[PATCH i-g-t 1/5] igt: Exercise creating context with shared GTT

From: Chris Wilson <chris@chris-wilson.co.uk>
To: intel-gfx@lists.freedesktop.org
Cc: igt-dev@lists.freedesktop.org
Subject: [PATCH i-g-t 1/5] igt: Exercise creating context with shared GTT
Date: Wed, 27 Jun 2018 15:46:35 +0100	[thread overview]
Message-ID: <20180627144639.27796-1-chris@chris-wilson.co.uk> (raw)

v2: Test each shared context is its own timeline and allows request
reordering between shared contexts.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 lib/i915/gem_context.c   |  66 +++
 lib/i915/gem_context.h   |  13 +
 tests/Makefile.sources   |   1 +
 tests/gem_ctx_shared.c   | 896 +++++++++++++++++++++++++++++++++++++++
 tests/gem_exec_whisper.c |  32 +-
 5 files changed, 999 insertions(+), 9 deletions(-)
 create mode 100644 tests/gem_ctx_shared.c

diff --git a/lib/i915/gem_context.c b/lib/i915/gem_context.c
index 669bd318c..02cf2ccf3 100644
--- a/lib/i915/gem_context.c
+++ b/lib/i915/gem_context.c
@@ -273,3 +273,69 @@ void gem_context_set_priority(int fd, uint32_t ctx_id, int prio)
 {
 	igt_assert(__gem_context_set_priority(fd, ctx_id, prio) == 0);
 }
+
+struct local_i915_gem_context_create_v2 {
+	uint32_t ctx_id;
+	uint32_t flags;
+	uint32_t share_ctx;
+	uint32_t pad;
+};
+
+#define LOCAL_IOCTL_I915_GEM_CONTEXT_CREATE       DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GEM_CONTEXT_CREATE, struct local_i915_gem_context_create_v2)
+
+int
+__gem_context_create_shared(int i915, uint32_t share, unsigned int flags,
+			    uint32_t *out)
+{
+	struct local_i915_gem_context_create_v2 arg = {
+		.flags = flags,
+		.share_ctx = share,
+	};
+	int err = 0;
+
+	if (igt_ioctl(i915, LOCAL_IOCTL_I915_GEM_CONTEXT_CREATE, &arg))
+		err = -errno;
+
+	*out = arg.ctx_id;
+
+	errno = 0;
+	return err;
+}
+
+static bool __gem_context_has(int i915, uint32_t flags)
+{
+	uint32_t ctx;
+
+	__gem_context_create_shared(i915, 0, flags, &ctx);
+	if (ctx)
+		gem_context_destroy(i915, ctx);
+
+	errno = 0;
+	return ctx;
+}
+
+bool gem_contexts_has_shared_gtt(int i915)
+{
+	return __gem_context_has(i915, I915_GEM_CONTEXT_SHARE_GTT);
+}
+
+bool gem_has_queues(int i915)
+{
+	return __gem_context_has(i915, I915_GEM_CONTEXT_SHARE_GTT);
+}
+
+uint32_t gem_context_create_shared(int i915, uint32_t share, unsigned int flags)
+{
+	uint32_t ctx;
+
+	igt_assert_eq(__gem_context_create_shared(i915, share, flags, &ctx), 0);
+
+	return ctx;
+}
+
+uint32_t gem_queue_create(int i915)
+{
+	return gem_context_create_shared(i915, 0,
+					 I915_GEM_CONTEXT_SHARE_GTT |
+					 I915_GEM_CONTEXT_SINGLE_TIMELINE);
+}
diff --git a/lib/i915/gem_context.h b/lib/i915/gem_context.h
index aef68dda6..5ed992a70 100644
--- a/lib/i915/gem_context.h
+++ b/lib/i915/gem_context.h
@@ -29,6 +29,19 @@ int __gem_context_create(int fd, uint32_t *ctx_id);
 void gem_context_destroy(int fd, uint32_t ctx_id);
 int __gem_context_destroy(int fd, uint32_t ctx_id);
 
+#define I915_GEM_CONTEXT_SHARE_GTT 0x1
+#define I915_GEM_CONTEXT_SINGLE_TIMELINE 0x2
+
+int __gem_context_create_shared(int i915,
+				uint32_t share, unsigned int flags,
+				uint32_t *out);
+uint32_t gem_context_create_shared(int i915,
+				   uint32_t share, unsigned int flags);
+uint32_t gem_queue_create(int i915);
+
+bool gem_contexts_has_shared_gtt(int i915);
+bool gem_has_queues(int i915);
+
 bool gem_has_contexts(int fd);
 void gem_require_contexts(int fd);
 void gem_context_require_bannable(int fd);
diff --git a/tests/Makefile.sources b/tests/Makefile.sources
index ad62611fb..bd0bd5439 100644
--- a/tests/Makefile.sources
+++ b/tests/Makefile.sources
@@ -60,6 +60,7 @@ TESTS_progs = \
 	gem_ctx_exec \
 	gem_ctx_isolation \
 	gem_ctx_param \
+	gem_ctx_shared \
 	gem_ctx_switch \
 	gem_ctx_thrash \
 	gem_double_irq_loop \
diff --git a/tests/gem_ctx_shared.c b/tests/gem_ctx_shared.c
new file mode 100644
index 000000000..9d3239f69
--- /dev/null
+++ b/tests/gem_ctx_shared.c
@@ -0,0 +1,896 @@
+/*
+ * Copyright © 2017 Intel Corporation
+ *
+ * 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.
