Re: [PATCH] vdpasim: support doorbell mapping

From: Longpeng (Mike, Cloud Infrastructure Service Product Dept.)
Date: Mon Nov 28 2022 - 08:21:30 EST




在 2022/11/28 20:20, Michael S. Tsirkin 写道:
On Mon, Nov 28, 2022 at 07:59:33PM +0800, Longpeng (Mike, Cloud Infrastructure Service Product Dept.) wrote:


在 2022/11/28 18:19, Michael S. Tsirkin 写道:
On Mon, Nov 28, 2022 at 04:19:30PM +0800, Longpeng (Mike, Cloud Infrastructure Service Product Dept.) wrote:


在 2022/11/28 12:05, Jason Wang 写道:
On Mon, Nov 28, 2022 at 10:56 AM Longpeng(Mike) <longpeng2@xxxxxxxxxx> wrote:

From: Longpeng <longpeng2@xxxxxxxxxx>

Support doorbell mapping for vdpasim devices, then we can test the notify
passthrough feature even if there's no real hardware on hand.

You can use vp_vdpa in L1 plus page_ver_vq in L0 to test it in L2.
That is how I test it.

Yes, using nested virtualization can work, but it's hard to deploy in my
working environment for some reasons, so I decided to emulate this
capability in vdpasim, it's much easier.


Allocates a dummy page which used to emulate the notify page of the device.
All values written to this page would be ignored, a periodic work will
check whether there're requests that need to process.

This seems tricky, it means the device is working even if there's no

Right. It just try to make the vdpasim device work properly, but the vdpasim
device is only used for testing, so maybe the tricky emulation is
acceptable?

Maybe. You can try enabling VIRTIO_F_NOTIFICATION_DATA and then
looking at the data written to figure out whether
you need to poll the vq.

We can try after the kernel supports the VIRTIO_F_NOTIFICATION_DATA feature,
while there is still a long way to go.

That would be up to you to implement ;) It's probably 10-20 lines of
code all in all.


Um...The kick operation won't trigger VMEXIT if we support notify mapping for the vdpasim device. So, does the vdpasim device has to poll the notify address and then decide which VQ to poll?
The pro of this approach is we can poll the notify address first instead of the unconditional poll the VQ, right?


kick. If we really want to do, we should try to use page fault handler
(probably by extending the config ops), but I'm not sure it's worth to
bother (or if we can find a use case for no simulator devices).

This need to modify the framework, it seems unworthy.


This cap is disabled as default, users can enable it as follow:
modprobe vdpa_sim notify_passthrough=true

Signed-off-by: Longpeng <longpeng2@xxxxxxxxxx>
---
drivers/vdpa/vdpa_sim/vdpa_sim.c | 71 ++++++++++++++++++++++++++--
drivers/vdpa/vdpa_sim/vdpa_sim.h | 5 +-
drivers/vdpa/vdpa_sim/vdpa_sim_blk.c | 5 +-
drivers/vdpa/vdpa_sim/vdpa_sim_net.c | 4 +-
4 files changed, 76 insertions(+), 9 deletions(-)

diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim.c b/drivers/vdpa/vdpa_sim/vdpa_sim.c
index 7438a89ce939..5c215b56b78b 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim.c
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim.c
@@ -14,6 +14,7 @@
#include <linux/slab.h>
#include <linux/sched.h>
#include <linux/dma-map-ops.h>
+#include <asm/set_memory.h>
#include <linux/vringh.h>
#include <linux/vdpa.h>
#include <linux/vhost_iotlb.h>
@@ -36,9 +37,15 @@ module_param(max_iotlb_entries, int, 0444);
MODULE_PARM_DESC(max_iotlb_entries,
"Maximum number of iotlb entries for each address space. 0 means unlimited. (default: 2048)");

+static bool notify_passthrough;
+module_param(notify_passthrough, bool, 0444);
+MODULE_PARM_DESC(notify_passthrough,
+ "Enable vq notify(doorbell) area mapping. (default: false)");
+
#define VDPASIM_QUEUE_ALIGN PAGE_SIZE
#define VDPASIM_QUEUE_MAX 256
#define VDPASIM_VENDOR_ID 0
+#define VDPASIM_VRING_POLL_PERIOD 100 /* ms */

static struct vdpasim *vdpa_to_sim(struct vdpa_device *vdpa)
{
@@ -276,7 +283,7 @@ struct vdpasim *vdpasim_create(struct vdpasim_dev_attr *dev_attr,
}

vdpasim->dev_attr = *dev_attr;
- INIT_WORK(&vdpasim->work, dev_attr->work_fn);
+ INIT_DELAYED_WORK(&vdpasim->vring_work, dev_attr->work_fn);
spin_lock_init(&vdpasim->lock);
spin_lock_init(&vdpasim->iommu_lock);

@@ -287,6 +294,15 @@ struct vdpasim *vdpasim_create(struct vdpasim_dev_attr *dev_attr,
set_dma_ops(dev, &vdpasim_dma_ops);
vdpasim->vdpa.mdev = dev_attr->mgmt_dev;

+ if (notify_passthrough) {
+ vdpasim->notify = __get_free_page(GFP_KERNEL | __GFP_ZERO);
+ if (!vdpasim->notify)
+ goto err_iommu;
+#ifdef CONFIG_X86
+ set_memory_uc(vdpasim->notify, 1);
+#endif

What's the reason for using uc memory?

