The patch set splits out 2 small locks from the original big gvt lock:
- vgpu_lock protects per-vGPU data and logic, especially the vGPU
trap emulation path.
- sched_lock protects gvt scheudler structure, context schedule logic
and vGPU's schedule data.
Use vgpu_lock to replace the gvt big lock. By doing this, the
mmio read/write trap path, vgpu virtual event emulation and other
vgpu related process, would be protected under per vgpu_lock.
v9:
- Change commit author since the patches are improved a lot compared
with original version.
Original author: Pei Zhang <pei.zhang@intel.com>
- Rebase to latest gvt-staging.
v8:
- Correct coding and comment style.
- Rebase to latest gvt-staging.
v7:
- Remove gtt_lock since already proteced by gvt_lock and vgpu_lock.
- Fix a typo in intel_gvt_deactivate_vgpu, unlock the wrong lock.
v6:
- Rebase to latest gvt-staging.
v5:
- Rebase to latest gvt-staging.
- intel_vgpu_page_track_handler should use vgpu_lock.
v4:
- Rebase to latest gvt-staging.
- Protect vgpu->active access with vgpu_lock.
- Do not wait gpu idle in vgpu_lock.
v3: update to latest code base
v2: add gvt->lock in function gvt_check_vblank_emulation
Performance comparison on Kabylake platform.
- Configuration:
Host: Ubuntu 16.04.
Guest 1 & 2: Ubuntu 16.04.
glmark2 score comparison:
- Configuration:
Host: glxgears.
Guests: glmark2.
+--------------------------------+-----------------+
| Setup | glmark2 score |
+--------------------------------+-----------------+
| unified lock, iommu=on | 58~62 (avg. 60) |
+--------------------------------+-----------------+
| unified lock, iommu=igfx_off | 57~61 (avg. 59) |
+--------------------------------+-----------------+
| per-logic lock, iommu=on | 60~68 (avg. 64) |
+--------------------------------+-----------------+
| per-logic lock, iommu=igfx_off | 61~67 (avg. 64) |
+--------------------------------+-----------------+
lock_stat comparison:
- Configuration:
Stop lock stat immediately after boot up.
Boot 2 VM Guests.
Run glmark2 in guests.
Start perf lock_stat for 20 seconds and stop again.
- Legend: c - contentions; w - waittime-avg
+------------+-----------------+-----------+---------------+------------+
| | gvt_lock |sched_lock | vgpu_lock | gtt_lock |
+ lock type; +-----------------+-----------+---------------+------------+
| iommu set | c | w | c | w | c | w | c | w |
+------------+-------+---------+----+------+------+--------+-----+------+
| unified; | 20697 | 839 |N/A | N/A | N/A | N/A | N/A | N/A |
| on | | | | | | | | |
+------------+-------+---------+----+------+------+--------+-----+------+
| unified; | 21838 | 658.15 |N/A | N/A | N/A | N/A | N/A | N/A |
| igfx_off | | | | | | | | |
+------------+-------+---------+----+------+------+--------+-----+------+
| per-logic; | 1553 | 1599.96 |9458|429.97| 5846 | 274.33 | 0 | 0.00 |
| on | | | | | | | | |
+------------+-------+---------+----+------+------+--------+-----+------+
| per-logic; | 1911 | 1678.32 |8335|445.16| 5451 | 244.80 | 0 | 0.00 |
| igfx_off | | | | | | | | |