+ *
+ */
+
+#include "igt.h"
+
+#include <unistd.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <fcntl.h>
+#include <inttypes.h>
+#include <errno.h>
+#include <sys/stat.h>
+#include <sys/ioctl.h>
+#include <sys/time.h>
+
+#include <drm.h>
+
+#include "igt_rand.h"
+#include "igt_vgem.h"
+#include "sync_file.h"
+
+#define LO 0
+#define HI 1
+#define NOISE 2
+
+#define MAX_PRIO LOCAL_I915_CONTEXT_MAX_USER_PRIORITY
+#define MIN_PRIO LOCAL_I915_CONTEXT_MIN_USER_PRIORITY
+
+static int priorities[] = {
+	[LO] = MIN_PRIO / 2,
+	[HI] = MAX_PRIO / 2,
+};
+
+#define BUSY_QLEN 8
+#define MAX_ELSP_QLEN 16
+
+IGT_TEST_DESCRIPTION("Test shared contexts.");
+
+static void create_shared_gtt(int i915, unsigned int flags)
+#define DETACHED 0x1
+{
+	const uint32_t bbe = MI_BATCH_BUFFER_END;
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = gem_create(i915, 4096),
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+	};
+	uint32_t parent, child;
+
+	gem_write(i915, obj.handle, 0, &bbe, sizeof(bbe));
+	gem_execbuf(i915, &execbuf);
+	gem_sync(i915, obj.handle);
+
+	child = flags & DETACHED ? gem_context_create(i915) : 0;
+	igt_until_timeout(2) {
+		parent = flags & DETACHED ? child : 0;
+		child = gem_context_create_shared(i915, parent,
+						  I915_GEM_CONTEXT_SHARE_GTT);
+		execbuf.rsvd1 = child;
+		gem_execbuf(i915, &execbuf);
+
+		if (flags & DETACHED) {
+			gem_context_destroy(i915, parent);
+			gem_execbuf(i915, &execbuf);
+		} else {
+			parent = child;
+			gem_context_destroy(i915, parent);
+		}
+
+		execbuf.rsvd1 = parent;
+		igt_assert_eq(__gem_execbuf(i915, &execbuf), -ENOENT);
+		igt_assert_eq(__gem_context_create_shared(i915, parent,
+							  I915_GEM_CONTEXT_SHARE_GTT,
+							  &parent), -ENOENT);
+	}
+	if (flags & DETACHED)
+		gem_context_destroy(i915, child);
+
+	gem_sync(i915, obj.handle);
+	gem_close(i915, obj.handle);
+}
+
+static void disjoint_timelines(int i915)
+{
+	const uint32_t bbe = MI_BATCH_BUFFER_END;
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = gem_create(i915, 4096),
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+	};
+	struct sync_fence_info parent, child;
+	struct sync_file_info sync_file_info = {
+		.num_fences = 1,
+	};
+
+	/*
+	 * Each context, although they share a vm, are expected to be
+	 * distinct timelines. A request queued to one context should be
+	 * independent of any shared contexts.
+	 *
+	 * This information is exposed via the sync_file->name which
+	 * includes the fence.context.
+	 */
+
+	gem_write(i915, obj.handle, 0, &bbe, sizeof(bbe));
+	gem_execbuf(i915, &execbuf);
+	gem_sync(i915, obj.handle);
+
+	execbuf.flags = I915_EXEC_FENCE_OUT;
+	gem_execbuf_wr(i915, &execbuf);
+	sync_file_info.sync_fence_info = to_user_pointer(&parent);
+	do_ioctl(execbuf.rsvd2 >> 32, SYNC_IOC_FILE_INFO, &sync_file_info);
+	close(execbuf.rsvd2 >> 32);
+
+	execbuf.rsvd1 =
+		gem_context_create_shared(i915, 0, I915_GEM_CONTEXT_SHARE_GTT);
+	gem_execbuf_wr(i915, &execbuf);
+	sync_file_info.sync_fence_info = to_user_pointer(&child);
+	do_ioctl(execbuf.rsvd2 >> 32, SYNC_IOC_FILE_INFO, &sync_file_info);
+	close(execbuf.rsvd2 >> 32);
+	gem_context_destroy(i915, execbuf.rsvd1);
+
+	igt_info("Parent fence: %s %s\n", parent.driver_name, parent.obj_name);
+	igt_info("Child fence: %s %s\n", child.driver_name, child.obj_name);
+
+	/* Driver should be the same, but on different timelines */
+	igt_assert(strcmp(parent.driver_name, child.driver_name) == 0);
+	igt_assert(strcmp(parent.obj_name, child.obj_name) != 0);
+
+	gem_sync(i915, obj.handle);
+	gem_close(i915, obj.handle);
+}
+
+static int reopen_driver(int fd)
+{
+	char path[256];
+
+	snprintf(path, sizeof(path), "/proc/self/fd/%d", fd);
+	fd = open(path, O_RDWR);
+	igt_assert_lte(0, fd);
+
+	return fd;
+}
+
+static void exhaust_shared_gtt(int i915, unsigned int flags)
+#define EXHAUST_LRC 0x1
+{
+	i915 = reopen_driver(i915);
+
+	igt_fork(pid, 1) {
+		const uint32_t bbe = MI_BATCH_BUFFER_END;
+		struct drm_i915_gem_exec_object2 obj = {
+			.handle = gem_create(i915, 4096)
+		};
+		struct drm_i915_gem_execbuffer2 execbuf = {
