
>>>      char tmp = *(volatile char *)(ptr + offset)
>>>
>>>      /* Don't optimize the read out. */
>>>      asm volatile ("" : "+r" (tmp));
>>> }
>>>
>> Thanks, good option, I'll change the code.
>>
>>> Compiles to
>>>
>>>      for (offset = 0; offset < bs->used_length;
>>>      316d:       48 8b 4b 30             mov 0x30(%rbx),%rcx
>>>      char *ptr = (char *) bs->host;
>>>      3171:       48 8b 73 18             mov 0x18(%rbx),%rsi
>>>      for (offset = 0; offset < bs->used_length;
>>>      3175:       48 85 c9                test   %rcx,%rcx
>>>      3178:       74 ce                   je     3148
>>> <ram_write_tracking_prepare+0x58>
>>>           offset += qemu_real_host_page_size) {
>>>      317a:       48 8b 05 00 00 00 00    mov 0x0(%rip),%rax        #
>>> 3181 <ram_write_tracking_prepare+0x91>
>>>      3181:       48 8b 38                mov    (%rax),%rdi
>>>      3184:       0f 1f 40 00             nopl   0x0(%rax)
>>>          char tmp = *(volatile char *)(ptr + offset);
>>>      3188:       48 8d 04 16             lea (%rsi,%rdx,1),%rax
>>>      318c:       0f b6 00                movzbl (%rax),%eax
>>>           offset += qemu_real_host_page_size) {
>>>      318f:       48 01 fa                add    %rdi,%rdx
>>>      for (offset = 0; offset < bs->used_length;
>>>      3192:       48 39 ca                cmp    %rcx,%rdx
>>>      3195:       72 f1                   jb     3188
>>> <ram_write_tracking_prepare+0x98>
>>>
>>>
>>> [1]
>>> https://programfan.github.io/blog/2015/04/27/prevent-gcc-optimize-away-code/ 
>>>
>>>
>>>
>>> I'll send patches soon to take care of virtio-mem via 
>>> RamDiscardManager -
>>> to skip populating the parts that are supposed to remain discarded and
>>> not migrated.
>>> Unfortunately, the RamDiscardManager patches are still stuck waiting 
>>> for
>>> acks ... and now we're in soft-freeze.
>>>
>> RamDiscardManager patches - do they also modify migration code?
>> I mean which part is responsible of not migrating discarded ranges.
>
> I haven't shared relevant patches yet that touch migration code. I'm 
> planning on doing that once the generic RamDiscardManager has all 
> relevant acks. I'll put you on cc.
>
Got it, thanks.

-- 
Andrey Gruzdev, Principal Engineer
Virtuozzo GmbH  +7-903-247-6397
                 virtuzzo.com

Re: [PATCH 1/3] migration: Fix missing qemu_fflush() on buffer file in bg_migration_thread

[David Hildenbrand]

On 18.03.21 18:46, Andrey Gruzdev wrote:
> Added missing qemu_fflush() on buffer file holding precopy device state.
> Increased initial QIOChannelBuffer allocation to 512KB to avoid reallocs.
> Typical configurations often require >200KB for device state and VMDESC.
> 
> Signed-off-by: Andrey Gruzdev <andrey.gruzdev@virtuozzo.com>
> ---
>   migration/migration.c | 4 +++-
>   1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/migration/migration.c b/migration/migration.c
> index 36768391b6..496cf6e17b 100644
> --- a/migration/migration.c
> +++ b/migration/migration.c
> @@ -3857,7 +3857,7 @@ static void *bg_migration_thread(void *opaque)
>        * with vCPUs running and, finally, write stashed non-RAM part of
>        * the vmstate from the buffer to the migration stream.
>        */
> -    s->bioc = qio_channel_buffer_new(128 * 1024);
> +    s->bioc = qio_channel_buffer_new(512 * 1024);

^ would that better be a separate patch? It sounds more like an 
improvement than a fix.

>       qio_channel_set_name(QIO_CHANNEL(s->bioc), "vmstate-buffer");
>       fb = qemu_fopen_channel_output(QIO_CHANNEL(s->bioc));
>       object_unref(OBJECT(s->bioc));
> @@ -3911,6 +3911,8 @@ static void *bg_migration_thread(void *opaque)
>       if (qemu_savevm_state_complete_precopy_non_iterable(fb, false, false)) {
>           goto fail;
>       }
> +    qemu_fflush(fb);

Fixes: ?

> +
>       /* Now initialize UFFD context and start tracking RAM writes */
>       if (ram_write_tracking_start()) {
>           goto fail;
> 


-- 
Thanks,

David / dhildenb

Re: [PATCH 1/3] migration: Fix missing qemu_fflush() on buffer file in bg_migration_thread

[Andrey Gruzdev]

On 19.03.2021 15:39, David Hildenbrand wrote:
> On 18.03.21 18:46, Andrey Gruzdev wrote:
>> Added missing qemu_fflush() on buffer file holding precopy device state.
>> Increased initial QIOChannelBuffer allocation to 512KB to avoid 
>> reallocs.
>> Typical configurations often require >200KB for device state and VMDESC.
>>
>> Signed-off-by: Andrey Gruzdev <andrey.gruzdev@virtuozzo.com>
>> ---
>>   migration/migration.c | 4 +++-
>>   1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/migration/migration.c b/migration/migration.c
>> index 36768391b6..496cf6e17b 100644
>> --- a/migration/migration.c
>> +++ b/migration/migration.c
>> @@ -3857,7 +3857,7 @@ static void *bg_migration_thread(void *opaque)
>>        * with vCPUs running and, finally, write stashed non-RAM part of
>>        * the vmstate from the buffer to the migration stream.
>>        */
>> -    s->bioc = qio_channel_buffer_new(128 * 1024);
>> +    s->bioc = qio_channel_buffer_new(512 * 1024);
>
> ^ would that better be a separate patch? It sounds more like an 
> improvement than a fix.
>
>> qio_channel_set_name(QIO_CHANNEL(s->bioc), "vmstate-buffer");
>>       fb = qemu_fopen_channel_output(QIO_CHANNEL(s->bioc));
>>       object_unref(OBJECT(s->bioc));
>> @@ -3911,6 +3911,8 @@ static void *bg_migration_thread(void *opaque)
>>       if (qemu_savevm_state_complete_precopy_non_iterable(fb, false, 
>> false)) {
>>           goto fail;
>>       }
>> +    qemu_fflush(fb);
>
> Fixes: ?
>
It fixes unflushed QEMUFile, not an improvement. If not flushed, the 
migrate_get_current()->bioc->data
would stay missing some bytes at the tail.
>> +
>>       /* Now initialize UFFD context and start tracking RAM writes */
>>       if (ram_write_tracking_start()) {
>>           goto fail;
>>
>
>

-- 
Andrey Gruzdev, Principal Engineer
Virtuozzo GmbH  +7-903-247-6397
                 virtuzzo.com