This series is mainly for testing live migration between 2 vdpa_sim_blk
devices.
The first patch is preparation and moves the buffer allocation into devices,
the second patch adds the `shared_buffer_mutex` parameter to vdpa_sim_blk to
use the same ramdisk for all devices.
Tested with QEMU v8.0.0-rc2 in this way:
modprobe vhost_vdpa
modprobe vdpa_sim_blk shared_backend=true
vdpa dev add mgmtdev vdpasim_blk name blk0
vdpa dev add mgmtdev vdpasim_blk name blk1
qemu-system-x86_64 -m 512M -smp 2 -M q35,accel=kvm,memory-backend=mem \
-object memory-backend-file,share=on,id=mem,size="512M",mem-path="/dev/shm"
...
-blockdev node-name=drive_src1,driver=virtio-blk-vhost-vdpa,path=/dev/vhost-vdpa-1,cache.direct=on \
-device virtio-blk-pci,id=src1,bootindex=2,drive=drive_src1 \
-incoming tcp:0:3333
qemu-system-x86_64 -m 512M -smp 2 -M q35,accel=kvm,memory-backend=mem \
-object memory-backend-file,share=on,id=mem,size="512M",mem-path="/dev/shm"
...
-blockdev node-name=drive_src1,driver=virtio-blk-vhost-vdpa,path=/dev/vhost-vdpa-0,cache.direct=on \
-device virtio-blk-pci,id=src1,bootindex=2,drive=drive_src1
(qemu) migrate -d tcp:0:3333
Stefano Garzarella (2):
vdpa_sim: move buffer allocation in the devices
vdpa_sim_blk: support shared backend
drivers/vdpa/vdpa_sim/vdpa_sim.h | 3 +-
drivers/vdpa/vdpa_sim/vdpa_sim.c | 7 +--
drivers/vdpa/vdpa_sim/vdpa_sim_blk.c | 83 +++++++++++++++++++++++++---
drivers/vdpa/vdpa_sim/vdpa_sim_net.c | 28 +++++++---
4 files changed, 100 insertions(+), 21 deletions(-)
--
2.39.2
Currently, the vdpa_sim core does not use the buffer, but only
allocates it.
The buffer is used by devices differently, and some future devices
may not use it. So let's move all its management inside the devices.
Add a new `free` device callback called to clean up the resources
allocated by the device.
Signed-off-by: Stefano Garzarella <[email protected]>
---
drivers/vdpa/vdpa_sim/vdpa_sim.h | 3 +--
drivers/vdpa/vdpa_sim/vdpa_sim.c | 7 ++---
drivers/vdpa/vdpa_sim/vdpa_sim_blk.c | 40 +++++++++++++++++++++++-----
drivers/vdpa/vdpa_sim/vdpa_sim_net.c | 28 ++++++++++++++-----
4 files changed, 57 insertions(+), 21 deletions(-)
diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim.h b/drivers/vdpa/vdpa_sim/vdpa_sim.h
index 3a42887d05d9..bb137e479763 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim.h
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim.h
@@ -39,7 +39,6 @@ struct vdpasim_dev_attr {
u64 supported_features;
size_t alloc_size;
size_t config_size;
- size_t buffer_size;
int nvqs;
u32 id;
u32 ngroups;
@@ -51,6 +50,7 @@ struct vdpasim_dev_attr {
int (*get_stats)(struct vdpasim *vdpasim, u16 idx,
struct sk_buff *msg,
struct netlink_ext_ack *extack);
+ void (*free)(struct vdpasim *vdpasim);
};
/* State of each vdpasim device */
@@ -67,7 +67,6 @@ struct vdpasim {
void *config;
