On 3/17/2024 8:22 PM, Jason Wang wrote:
On Sat, Mar 16, 2024 at 2:45 AM Si-Wei Liu <si-wei....@oracle.com> wrote:
On 3/14/2024 9:03 PM, Jason Wang wrote:
On Fri, Mar 15, 2024 at 5:39 AM Si-Wei Liu <si-wei....@oracle.com> wrote:
On setups with one or more virtio-net devices with vhost on,
dirty tracking iteration increases cost the bigger the number
amount of queues are set up e.g. on idle guests migration the
following is observed with virtio-net with vhost=on:
48 queues -> 78.11% [.] vhost_dev_sync_region.isra.13
8 queues -> 40.50% [.] vhost_dev_sync_region.isra.13
1 queue -> 6.89% [.] vhost_dev_sync_region.isra.13
2 devices, 1 queue -> 18.60% [.] vhost_dev_sync_region.isra.14
With high memory rates the symptom is lack of convergence as soon
as it has a vhost device with a sufficiently high number of queues,
the sufficient number of vhost devices.
On every migration iteration (every 100msecs) it will redundantly
query the *shared log* the number of queues configured with vhost
that exist in the guest. For the virtqueue data, this is necessary,
but not for the memory sections which are the same. So essentially
we end up scanning the dirty log too often.
To fix that, select a vhost device responsible for scanning the
log with regards to memory sections dirty tracking. It is selected
when we enable the logger (during migration) and cleared when we
disable the logger. If the vhost logger device goes away for some
reason, the logger will be re-selected from the rest of vhost
devices.
After making mem-section logger a singleton instance, constant cost
of 7%-9% (like the 1 queue report) will be seen, no matter how many
queues or how many vhost devices are configured:
48 queues -> 8.71% [.] vhost_dev_sync_region.isra.13
2 devices, 8 queues -> 7.97% [.] vhost_dev_sync_region.isra.14
Co-developed-by: Joao Martins <joao.m.mart...@oracle.com>
Signed-off-by: Joao Martins <joao.m.mart...@oracle.com>
Signed-off-by: Si-Wei Liu <si-wei....@oracle.com>
---
v3 -> v4:
- add comment to clarify effect on cache locality and
performance
v2 -> v3:
- add after-fix benchmark to commit log
- rename vhost_log_dev_enabled to vhost_dev_should_log
- remove unneeded comparisons for backend_type
- use QLIST array instead of single flat list to store vhost
logger devices
- simplify logger election logic
---
hw/virtio/vhost.c | 67
++++++++++++++++++++++++++++++++++++++++++-----
include/hw/virtio/vhost.h | 1 +
2 files changed, 62 insertions(+), 6 deletions(-)
diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c
index 612f4db..58522f1 100644
--- a/hw/virtio/vhost.c
+++ b/hw/virtio/vhost.c
@@ -45,6 +45,7 @@
static struct vhost_log *vhost_log[VHOST_BACKEND_TYPE_MAX];
static struct vhost_log *vhost_log_shm[VHOST_BACKEND_TYPE_MAX];
+static QLIST_HEAD(, vhost_dev) vhost_log_devs[VHOST_BACKEND_TYPE_MAX];
/* Memslots used by backends that support private memslots (without an fd).
*/
static unsigned int used_memslots;
@@ -149,6 +150,47 @@ bool vhost_dev_has_iommu(struct vhost_dev *dev)
}
}
+static inline bool vhost_dev_should_log(struct vhost_dev *dev)
+{
+ assert(dev->vhost_ops);
+ assert(dev->vhost_ops->backend_type > VHOST_BACKEND_TYPE_NONE);
+ assert(dev->vhost_ops->backend_type < VHOST_BACKEND_TYPE_MAX);
+
+ return dev == QLIST_FIRST(&vhost_log_devs[dev->vhost_ops->backend_type]);
A dumb question, why not simple check
dev->log == vhost_log_shm[dev->vhost_ops->backend_type]
Because we are not sure if the logger comes from vhost_log_shm[] or
vhost_log[]. Don't want to complicate the check here by calling into
vhost_dev_log_is_shared() everytime when the .log_sync() is called.
It has very low overhead, isn't it?
Whether this has low overhead will have to depend on the specific
backend's implementation for .vhost_requires_shm_log(), which the common
vhost layer should not assume upon or rely on the current implementation.
static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
{
return dev->vhost_ops->vhost_requires_shm_log &&
dev->vhost_ops->vhost_requires_shm_log(dev);
}
And it helps to simplify the logic.
Generally yes, but when it comes to hot path operations the performance
consideration could override this principle. I think there's no harm to
check against logger device cached in vhost layer itself, and the
current patch does not create a lot of complexity or performance side
effect (actually I think the conditional should be very straightforward
to turn into just a couple of assembly compare and branch instructions
rather than indirection through another jmp call).
-Siwei
Thanks
-Siwei
?
Thanks