+			.buffers_ptr = to_user_pointer(&obj),
+			.buffer_count = 1,
+		};
+		uint32_t parent, child;
+		unsigned long count = 0;
+		int err;
+
+		gem_write(i915, obj.handle, 0, &bbe, sizeof(bbe));
+
+		child = 0;
+		for (;;) {
+			parent = child;
+			err = __gem_context_create_shared(i915, parent,
+							  I915_GEM_CONTEXT_SHARE_GTT,
+							  &child);
+			if (err)
+				break;
+
+			if (flags & EXHAUST_LRC) {
+				execbuf.rsvd1 = child;
+				err = __gem_execbuf(i915, &execbuf);
+				if (err)
+					break;
+			}
+
+			count++;
+		}
+		gem_sync(i915, obj.handle);
+
+		igt_info("Created %lu shared contexts, before %d (%s)\n",
+			 count, err, strerror(-err));
+	}
+	close(i915);
+	igt_waitchildren();
+}
+
+static void exec_shared_gtt(int i915, unsigned int ring)
+{
+	const int gen = intel_gen(intel_get_drm_devid(i915));
+	const uint32_t bbe = MI_BATCH_BUFFER_END;
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = gem_create(i915, 4096)
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.flags = ring,
+	};
+	uint32_t scratch = obj.handle;
+	uint32_t batch[16];
+	int i;
+
+	gem_require_ring(i915, ring);
+	igt_require(gem_can_store_dword(i915, ring));
+
+	/* Load object into place in the GTT */
+	gem_write(i915, obj.handle, 0, &bbe, sizeof(bbe));
+	gem_execbuf(i915, &execbuf);
+
+	/* Presume nothing causes an eviction in the meantime */
+
+	obj.handle = gem_create(i915, 4096);
+
+	i = 0;
+	batch[i] = MI_STORE_DWORD_IMM | (gen < 6 ? 1 << 22 : 0);
+	if (gen >= 8) {
+		batch[++i] = obj.offset;
+		batch[++i] = 0;
+	} else if (gen >= 4) {
+		batch[++i] = 0;
+		batch[++i] = obj.offset;
+	} else {
+		batch[i]--;
+		batch[++i] = obj.offset;
+	}
+	batch[++i] = 0xc0ffee;
+	batch[++i] = MI_BATCH_BUFFER_END;
+	gem_write(i915, obj.handle, 0, batch, sizeof(batch));
+
+	obj.offset += 4096; /* make sure we don't cause an eviction! */
+	obj.flags |= EXEC_OBJECT_PINNED;
+	execbuf.rsvd1 = gem_context_create_shared(i915, 0,
+						  I915_GEM_CONTEXT_SHARE_GTT);
+	if (gen > 3 && gen < 6)
+		execbuf.flags |= I915_EXEC_SECURE;
+
+	gem_execbuf(i915, &execbuf);
+	gem_context_destroy(i915, execbuf.rsvd1);
+	gem_sync(i915, obj.handle); /* write hazard lies */
+	gem_close(i915, obj.handle);
+
+	gem_read(i915, scratch, 0, batch, sizeof(uint32_t));
+	gem_close(i915, scratch);
+
+	igt_assert_eq_u32(*batch, 0xc0ffee);
+}
+
+static int nop_sync(int i915, uint32_t ctx, unsigned int ring, int64_t timeout)
+{
+	const uint32_t bbe = MI_BATCH_BUFFER_END;
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = gem_create(i915, 4096),
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.flags = ring,
+		.rsvd1 = ctx,
+	};
+	int err;
+
+	gem_write(i915, obj.handle, 0, &bbe, sizeof(bbe));
+	gem_execbuf(i915, &execbuf);
+	err = gem_wait(i915, obj.handle, &timeout);
+	gem_close(i915, obj.handle);
+
+	return err;
+}
+
+static bool has_single_timeline(int i915)
+{
+	uint32_t ctx;
+
+	__gem_context_create_shared(i915, 0,
+				    I915_GEM_CONTEXT_SINGLE_TIMELINE, &ctx);
+	if (ctx)
+		gem_context_destroy(i915, ctx);
+
+	return ctx != 0;
+}
+
+static bool ignore_engine(unsigned engine)
+{
+	if (engine == 0)
+		return true;
+
+	if (engine == I915_EXEC_BSD)
+		return true;
+
+	return false;
+}
+
+static void single_timeline(int i915)
+{
+	const uint32_t bbe = MI_BATCH_BUFFER_END;
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = gem_create(i915, 4096),
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+	};
+	struct sync_fence_info rings[16];
+	struct sync_file_info sync_file_info = {
+		.num_fences = 1,
+	};
+	unsigned int engine;
+	int n;
+
+	igt_require(has_single_timeline(i915));
+
+	gem_write(i915, obj.handle, 0, &bbe, sizeof(bbe));
+	gem_execbuf(i915, &execbuf);
+	gem_sync(i915, obj.handle);
+
+	/*
+	 * For a "single timeline" context, each ring is on the common
+	 * timeline, unlike a normal context where each ring has an
+	 * independent timeline. That is no matter which engine we submit
+	 * to, it reports the same timeline name and fence context. However,
+	 * the fence context is not reported through the sync_fence_info.