struct vhost_iotlb *iommu;
bool *iommu_pt;
- void *buffer;
u32 status;
u32 generation;
u64 features;
diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim.c b/drivers/vdpa/vdpa_sim/vdpa_sim.c
index 2c706bb18897..d343af4fa60e 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim.c
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim.c
@@ -261,10 +261,6 @@ struct vdpasim *vdpasim_create(struct vdpasim_dev_attr *dev_attr,
for (i = 0; i < vdpasim->dev_attr.nas; i++)
vhost_iotlb_init(&vdpasim->iommu[i], max_iotlb_entries, 0);
- vdpasim->buffer = kvmalloc(dev_attr->buffer_size, GFP_KERNEL);
- if (!vdpasim->buffer)
- goto err_iommu;
-
for (i = 0; i < dev_attr->nvqs; i++)
vringh_set_iotlb(&vdpasim->vqs[i].vring, &vdpasim->iommu[0],
&vdpasim->iommu_lock);
@@ -714,7 +710,8 @@ static void vdpasim_free(struct vdpa_device *vdpa)
vringh_kiov_cleanup(&vdpasim->vqs[i].in_iov);
}
- kvfree(vdpasim->buffer);
+ vdpasim->dev_attr.free(vdpasim);
+
for (i = 0; i < vdpasim->dev_attr.nas; i++)
vhost_iotlb_reset(&vdpasim->iommu[i]);
kfree(vdpasim->iommu);
diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c b/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c
index 568119e1553f..c996e750dc02 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c
@@ -43,6 +43,16 @@
#define VDPASIM_BLK_AS_NUM 1
#define VDPASIM_BLK_GROUP_NUM 1
+struct vdpasim_blk {
+ struct vdpasim vdpasim;
+ void *buffer;
+};
+
+static struct vdpasim_blk *sim_to_blk(struct vdpasim *vdpasim)
+{
+ return container_of(vdpasim, struct vdpasim_blk, vdpasim);
+}
+
static char vdpasim_blk_id[VIRTIO_BLK_ID_BYTES] = "vdpa_blk_sim";
static bool vdpasim_blk_check_range(struct vdpasim *vdpasim, u64 start_sector,
@@ -78,6 +88,7 @@ static bool vdpasim_blk_check_range(struct vdpasim *vdpasim, u64 start_sector,
static bool vdpasim_blk_handle_req(struct vdpasim *vdpasim,
struct vdpasim_virtqueue *vq)
{
+ struct vdpasim_blk *blk = sim_to_blk(vdpasim);
size_t pushed = 0, to_pull, to_push;
struct virtio_blk_outhdr hdr;
bool handled = false;
@@ -144,8 +155,7 @@ static bool vdpasim_blk_handle_req(struct vdpasim *vdpasim,
}
bytes = vringh_iov_push_iotlb(&vq->vring, &vq->in_iov,
- vdpasim->buffer + offset,
- to_push);
+ blk->buffer + offset, to_push);
if (bytes < 0) {
dev_dbg(&vdpasim->vdpa.dev,
"vringh_iov_push_iotlb() error: %zd offset: 0x%llx len: 0x%zx\n",
@@ -166,8 +176,7 @@ static bool vdpasim_blk_handle_req(struct vdpasim *vdpasim,
}
bytes = vringh_iov_pull_iotlb(&vq->vring, &vq->out_iov,
- vdpasim->buffer + offset,
- to_pull);
+ blk->buffer + offset, to_pull);
if (bytes < 0) {
dev_dbg(&vdpasim->vdpa.dev,
"vringh_iov_pull_iotlb() error: %zd offset: 0x%llx len: 0x%zx\n",
@@ -247,7 +256,7 @@ static bool vdpasim_blk_handle_req(struct vdpasim *vdpasim,
}
if (type == VIRTIO_BLK_T_WRITE_ZEROES) {
- memset(vdpasim->buffer + offset, 0,
+ memset(blk->buffer + offset, 0,
