提交 28c4c6ca 编写于 作者: Z Zhi Wang 提交者: Zhenyu Wang

drm/i915/gvt: vGPU workload submission

This patch introduces the vGPU workload submission logics.

Under virtualization environment, guest will submit workload through
virtual execlist submit port. The submitted workload load will be wrapped
into an gvt workload which will be picked by GVT workload scheduler and
executed on host i915 later.
Signed-off-by: NZhi Wang <zhi.a.wang@intel.com>
Signed-off-by: NZhenyu Wang <zhenyuw@linux.intel.com>
上级 8453d674
......@@ -167,7 +167,7 @@ static void emulate_csb_update(struct intel_vgpu_execlist *execlist,
ring_id_to_context_switch_event(execlist->ring_id));
}
int emulate_execlist_ctx_schedule_out(
static int emulate_execlist_ctx_schedule_out(
struct intel_vgpu_execlist *execlist,
struct execlist_ctx_descriptor_format *ctx)
{
......@@ -260,7 +260,7 @@ static struct intel_vgpu_execlist_slot *get_next_execlist_slot(
return &execlist->slot[status.execlist_write_pointer];
}
int emulate_execlist_schedule_in(struct intel_vgpu_execlist *execlist,
static int emulate_execlist_schedule_in(struct intel_vgpu_execlist *execlist,
struct execlist_ctx_descriptor_format ctx[2])
{
struct intel_vgpu_execlist_slot *running = execlist->running_slot;
......@@ -353,6 +353,279 @@ int emulate_execlist_schedule_in(struct intel_vgpu_execlist *execlist,
return 0;
}
static void free_workload(struct intel_vgpu_workload *workload)
{
intel_vgpu_unpin_mm(workload->shadow_mm);
intel_gvt_mm_unreference(workload->shadow_mm);
kmem_cache_free(workload->vgpu->workloads, workload);
}
#define get_desc_from_elsp_dwords(ed, i) \
((struct execlist_ctx_descriptor_format *)&((ed)->data[i * 2]))
static int prepare_execlist_workload(struct intel_vgpu_workload *workload)
{
struct intel_vgpu *vgpu = workload->vgpu;
struct execlist_ctx_descriptor_format ctx[2];
int ring_id = workload->ring_id;
intel_vgpu_pin_mm(workload->shadow_mm);
intel_vgpu_sync_oos_pages(workload->vgpu);
intel_vgpu_flush_post_shadow(workload->vgpu);
if (!workload->emulate_schedule_in)
return 0;
ctx[0] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 1);
ctx[1] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 0);
return emulate_execlist_schedule_in(&vgpu->execlist[ring_id], ctx);
}
static int complete_execlist_workload(struct intel_vgpu_workload *workload)
{
struct intel_vgpu *vgpu = workload->vgpu;
struct intel_vgpu_execlist *execlist =
&vgpu->execlist[workload->ring_id];
struct intel_vgpu_workload *next_workload;
struct list_head *next = workload_q_head(vgpu, workload->ring_id)->next;
bool lite_restore = false;
int ret;
gvt_dbg_el("complete workload %p status %d\n", workload,
workload->status);
if (workload->status)
goto out;
if (!list_empty(workload_q_head(vgpu, workload->ring_id))) {
struct execlist_ctx_descriptor_format *this_desc, *next_desc;
next_workload = container_of(next,
struct intel_vgpu_workload, list);
this_desc = &workload->ctx_desc;
next_desc = &next_workload->ctx_desc;