The vma->vm_page_prot of notify mapping is pgprot_noncached (see
vhost_vdpa_fault) but the vdpasim->notify is WB, so we should set its
memtype to UC here and set it back to WB when releasing the device.

You never look at this memory though. Why does it matter whether
it's UC or WB?


The warning in trace_pfn_remap() would be triggered.

For example:

x86/PAT: CPU 16/KVM:17819 map pfn RAM range req uncached-minus for [mem
0x5151f3000-0x5151f3fff], got write-back


+ }
+
vdpasim->config = kzalloc(dev_attr->config_size, GFP_KERNEL);
if (!vdpasim->config)
goto err_iommu;
@@ -357,8 +373,11 @@ static void vdpasim_kick_vq(struct vdpa_device *vdpa, u16 idx)
struct vdpasim *vdpasim = vdpa_to_sim(vdpa);
struct vdpasim_virtqueue *vq = &vdpasim->vqs[idx];

+ if (notify_passthrough)
+ return;

So we should keep the two paths to be used at the same time. Userspace
can choose to not map doorbells?

It can work even if the userspace does not to map doorbells (e.g start
without page-per-vq=on), because the device will periodic check its vqs.

Thanks

+
if (vq->ready)
- schedule_work(&vdpasim->work);
+ schedule_work(&vdpasim->vring_work.work);
}

static void vdpasim_set_vq_cb(struct vdpa_device *vdpa, u16 idx,
@@ -495,6 +514,18 @@ static u8 vdpasim_get_status(struct vdpa_device *vdpa)
return status;
}

+static void vdpasim_set_vring_work(struct vdpasim *vdpasim, bool start)
+{
+ if (!notify_passthrough)
+ return;
+
+ if (start)
+ schedule_delayed_work(&vdpasim->vring_work,
+ msecs_to_jiffies(VDPASIM_VRING_POLL_PERIOD));
+ else
+ cancel_delayed_work_sync(&vdpasim->vring_work);
+}
+
static void vdpasim_set_status(struct vdpa_device *vdpa, u8 status)
{
struct vdpasim *vdpasim = vdpa_to_sim(vdpa);
@@ -502,12 +533,16 @@ static void vdpasim_set_status(struct vdpa_device *vdpa, u8 status)
spin_lock(&vdpasim->lock);
vdpasim->status = status;
spin_unlock(&vdpasim->lock);
+
+ vdpasim_set_vring_work(vdpasim, status & VIRTIO_CONFIG_S_DRIVER_OK);
}

static int vdpasim_reset(struct vdpa_device *vdpa, bool clear)
{
struct vdpasim *vdpasim = vdpa_to_sim(vdpa);

+ vdpasim_set_vring_work(vdpasim, false);
+
spin_lock(&vdpasim->lock);
vdpasim->status = 0;
vdpasim_do_reset(vdpasim);
@@ -672,12 +707,24 @@ static int vdpasim_dma_unmap(struct vdpa_device *vdpa, unsigned int asid,
return 0;
}

+static struct vdpa_notification_area
+vdpasim_get_vq_notification(struct vdpa_device *vdpa, u16 qid)
+{
+ struct vdpasim *vdpasim = vdpa_to_sim(vdpa);
+ struct vdpa_notification_area notify;
+
+ notify.addr = virt_to_phys((void *)vdpasim->notify);
+ notify.size = PAGE_SIZE;
+
+ return notify;
+}
+
static void vdpasim_free(struct vdpa_device *vdpa)
{
struct vdpasim *vdpasim = vdpa_to_sim(vdpa);
int i;

