]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
drm/i915/gvt: vGPU execlist virtualization
authorZhi Wang <zhi.a.wang@intel.com>
Sun, 1 May 2016 06:48:25 +0000 (02:48 -0400)
committerZhenyu Wang <zhenyuw@linux.intel.com>
Fri, 14 Oct 2016 10:14:28 +0000 (18:14 +0800)
This patch introduces the vGPU execlist virtualization.

Under virtulization environment, HW execlist interface are fully emulated
including virtual CSB emulation, virtual execlist emulation. The framework
will emulate the virtual CSB according to the guest workload running status

Signed-off-by: Zhi Wang <zhi.a.wang@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
drivers/gpu/drm/i915/gvt/Makefile
drivers/gpu/drm/i915/gvt/debug.h
drivers/gpu/drm/i915/gvt/execlist.c [new file with mode: 0644]
drivers/gpu/drm/i915/gvt/execlist.h [new file with mode: 0644]
drivers/gpu/drm/i915/gvt/gvt.h
drivers/gpu/drm/i915/gvt/vgpu.c

index f87cd7fe95743c1fcd8ec46ad9f313025a3ea644..f15197390d768fe40658271f91ea6bc0fbcd07bf 100644 (file)
@@ -1,6 +1,7 @@
 GVT_DIR := gvt
 GVT_SOURCE := gvt.o aperture_gm.o handlers.o vgpu.o trace_points.o firmware.o \
-       interrupt.o gtt.o cfg_space.o opregion.o mmio.o display.o edid.o
+       interrupt.o gtt.o cfg_space.o opregion.o mmio.o display.o edid.o \
+       execlist.o
 
 ccflags-y                      += -I$(src) -I$(src)/$(GVT_DIR) -Wall
 i915-y                        += $(addprefix $(GVT_DIR)/, $(GVT_SOURCE))
index 5c21c585ac00469e75510d645ce714d4dd522d49..a7e0e7956ee971c8a9c4370061faf3b299989e7b 100644 (file)
@@ -42,4 +42,7 @@
 #define gvt_dbg_dpy(fmt, args...) \
        DRM_DEBUG_DRIVER("gvt: dpy: "fmt, ##args)
 