+------------+-------+---------+----+------+------+--------+-----+------+
Signed-off-by: Pei Zhang <pei.zhang@intel.com>
Signed-off-by: Colin Xu <colin.xu@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
struct intel_gvt_irq *irq = &gvt->irq;
struct intel_vgpu *vgpu;
int pipe, id;
+ int found = false;
- if (WARN_ON(!mutex_is_locked(&gvt->lock)))
- return;
-
+ mutex_lock(&gvt->lock);
for_each_active_vgpu(gvt, vgpu, id) {
for (pipe = 0; pipe < I915_MAX_PIPES; pipe++) {
- if (pipe_is_enabled(vgpu, pipe))
- goto out;
+ if (pipe_is_enabled(vgpu, pipe)) {
+ found = true;
+ break;
+ }
}
+ if (found)
+ break;
}
/* all the pipes are disabled */
- hrtimer_cancel(&irq->vblank_timer.timer);
- return;
-
-out:
- hrtimer_start(&irq->vblank_timer.timer,
- ktime_add_ns(ktime_get(), irq->vblank_timer.period),
- HRTIMER_MODE_ABS);
-
+ if (!found)
+ hrtimer_cancel(&irq->vblank_timer.timer);
+ else
+ hrtimer_start(&irq->vblank_timer.timer,
+ ktime_add_ns(ktime_get(), irq->vblank_timer.period),
+ HRTIMER_MODE_ABS);
+ mutex_unlock(&gvt->lock);
}
static void emulate_vblank_on_pipe(struct intel_vgpu *vgpu, int pipe)
{
int pipe;
+ mutex_lock(&vgpu->vgpu_lock);
for_each_pipe(vgpu->gvt->dev_priv, pipe)
emulate_vblank_on_pipe(vgpu, pipe);
+ mutex_unlock(&vgpu->vgpu_lock);
}
/**
struct intel_vgpu *vgpu;
int id;
- if (WARN_ON(!mutex_is_locked(&gvt->lock)))
- return;
-
+ mutex_lock(&gvt->lock);
for_each_active_vgpu(gvt, vgpu, id)
emulate_vblank(vgpu);
+ mutex_unlock(&gvt->lock);
}
/**
continue;
if (test_and_clear_bit(INTEL_GVT_REQUEST_EMULATE_VBLANK,
- (void *)&gvt->service_request)) {
- mutex_lock(&gvt->lock);
+ (void *)&gvt->service_request))
intel_gvt_emulate_vblank(gvt);
- mutex_unlock(&gvt->lock);
- }
if (test_bit(INTEL_GVT_REQUEST_SCHED,
(void *)&gvt->service_request) ||
struct intel_vgpu {
struct intel_gvt *gvt;
+ struct mutex vgpu_lock;
int id;
unsigned long handle; /* vGPU handle used by hypervisor MPT modules */
bool active;
};
struct intel_gvt {
+ /* GVT scope lock, protect GVT itself, and all resource currently
+ * not yet protected by special locks(vgpu and scheduler lock).
+ */
struct mutex lock;
struct drm_i915_private *dev_priv;
struct idr vgpu_idr; /* vGPU IDR pool */
struct task_struct *service_thread;
wait_queue_head_t service_thread_wq;
+
+ /* service_request is always used in bit operation, we should always
+ * use it with atomic bit ops so that no need to use gvt big lock.
+ */
unsigned long service_request;
struct {
}
}
+ /* vgpu_lock already hold by emulate mmio r/w */
intel_gvt_reset_vgpu_locked(vgpu, false, engine_mask);
/* sw will wait for the device to ack the reset request */
vgpu_vreg(vgpu, offset) |= I965_PIPECONF_ACTIVE;
else
vgpu_vreg(vgpu, offset) &= ~I965_PIPECONF_ACTIVE;
+ /* vgpu_lock already hold by emulate mmio r/w */
+ mutex_unlock(&vgpu->vgpu_lock);
intel_gvt_check_vblank_emulation(vgpu->gvt);
+ mutex_lock(&vgpu->vgpu_lock);
return 0;
}
return;
gvt = vgpu->gvt;
- mutex_lock(&gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
offset = intel_vgpu_gpa_to_mmio_offset(vgpu, pa);