- cancel_work_sync(&vdpasim->work);
+ cancel_delayed_work_sync(&vdpasim->vring_work);

for (i = 0; i < vdpasim->dev_attr.nvqs; i++) {
vringh_kiov_cleanup(&vdpasim->vqs[i].out_iov);
@@ -693,7 +740,23 @@ static void vdpasim_free(struct vdpa_device *vdpa)
vhost_iotlb_free(vdpasim->iommu);
kfree(vdpasim->vqs);
kfree(vdpasim->config);
+ if (vdpasim->notify) {
+#ifdef CONFIG_X86
+ set_memory_wb(vdpasim->notify, 1);
+#endif
+ free_page(vdpasim->notify);
+ }
+}
+
+void vdpasim_schedule_work(struct vdpasim *vdpasim, bool sched_now)
+{
+ if (sched_now)
+ schedule_work(&vdpasim->vring_work.work);
+ else if (notify_passthrough)
+ schedule_delayed_work(&vdpasim->vring_work,
+ msecs_to_jiffies(VDPASIM_VRING_POLL_PERIOD));
}
+EXPORT_SYMBOL_GPL(vdpasim_schedule_work);

static const struct vdpa_config_ops vdpasim_config_ops = {
.set_vq_address = vdpasim_set_vq_address,
@@ -704,6 +767,7 @@ static const struct vdpa_config_ops vdpasim_config_ops = {
.get_vq_ready = vdpasim_get_vq_ready,
.set_vq_state = vdpasim_set_vq_state,
.get_vq_state = vdpasim_get_vq_state,
+ .get_vq_notification = vdpasim_get_vq_notification,
.get_vq_align = vdpasim_get_vq_align,
.get_vq_group = vdpasim_get_vq_group,
.get_device_features = vdpasim_get_device_features,
@@ -737,6 +801,7 @@ static const struct vdpa_config_ops vdpasim_batch_config_ops = {
.get_vq_ready = vdpasim_get_vq_ready,
.set_vq_state = vdpasim_set_vq_state,
.get_vq_state = vdpasim_get_vq_state,
+ .get_vq_notification = vdpasim_get_vq_notification,
.get_vq_align = vdpasim_get_vq_align,
.get_vq_group = vdpasim_get_vq_group,
.get_device_features = vdpasim_get_device_features,
diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim.h b/drivers/vdpa/vdpa_sim/vdpa_sim.h
index 0e78737dcc16..da0866834918 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim.h
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim.h
@@ -53,7 +53,7 @@ struct vdpasim_dev_attr {
struct vdpasim {
struct vdpa_device vdpa;
struct vdpasim_virtqueue *vqs;
- struct work_struct work;
+ struct delayed_work vring_work;
struct vdpasim_dev_attr dev_attr;
/* spinlock to synchronize virtqueue state */
spinlock_t lock;
@@ -69,10 +69,13 @@ struct vdpasim {
bool running;
/* spinlock to synchronize iommu table */
spinlock_t iommu_lock;
+ /* dummy notify page */
+ unsigned long notify;
};

struct vdpasim *vdpasim_create(struct vdpasim_dev_attr *attr,
const struct vdpa_dev_set_config *config);
+void vdpasim_schedule_work(struct vdpasim *vdpasim, bool sched_now);

/* TODO: cross-endian support */
static inline bool vdpasim_is_little_endian(struct vdpasim *vdpasim)
diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c b/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c
index c6db1a1baf76..8a640ea82284 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim_blk.c
@@ -288,7 +288,7 @@ static bool vdpasim_blk_handle_req(struct vdpasim *vdpasim,

static void vdpasim_blk_work(struct work_struct *work)
{
- struct vdpasim *vdpasim = container_of(work, struct vdpasim, work);
+ struct vdpasim *vdpasim = container_of(work, struct vdpasim, vring_work.work);
bool reschedule = false;
int i;

@@ -325,8 +325,7 @@ static void vdpasim_blk_work(struct work_struct *work)
out:
spin_unlock(&vdpasim->lock);

- if (reschedule)
- schedule_work(&vdpasim->work);
+ vdpasim_schedule_work(vdpasim, reschedule);
}

static void vdpasim_blk_get_config(struct vdpasim *vdpasim, void *config)
diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim_net.c b/drivers/vdpa/vdpa_sim/vdpa_sim_net.c
index c3cb225ea469..8b998952384b 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim_net.c
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim_net.c
@@ -145,7 +145,7 @@ static void vdpasim_handle_cvq(struct vdpasim *vdpasim)

static void vdpasim_net_work(struct work_struct *work)
{
- struct vdpasim *vdpasim = container_of(work, struct vdpasim, work);
+ struct vdpasim *vdpasim = container_of(work, struct vdpasim, vring_work.work);
struct vdpasim_virtqueue *txq = &vdpasim->vqs[1];
struct vdpasim_virtqueue *rxq = &vdpasim->vqs[0];
ssize_t read, write;
@@ -196,7 +196,7 @@ static void vdpasim_net_work(struct work_struct *work)
vdpasim_net_complete(rxq, write);

if (++pkts > 4) {
- schedule_work(&vdpasim->work);
+ vdpasim_schedule_work(vdpasim, true);
goto out;
}
}
--
2.23.0


.

.

.