+#define gvt_dbg_el(fmt, args...) \
+       DRM_DEBUG_DRIVER("gvt: el: "fmt, ##args)
+
 #endif
diff --git a/drivers/gpu/drm/i915/gvt/execlist.c b/drivers/gpu/drm/i915/gvt/execlist.c
new file mode 100644 (file)
index 0000000..3c38e2a
--- /dev/null
@@ -0,0 +1,386 @@
+/*
+ * 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.
+ *
+ * Authors:
+ *    Zhiyuan Lv <zhiyuan.lv@intel.com>
+ *    Zhi Wang <zhi.a.wang@intel.com>
+ *
+ * Contributors:
+ *    Min He <min.he@intel.com>
+ *    Bing Niu <bing.niu@intel.com>
+ *    Ping Gao <ping.a.gao@intel.com>
+ *    Tina Zhang <tina.zhang@intel.com>
+ *
+ */
+
+#include "i915_drv.h"
+
+#define _EL_OFFSET_STATUS       0x234
+#define _EL_OFFSET_STATUS_BUF   0x370
+#define _EL_OFFSET_STATUS_PTR   0x3A0
+
+#define execlist_ring_mmio(gvt, ring_id, offset) \
+       (gvt->dev_priv->engine[ring_id].mmio_base + (offset))
+
+#define valid_context(ctx) ((ctx)->valid)
+#define same_context(a, b) (((a)->context_id == (b)->context_id) && \
+               ((a)->lrca == (b)->lrca))
+
+static int context_switch_events[] = {
+       [RCS] = RCS_AS_CONTEXT_SWITCH,
+       [BCS] = BCS_AS_CONTEXT_SWITCH,
+       [VCS] = VCS_AS_CONTEXT_SWITCH,
+       [VCS2] = VCS2_AS_CONTEXT_SWITCH,
+       [VECS] = VECS_AS_CONTEXT_SWITCH,
+};
+
+static int ring_id_to_context_switch_event(int ring_id)
+{
+       if (WARN_ON(ring_id < RCS && ring_id >
+                               ARRAY_SIZE(context_switch_events)))
+               return -EINVAL;
+
+       return context_switch_events[ring_id];
+}
+
+static void switch_virtual_execlist_slot(struct intel_vgpu_execlist *execlist)
+{
+       gvt_dbg_el("[before] running slot %d/context %x pending slot %d\n",
+                       execlist->running_slot ?
+                       execlist->running_slot->index : -1,
+                       execlist->running_context ?
+                       execlist->running_context->context_id : 0,
+                       execlist->pending_slot ?
+                       execlist->pending_slot->index : -1);
+
+       execlist->running_slot = execlist->pending_slot;
+       execlist->pending_slot = NULL;
+       execlist->running_context = execlist->running_context ?
+               &execlist->running_slot->ctx[0] : NULL;
+
+       gvt_dbg_el("[after] running slot %d/context %x pending slot %d\n",
+                       execlist->running_slot ?
+                       execlist->running_slot->index : -1,
+                       execlist->running_context ?
+                       execlist->running_context->context_id : 0,
+                       execlist->pending_slot ?
+                       execlist->pending_slot->index : -1);
+}
+
+static void emulate_execlist_status(struct intel_vgpu_execlist *execlist)
+{
+       struct intel_vgpu_execlist_slot *running = execlist->running_slot;
+       struct intel_vgpu_execlist_slot *pending = execlist->pending_slot;
+       struct execlist_ctx_descriptor_format *desc = execlist->running_context;
+       struct intel_vgpu *vgpu = execlist->vgpu;
+       struct execlist_status_format status;
+       int ring_id = execlist->ring_id;
+       u32 status_reg = execlist_ring_mmio(vgpu->gvt,
+                       ring_id, _EL_OFFSET_STATUS);
+
+       status.ldw = vgpu_vreg(vgpu, status_reg);
+       status.udw = vgpu_vreg(vgpu, status_reg + 4);
+
+       if (running) {
+               status.current_execlist_pointer = !!running->index;
+               status.execlist_write_pointer = !!!running->index;
+               status.execlist_0_active = status.execlist_0_valid =
+                       !!!(running->index);
+               status.execlist_1_active = status.execlist_1_valid =
+                       !!(running->index);
+       } else {
+               status.context_id = 0;
+               status.execlist_0_active = status.execlist_0_valid = 0;
+               status.execlist_1_active = status.execlist_1_valid = 0;
+       }
+
+       status.context_id = desc ? desc->context_id : 0;
+       status.execlist_queue_full = !!(pending);
+
+       vgpu_vreg(vgpu, status_reg) = status.ldw;
+       vgpu_vreg(vgpu, status_reg + 4) = status.udw;
+
+       gvt_dbg_el("vgpu%d: status reg offset %x ldw %x udw %x\n",
+               vgpu->id, status_reg, status.ldw, status.udw);
+}
+
+static void emulate_csb_update(struct intel_vgpu_execlist *execlist,
+               struct execlist_context_status_format *status,
+               bool trigger_interrupt_later)
+{
+       struct intel_vgpu *vgpu = execlist->vgpu;
+       int ring_id = execlist->ring_id;
+       struct execlist_context_status_pointer_format ctx_status_ptr;
+       u32 write_pointer;