lite_restore = same_context(this_desc, next_desc);
}
if (lite_restore) {
gvt_dbg_el("next context == current - no schedule-out\n");
free_workload(workload);
return 0;
}
ret = emulate_execlist_ctx_schedule_out(execlist, &workload->ctx_desc);
if (ret)
goto err;
out:
free_workload(workload);
return 0;
err:
free_workload(workload);
return ret;
}
#define RING_CTX_OFF(x) \
offsetof(struct execlist_ring_context, x)
static void read_guest_pdps(struct intel_vgpu *vgpu,
u64 ring_context_gpa, u32 pdp[8])
{
u64 gpa;
int i;
gpa = ring_context_gpa + RING_CTX_OFF(pdp3_UDW.val);
for (i = 0; i < 8; i++)
intel_gvt_hypervisor_read_gpa(vgpu,
gpa + i * 8, &pdp[7 - i], 4);
}
static int prepare_mm(struct intel_vgpu_workload *workload)
{
struct execlist_ctx_descriptor_format *desc = &workload->ctx_desc;
struct intel_vgpu_mm *mm;
int page_table_level;
u32 pdp[8];
if (desc->addressing_mode == 1) { /* legacy 32-bit */
page_table_level = 3;
} else if (desc->addressing_mode == 3) { /* legacy 64 bit */
page_table_level = 4;
} else {
gvt_err("Advanced Context mode(SVM) is not supported!\n");
return -EINVAL;
}
read_guest_pdps(workload->vgpu, workload->ring_context_gpa, pdp);
mm = intel_vgpu_find_ppgtt_mm(workload->vgpu, page_table_level, pdp);
if (mm) {
intel_gvt_mm_reference(mm);
} else {
mm = intel_vgpu_create_mm(workload->vgpu, INTEL_GVT_MM_PPGTT,
pdp, page_table_level, 0);
if (IS_ERR(mm)) {
gvt_err("fail to create mm object.\n");
return PTR_ERR(mm);
}
}
workload->shadow_mm = mm;
return 0;
}
#define get_last_workload(q) \
(list_empty(q) ? NULL : container_of(q->prev, \
struct intel_vgpu_workload, list))
bool submit_context(struct intel_vgpu *vgpu, int ring_id,
struct execlist_ctx_descriptor_format *desc,
bool emulate_schedule_in)
{
struct list_head *q = workload_q_head(vgpu, ring_id);
struct intel_vgpu_workload *last_workload = get_last_workload(q);
struct intel_vgpu_workload *workload = NULL;
u64 ring_context_gpa;
u32 head, tail, start, ctl, ctx_ctl;
int ret;
ring_context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
(u32)((desc->lrca + 1) << GTT_PAGE_SHIFT));
if (ring_context_gpa == INTEL_GVT_INVALID_ADDR) {
gvt_err("invalid guest context LRCA: %x\n", desc->lrca);
return -EINVAL;
}
intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
RING_CTX_OFF(ring_header.val), &head, 4);
intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
RING_CTX_OFF(ring_tail.val), &tail, 4);
head &= RB_HEAD_OFF_MASK;
tail &= RB_TAIL_OFF_MASK;
if (last_workload && same_context(&last_workload->ctx_desc, desc)) {
gvt_dbg_el("ring id %d cur workload == last\n", ring_id);
gvt_dbg_el("ctx head %x real head %lx\n", head,
last_workload->rb_tail);
/*
* cannot use guest context head pointer here,
* as it might not be updated at this time
*/
head = last_workload->rb_tail;
}
gvt_dbg_el("ring id %d begin a new workload\n", ring_id);
workload = kmem_cache_zalloc(vgpu->workloads, GFP_KERNEL);
if (!workload)
return -ENOMEM;
/* record some ring buffer register values for scan and shadow */
intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
RING_CTX_OFF(rb_start.val), &start, 4);
intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
RING_CTX_OFF(rb_ctrl.val), &ctl, 4);
intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
RING_CTX_OFF(ctx_ctrl.val), &ctx_ctl, 4);
INIT_LIST_HEAD(&workload->list);
init_waitqueue_head(&workload->shadow_ctx_status_wq);
atomic_set(&workload->shadow_ctx_active, 0);
workload->vgpu = vgpu;
workload->ring_id = ring_id;
workload->ctx_desc = *desc;
workload->ring_context_gpa = ring_context_gpa;
workload->rb_head = head;
workload->rb_tail = tail;
workload->rb_start = start;
workload->rb_ctl = ctl;
workload->prepare = prepare_execlist_workload;
workload->complete = complete_execlist_workload;
workload->status = -EINPROGRESS;
workload->emulate_schedule_in = emulate_schedule_in;
if (emulate_schedule_in)
memcpy(&workload->elsp_dwords,
&vgpu->execlist[ring_id].elsp_dwords,
sizeof(workload->elsp_dwords));
gvt_dbg_el("workload %p ring id %d head %x tail %x start %x ctl %x\n",
workload, ring_id, head, tail, start, ctl);
gvt_dbg_el("workload %p emulate schedule_in %d\n", workload,
emulate_schedule_in);
ret = prepare_mm(workload);
if (ret) {
kmem_cache_free(vgpu->workloads, workload);
return ret;
}
queue_workload(workload);
return 0;
}
int intel_vgpu_submit_execlist(struct intel_vgpu *vgpu, int ring_id)
{
struct intel_vgpu_execlist *execlist = &vgpu->execlist[ring_id];
struct execlist_ctx_descriptor_format *desc[2], valid_desc[2];
unsigned long valid_desc_bitmap = 0;
bool emulate_schedule_in = true;
int ret;
int i;
memset(valid_desc, 0, sizeof(valid_desc));
desc[0] = get_desc_from_elsp_dwords(&execlist->elsp_dwords, 1);
desc[1] = get_desc_from_elsp_dwords(&execlist->elsp_dwords, 0);
for (i = 0; i < 2; i++) {
if (!desc[i]->valid)
continue;
if (!desc[i]->privilege_access) {
gvt_err("vgpu%d: unexpected GGTT elsp submission\n",
vgpu->id);
return -EINVAL;
}
/* TODO: add another guest context checks here. */
set_bit(i, &valid_desc_bitmap);
valid_desc[i] = *desc[i];
}
if (!valid_desc_bitmap) {
gvt_err("vgpu%d: no valid desc in a elsp submission\n",
vgpu->id);
return -EINVAL;
}
if (!test_bit(0, (void *)&valid_desc_bitmap) &&
test_bit(1, (void *)&valid_desc_bitmap)) {
gvt_err("vgpu%d: weird elsp submission, desc 0 is not valid\n",
vgpu->id);
return -EINVAL;
}
/* submit workload */
for_each_set_bit(i, (void *)&valid_desc_bitmap, 2) {
ret = submit_context(vgpu, ring_id, &valid_desc[i],
emulate_schedule_in);
if (ret) {
gvt_err("vgpu%d: fail to schedule workload\n",
vgpu->id);
return ret;
}
emulate_schedule_in = false;
}
return 0;
}
static void init_vgpu_execlist(struct intel_vgpu *vgpu, int ring_id)
{
struct intel_vgpu_execlist *execlist = &vgpu->execlist[ring_id];
......@@ -374,13 +647,28 @@ static void init_vgpu_execlist(struct intel_vgpu *vgpu, int ring_id)
vgpu_vreg(vgpu, ctx_status_ptr_reg) = ctx_status_ptr.dw;
}
void intel_vgpu_clean_execlist(struct intel_vgpu *vgpu)
{