+	 */
+	execbuf.rsvd1 =
+		gem_context_create_shared(i915, 0,
+					  I915_GEM_CONTEXT_SINGLE_TIMELINE);
+	execbuf.flags = I915_EXEC_FENCE_OUT;
+	n = 0;
+	for_each_engine(i915, engine) {
+		gem_execbuf_wr(i915, &execbuf);
+		sync_file_info.sync_fence_info = to_user_pointer(&rings[n]);
+		do_ioctl(execbuf.rsvd2 >> 32, SYNC_IOC_FILE_INFO, &sync_file_info);
+		close(execbuf.rsvd2 >> 32);
+
+		igt_info("ring[%d] fence: %s %s\n",
+			 n, rings[n].driver_name, rings[n].obj_name);
+		n++;
+	}
+	gem_sync(i915, obj.handle);
+	gem_close(i915, obj.handle);
+
+	for (int i = 1; i < n; i++) {
+		igt_assert(!strcmp(rings[0].driver_name, rings[i].driver_name));
+		igt_assert(!strcmp(rings[0].obj_name, rings[i].obj_name));
+	}
+}
+
+static void exec_single_timeline(int i915, unsigned int ring)
+{
+	unsigned int other;
+	igt_spin_t *spin;
+	uint32_t ctx;
+
+	gem_require_ring(i915, ring);
+	igt_require(has_single_timeline(i915));
+
+	/*
+	 * On an ordinary context, a blockage on one ring doesn't prevent
+	 * execution on an other.
+	 */
+	ctx = 0;
+	spin = NULL;
+	for_each_engine(i915, other) {
+		if (other == ring || ignore_engine(other))
+			continue;
+
+		if (spin == NULL) {
+			spin = __igt_spin_batch_new(i915, ctx, other, 0);
+		} else {
+			struct drm_i915_gem_exec_object2 obj = {
+				.handle = spin->handle,
+			};
+			struct drm_i915_gem_execbuffer2 execbuf = {
+				.buffers_ptr = to_user_pointer(&obj),
+				.buffer_count = 1,
+				.flags = other,
+				.rsvd1 = ctx,
+			};
+			gem_execbuf(i915, &execbuf);
+		}
+	}
+	igt_require(spin);
+	igt_assert_eq(nop_sync(i915, ctx, ring, NSEC_PER_SEC), 0);
+	igt_spin_batch_free(i915, spin);
+
+	/*
+	 * But if we create a context with just a single shared timeline,
+	 * then it will block waiting for the earlier requests on the
+	 * other engines.
+	 */
+	ctx = gem_context_create_shared(i915, 0,
+					I915_GEM_CONTEXT_SINGLE_TIMELINE);
+	spin = NULL;
+	for_each_engine(i915, other) {
+		if (other == ring || ignore_engine(other))
+			continue;
+
+		if (spin == NULL) {
+			spin = __igt_spin_batch_new(i915, ctx, other, 0);
+		} else {
+			struct drm_i915_gem_exec_object2 obj = {
+				.handle = spin->handle,
+			};
+			struct drm_i915_gem_execbuffer2 execbuf = {
+				.buffers_ptr = to_user_pointer(&obj),
+				.buffer_count = 1,
+				.flags = other,
+				.rsvd1 = ctx,
+			};
+			gem_execbuf(i915, &execbuf);
+		}
+	}
+	igt_assert(spin);
+	igt_assert_eq(nop_sync(i915, ctx, ring, NSEC_PER_SEC), -ETIME);
+	igt_spin_batch_free(i915, spin);
+}
+
+static void store_dword(int i915, uint32_t ctx, unsigned ring,
+			uint32_t target, uint32_t offset, uint32_t value,
+			uint32_t cork, unsigned write_domain)
+{
+	const int gen = intel_gen(intel_get_drm_devid(i915));
+	struct drm_i915_gem_exec_object2 obj[3];
+	struct drm_i915_gem_relocation_entry reloc;
+	struct drm_i915_gem_execbuffer2 execbuf;
+	uint32_t batch[16];
+	int i;
+
+	memset(&execbuf, 0, sizeof(execbuf));
+	execbuf.buffers_ptr = to_user_pointer(obj + !cork);
+	execbuf.buffer_count = 2 + !!cork;
+	execbuf.flags = ring;
+	if (gen < 6)
+		execbuf.flags |= I915_EXEC_SECURE;
+	execbuf.rsvd1 = ctx;
+
+	memset(obj, 0, sizeof(obj));
+	obj[0].handle = cork;