+       u32 ctx_status_ptr_reg, ctx_status_buf_reg, offset;
+
+       ctx_status_ptr_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
+                       _EL_OFFSET_STATUS_PTR);
+       ctx_status_buf_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
+                       _EL_OFFSET_STATUS_BUF);
+
+       ctx_status_ptr.dw = vgpu_vreg(vgpu, ctx_status_ptr_reg);
+
+       write_pointer = ctx_status_ptr.write_ptr;
+
+       if (write_pointer == 0x7)
+               write_pointer = 0;
+       else {
+               ++write_pointer;
+               write_pointer %= 0x6;
+       }
+
+       offset = ctx_status_buf_reg + write_pointer * 8;
+
+       vgpu_vreg(vgpu, offset) = status->ldw;
+       vgpu_vreg(vgpu, offset + 4) = status->udw;
+
+       ctx_status_ptr.write_ptr = write_pointer;
+       vgpu_vreg(vgpu, ctx_status_ptr_reg) = ctx_status_ptr.dw;
+
+       gvt_dbg_el("vgpu%d: w pointer %u reg %x csb l %x csb h %x\n",
+               vgpu->id, write_pointer, offset, status->ldw, status->udw);
+
+       if (trigger_interrupt_later)
+               return;
+
+       intel_vgpu_trigger_virtual_event(vgpu,
+                       ring_id_to_context_switch_event(execlist->ring_id));
+}
+
+int emulate_execlist_ctx_schedule_out(
+               struct intel_vgpu_execlist *execlist,
+               struct execlist_ctx_descriptor_format *ctx)
+{
+       struct intel_vgpu_execlist_slot *running = execlist->running_slot;
+       struct intel_vgpu_execlist_slot *pending = execlist->pending_slot;
+       struct execlist_ctx_descriptor_format *ctx0 = &running->ctx[0];
+       struct execlist_ctx_descriptor_format *ctx1 = &running->ctx[1];
+       struct execlist_context_status_format status;
+
+       memset(&status, 0, sizeof(status));
+
+       gvt_dbg_el("schedule out context id %x\n", ctx->context_id);
+
+       if (WARN_ON(!same_context(ctx, execlist->running_context))) {
+               gvt_err("schedule out context is not running context,"
+                               "ctx id %x running ctx id %x\n",
+                               ctx->context_id,
+                               execlist->running_context->context_id);
+               return -EINVAL;
+       }
+
+       /* ctx1 is valid, ctx0/ctx is scheduled-out -> element switch */
+       if (valid_context(ctx1) && same_context(ctx0, ctx)) {
+               gvt_dbg_el("ctx 1 valid, ctx/ctx 0 is scheduled-out\n");
+
+               execlist->running_context = ctx1;
+
+               emulate_execlist_status(execlist);
+
+               status.context_complete = status.element_switch = 1;
+               status.context_id = ctx->context_id;
+
+               emulate_csb_update(execlist, &status, false);
+               /*
+                * ctx1 is not valid, ctx == ctx0
+                * ctx1 is valid, ctx1 == ctx
+                *      --> last element is finished
+                * emulate:
+                *      active-to-idle if there is *no* pending execlist
+                *      context-complete if there *is* pending execlist
+                */
+       } else if ((!valid_context(ctx1) && same_context(ctx0, ctx))
+                       || (valid_context(ctx1) && same_context(ctx1, ctx))) {
+               gvt_dbg_el("need to switch virtual execlist slot\n");
+
+               switch_virtual_execlist_slot(execlist);
+
+               emulate_execlist_status(execlist);
+
+               status.context_complete = status.active_to_idle = 1;
+               status.context_id = ctx->context_id;
+
+               if (!pending) {
+                       emulate_csb_update(execlist, &status, false);
+               } else {
+                       emulate_csb_update(execlist, &status, true);
+
+                       memset(&status, 0, sizeof(status));
+
+                       status.idle_to_active = 1;
+                       status.context_id = 0;
+
+                       emulate_csb_update(execlist, &status, false);
+               }
+       } else {
+               WARN_ON(1);
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static struct intel_vgpu_execlist_slot *get_next_execlist_slot(
+               struct intel_vgpu_execlist *execlist)
+{
+       struct intel_vgpu *vgpu = execlist->vgpu;
+       int ring_id = execlist->ring_id;
+       u32 status_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
+                       _EL_OFFSET_STATUS);
+       struct execlist_status_format status;
+
+       status.ldw = vgpu_vreg(vgpu, status_reg);
+       status.udw = vgpu_vreg(vgpu, status_reg + 4);
+
+       if (status.execlist_queue_full) {