num_sectors << SECTOR_SHIFT);
}
@@ -353,6 +362,13 @@ static void vdpasim_blk_get_config(struct vdpasim *vdpasim, void *config)
}
+static void vdpasim_blk_free(struct vdpasim *vdpasim)
+{
+ struct vdpasim_blk *blk = sim_to_blk(vdpasim);
+
+ kvfree(blk->buffer);
+}
+
static void vdpasim_blk_mgmtdev_release(struct device *dev)
{
}
@@ -366,6 +382,7 @@ static int vdpasim_blk_dev_add(struct vdpa_mgmt_dev *mdev, const char *name,
const struct vdpa_dev_set_config *config)
{
struct vdpasim_dev_attr dev_attr = {};
+ struct vdpasim_blk *blk;
struct vdpasim *simdev;
int ret;
@@ -376,16 +393,25 @@ static int vdpasim_blk_dev_add(struct vdpa_mgmt_dev *mdev, const char *name,
dev_attr.nvqs = VDPASIM_BLK_VQ_NUM;
dev_attr.ngroups = VDPASIM_BLK_GROUP_NUM;
dev_attr.nas = VDPASIM_BLK_AS_NUM;
- dev_attr.alloc_size = sizeof(struct vdpasim);
+ dev_attr.alloc_size = sizeof(struct vdpasim_blk);
dev_attr.config_size = sizeof(struct virtio_blk_config);
dev_attr.get_config = vdpasim_blk_get_config;
dev_attr.work_fn = vdpasim_blk_work;
- dev_attr.buffer_size = VDPASIM_BLK_CAPACITY << SECTOR_SHIFT;
+ dev_attr.free = vdpasim_blk_free;
simdev = vdpasim_create(&dev_attr, config);
if (IS_ERR(simdev))
return PTR_ERR(simdev);
+ blk = sim_to_blk(simdev);
+
+ blk->buffer = kvmalloc(VDPASIM_BLK_CAPACITY << SECTOR_SHIFT,
+ GFP_KERNEL);
+ if (!blk->buffer) {
+ ret = -ENOMEM;
+ goto put_dev;
+ }
+
ret = _vdpa_register_device(&simdev->vdpa, VDPASIM_BLK_VQ_NUM);
if (ret)
goto put_dev;
diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim_net.c b/drivers/vdpa/vdpa_sim/vdpa_sim_net.c
index 55920502f76b..cfe962911804 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim_net.c
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim_net.c
@@ -58,6 +58,7 @@ struct vdpasim_net{
struct vdpasim_dataq_stats tx_stats;
struct vdpasim_dataq_stats rx_stats;
struct vdpasim_cq_stats cq_stats;
+ void *buffer;
};
static struct vdpasim_net *sim_to_net(struct vdpasim *vdpasim)
@@ -87,14 +88,15 @@ static bool receive_filter(struct vdpasim *vdpasim, size_t len)
size_t hdr_len = modern ? sizeof(struct virtio_net_hdr_v1) :
sizeof(struct virtio_net_hdr);
struct virtio_net_config *vio_config = vdpasim->config;
+ struct vdpasim_net *net = sim_to_net(vdpasim);
if (len < ETH_ALEN + hdr_len)
return false;
- if (is_broadcast_ether_addr(vdpasim->buffer + hdr_len) ||
- is_multicast_ether_addr(vdpasim->buffer + hdr_len))
+ if (is_broadcast_ether_addr(net->buffer + hdr_len) ||
+ is_multicast_ether_addr(net->buffer + hdr_len))
return true;
- if (!strncmp(vdpasim->buffer + hdr_len, vio_config->mac, ETH_ALEN))
+ if (!strncmp(net->buffer + hdr_len, vio_config->mac, ETH_ALEN))
return true;
return false;
@@ -225,8 +227,7 @@ static void vdpasim_net_work(struct vdpasim *vdpasim)
++tx_pkts;
read = vringh_iov_pull_iotlb(&txq->vring, &txq->out_iov,
- vdpasim->buffer,
