kvm: fix slot flags sync between Qemu and KVM
diff mbox

Message ID 5524CC0E.8020208@linux.intel.com
State New, archived
Headers show

Commit Message

Xiao Guangrong April 8, 2015, 6:34 a.m. UTC
We noticed that KVM keeps tracking dirty for the memslots when
live migration failed which causes bad performance due to huge
page mapping disallowed for this kind of memslot

It is caused by slot flags does not properly sync-ed between Qemu
and KVM. Current code doing slot update depends on slot->flags
which hopes to omit unnecessary ioctl. However, slot->flags only
reflects the stauts of corresponding memory region, vmsave and
live migration do dirty tracking which overset
KVM_MEM_LOG_DIRTY_PAGES for the slot. That causes the slot status
recorded in the flags does not exactly match the stauts in kernel.

We fixed it by introducing slot->is_dirty_logging which indicates
the dirty status in kernel so that it helps us to sync the status
between userspace and kernel

Wanpeng Li <wanpeng.li@linux.intel.com>
Signed-off-by: Xiao Guangrong <guangrong.xiao@linux.intel.com>
---
  kvm-all.c | 19 ++++++++++++++++++-
  1 file changed, 18 insertions(+), 1 deletion(-)

Comments

Xiao Guangrong April 8, 2015, 6:37 a.m. UTC | #1
On 04/08/2015 02:34 PM, Xiao Guangrong wrote:
> We noticed that KVM keeps tracking dirty for the memslots when
> live migration failed which causes bad performance due to huge
> page mapping disallowed for this kind of memslot
>
> It is caused by slot flags does not properly sync-ed between Qemu
> and KVM. Current code doing slot update depends on slot->flags
> which hopes to omit unnecessary ioctl. However, slot->flags only
> reflects the stauts of corresponding memory region, vmsave and
> live migration do dirty tracking which overset
> KVM_MEM_LOG_DIRTY_PAGES for the slot. That causes the slot status
> recorded in the flags does not exactly match the stauts in kernel.
>
> We fixed it by introducing slot->is_dirty_logging which indicates
> the dirty status in kernel so that it helps us to sync the status
> between userspace and kernel
>
> Wanpeng Li <wanpeng.li@linux.intel.com>

Sorry for the typo :( , this should be:

Reported-by: Wanpeng Li <wanpeng.li@linux.intel.com>
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Paolo Bonzini April 8, 2015, 10:46 a.m. UTC | #2
On 08/04/2015 08:34, Xiao Guangrong wrote:
> We noticed that KVM keeps tracking dirty for the memslots when
> live migration failed which causes bad performance due to huge
> page mapping disallowed for this kind of memslot
> 
> It is caused by slot flags does not properly sync-ed between Qemu
> and KVM. Current code doing slot update depends on slot->flags
> which hopes to omit unnecessary ioctl. However, slot->flags only
> reflects the stauts of corresponding memory region, vmsave and
> live migration do dirty tracking which overset
> KVM_MEM_LOG_DIRTY_PAGES for the slot. That causes the slot status
> recorded in the flags does not exactly match the stauts in kernel.
> 
> We fixed it by introducing slot->is_dirty_logging which indicates
> the dirty status in kernel so that it helps us to sync the status
> between userspace and kernel
> 
> Wanpeng Li <wanpeng.li@linux.intel.com>
> Signed-off-by: Xiao Guangrong <guangrong.xiao@linux.intel.com>

Hi Xiao,

the patch looks good.

However, I am planning to remove s->migration_log completely from QEMU
2.4 and have slot->flags also track the migration state.  This has the
side effect of fixing this bug.  I'll Cc you on the patches when I post
them (next week probably).

Thanks!