+	obj[1].handle = target;
+	obj[2].handle = gem_create(i915, 4096);
+
+	memset(&reloc, 0, sizeof(reloc));
+	reloc.target_handle = obj[1].handle;
+	reloc.presumed_offset = 0;
+	reloc.offset = sizeof(uint32_t);
+	reloc.delta = offset;
+	reloc.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
+	reloc.write_domain = write_domain;
+	obj[2].relocs_ptr = to_user_pointer(&reloc);
+	obj[2].relocation_count = 1;
+
+	i = 0;
+	batch[i] = MI_STORE_DWORD_IMM | (gen < 6 ? 1 << 22 : 0);
+	if (gen >= 8) {
+		batch[++i] = offset;
+		batch[++i] = 0;
+	} else if (gen >= 4) {
+		batch[++i] = 0;
+		batch[++i] = offset;
+		reloc.offset += sizeof(uint32_t);
+	} else {
+		batch[i]--;
+		batch[++i] = offset;
+	}
+	batch[++i] = value;
+	batch[++i] = MI_BATCH_BUFFER_END;
+	gem_write(i915, obj[2].handle, 0, batch, sizeof(batch));
+	gem_execbuf(i915, &execbuf);
+	gem_close(i915, obj[2].handle);
+}
+
+struct cork {
+	int device;
+	uint32_t handle;
+	uint32_t fence;
+};
+
+static void plug(int i915, struct cork *c)
+{
+	struct vgem_bo bo;
+	int dmabuf;
+
+	c->device = drm_open_driver(DRIVER_VGEM);
+
+	bo.width = bo.height = 1;
+	bo.bpp = 4;
+	vgem_create(c->device, &bo);
+	c->fence = vgem_fence_attach(c->device, &bo, VGEM_FENCE_WRITE);
+
+	dmabuf = prime_handle_to_fd(c->device, bo.handle);
+	c->handle = prime_fd_to_handle(i915, dmabuf);
+	close(dmabuf);
+}
+
+static void unplug(struct cork *c)
+{
+	vgem_fence_signal(c->device, c->fence);
+	close(c->device);
+}
+
+static uint32_t create_highest_priority(int i915)
+{
+	uint32_t ctx = gem_context_create(i915);
+
+	/*
+	 * If there is no priority support, all contexts will have equal
+	 * priority (and therefore the max user priority), so no context
+	 * can overtake us, and we effectively can form a plug.
+	 */
+	__gem_context_set_priority(i915, ctx, MAX_PRIO);
+
+	return ctx;
+}
+
+static void unplug_show_queue(int i915, struct cork *c, unsigned int engine)
+{
+	igt_spin_t *spin[BUSY_QLEN];
+
+	for (int n = 0; n < ARRAY_SIZE(spin); n++) {
+		uint32_t ctx = create_highest_priority(i915);
+		spin[n] = __igt_spin_batch_new(i915, ctx, engine, 0);
+		gem_context_destroy(i915, ctx);
+	}
+
+	unplug(c); /* batches will now be queued on the engine */
+	igt_debugfs_dump(i915, "i915_engine_info");
+
+	for (int n = 0; n < ARRAY_SIZE(spin); n++)
+		igt_spin_batch_free(i915, spin[n]);
+}
+
+static uint32_t store_timestamp(int i915, uint32_t ctx, unsigned ring)
+{
+	const bool r64b = intel_gen(intel_get_drm_devid(i915)) >= 8;
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = gem_create(i915, 4096),
+		.relocation_count = 1,
+	};
+	struct drm_i915_gem_relocation_entry reloc = {
+		.target_handle = obj.handle,
+		.offset = 2 * sizeof(uint32_t),
+		.delta = 4092,
+		.read_domains = I915_GEM_DOMAIN_INSTRUCTION,
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.flags = ring,
+		.rsvd1 = ctx,
+	};
+	uint32_t batch[] = {
+		0x24 << 23 | (1 + r64b), /* SRM */
+		0x2358,
+		4092,
+		0,
+		MI_BATCH_BUFFER_END
+	};
+
+	igt_require(intel_gen(intel_get_drm_devid(i915)) >= 7);
+
+	gem_write(i915, obj.handle, 0, batch, sizeof(batch));
+	obj.relocs_ptr = to_user_pointer(&reloc);
+
+	gem_execbuf(i915, &execbuf);