- PAGE_SIZE);
+ net->buffer, PAGE_SIZE);
tx_bytes += read;
@@ -245,7 +246,7 @@ static void vdpasim_net_work(struct vdpasim *vdpasim)
}
write = vringh_iov_push_iotlb(&rxq->vring, &rxq->in_iov,
- vdpasim->buffer, read);
+ net->buffer, read);
if (write <= 0) {
++rx_errors;
break;
@@ -427,6 +428,13 @@ static void vdpasim_net_setup_config(struct vdpasim *vdpasim,
vio_config->mtu = cpu_to_vdpasim16(vdpasim, 1500);
}
+static void vdpasim_net_free(struct vdpasim *vdpasim)
+{
+ struct vdpasim_net *net = sim_to_net(vdpasim);
+
+ kvfree(net->buffer);
+}
+
static void vdpasim_net_mgmtdev_release(struct device *dev)
{
}
@@ -456,7 +464,7 @@ static int vdpasim_net_dev_add(struct vdpa_mgmt_dev *mdev, const char *name,
dev_attr.get_config = vdpasim_net_get_config;
dev_attr.work_fn = vdpasim_net_work;
dev_attr.get_stats = vdpasim_net_get_stats;
- dev_attr.buffer_size = PAGE_SIZE;
+ dev_attr.free = vdpasim_net_free;
simdev = vdpasim_create(&dev_attr, config);
if (IS_ERR(simdev))
@@ -470,6 +478,12 @@ static int vdpasim_net_dev_add(struct vdpa_mgmt_dev *mdev, const char *name,
u64_stats_init(&net->rx_stats.syncp);
u64_stats_init(&net->cq_stats.syncp);
+ net->buffer = kvmalloc(PAGE_SIZE, GFP_KERNEL);
+ if (!net->buffer) {
+ ret = -ENOMEM;
+ goto reg_err;
+ }
+
/*
* Initialization must be completed before this call, since it can
* connect the device to the vDPA bus, so requests can arrive after
--
2.39.2
The vdpa_sim_blk simulator uses a ramdisk as the backend. To test live
migration, we need two devices that share the backend to have the data
synchronized with each other.
Add a new module parameter to make the buffer shared between all devices.
The shared_buffer_mutex is used just to ensure that each operation is
atomic, but it is up to the user to use the devices knowing that the
underlying ramdisk is shared.
For example, when we do a migration, the VMM (e.g., QEMU) will guarantee
to write to the destination device, only after completing operations with
the source device.
Signed-off-by: Stefano Garzarella <[email protected]>
---
drivers/vdpa/vdpa_sim/vdpa_sim_blk.c | 57 ++++++++++++++++++++++++----
1 file changed, 50 insertions(+), 7 deletions(-)
diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c b/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c
index c996e750dc02..00d7d72713be 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c
@@ -46,6 +46,7 @@
struct vdpasim_blk {
struct vdpasim vdpasim;
void *buffer;
+ bool shared_backend;
};
static struct vdpasim_blk *sim_to_blk(struct vdpasim *vdpasim)
@@ -55,6 +56,26 @@ static struct vdpasim_blk *sim_to_blk(struct vdpasim *vdpasim)
static char vdpasim_blk_id[VIRTIO_BLK_ID_BYTES] = "vdpa_blk_sim";
+static bool shared_backend;
+module_param(shared_backend, bool, 0444);
+MODULE_PARM_DESC(shared_backend, "Enable the shared backend between virtio-blk devices");
+