if (reg_is_mmio(gvt, offset)) {
if (read)
memcpy(pt, p_data, bytes);
}
- mutex_unlock(&gvt->lock);
+ mutex_unlock(&vgpu->vgpu_lock);
}
/**
failsafe_emulate_mmio_rw(vgpu, pa, p_data, bytes, true);
return 0;
}
- mutex_lock(&gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
offset = intel_vgpu_gpa_to_mmio_offset(vgpu, pa);
gvt_vgpu_err("fail to emulate MMIO read %08x len %d\n",
offset, bytes);
out:
- mutex_unlock(&gvt->lock);
+ mutex_unlock(&vgpu->vgpu_lock);
return ret;
}
return 0;
}
- mutex_lock(&gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
offset = intel_vgpu_gpa_to_mmio_offset(vgpu, pa);
gvt_vgpu_err("fail to emulate MMIO write %08x len %d\n", offset,
bytes);
out:
- mutex_unlock(&gvt->lock);
+ mutex_unlock(&vgpu->vgpu_lock);
return ret;
}
int intel_vgpu_page_track_handler(struct intel_vgpu *vgpu, u64 gpa,
void *data, unsigned int bytes)
{
- struct intel_gvt *gvt = vgpu->gvt;
struct intel_vgpu_page_track *page_track;
int ret = 0;
- mutex_lock(&gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
page_track = intel_vgpu_find_page_track(vgpu, gpa >> PAGE_SHIFT);
if (!page_track) {
}
out:
- mutex_unlock(&gvt->lock);
+ mutex_unlock(&vgpu->vgpu_lock);
return ret;
}
gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
ring_id, workload);
+ mutex_lock(&vgpu->vgpu_lock);
mutex_lock(&dev_priv->drm.struct_mutex);
ret = intel_gvt_scan_and_shadow_workload(workload);
}
mutex_unlock(&dev_priv->drm.struct_mutex);
+ mutex_unlock(&vgpu->vgpu_lock);
return ret;
}
int event;
mutex_lock(&gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
/* For the workload w/ request, needs to wait for the context
* switch to make sure request is completed.
* For the workload w/o request, directly complete the workload.
*/
if (workload->req) {
- struct drm_i915_private *dev_priv =
- workload->vgpu->gvt->dev_priv;
+ struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
struct intel_engine_cs *engine =
dev_priv->engine[workload->ring_id];
wait_event(workload->shadow_ctx_status_wq,
if (gvt->scheduler.need_reschedule)
intel_gvt_request_service(gvt, INTEL_GVT_REQUEST_EVENT_SCHED);
+ mutex_unlock(&vgpu->vgpu_lock);
mutex_unlock(&gvt->lock);
}
intel_uncore_forcewake_get(gvt->dev_priv,
FORCEWAKE_ALL);
- mutex_lock(&gvt->lock);
ret = dispatch_workload(workload);
- mutex_unlock(&gvt->lock);
if (ret) {
vgpu = workload->vgpu;
*/
void intel_gvt_deactivate_vgpu(struct intel_vgpu *vgpu)
{
- struct intel_gvt *gvt = vgpu->gvt;
-
- mutex_lock(&gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
vgpu->active = false;
if (atomic_read(&vgpu->submission.running_workload_num)) {
- mutex_unlock(&gvt->lock);
+ mutex_unlock(&vgpu->vgpu_lock);
intel_gvt_wait_vgpu_idle(vgpu);
- mutex_lock(&gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
}
intel_vgpu_stop_schedule(vgpu);
intel_vgpu_dmabuf_cleanup(vgpu);
- mutex_unlock(&gvt->lock);
+ mutex_unlock(&vgpu->vgpu_lock);
}
/**
{
struct intel_gvt *gvt = vgpu->gvt;
- mutex_lock(&gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
WARN(vgpu->active, "vGPU is still active!\n");