kmem_cache_destroy(vgpu->workloads);
}
int intel_vgpu_init_execlist(struct intel_vgpu *vgpu)
{
int i;
/* each ring has a virtual execlist engine */
for (i = 0; i < I915_NUM_ENGINES; i++)
for (i = 0; i < I915_NUM_ENGINES; i++) {
init_vgpu_execlist(vgpu, i);
INIT_LIST_HEAD(&vgpu->workload_q_head[i]);
}
vgpu->workloads = kmem_cache_create("gvt-g vgpu workload",
sizeof(struct intel_vgpu_workload), 0,
SLAB_HWCACHE_ALIGN,
NULL);
if (!vgpu->workloads)
return -ENOMEM;
return 0;
}
......@@ -118,6 +118,49 @@ struct execlist_context_status_format {
};
};
struct execlist_mmio_pair {
u32 addr;
u32 val;
};
/* The first 52 dwords in register state context */
struct execlist_ring_context {
u32 nop1;
u32 lri_cmd_1;
struct execlist_mmio_pair ctx_ctrl;
struct execlist_mmio_pair ring_header;
struct execlist_mmio_pair ring_tail;
struct execlist_mmio_pair rb_start;
struct execlist_mmio_pair rb_ctrl;
struct execlist_mmio_pair bb_cur_head_UDW;
struct execlist_mmio_pair bb_cur_head_LDW;
struct execlist_mmio_pair bb_state;
struct execlist_mmio_pair second_bb_addr_UDW;
struct execlist_mmio_pair second_bb_addr_LDW;
struct execlist_mmio_pair second_bb_state;
struct execlist_mmio_pair bb_per_ctx_ptr;
struct execlist_mmio_pair rcs_indirect_ctx;
struct execlist_mmio_pair rcs_indirect_ctx_offset;
u32 nop2;
u32 nop3;
u32 nop4;
u32 lri_cmd_2;
struct execlist_mmio_pair ctx_timestamp;
struct execlist_mmio_pair pdp3_UDW;
struct execlist_mmio_pair pdp3_LDW;
struct execlist_mmio_pair pdp2_UDW;
struct execlist_mmio_pair pdp2_LDW;
struct execlist_mmio_pair pdp1_UDW;
struct execlist_mmio_pair pdp1_LDW;
struct execlist_mmio_pair pdp0_UDW;
struct execlist_mmio_pair pdp0_LDW;
};
struct intel_vgpu_elsp_dwords {
u32 data[4];
u32 index;
};
struct intel_vgpu_execlist_slot {
struct execlist_ctx_descriptor_format ctx[2];
u32 index;
......@@ -130,8 +173,13 @@ struct intel_vgpu_execlist {
struct execlist_ctx_descriptor_format *running_context;
int ring_id;
struct intel_vgpu *vgpu;
struct intel_vgpu_elsp_dwords elsp_dwords;
};
void intel_vgpu_clean_execlist(struct intel_vgpu *vgpu);
int intel_vgpu_init_execlist(struct intel_vgpu *vgpu);
int intel_vgpu_submit_execlist(struct intel_vgpu *vgpu, int ring_id);
#endif /*_GVT_EXECLIST_H_*/
......@@ -42,6 +42,7 @@
#include "display.h"
#include "edid.h"
#include "execlist.h"
#include "scheduler.h"
#define GVT_MAX_VGPU 8
......@@ -147,8 +148,9 @@ struct intel_vgpu {
struct intel_vgpu_gtt gtt;
struct intel_vgpu_opregion opregion;
struct intel_vgpu_display display;
/* TODO: move the declaration of intel_gvt.h to a proper place. */
struct intel_vgpu_execlist execlist[I915_NUM_ENGINES];
struct list_head workload_q_head[I915_NUM_ENGINES];
struct kmem_cache *workloads;
};
struct intel_gvt_gm {
......@@ -193,6 +195,7 @@ struct intel_gvt {
struct intel_gvt_irq irq;
struct intel_gvt_gtt gtt;
struct intel_gvt_opregion opregion;
struct intel_gvt_workload_scheduler scheduler;
struct task_struct *service_thread;
wait_queue_head_t service_thread_wq;
......