+
+	return obj.handle;
+}
+
+static void independent(int i915, unsigned ring, unsigned flags)
+{
+	uint32_t handle[2];
+	igt_spin_t *spin;
+
+	spin = igt_spin_batch_new(i915, 0, ring, 0);
+	for (int i = 0; i < 16; i++){
+		struct drm_i915_gem_exec_object2 obj = {
+			.handle = spin->handle
+		};
+		struct drm_i915_gem_execbuffer2 eb = {
+			.buffer_count = 1,
+			.buffers_ptr = to_user_pointer(&obj),
+			.flags = ring,
+			.rsvd1 = gem_context_create(i915),
+		};
+		gem_context_set_priority(i915, eb.rsvd1, MAX_PRIO);
+		gem_execbuf(i915, &eb);
+		gem_context_destroy(i915, eb.rsvd1);
+	}
+
+	for (int i = LO; i <= HI; i++) {
+		uint32_t ctx = gem_queue_create(i915);
+		gem_context_set_priority(i915, ctx, priorities[i]);
+		handle[i] = store_timestamp(i915, ctx, ring);
+		gem_context_destroy(i915, ctx);
+	}
+
+	igt_spin_batch_free(i915, spin);
+
+	for (int i = LO; i <= HI; i++) {
+		uint32_t *ptr;
+
+		ptr = gem_mmap__gtt(i915, handle[i], 4096, PROT_READ);
+		gem_set_domain(i915, handle[i], /* no write hazard lies! */
+			       I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
+		gem_close(i915, handle[i]);
+
+		handle[i] = ptr[1023];
+		munmap(ptr, 4096);
+
+		igt_debug("ctx[%d] .prio=%d, timestamp=%u\n",
+			  i, priorities[i], handle[i]);
+	}
+
+	igt_assert((int32_t)(handle[HI] - handle[LO]) < 0);
+}
+
+static void reorder(int i915, unsigned ring, unsigned flags)
+#define EQUAL 1
+{
+	struct cork cork;
+	uint32_t scratch;
+	uint32_t *ptr;
+	uint32_t ctx[2];
+
+	ctx[LO] = gem_queue_create(i915);
+	gem_context_set_priority(i915, ctx[LO], MIN_PRIO);
+
+	ctx[HI] = gem_queue_create(i915);
+	gem_context_set_priority(i915, ctx[HI], flags & EQUAL ? MIN_PRIO : 0);
+
+	scratch = gem_create(i915, 4096);
+	plug(i915, &cork);
+
+	/* We expect the high priority context to be executed first, and
+	 * so the final result will be value from the low priority context.
+	 */
+	store_dword(i915, ctx[LO], ring, scratch, 0, ctx[LO], cork.handle, 0);
+	store_dword(i915, ctx[HI], ring, scratch, 0, ctx[HI], cork.handle, 0);
+
+	unplug_show_queue(i915, &cork, ring);
+
+	gem_context_destroy(i915, ctx[LO]);
+	gem_context_destroy(i915, ctx[HI]);
+
+	ptr = gem_mmap__gtt(i915, scratch, 4096, PROT_READ);
+	gem_set_domain(i915, scratch, /* no write hazard lies! */
+		       I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
+	gem_close(i915, scratch);
+
+	if (flags & EQUAL) /* equal priority, result will be fifo */
+		igt_assert_eq_u32(ptr[0], ctx[HI]);
+	else
+		igt_assert_eq_u32(ptr[0], ctx[LO]);
+	munmap(ptr, 4096);
+}
+
+static void promotion(int i915, unsigned ring)
+{
+	struct cork cork;
+	uint32_t result, dep;
+	uint32_t *ptr;
+	uint32_t ctx[3];
+
+	ctx[LO] = gem_queue_create(i915);
+	gem_context_set_priority(i915, ctx[LO], MIN_PRIO);
+
+	ctx[HI] = gem_queue_create(i915);
+	gem_context_set_priority(i915, ctx[HI], 0);
+
+	ctx[NOISE] = gem_queue_create(i915);
+	gem_context_set_priority(i915, ctx[NOISE], MIN_PRIO/2);
+
+	result = gem_create(i915, 4096);
+	dep = gem_create(i915, 4096);
+
+	plug(i915, &cork);
+
+	/* Expect that HI promotes LO, so the order will be LO, HI, NOISE.
+	 *
+	 * fifo would be NOISE, LO, HI.