intel_gvt_debugfs_remove_vgpu(vgpu);
- idr_remove(&gvt->vgpu_idr, vgpu->id);
- if (idr_is_empty(&gvt->vgpu_idr))
- intel_gvt_clean_irq(gvt);
intel_vgpu_clean_sched_policy(vgpu);
intel_vgpu_clean_submission(vgpu);
intel_vgpu_clean_display(vgpu);
intel_vgpu_free_resource(vgpu);
intel_vgpu_clean_mmio(vgpu);
intel_vgpu_dmabuf_cleanup(vgpu);
- vfree(vgpu);
+ mutex_unlock(&vgpu->vgpu_lock);
+ mutex_lock(&gvt->lock);
+ idr_remove(&gvt->vgpu_idr, vgpu->id);
+ if (idr_is_empty(&gvt->vgpu_idr))
+ intel_gvt_clean_irq(gvt);
intel_gvt_update_vgpu_types(gvt);
mutex_unlock(&gvt->lock);
+
+ vfree(vgpu);
}
#define IDLE_VGPU_IDR 0
vgpu->id = IDLE_VGPU_IDR;
vgpu->gvt = gvt;
+ mutex_init(&vgpu->vgpu_lock);
for (i = 0; i < I915_NUM_ENGINES; i++)
INIT_LIST_HEAD(&vgpu->submission.workload_q_head[i]);
*/
void intel_gvt_destroy_idle_vgpu(struct intel_vgpu *vgpu)
{
+ mutex_lock(&vgpu->vgpu_lock);
intel_vgpu_clean_sched_policy(vgpu);
+ mutex_unlock(&vgpu->vgpu_lock);
+
vfree(vgpu);
}
if (!vgpu)
return ERR_PTR(-ENOMEM);
- mutex_lock(&gvt->lock);
-
ret = idr_alloc(&gvt->vgpu_idr, vgpu, IDLE_VGPU_IDR + 1, GVT_MAX_VGPU,
GFP_KERNEL);
if (ret < 0)
vgpu->handle = param->handle;
vgpu->gvt = gvt;
vgpu->sched_ctl.weight = param->weight;
+ mutex_init(&vgpu->vgpu_lock);
INIT_LIST_HEAD(&vgpu->dmabuf_obj_list_head);
INIT_RADIX_TREE(&vgpu->page_track_tree, GFP_KERNEL);
idr_init(&vgpu->object_idr);
if (ret)
goto out_clean_sched_policy;
- mutex_unlock(&gvt->lock);
-
return vgpu;
out_clean_sched_policy:
idr_remove(&gvt->vgpu_idr, vgpu->id);
out_free_vgpu:
vfree(vgpu);
- mutex_unlock(&gvt->lock);
return ERR_PTR(ret);
}
param.low_gm_sz = BYTES_TO_MB(param.low_gm_sz);
param.high_gm_sz = BYTES_TO_MB(param.high_gm_sz);
+ mutex_lock(&gvt->lock);
vgpu = __intel_gvt_create_vgpu(gvt, ¶m);
- if (IS_ERR(vgpu))
- return vgpu;
-
- /* calculate left instance change for types */
- intel_gvt_update_vgpu_types(gvt);
+ if (!IS_ERR(vgpu))
+ /* calculate left instance change for types */
+ intel_gvt_update_vgpu_types(gvt);
+ mutex_unlock(&gvt->lock);
return vgpu;
}
* @engine_mask: engines to reset for GT reset
*
* This function is called when user wants to reset a virtual GPU through
- * device model reset or GT reset. The caller should hold the gvt lock.
+ * device model reset or GT reset. The caller should hold the vgpu lock.
*
* vGPU Device Model Level Reset (DMLR) simulates the PCI level reset to reset
* the whole vGPU to default state as when it is created. This vGPU function
* scheduler when the reset is triggered by current vgpu.
*/
if (scheduler->current_vgpu == NULL) {
- mutex_unlock(&gvt->lock);
+ mutex_unlock(&vgpu->vgpu_lock);
intel_gvt_wait_vgpu_idle(vgpu);
- mutex_lock(&gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
}
intel_vgpu_reset_submission(vgpu, resetting_eng);
*/
void intel_gvt_reset_vgpu(struct intel_vgpu *vgpu)
{
- mutex_lock(&vgpu->gvt->lock);
+ mutex_lock(&vgpu->vgpu_lock);
intel_gvt_reset_vgpu_locked(vgpu, true, 0);
- mutex_unlock(&vgpu->gvt->lock);
+ mutex_unlock(&vgpu->vgpu_lock);
}
projects / linux.git / commitdiff