@@ -312,6 +312,266 @@ static bool emulate_execlist_schedule_in(struct gvt_virtual_execlist_info *info,
return true;
}
+static bool execlist_workload_complete(struct gvt_workload *workload)
+{
+ struct vgt_device *vgt = workload->vgt;
+ struct gvt_virtual_execlist_info *info =
+ &vgt->virtual_execlist_info[workload->ring_id];
+ struct gvt_workload *next_workload;
+ struct list_head *next = workload_q_head(vgt, workload->ring_id);
+ bool lite_restore = false;
+
+ gvt_dbg_el("complete workload %p status %d", workload, workload->status);
+
+ if (workload->status)
+ goto out;
+
+ if (!list_empty(workload_q_head(vgt, workload->ring_id))) {
+ struct execlist_ctx_descriptor_format *this_desc, *next_desc;
+
+ next_workload = container_of(next, struct gvt_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 workload context is same as current - no schedule-out");
+ goto out;
+ }
+
+ if (!emulate_execlist_ctx_schedule_out(info, &workload->ctx_desc)) {
+ kfree(workload);
+ return false;
+ }
+
+out:
+ gvt_destroy_mm(workload->shadow_mm);
+ kfree(workload);
+ return true;
+}
+
+void gvt_get_context_pdp_root_pointer(struct vgt_device *vgt,
+ struct execlist_ring_context *ring_context,
+ u32 pdp[8])
+{
+ struct gvt_execlist_mmio_pair *pdp_pair = &ring_context->pdp3_UDW;
+ u32 v;
+ int i;
+
+ for (i = 0; i < 8; i++) {
+ hypervisor_read_va(vgt, &pdp_pair[i].val, &v, 4, 1);
+ pdp[7 - i] = v;
+ }
+}
+
+void gvt_set_context_pdp_root_pointer(struct vgt_device *vgt,
+ struct execlist_ring_context *ring_context,
+ u32 pdp[8])
+{
+ struct gvt_execlist_mmio_pair *pdp_pair = &ring_context->pdp3_UDW;
+ int i;
+
+ for (i = 0; i < 8; i++)
+ pdp_pair[i].val = pdp[7 - i];
+}
+
+static struct execlist_ring_context *get_ring_context(struct vgt_device *vgt,
+ u32 lrca)
+{
+ struct execlist_ring_context *context;
+ u32 gma = (lrca + 1) << GTT_PAGE_SHIFT;
+
+ context = (struct execlist_ring_context *)
+ gvt_gma_to_va(vgt->gtt.ggtt_mm, gma);
+
+ return context;
+}
+
+static bool prepare_workload(struct gvt_workload *workload)
+{
+ struct execlist_ctx_descriptor_format *desc = &workload->ctx_desc;
+ struct gvt_mm *mm;
+ gtt_type_t root_entry_type;
+ int page_table_level;
+ u32 pdp[8];
+
+ if (desc->addressing_mode == 1) { /* legacy 32-bit */
+ page_table_level = 3;
+ root_entry_type = GTT_TYPE_PPGTT_ROOT_L3_ENTRY;
+ } else if (desc->addressing_mode == 3) { /* legacy 64 bit */
+ page_table_level = 4;
+ root_entry_type = GTT_TYPE_PPGTT_ROOT_L4_ENTRY;
+ } else {
+ gvt_err("Advanced Context mode(SVM) is not supported!\n");
+ return false;
+ }
+
+ gvt_get_context_pdp_root_pointer(workload->vgt, workload->ring_context, pdp);
+
+ mm = gvt_create_mm(workload->vgt, GVT_MM_PPGTT, root_entry_type,
+ pdp, page_table_level, 0);
+ if (!mm) {
+ gvt_err("fail to create mm object.\n");
+ return false;
+ }
+
+ workload->shadow_mm = mm;
+
+ return true;
+}
+
+bool submit_context(struct vgt_device *vgt, int ring_id,
+ struct execlist_ctx_descriptor_format *desc)
+{
+ struct list_head *q = workload_q_head(vgt, ring_id);
+ struct gvt_workload *last_workload = list_empty(q) ? NULL :
+ container_of(q->prev, struct gvt_workload, list);
+ struct gvt_workload *workload = NULL;
+
+ struct execlist_ring_context *ring_context = get_ring_context(
+ vgt, desc->lrca);
+
+ u32 head, tail, start, ctl;
+
+ if (!ring_context) {
+ gvt_err("invalid guest context LRCA: %x", desc->lrca);
+ return false;
+ }
+
+ hypervisor_read_va(vgt, &ring_context->ring_header.val,
+ &head, 4, 1);
+
+ hypervisor_read_va(vgt, &ring_context->ring_tail.val,
+ &tail, 4, 1);
+
+ 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 same workload as last workload", ring_id);
+ if (last_workload->dispatched) {
+ gvt_dbg_el("ring id %d last workload has been dispatched",
+ ring_id);
+ gvt_dbg_el("ctx head %x real head %lx",
+ 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;
+ } else {
+ gvt_dbg_el("ring id %d merged into last workload", ring_id);
+ /*
+ * if last workload hasn't been dispatched (scanned + shadowed),
+ * and the context for current submission is just the same as last
+ * workload context, then we can merge this submission into
+ * last workload.