+	 * strict priority would be  HI, NOISE, LO
+	 */
+	store_dword(i915, ctx[NOISE], ring, result, 0, ctx[NOISE], cork.handle, 0);
+	store_dword(i915, ctx[LO], ring, result, 0, ctx[LO], cork.handle, 0);
+
+	/* link LO <-> HI via a dependency on another buffer */
+	store_dword(i915, ctx[LO], ring, dep, 0, ctx[LO], 0, I915_GEM_DOMAIN_INSTRUCTION);
+	store_dword(i915, ctx[HI], ring, dep, 0, ctx[HI], 0, 0);
+
+	store_dword(i915, ctx[HI], ring, result, 0, ctx[HI], 0, 0);
+
+	unplug_show_queue(i915, &cork, ring);
+
+	gem_context_destroy(i915, ctx[NOISE]);
+	gem_context_destroy(i915, ctx[LO]);
+	gem_context_destroy(i915, ctx[HI]);
+
+	ptr = gem_mmap__gtt(i915, dep, 4096, PROT_READ);
+	gem_set_domain(i915, dep, /* no write hazard lies! */
+			I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
+	gem_close(i915, dep);
+
+	igt_assert_eq_u32(ptr[0], ctx[HI]);
+	munmap(ptr, 4096);
+
+	ptr = gem_mmap__gtt(i915, result, 4096, PROT_READ);
+	gem_set_domain(i915, result, /* no write hazard lies! */
+			I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
+	gem_close(i915, result);
+
+	igt_assert_eq_u32(ptr[0], ctx[NOISE]);
+	munmap(ptr, 4096);
+}
+
+static void smoketest(int i915, unsigned ring, unsigned timeout)
+{
+	const int ncpus = sysconf(_SC_NPROCESSORS_ONLN);
+	unsigned engines[16];
+	unsigned nengine;
+	unsigned engine;
+	uint32_t scratch;
+	uint32_t *ptr;
+
+	nengine = 0;
+	for_each_engine(i915, engine) {
+		if (ignore_engine(engine))
+			continue;
+
+		engines[nengine++] = engine;
+	}
+	igt_require(nengine);
+
+	scratch = gem_create(i915, 4096);
+	igt_fork(child, ncpus) {
+		unsigned long count = 0;
+		uint32_t ctx;
+
+		hars_petruska_f54_1_random_perturb(child);
+
+		ctx = gem_queue_create(i915);
+		igt_until_timeout(timeout) {
+			int prio;
+
+			prio = hars_petruska_f54_1_random_unsafe_max(MAX_PRIO - MIN_PRIO) + MIN_PRIO;
+			gem_context_set_priority(i915, ctx, prio);
+
+			engine = engines[hars_petruska_f54_1_random_unsafe_max(nengine)];
+			store_dword(i915, ctx, engine, scratch,
+				    8*child + 0, ~child,
+				    0, 0);
+			for (unsigned int step = 0; step < 8; step++)
+				store_dword(i915, ctx, engine, scratch,
+					    8*child + 4, count++,
+					    0, 0);
+		}
+		gem_context_destroy(i915, ctx);
+	}
+	igt_waitchildren();
+
+	ptr = gem_mmap__gtt(i915, scratch, 4096, PROT_READ);
+	gem_set_domain(i915, scratch, /* no write hazard lies! */
+			I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
+	gem_close(i915, scratch);
+
+	for (unsigned n = 0; n < ncpus; n++) {
+		igt_assert_eq_u32(ptr[2*n], ~n);
+		/*
+		 * Note this count is approximate due to unconstrained
+		 * ordering of the dword writes between engines.
+		 *
+		 * Take the result with a pinch of salt.
+		 */
+		igt_info("Child[%d] completed %u cycles\n",  n, ptr[2*n+1]);
+	}
+	munmap(ptr, 4096);
+}
+
+igt_main
+{
+	const struct intel_execution_engine *e;
+	int i915 = -1;
+
+	igt_fixture {
+		i915 = drm_open_driver(DRIVER_INTEL);
+		igt_require_gem(i915);
+	}
+
+	igt_subtest_group {
+		igt_fixture {
+			igt_require(gem_contexts_has_shared_gtt(i915));
+			igt_fork_hang_detector(i915);
+		}
+
+		igt_subtest("create-shared-gtt")
+			create_shared_gtt(i915, 0);
+
+		igt_subtest("detached-shared-gtt")
+			create_shared_gtt(i915, DETACHED);
+
+		igt_subtest("disjoint-timelines")
+			disjoint_timelines(i915);
+
+		igt_subtest("single-timeline")
+			single_timeline(i915);
+
+		igt_subtest("exhaust-shared-gtt")
+			exhaust_shared_gtt(i915, 0);
+
+		igt_subtest("exhaust-shared-gtt-lrc")
+			exhaust_shared_gtt(i915, EXHAUST_LRC);
+
+		for (e = intel_execution_engines; e->name; e++) {
+			igt_subtest_f("exec-shared-gtt-%s", e->name)
+				exec_shared_gtt(i915, e->exec_id | e->flags);
+
+			if (!ignore_engine(e->exec_id | e->flags)) {
+				igt_subtest_f("exec-single-timeline-%s",
+					      e->name)
+					exec_single_timeline(i915,
+							     e->exec_id | e->flags);
+			}
+
+			/*
+			 * Check that the shared contexts operate independently,
+			 * that is requests on one ("queue") can be scheduled
+			 * around another queue. We only check the basics here,
+			 * enough to reduce the queue into just another context,
+			 * and so rely on gem_exec_schedule to prove the rest.