+               gvt_err("virtual execlist slots are full\n");
+               return NULL;
+       }
+
+       return &execlist->slot[status.execlist_write_pointer];
+}
+
+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;
+       struct intel_vgpu_execlist_slot *slot =
+               get_next_execlist_slot(execlist);
+
+       struct execlist_ctx_descriptor_format *ctx0, *ctx1;
+       struct execlist_context_status_format status;
+
+       gvt_dbg_el("emulate schedule-in\n");
+
+       if (!slot) {
+               gvt_err("no available execlist slot\n");
+               return -EINVAL;
+       }
+
+       memset(&status, 0, sizeof(status));
+       memset(slot->ctx, 0, sizeof(slot->ctx));
+
+       slot->ctx[0] = ctx[0];
+       slot->ctx[1] = ctx[1];
+
+       gvt_dbg_el("alloc slot index %d ctx 0 %x ctx 1 %x\n",
+                       slot->index, ctx[0].context_id,
+                       ctx[1].context_id);
+
+       /*
+        * no running execlist, make this write bundle as running execlist
+        * -> idle-to-active
+        */
+       if (!running) {
+               gvt_dbg_el("no current running execlist\n");
+
+               execlist->running_slot = slot;
+               execlist->pending_slot = NULL;
+               execlist->running_context = &slot->ctx[0];
+
+               gvt_dbg_el("running slot index %d running context %x\n",
+                               execlist->running_slot->index,
+                               execlist->running_context->context_id);
+
+               emulate_execlist_status(execlist);
+
+               status.idle_to_active = 1;
+               status.context_id = 0;
+
+               emulate_csb_update(execlist, &status, false);
+               return 0;
+       }
+
+       ctx0 = &running->ctx[0];
+       ctx1 = &running->ctx[1];
+
+       gvt_dbg_el("current running slot index %d ctx 0 %x ctx 1 %x\n",
+               running->index, ctx0->context_id, ctx1->context_id);
+
+       /*
+        * already has an running execlist
+        *      a. running ctx1 is valid,
+        *         ctx0 is finished, and running ctx1 == new execlist ctx[0]
+        *      b. running ctx1 is not valid,
+        *         ctx0 == new execlist ctx[0]
+        * ----> lite-restore + preempted
+        */
+       if ((valid_context(ctx1) && same_context(ctx1, &slot->ctx[0]) &&
+               /* condition a */
+               (!same_context(ctx0, execlist->running_context))) ||
+                       (!valid_context(ctx1) &&
+                        same_context(ctx0, &slot->ctx[0]))) { /* condition b */
+               gvt_dbg_el("need to switch virtual execlist slot\n");
+
+               execlist->pending_slot = slot;
+               switch_virtual_execlist_slot(execlist);
+
+               emulate_execlist_status(execlist);
+
+               status.lite_restore = status.preempted = 1;
+               status.context_id = ctx[0].context_id;
+
+               emulate_csb_update(execlist, &status, false);
+       } else {
+               gvt_dbg_el("emulate as pending slot\n");
+               /*
+                * otherwise
+                * --> emulate pending execlist exist + but no preemption case
+                */
+               execlist->pending_slot = slot;
+               emulate_execlist_status(execlist);
+       }
+       return 0;
+}
+
+static void init_vgpu_execlist(struct intel_vgpu *vgpu, int ring_id)
+{
+       struct intel_vgpu_execlist *execlist = &vgpu->execlist[ring_id];
+       struct execlist_context_status_pointer_format ctx_status_ptr;
+       u32 ctx_status_ptr_reg;
+
+       memset(execlist, 0, sizeof(*execlist));
+
+       execlist->vgpu = vgpu;
+       execlist->ring_id = ring_id;
+       execlist->slot[0].index = 0;
+       execlist->slot[1].index = 1;
+
+       ctx_status_ptr_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
+                       _EL_OFFSET_STATUS_PTR);
+
+       ctx_status_ptr.dw = vgpu_vreg(vgpu, ctx_status_ptr_reg);
+       ctx_status_ptr.read_ptr = ctx_status_ptr.write_ptr = 0x7;
+       vgpu_vreg(vgpu, ctx_status_ptr_reg) = ctx_status_ptr.dw;
+}
+
+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++)
+               init_vgpu_execlist(vgpu, i);
+
+       return 0;
+}
diff --git a/drivers/gpu/drm/i915/gvt/execlist.h b/drivers/gpu/drm/i915/gvt/execlist.h
new file mode 100644 (file)
index 0000000..428f543
--- /dev/null
@@ -0,0 +1,137 @@
+/*
+ * 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.
+ *
+ * Authors:
+ *    Zhiyuan Lv <zhiyuan.lv@intel.com>
+ *    Zhi Wang <zhi.a.wang@intel.com>
+ *
+ * Contributors:
+ *    Min He <min.he@intel.com>
+ *    Bing Niu <bing.niu@intel.com>