+ */
+ last_workload->rb_tail = tail;
+ return true;
+ }
+ }
+
+ gvt_dbg_el("ring id %d begin a new workload", ring_id);
+
+ workload = kzalloc(sizeof(*workload), GFP_KERNEL);
+ if (!workload) {
+ gvt_err("fail to allocate memory for workload");
+ return false;
+ }
+
+ /* record some ring buffer register values for scan and shadow */
+ hypervisor_read_va(vgt, &ring_context->rb_start.val,
+ &start, 4, 1);
+ hypervisor_read_va(vgt, &ring_context->rb_ctrl.val,
+ &ctl, 4, 1);
+
+ INIT_LIST_HEAD(&workload->list);
+
+ init_waitqueue_head(&workload->shadow_ctx_status_wq);
+ atomic_set(&workload->shadow_ctx_active, 0);
+
+ workload->vgt = vgt;
+ workload->ring_id = ring_id;
+ workload->ctx_desc = *desc;
+ workload->ring_context = ring_context;
+ workload->rb_head = head;
+ workload->rb_tail = tail;
+ workload->rb_start = start;
+ workload->rb_ctl = ctl;
+ workload->complete = execlist_workload_complete;
+ workload->status = -EINPROGRESS;
+
+ gvt_dbg_el("workload %p ring id %d head %x tail %x start %x ctl %x",
+ workload, ring_id, head, tail, start, ctl);
+
+ if (!prepare_workload(workload)) {
+ kfree(workload);
+ return false;
+ }
+
+ queue_workload(workload);
+
+ return true;
+}
+
+bool gvt_execlist_elsp_submit(struct vgt_device *vgt, int ring_id)
+{
+ struct gvt_virtual_execlist_info *info =
+ &vgt->virtual_execlist_info[ring_id];
+ struct execlist_ctx_descriptor_format *desc[2], valid_desc[2];
+ unsigned long valid_desc_bitmap = 0;
+ int i;
+
+ memset(valid_desc, 0, sizeof(valid_desc));
+
+ desc[0] = (struct execlist_ctx_descriptor_format *)&info->bundle.data[2];
+ desc[1] = (struct execlist_ctx_descriptor_format *)&info->bundle.data[0];
+
+ for (i = 0; i < 2; i++) {
+ if (!desc[i]->valid)
+ continue;
+
+ if (!desc[i]->privilege_access) {
+ gvt_err("[vgt %d] unexpected GGTT elsp submission", vgt->id);
+ return false;
+ }
+
+ /* TODO: add another guest context checks here. */
+ set_bit(i, &valid_desc_bitmap);
+ valid_desc[i] = *desc[i];
+ }
+
+ if (!valid_desc_bitmap) {
+ gvt_err("[vgt %d] no valid desc in a elsp submission",
+ vgt->id);
+ return false;
+ }
+
+ if (!test_bit(0, (void *)&valid_desc_bitmap) &&
+ test_bit(1, (void *)&valid_desc_bitmap)) {
+ gvt_err("[vgt %d] weird elsp submission, desc 0 is not valid",
+ vgt->id);
+ return false;
+ }
+
+ if (!emulate_execlist_schedule_in(info, valid_desc)) {
+ gvt_err("[vgt %d] fail to emulate execlist schedule-in", vgt->id);
+ return false;
+ }
+
+ /* submit workload */
+ for_each_set_bit(i, (void *)&valid_desc_bitmap, 2) {
+ if (!submit_context(vgt, ring_id, &valid_desc[i])) {
+ gvt_err("[vgt %d] fail to schedule workload", vgt->id);
+ return false;
+ }
+ }
+ return true;
+}
+
static bool init_virtual_execlist_info(struct vgt_device *vgt,
int ring_id, struct gvt_virtual_execlist_info *info)
{
@@ -325,6 +585,8 @@ static bool init_virtual_execlist_info(struct vgt_device *vgt,
info->execlist[0].index = 0;
info->execlist[1].index = 1;
+ INIT_LIST_HEAD(&info->workload_q_head);
+
ctx_status_ptr_reg = execlist_ring_mmio(info->ring_id,
