vdpa/virtio: Optimize device close time #10
Conversation
kailiangz1
force-pushed
the
dev_close_async
branch
2 times, most recently
from
5b94962
to
b6bd8e6
Compare
Move DMA unmap and state resoter work to work thread Signed-off-by: Kailiang Zhou <kailiangz@nvidia.com>
kailiangz1
force-pushed
the
dev_close_async
branch
from
a2151db
to
614ee88
Compare
drivers/vdpa/virtio/virtio_vdpa.c
Outdated
| } | ||
| static inline bool | ||
| virtio_vdpa_find_mem_reg(struct rte_vhost_mem_region *key,struct rte_vhost_memory *mem) |
There was a problem hiding this comment.
white space before struct rte_vhost_memory
| uint32_t i = 0; | ||
| struct rte_vhost_mem_region *reg; | ||
|
|
||
| if (mem == NULL) |
There was a problem hiding this comment.
calller cannot pass null. please remove this check.
There was a problem hiding this comment.
it can pass null when no mem map
There was a problem hiding this comment.
@kailiangz1 , who can pass null? vhost from qemu side?
There was a problem hiding this comment.
In line 669, if previous no mem map, it will be NULL
drivers/vdpa/virtio/virtio_vdpa.c
Outdated
| static inline bool | ||
| virtio_vdpa_find_mem_reg(struct rte_vhost_mem_region *key,struct rte_vhost_memory *mem) | ||
| { | ||
| uint32_t i = 0; |
There was a problem hiding this comment.
no need to zero initialize here, below for loop already doe it.
drivers/vdpa/virtio/virtio_vdpa.c
Outdated
| } | ||
| static inline bool | ||
| virtio_vdpa_find_mem_reg(struct rte_vhost_mem_region *key,struct rte_vhost_memory *mem) |
There was a problem hiding this comment.
make key and mem const pointer as this search routine is not going to modify anything.
drivers/vdpa/virtio/virtio_vdpa.c
Outdated
| if (priv->mem) { | ||
| for (i = 0; i < priv->mem->nregions; i++) { | ||
| reg = &priv->mem->regions[i]; | ||
| if (virtio_vdpa_find_mem_reg(reg, cur_mem)) { |
There was a problem hiding this comment.
parsing the code in the else to do the actual work is very confusing.
do it as,
ret = virtio_vdpa_find_mem_reg();
if (ret == 0) {
unmap code()
} else {
debug print
}
| DRV_LOG(ERR, "%s DMA unmap failed ret:%d", | ||
| priv->vdev->device->name, ret); | ||
| /* Map reagion dind't exsit in privious */ | ||
| for (i = 0; i < cur_mem->nregions; i++) { |
There was a problem hiding this comment.
same comment as above.
There was a problem hiding this comment.
and also have the error unwinding routine label to unmap or map the memory if the for loop fails.
There was a problem hiding this comment.
it is not necessary to unwinding because mem map is to map the mem, if it exist, we will no map them. caller can call again to recover. Map exist mem no harm.
drivers/vdpa/virtio/virtio_vdpa.c
Outdated
| if (ret < 0) { | ||
| DRV_LOG(ERR, "%s DMA unmap failed ret:%d", | ||
| priv->vdev->device->name, ret); | ||
| /* Map reagion dind't exsit in privious */ |
There was a problem hiding this comment.
comment should be,
/* Map the region if it doesn't exist yet */
| priv->vdev->device->name, ret); | ||
| goto exit; | ||
| /* Unmap reagion dind't exsit in current */ | ||
| if (priv->mem) { |
There was a problem hiding this comment.
better to have two different helper routines for map and unmap and call them here.
lib/vhost/vhost_user.c
Outdated
| @@ -1445,6 +1445,9 @@ vhost_user_set_mem_table(struct virtio_net **pdev, | |||
| if (dev->async_copy && rte_vfio_is_enabled("vfio")) | |||
| async_dma_map(dev, true); | |||
|
|
|||
| if ((dev->vdpa_dev) && (dev->vdpa_dev->ops->set_mem_table)) | |||
| dev->vdpa_dev->ops->set_mem_table(dev->vdpa_dev); | |||
There was a problem hiding this comment.
good work to have explicit call, instead of doing implicit in the dev_config.
There was a problem hiding this comment.
I am not understand. Can you show me the code example?
kailiangz1
force-pushed
the
dev_close_async
branch
2 times, most recently
from
d3db83c
to
638afe8
Compare
Add new set mem operation, save dma map time when device config Signed-off-by: Kailiang Zhou <kailiangz@nvidia.com>
kailiangz1
force-pushed
the
dev_close_async
branch
from
638afe8
to
b31739b
Compare
Move DMA unmap and state resoter work to work thread
Signed-off-by: Kailiang Zhou kailiangz@nvidia.com