+ *    Ping Gao <ping.a.gao@intel.com>
+ *    Tina Zhang <tina.zhang@intel.com>
+ *
+ */
+
+#ifndef _GVT_EXECLIST_H_
+#define _GVT_EXECLIST_H_
+
+struct execlist_ctx_descriptor_format {
+       union {
+               u32 udw;
+               u32 context_id;
+       };
+       union {
+               u32 ldw;
+               struct {
+                       u32 valid                  : 1;
+                       u32 force_pd_restore       : 1;
+                       u32 force_restore          : 1;
+                       u32 addressing_mode        : 2;
+                       u32 llc_coherency          : 1;
+                       u32 fault_handling         : 2;
+                       u32 privilege_access       : 1;
+                       u32 reserved               : 3;
+                       u32 lrca                   : 20;
+               };
+       };
+};
+
+struct execlist_status_format {
+       union {
+               u32 ldw;
+               struct {
+                       u32 current_execlist_pointer       :1;
+                       u32 execlist_write_pointer         :1;
+                       u32 execlist_queue_full            :1;
+                       u32 execlist_1_valid               :1;
+                       u32 execlist_0_valid               :1;
+                       u32 last_ctx_switch_reason         :9;
+                       u32 current_active_elm_status      :2;
+                       u32 arbitration_enable             :1;
+                       u32 execlist_1_active              :1;
+                       u32 execlist_0_active              :1;
+                       u32 reserved                       :13;
+               };
+       };
+       union {
+               u32 udw;
+               u32 context_id;
+       };
+};
+
+struct execlist_context_status_pointer_format {
+       union {
+               u32 dw;
+               struct {
+                       u32 write_ptr              :3;
+                       u32 reserved               :5;
+                       u32 read_ptr               :3;
+                       u32 reserved2              :5;
+                       u32 mask                   :16;
+               };
+       };
+};
+
+struct execlist_context_status_format {
+       union {
+               u32 ldw;
+               struct {
+                       u32 idle_to_active         :1;
+                       u32 preempted              :1;
+                       u32 element_switch         :1;
+                       u32 active_to_idle         :1;
+                       u32 context_complete       :1;
+                       u32 wait_on_sync_flip      :1;
+                       u32 wait_on_vblank         :1;
+                       u32 wait_on_semaphore      :1;
+                       u32 wait_on_scanline       :1;
+                       u32 reserved               :2;
+                       u32 semaphore_wait_mode    :1;
+                       u32 display_plane          :3;
+                       u32 lite_restore           :1;
+                       u32 reserved_2             :16;
+               };
+       };
+       union {
+               u32 udw;
+               u32 context_id;
+       };
+};
+
+struct intel_vgpu_execlist_slot {
+       struct execlist_ctx_descriptor_format ctx[2];
+       u32 index;
+};
+
+struct intel_vgpu_execlist {
+       struct intel_vgpu_execlist_slot slot[2];
+       struct intel_vgpu_execlist_slot *running_slot;
+       struct intel_vgpu_execlist_slot *pending_slot;
+       struct execlist_ctx_descriptor_format *running_context;
+       int ring_id;
+       struct intel_vgpu *vgpu;
+};
+
+int intel_vgpu_init_execlist(struct intel_vgpu *vgpu);
+
+#endif /*_GVT_EXECLIST_H_*/
index 1619881dbd51dfb3ae3b03779419d312c25afae6..82f932a24e7d41a2f3924ad65800813757cd0819 100644 (file)
@@ -41,6 +41,7 @@
 #include "gtt.h"
 #include "display.h"
 #include "edid.h"
+#include "execlist.h"
 
 #define GVT_MAX_VGPU 8
 
@@ -146,6 +147,8 @@ 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 intel_gvt_gm {
index 47ed0a085e4c9667a648cf28f9f3cb8199cb89e0..e806b0e4035a5f02e1565cdebf7417ee5dc1729e 100644 (file)
@@ -221,11 +221,17 @@ struct intel_vgpu *intel_gvt_create_vgpu(struct intel_gvt *gvt,
        if (ret)
                goto out_clean_opregion;
 
+       ret = intel_vgpu_init_execlist(vgpu);
+       if (ret)
+               goto out_clean_display;
+
        vgpu->active = true;
        mutex_unlock(&gvt->lock);
 
        return vgpu;
 
+out_clean_display:
+       intel_vgpu_clean_display(vgpu);
 out_clean_opregion:
        intel_vgpu_clean_opregion(vgpu);
 out_clean_gtt: