[PATCH v5] drm/i915/gvt: save RING_HEAD into vreg when vgpu switched out

[Zhang, Xiaolin]

On 05/31/2019 04:59 PM, Zhenyu Wang wrote:
> On 2019.05.31 08:28:10 +0000, Li, Weinan Z wrote:
>>> -----Original Message-----
>>> From: intel-gvt-dev [mailto:intel-gvt-dev-bounces at lists.freedesktop.org] On
>>> Behalf Of Zhenyu Wang
>>> Sent: Friday, May 31, 2019 1:42 PM
>>> To: Zhang, Xiaolin <xiaolin.zhang at intel.com>
>>> Cc: intel-gvt-dev at lists.freedesktop.org
>>> Subject: Re: [PATCH v5] drm/i915/gvt: save RING_HEAD into vreg when vgpu
>>> switched out
>>>
>>> On 2019.05.31 13:18:09 +0800, Xiaolin Zhang wrote:
>>>> Save RING_HEAD into vgpu reg when vgpu switched out and report it's
>>>> value back to guest.
>>>>
>>>> v5: ring head wrap count support.
>>>> v4: updated HEAD/TAIL with guest value, not host value. (Yan Zhao)
>>>> v3: save RING HEAD/TAIL vgpu reg in save_ring_hw_state. (Zhenyu Wang)
>>>> v2: save RING_TAIL as well during vgpu mmio switch to meet
>>>> ring_is_idle condition. (Fred Gao)
>>>> v1: based on input from Weinan. (Weinan Li)
>>>>
>>>> [zhenyuw: Include this fix for possible future guest kernel that would
>>>> utilize RING_HEAD for hangcheck.]
>>>>
>>>> Reviewed-by: Zhenyu Wang <zhenyuw at linux.intel.com>
>>>> Signed-off-by: Xiaolin Zhang <xiaolin.zhang at intel.com>
>>>> Signed-off-by: Zhenyu Wang <zhenyuw at linux.intel.com>
>>>> ---
>>>>  drivers/gpu/drm/i915/gvt/reg.h       |  3 +++
>>>>  drivers/gpu/drm/i915/gvt/scheduler.c | 28
>>>> ++++++++++++++++++++++++++++
>>> drivers/gpu/drm/i915/gvt/scheduler.h |
>>>> 1 +
>>>>  3 files changed, 32 insertions(+)
>>>>
>>>> diff --git a/drivers/gpu/drm/i915/gvt/reg.h
>>>> b/drivers/gpu/drm/i915/gvt/reg.h index 33aaa14..b210e9a 100644
>>>> --- a/drivers/gpu/drm/i915/gvt/reg.h
>>>> +++ b/drivers/gpu/drm/i915/gvt/reg.h
>>>> @@ -102,6 +102,9 @@
>>>>  #define FORCEWAKE_ACK_MEDIA_GEN9_REG 0x0D88  #define
>>>> FORCEWAKE_ACK_HSW_REG 0x130044
>>>>
>>>> +#define RB_HEAD_WRAP_CNT_MAX	((1 << 11) - 1)
>>>> +#define RB_HEAD_WRAP_CNT_OFF	21
>>>> +#define RB_HEAD_WRAP_CNT_MASK	(0xFFE00000)
>>>>  #define RB_HEAD_OFF_MASK	((1U << 21) - (1U << 2))
>>>>  #define RB_TAIL_OFF_MASK	((1U << 21) - (1U << 3))
>>>>  #define RB_TAIL_SIZE_MASK	((1U << 21) - (1U << 12))
>>>> diff --git a/drivers/gpu/drm/i915/gvt/scheduler.c
>>>> b/drivers/gpu/drm/i915/gvt/scheduler.c
>>>> index 38897d2..e6523e8 100644
>>>> --- a/drivers/gpu/drm/i915/gvt/scheduler.c
>>>> +++ b/drivers/gpu/drm/i915/gvt/scheduler.c
>>>> @@ -793,10 +793,34 @@ static void update_guest_context(struct
>>> intel_vgpu_workload *workload)
>>>>  	void *src;
>>>>  	unsigned long context_gpa, context_page_num;
>>>>  	int i;
>>>> +	struct drm_i915_private *dev_priv = gvt->dev_priv;
>>>> +	i915_reg_t reg;
>>>> +	u32 ring_base;
>>>> +	u32 head, tail;
>>>> +	u16 wrap_count;
>>>>
>>>>  	gvt_dbg_sched("ring id %d workload lrca %x\n", rq->engine->id,
>>>>  		      workload->ctx_desc.lrca);
>>>>
>>>> +	head = workload->rb_head;
>>>> +	tail = workload->rb_tail;
>>>> +	wrap_count = workload->guest_rb_head &
>>> RB_HEAD_WRAP_CNT_MASK;
>>>> +
>>> Looks missed the shift for real counter here.