......@@ -128,6 +128,18 @@ static int new_mmio_info(struct intel_gvt *gvt,
return 0;
}
static int render_mmio_to_ring_id(struct intel_gvt *gvt, unsigned int reg)
{
int i;
reg &= ~GENMASK(11, 0);
for (i = 0; i < I915_NUM_ENGINES; i++) {
if (gvt->dev_priv->engine[i].mmio_base == reg)
return i;
}
return -1;
}
#define offset_to_fence_num(offset) \
((offset - i915_mmio_reg_offset(FENCE_REG_GEN6_LO(0))) >> 3)
......@@ -1262,6 +1274,28 @@ static int ring_timestamp_mmio_read(struct intel_vgpu *vgpu,
return intel_vgpu_default_mmio_read(vgpu, offset, p_data, bytes);
}
static int elsp_mmio_write(struct intel_vgpu *vgpu, unsigned int offset,
void *p_data, unsigned int bytes)
{
int ring_id = render_mmio_to_ring_id(vgpu->gvt, offset);
struct intel_vgpu_execlist *execlist;
u32 data = *(u32 *)p_data;
int ret;
if (WARN_ON(ring_id < 0))
return -EINVAL;
execlist = &vgpu->execlist[ring_id];
execlist->elsp_dwords.data[execlist->elsp_dwords.index] = data;
if (execlist->elsp_dwords.index == 3)
ret = intel_vgpu_submit_execlist(vgpu, ring_id);
++execlist->elsp_dwords.index;
execlist->elsp_dwords.index &= 0x3;
return 0;
}
#define MMIO_F(reg, s, f, am, rm, d, r, w) do { \
ret = new_mmio_info(gvt, INTEL_GVT_MMIO_OFFSET(reg), \
f, s, am, rm, d, r, w); \
......@@ -2169,8 +2203,8 @@ static int init_broadwell_mmio_info(struct intel_gvt *gvt)
MMIO_RING_D(RING_ACTHD_UDW, D_BDW_PLUS);
#define RING_REG(base) (base + 0x230)
MMIO_RING_DFH(RING_REG, D_BDW_PLUS, 0, NULL, NULL);
MMIO_DH(RING_REG(GEN8_BSD2_RING_BASE), D_BDW_PLUS, NULL, NULL);
MMIO_RING_DFH(RING_REG, D_BDW_PLUS, 0, NULL, elsp_mmio_write);
MMIO_DH(RING_REG(GEN8_BSD2_RING_BASE), D_BDW_PLUS, NULL, elsp_mmio_write);
#undef RING_REG
#define RING_REG(base) (base + 0x234)
......
......@@ -72,4 +72,9 @@
#define FORCEWAKE_ACK_MEDIA_GEN9_REG 0x0D88
#define FORCEWAKE_ACK_HSW_REG 0x130044
#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))
#define _RING_CTL_BUF_SIZE(ctl) (((ctl) & RB_TAIL_SIZE_MASK) + GTT_PAGE_SIZE)
#endif
/*
* Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#ifndef _GVT_SCHEDULER_H_
#define _GVT_SCHEDULER_H_
struct intel_gvt_workload_scheduler {
struct list_head workload_q_head[I915_NUM_ENGINES];
};
struct intel_vgpu_workload {
struct intel_vgpu *vgpu;
int ring_id;
struct drm_i915_gem_request *req;
/* if this workload has been dispatched to i915? */
bool dispatched;
int status;
struct intel_vgpu_mm *shadow_mm;
/* different submission model may need different handler */
int (*prepare)(struct intel_vgpu_workload *);
int (*complete)(struct intel_vgpu_workload *);
struct list_head list;
/* execlist context information */
struct execlist_ctx_descriptor_format ctx_desc;
struct execlist_ring_context *ring_context;
unsigned long rb_head, rb_tail, rb_ctl, rb_start;
struct intel_vgpu_elsp_dwords elsp_dwords;
bool emulate_schedule_in;
atomic_t shadow_ctx_active;
wait_queue_head_t shadow_ctx_status_wq;
u64 ring_context_gpa;
};
#define workload_q_head(vgpu, ring_id) \
(&(vgpu->workload_q_head[ring_id]))
#define queue_workload(workload) \
list_add_tail(&workload->list, \
workload_q_head(workload->vgpu, workload->ring_id))
#endif
......@@ -146,6 +146,7 @@ void intel_gvt_destroy_vgpu(struct intel_vgpu *vgpu)
vgpu->active = false;
idr_remove(&gvt->vgpu_idr, vgpu->id);
intel_vgpu_clean_execlist(vgpu);
intel_vgpu_clean_display(vgpu);
intel_vgpu_clean_opregion(vgpu);
intel_vgpu_clean_gtt(vgpu);
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册