_EL_OFFSET_STATUS_PTR);
@@ -339,6 +601,8 @@ bool gvt_init_virtual_execlist_info(struct vgt_device *vgt)
{
int i;
+ atomic_set(&vgt->running_workload_num, 0);
+
/* each ring has a virtual execlist engine */
for (i = 0; i < I915_NUM_RINGS; i++)
init_virtual_execlist_info(vgt,
@@ -107,20 +107,73 @@ struct execlist_context_status_format {
};
};
+struct gvt_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 gvt_execlist_mmio_pair ctx_ctrl;
+ struct gvt_execlist_mmio_pair ring_header;
+ struct gvt_execlist_mmio_pair ring_tail;
+ struct gvt_execlist_mmio_pair rb_start;
+ struct gvt_execlist_mmio_pair rb_ctrl;
+ struct gvt_execlist_mmio_pair bb_cur_head_UDW;
+ struct gvt_execlist_mmio_pair bb_cur_head_LDW;
+ struct gvt_execlist_mmio_pair bb_state;
+ struct gvt_execlist_mmio_pair second_bb_addr_UDW;
+ struct gvt_execlist_mmio_pair second_bb_addr_LDW;
+ struct gvt_execlist_mmio_pair second_bb_state;
+ struct gvt_execlist_mmio_pair bb_per_ctx_ptr;
+ struct gvt_execlist_mmio_pair rcs_indirect_ctx;
+ struct gvt_execlist_mmio_pair rcs_indirect_ctx_offset;
+ u32 nop2;
+ u32 nop3;
+ u32 nop4;
+ u32 lri_cmd_2;
+ struct gvt_execlist_mmio_pair ctx_timestamp;
+ struct gvt_execlist_mmio_pair pdp3_UDW;
+ struct gvt_execlist_mmio_pair pdp3_LDW;
+ struct gvt_execlist_mmio_pair pdp2_UDW;
+ struct gvt_execlist_mmio_pair pdp2_LDW;
+ struct gvt_execlist_mmio_pair pdp1_UDW;
+ struct gvt_execlist_mmio_pair pdp1_LDW;
+ struct gvt_execlist_mmio_pair pdp0_UDW;
+ struct gvt_execlist_mmio_pair pdp0_LDW;
+};
+
struct gvt_execlist_state {
struct execlist_ctx_descriptor_format ctx[2];
u32 index;
};
+struct gvt_execlist_write_bundle {
+ u32 data[4];
+ u32 index;
+};
+
struct gvt_virtual_execlist_info {
struct gvt_execlist_state execlist[2];
struct gvt_execlist_state *running_execlist;
struct gvt_execlist_state *pending_execlist;
struct execlist_ctx_descriptor_format *running_context;
+ struct gvt_execlist_write_bundle bundle;
int ring_id;
struct vgt_device *vgt;
+ struct list_head workload_q_head;
};
bool gvt_init_virtual_execlist_info(struct vgt_device *vgt);
+void gvt_get_context_pdp_root_pointer(struct vgt_device *vgt,
+ struct execlist_ring_context *ring_context,
+ u32 pdp[8]);
+
+void gvt_set_context_pdp_root_pointer(struct vgt_device *vgt,
+ struct execlist_ring_context *ring_context,
+ u32 pdp[8]);
+
#endif /*_GVT_EXECLIST_H_*/
@@ -43,6 +43,7 @@
#include "edid.h"
#include "display.h"
#include "execlist.h"
+#include "scheduler.h"
#define GVT_MAX_VGPU 8
@@ -151,6 +152,7 @@ struct vgt_device {
atomic_t active;
struct gvt_virtual_device_state state;
struct gvt_virtual_execlist_info virtual_execlist_info[I915_NUM_RINGS];
+ atomic_t running_workload_num;
struct gvt_statistics stat;
struct gvt_vgtt_info gtt;
void *hypervisor_data;
@@ -698,6 +700,8 @@ void gvt_emulate_display_events(struct pgt_device *pdev);
bool gvt_setup_control_interface(struct pgt_device *pdev);
void gvt_clean_control_interface(struct pgt_device *pdev);
+bool gvt_execlist_elsp_submit(struct vgt_device *vgt, int ring_id);
+
#include "mpt.h"
#endif
@@ -1100,6 +1100,26 @@ bool fpga_dbg_write(struct vgt_device *vgt, unsigned int reg,
return gvt_default_mmio_write(vgt, reg, &v, bytes);
}
+static bool elsp_write(struct vgt_device *vgt, unsigned int offset,
+ void *p_data, unsigned int bytes)
+{
+ int ring_id = gvt_render_mmio_to_ring_id(offset);
+ struct gvt_virtual_execlist_info *info =
+ &vgt->virtual_execlist_info[ring_id];
+ u32 data = *(u32 *)p_data;
+ bool rc = true;
+
+ info->bundle.data[info->bundle.index] = data;
+
+ if (info->bundle.index == 3)
+ rc = gvt_execlist_elsp_submit(vgt, ring_id);
+
+ ++info->bundle.index;
+ info->bundle.index &= 0x3;
+
+ return rc;
+}
+
struct gvt_reg_info gvt_general_reg_info[] = {
/* Interrupt registers - GT */
{_RING_IMR(RENDER_RING_BASE), 4, F_RDR, 0, D_ALL, NULL, gvt_reg_imr_handler},
@@ -1912,15 +1932,15 @@ struct gvt_reg_info gvt_broadwell_reg_info[] = {
{_REG_VCS2_TIMESTAMP, 8, F_PT, 0, D_BDW_PLUS, NULL, NULL},
{_REG_RCS_EXECLIST_SUBMITPORT, 4, F_VIRT, 0, D_BDW_PLUS,
- mmio_not_allow_read, NULL},
+ mmio_not_allow_read, elsp_write},
{_REG_VCS_EXECLIST_SUBMITPORT, 4, F_VIRT, 0, D_BDW_PLUS,
- mmio_not_allow_read, NULL},
+ mmio_not_allow_read, elsp_write},
{_REG_VECS_EXECLIST_SUBMITPORT, 4, F_VIRT, 0, D_BDW_PLUS,
- mmio_not_allow_read, NULL},
+ mmio_not_allow_read, elsp_write},
{_REG_VCS2_EXECLIST_SUBMITPORT, 4, F_VIRT, 0, D_BDW_PLUS,
- mmio_not_allow_read, NULL},
+ mmio_not_allow_read, elsp_write},
{_REG_BCS_EXECLIST_SUBMITPORT, 4, F_VIRT, 0, D_BDW_PLUS,
- mmio_not_allow_read, NULL},
+ mmio_not_allow_read, elsp_write},
{_REG_RCS_EXECLIST_STATUS, 8, F_RDR, 0, D_BDW_PLUS, NULL,
mmio_not_allow_write},
new file mode 100644
@@ -0,0 +1,59 @@
+/*
+ * 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 gvt_workload {
+ struct vgt_device *vgt;
+ int ring_id;
+
+ struct drm_i915_gem_request *req;
+
+ /* if this workload has been dispatched to i915? */
+ bool dispatched;
+ int status;
+ atomic_t shadow_ctx_active;
+ wait_queue_head_t shadow_ctx_status_wq;
+
+ /* 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 gvt_mm *shadow_mm;
+
+ /* different submission model may need different complete handler */
+ bool (*complete)(struct gvt_workload *);
+
+ struct list_head list;
+};
+
+#define workload_q_head(vgt, ring_id) \
+ (&(vgt->virtual_execlist_info[ring_id].workload_q_head))
+
+#define queue_workload(workload) \
+ list_add_tail(&workload->list, \
+ workload_q_head(workload->vgt, workload->ring_id))
+
+#endif
This patch introduces the vGPU execlist 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: Zhi Wang <zhi.a.wang@intel.com> --- drivers/gpu/drm/i915/gvt/execlist.c | 264 +++++++++++++++++++++++++++++++++++ drivers/gpu/drm/i915/gvt/execlist.h | 53 +++++++ drivers/gpu/drm/i915/gvt/gvt.h | 4 + drivers/gpu/drm/i915/gvt/handlers.c | 30 +++- drivers/gpu/drm/i915/gvt/scheduler.h | 59 ++++++++ 5 files changed, 405 insertions(+), 5 deletions(-) create mode 100644 drivers/gpu/drm/i915/gvt/scheduler.h