+			 */
+			igt_subtest_group {
+				igt_fixture {
+					gem_require_ring(i915, e->exec_id | e->flags);
+					igt_require(gem_can_store_dword(i915, e->exec_id) | e->flags);
+					igt_require(gem_scheduler_enabled(i915));
+					igt_require(gem_scheduler_has_ctx_priority(i915));
+				}
+
+				igt_subtest_f("Q-independent-%s", e->name)
+					independent(i915, e->exec_id | e->flags, 0);
+
+				igt_subtest_f("Q-in-order-%s", e->name)
+					reorder(i915, e->exec_id | e->flags, EQUAL);
+
+				igt_subtest_f("Q-out-order-%s", e->name)
+					reorder(i915, e->exec_id | e->flags, 0);
+
+				igt_subtest_f("Q-promotion-%s", e->name)
+					promotion(i915, e->exec_id | e->flags);
+
+				igt_subtest_f("Q-smoketest-%s", e->name)
+					smoketest(i915, e->exec_id | e->flags, 5);
+			}
+		}
+
+		igt_subtest("Q-smoketest-all") {
+			igt_require(gem_scheduler_enabled(i915));
+			igt_require(gem_scheduler_has_ctx_priority(i915));
+			smoketest(i915, -1, 30);
+		}
+
+		igt_fixture {
+			igt_stop_hang_detector();
+		}
+	}
+}
diff --git a/tests/gem_exec_whisper.c b/tests/gem_exec_whisper.c
index 81303f847..a41614a7b 100644
--- a/tests/gem_exec_whisper.c
+++ b/tests/gem_exec_whisper.c
@@ -87,6 +87,7 @@ static void verify_reloc(int fd, uint32_t handle,
 #define HANG 0x20
 #define SYNC 0x40
 #define PRIORITY 0x80
+#define QUEUES 0x100
 
 struct hang {
 	struct drm_i915_gem_exec_object2 obj;
@@ -171,7 +172,7 @@ static void ctx_set_random_priority(int fd, uint32_t ctx)
 {
 	int prio = hars_petruska_f54_1_random_unsafe_max(1024) - 512;
 	gem_context_set_priority(fd, ctx, prio);
-};
+}
 
 static void whisper(int fd, unsigned engine, unsigned flags)
 {
@@ -223,6 +224,9 @@ static void whisper(int fd, unsigned engine, unsigned flags)
 	if (flags & CONTEXTS)
 		gem_require_contexts(fd);
 
+	if (flags & QUEUES)
+		igt_require(gem_has_queues(fd));
+
 	if (flags & HANG)
 		init_hang(&hang);
 
@@ -284,6 +288,10 @@ static void whisper(int fd, unsigned engine, unsigned flags)
 			for (n = 0; n < 64; n++)
 				contexts[n] = gem_context_create(fd);
 		}
+		if (flags & QUEUES) {
+			for (n = 0; n < 64; n++)
+				contexts[n] = gem_queue_create(fd);
+		}
 		if (flags & FDS) {
 			for (n = 0; n < 64; n++)
 				fds[n] = drm_open_driver(DRIVER_INTEL);
@@ -396,7 +404,7 @@ static void whisper(int fd, unsigned engine, unsigned flags)
 						execbuf.flags &= ~ENGINE_MASK;
 						execbuf.flags |= engines[rand() % nengine];
 					}
-					if (flags & CONTEXTS) {
+					if (flags & (CONTEXTS | QUEUES)) {
 						execbuf.rsvd1 = contexts[rand() % 64];
 						if (flags & PRIORITY)
 							ctx_set_random_priority(this_fd, execbuf.rsvd1);
@@ -475,7 +483,7 @@ static void whisper(int fd, unsigned engine, unsigned flags)
 			for (n = 0; n < 64; n++)
 				close(fds[n]);
 		}
-		if (flags & CONTEXTS) {
+		if (flags & (CONTEXTS | QUEUES)) {
 			for (n = 0; n < 64; n++)
 				gem_context_destroy(fd, contexts[n]);
 		}
@@ -507,18 +515,24 @@ igt_main
 		{ "chain-forked", CHAIN | FORKED },
 		{ "chain-interruptible", CHAIN | INTERRUPTIBLE },
 		{ "chain-sync", CHAIN | SYNC },
-		{ "contexts", CONTEXTS },
-		{ "contexts-interruptible", CONTEXTS | INTERRUPTIBLE},
-		{ "contexts-forked", CONTEXTS | FORKED},
-		{ "contexts-priority", CONTEXTS | FORKED | PRIORITY },
-		{ "contexts-chain", CONTEXTS | CHAIN },
-		{ "contexts-sync", CONTEXTS | SYNC },
 		{ "fds", FDS },
 		{ "fds-interruptible", FDS | INTERRUPTIBLE},
 		{ "fds-forked", FDS | FORKED},
 		{ "fds-priority", FDS | FORKED | PRIORITY },
 		{ "fds-chain", FDS | CHAIN},
 		{ "fds-sync", FDS | SYNC},
+		{ "contexts", CONTEXTS },
+		{ "contexts-interruptible", CONTEXTS | INTERRUPTIBLE},
+		{ "contexts-forked", CONTEXTS | FORKED},
+		{ "contexts-priority", CONTEXTS | FORKED | PRIORITY },
+		{ "contexts-chain", CONTEXTS | CHAIN },
+		{ "contexts-sync", CONTEXTS | SYNC },
+		{ "queues", QUEUES },
+		{ "queues-interruptible", QUEUES | INTERRUPTIBLE},
+		{ "queues-forked", QUEUES | FORKED},
+		{ "queues-priority", QUEUES | FORKED | PRIORITY },
+		{ "queues-chain", QUEUES | CHAIN },
+		{ "queues-sync", QUEUES | SYNC },
 		{ NULL }
 	};
 	int fd;
-- 
2.18.0

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