Paolo

> ---
>  kvm-all.c | 19 ++++++++++++++++++-
>  1 file changed, 18 insertions(+), 1 deletion(-)
> 
> diff --git a/kvm-all.c b/kvm-all.c
> index dd44f8c..69fa233 100644
> --- a/kvm-all.c
> +++ b/kvm-all.c
> @@ -60,6 +60,15 @@
> 
>  #define KVM_MSI_HASHTAB_SIZE    256
> 
> +/*
> + * @flags only reflects the stauts of corresponding memory region,
> however,
> + * vmsave and live migration do dirty tracking which overset
> + * KVM_MEM_LOG_DIRTY_PAGES for the slot. That causes the slot status
> recorded
> + * in @flags does not exactly match the stauts in kernel.
> + *
> + * @is_dirty_logging indicating the dirty status in kernel helps us to
> sync
> + * the status between userspace and kernel.
> + */
>  typedef struct KVMSlot
>  {
>      hwaddr start_addr;
> @@ -67,6 +76,7 @@ typedef struct KVMSlot
>      void *ram;
>      int slot;
>      int flags;
> +    bool is_dirty_logging;
>  } KVMSlot;
> 
>  typedef struct kvm_dirty_log KVMDirtyLog;
> @@ -245,6 +255,7 @@ static int kvm_set_user_memory_region(KVMState *s,
> KVMSlot *slot)
>          kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
>      }
>      mem.memory_size = slot->memory_size;
> +    slot->is_dirty_logging = !!(mem.flags & KVM_MEM_LOG_DIRTY_PAGES);
>      return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
>  }
> 
> @@ -312,6 +323,7 @@ static int kvm_slot_dirty_pages_log_change(KVMSlot
> *mem, bool log_dirty)
>      int old_flags;
> 
>      old_flags = mem->flags;
> +    old_flags |= mem->is_dirty_logging ? KVM_MEM_LOG_DIRTY_PAGES : 0;
> 
>      flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty, false);
>      mem->flags = flags;
> @@ -376,12 +388,17 @@ static int kvm_set_migration_log(bool enable)
>      s->migration_log = enable;
> 
>      for (i = 0; i < s->nr_slots; i++) {
> +        int dirty_enable;
> +
>          mem = &s->slots[i];
> 
>          if (!mem->memory_size) {
>              continue;
>          }
> -        if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
> +
> +        /* Keep the dirty bit if it is tracked by the memory region. */
> +        dirty_enable = enable | (mem->flags & KVM_MEM_LOG_DIRTY_PAGES);
> +        if (mem->is_dirty_logging == dirty_enable) {
>              continue;
>          }
>          err = kvm_set_user_memory_region(s, mem);
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Xiao Guangrong April 9, 2015, 1:24 a.m. UTC | #3
On 04/08/2015 06:46 PM, Paolo Bonzini wrote:
>
>
> On 08/04/2015 08:34, Xiao Guangrong wrote:
>> We noticed that KVM keeps tracking dirty for the memslots when
>> live migration failed which causes bad performance due to huge
>> page mapping disallowed for this kind of memslot
>>
>> It is caused by slot flags does not properly sync-ed between Qemu
>> and KVM. Current code doing slot update depends on slot->flags
>> which hopes to omit unnecessary ioctl. However, slot->flags only
>> reflects the stauts of corresponding memory region, vmsave and
>> live migration do dirty tracking which overset
>> KVM_MEM_LOG_DIRTY_PAGES for the slot. That causes the slot status
>> recorded in the flags does not exactly match the stauts in kernel.
>>
>> We fixed it by introducing slot->is_dirty_logging which indicates
>> the dirty status in kernel so that it helps us to sync the status
>> between userspace and kernel
>>
>> Wanpeng Li <wanpeng.li@linux.intel.com>
>> Signed-off-by: Xiao Guangrong <guangrong.xiao@linux.intel.com>
>
> Hi Xiao,
>
> the patch looks good.
>
> However, I am planning to remove s->migration_log completely from QEMU
> 2.4 and have slot->flags also track the migration state.  This has the
> side effect of fixing this bug.  I'll Cc you on the patches when I post
> them (next week probably).

Good to know it, look forward to your patches. Thank you, Paolo!
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch
diff mbox

diff --git a/kvm-all.c b/kvm-all.c
index dd44f8c..69fa233 100644
--- a/kvm-all.c
+++ b/kvm-all.c
@@ -60,6 +60,15 @@ 

  #define KVM_MSI_HASHTAB_SIZE    256

+/*
+ * @flags only reflects the stauts of corresponding memory region, however,
+ * vmsave and live migration do dirty tracking which overset
+ * KVM_MEM_LOG_DIRTY_PAGES for the slot. That causes the slot status recorded
+ * in @flags does not exactly match the stauts in kernel.
+ *
+ * @is_dirty_logging indicating the dirty status in kernel helps us to sync
+ * the status between userspace and kernel.
+ */
  typedef struct KVMSlot
  {
      hwaddr start_addr;
@@ -67,6 +76,7 @@  typedef struct KVMSlot
      void *ram;
      int slot;
      int flags;
+    bool is_dirty_logging;
  } KVMSlot;

  typedef struct kvm_dirty_log KVMDirtyLog;
@@ -245,6 +255,7 @@  static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
          kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
      }
      mem.memory_size = slot->memory_size;
+    slot->is_dirty_logging = !!(mem.flags & KVM_MEM_LOG_DIRTY_PAGES);
      return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
  }

@@ -312,6 +323,7 @@  static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
      int old_flags;

      old_flags = mem->flags;
+    old_flags |= mem->is_dirty_logging ? KVM_MEM_LOG_DIRTY_PAGES : 0;

      flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty, false);
      mem->flags = flags;
@@ -376,12 +388,17 @@  static int kvm_set_migration_log(bool enable)
      s->migration_log = enable;

      for (i = 0; i < s->nr_slots; i++) {
+        int dirty_enable;
+
          mem = &s->slots[i];

          if (!mem->memory_size) {
              continue;
          }
-        if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
+
+        /* Keep the dirty bit if it is tracked by the memory region. */
+        dirty_enable = enable | (mem->flags & KVM_MEM_LOG_DIRTY_PAGES);
+        if (mem->is_dirty_logging == dirty_enable) {
              continue;
          }
          err = kvm_set_user_memory_region(s, mem);