+static void *shared_buffer;
+/* mutex to synchronize shared_buffer access */
+static DEFINE_MUTEX(shared_buffer_mutex);
+
+static void vdpasim_blk_buffer_lock(struct vdpasim_blk *blk)
+{
+ if (blk->shared_backend)
+ mutex_lock(&shared_buffer_mutex);
+}
+
+static void vdpasim_blk_buffer_unlock(struct vdpasim_blk *blk)
+{
+ if (blk->shared_backend)
+ mutex_unlock(&shared_buffer_mutex);
+}
+
static bool vdpasim_blk_check_range(struct vdpasim *vdpasim, u64 start_sector,
u64 num_sectors, u64 max_sectors)
{
@@ -154,8 +175,10 @@ static bool vdpasim_blk_handle_req(struct vdpasim *vdpasim,
break;
}
+ vdpasim_blk_buffer_lock(blk);
bytes = vringh_iov_push_iotlb(&vq->vring, &vq->in_iov,
blk->buffer + offset, to_push);
+ vdpasim_blk_buffer_unlock(blk);
if (bytes < 0) {
dev_dbg(&vdpasim->vdpa.dev,
"vringh_iov_push_iotlb() error: %zd offset: 0x%llx len: 0x%zx\n",
@@ -175,8 +198,10 @@ static bool vdpasim_blk_handle_req(struct vdpasim *vdpasim,
break;
}
+ vdpasim_blk_buffer_lock(blk);
bytes = vringh_iov_pull_iotlb(&vq->vring, &vq->out_iov,
blk->buffer + offset, to_pull);
+ vdpasim_blk_buffer_unlock(blk);
if (bytes < 0) {
dev_dbg(&vdpasim->vdpa.dev,
"vringh_iov_pull_iotlb() error: %zd offset: 0x%llx len: 0x%zx\n",
@@ -256,8 +281,10 @@ static bool vdpasim_blk_handle_req(struct vdpasim *vdpasim,
}
if (type == VIRTIO_BLK_T_WRITE_ZEROES) {
+ vdpasim_blk_buffer_lock(blk);
memset(blk->buffer + offset, 0,
num_sectors << SECTOR_SHIFT);
+ vdpasim_blk_buffer_unlock(blk);
}
break;
@@ -366,7 +393,8 @@ static void vdpasim_blk_free(struct vdpasim *vdpasim)
{
struct vdpasim_blk *blk = sim_to_blk(vdpasim);
- kvfree(blk->buffer);
+ if (!blk->shared_backend)
+ kvfree(blk->buffer);
}
static void vdpasim_blk_mgmtdev_release(struct device *dev)
@@ -404,12 +432,17 @@ static int vdpasim_blk_dev_add(struct vdpa_mgmt_dev *mdev, const char *name,
return PTR_ERR(simdev);
blk = sim_to_blk(simdev);
-
- blk->buffer = kvmalloc(VDPASIM_BLK_CAPACITY << SECTOR_SHIFT,
- GFP_KERNEL);
- if (!blk->buffer) {
- ret = -ENOMEM;
- goto put_dev;
+ blk->shared_backend = shared_backend;
+
+ if (blk->shared_backend) {
+ blk->buffer = shared_buffer;
+ } else {
+ blk->buffer = kvmalloc(VDPASIM_BLK_CAPACITY << SECTOR_SHIFT,
+ GFP_KERNEL);
+ if (!blk->buffer) {
+ ret = -ENOMEM;
+ goto put_dev;
+ }
}
ret = _vdpa_register_device(&simdev->vdpa, VDPASIM_BLK_VQ_NUM);
@@ -461,6 +494,15 @@ static int __init vdpasim_blk_init(void)
if (ret)
goto parent_err;
+ if (shared_backend) {
+ shared_buffer = kvmalloc(VDPASIM_BLK_CAPACITY << SECTOR_SHIFT,
+ GFP_KERNEL);
+ if (!shared_buffer) {
+ ret = -ENOMEM;
+ goto parent_err;
+ }
+ }
+
return 0;
parent_err:
@@ -470,6 +512,7 @@ static int __init vdpasim_blk_init(void)
static void __exit vdpasim_blk_exit(void)
{
+ kvfree(shared_buffer);
vdpa_mgmtdev_unregister(&mgmt_dev);
device_unregister(&vdpasim_blk_mgmtdev);
}
--
2.39.2