>>>
>>>> +	if (tail < head) {
>>>> +		if (wrap_count == RB_HEAD_WRAP_CNT_MAX)
>>>> +			wrap_count = 0;
>>>> +		else
>>>> +			wrap_count += 1;
>>>> +	}
>>>> +
>>>> +	head = (wrap_count << RB_HEAD_WRAP_CNT_OFF) | tail;
>>>> +
>>>> +	ring_base = dev_priv->engine[workload->ring_id]->mmio_base;
>>>> +	reg = RING_TAIL(ring_base);
>>>> +	vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) = tail;
>>>> +	reg = RING_HEAD(ring_base);
>>>> +	vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) = head;
>>>> +
>>> Can write as
>>>
>>>         vgpu_vreg_t(vgpu, RING_TAIL(ring_base)) = tail;
>>>         vgpu_vreg_t(vgpu, RING_HEAD(ring_base)) = head;
>>>
>>>
>> Just update the vreg during "update_guest_context()" is not enough, thinking there is one workload was preempted out by host many times.
>>
> For preemption case which need extra track of host vs. guest ring
> to reflect guest state, this one is always needed for normal finish.
> Yeah, need further change to update guest state in time.
per talk with Weinan, it looks like if we put these code change to
switch_mmio, it can handle preempt case very well.
>>>>  	context_page_num = rq->engine->context_size;
>>>>  	context_page_num = context_page_num >> PAGE_SHIFT;
>>>>
>>>> @@ -1418,6 +1442,7 @@ intel_vgpu_create_workload(struct intel_vgpu
>>> *vgpu, int ring_id,
>>>>  	struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
>>>>  	u64 ring_context_gpa;
>>>>  	u32 head, tail, start, ctl, ctx_ctl, per_ctx, indirect_ctx;
>>>> +	u32 guest_head;
>>>>  	int ret;
>>>>
>>>>  	ring_context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
>>>> @@ -1433,6 +1458,8 @@ intel_vgpu_create_workload(struct intel_vgpu
>>> *vgpu, int ring_id,
>>>>  	intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
>>>>  			RING_CTX_OFF(ring_tail.val), &tail, 4);
>>>>
>>>> +	guest_head = head;
>>>> +
>>>>  	head &= RB_HEAD_OFF_MASK;
>>>>  	tail &= RB_TAIL_OFF_MASK;
>>>>
>>>> @@ -1465,6 +1492,7 @@ intel_vgpu_create_workload(struct intel_vgpu
>>> *vgpu, int ring_id,
>>>>  	workload->ctx_desc = *desc;
>>>>  	workload->ring_context_gpa = ring_context_gpa;
>>>>  	workload->rb_head = head;
>>>> +	workload->guest_rb_head = guest_head;
>>>>  	workload->rb_tail = tail;
>>>>  	workload->rb_start = start;
>>>>  	workload->rb_ctl = ctl;
>>>> diff --git a/drivers/gpu/drm/i915/gvt/scheduler.h
>>>> b/drivers/gpu/drm/i915/gvt/scheduler.h
>>>> index 90c6756..c50d14a 100644
>>>> --- a/drivers/gpu/drm/i915/gvt/scheduler.h
>>>> +++ b/drivers/gpu/drm/i915/gvt/scheduler.h
>>>> @@ -100,6 +100,7 @@ struct intel_vgpu_workload {
>>>>  	struct execlist_ctx_descriptor_format ctx_desc;
>>>>  	struct execlist_ring_context *ring_context;
>>>>  	unsigned long rb_head, rb_tail, rb_ctl, rb_start, rb_len;
>>>> +	unsigned long guest_rb_head;
>>>>  	bool restore_inhibit;
>>>>  	struct intel_vgpu_elsp_dwords elsp_dwords;
>>>>  	bool emulate_schedule_in;
>>>> --
>>>> 2.7.4
>>>>
>>> --
>>> Open Source Technology Center, Intel ltd.
>>>
>>> $gpg --keyserver wwwkeys.pgp.net --recv-keys 4D781827