Skip to contentSkip to conversation

Introduce a ww transaction utility

RFC PATCH Superseded Thomas Hellström (Intel) 2 participants 4 messages Sep 17, 2020

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover Introduce a ww transaction utility

A ww transaction utility intended to help removing the obj->mm.lock from the driver and introduce ww transactions in a robust way.

Patch 1/2 breaks the current i915 utilities out to separate files Patch 2/2 introduces the ww transaction utility

A similar utility could easily be introduced in the core ww_mutex code, allowing even for cross-driver ww transactions, and the template argument could then allow for per-ww-class derived ww_acquire_mutex types. To facilitate a core implementation, (since we can never guarantee the contended lock stays alive), we'd need a

void ww_mutex_relax(struct ww_acquire_ctx *)

and its interruptible variant that does the equivalent of locking and unlocking the contended mutex.

With this driver implementation, we can extend the code to take a reference on the object containing the contended lock to make sure it stays alive.

Finally, the drawback of the current implementation is the use of the hash table and corresponding performance cost, but as mentioned in patch 2, a core variant could probably do this in a much more efficient way.

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#1 lore raw

RFC PATCH 1/2 drm/i915: Break out dma_resv ww locking utilities to separate files

From: Thomas Hellström <[email protected]>

As we're about to add more ww-related functionality, break out the dma_resv ww locking utilities to their own files

---
 drivers/gpu/drm/i915/Makefile               |  1 +
 drivers/gpu/drm/i915/gem/i915_gem_object.h  |  1 +
 drivers/gpu/drm/i915/gt/intel_renderstate.h |  1 +
 drivers/gpu/drm/i915/i915_gem.c             | 64 ------------------
 drivers/gpu/drm/i915/i915_gem.h             | 15 -----
 drivers/gpu/drm/i915/i915_gem_ww.c          | 72 +++++++++++++++++++++
 drivers/gpu/drm/i915/i915_gem_ww.h          | 23 +++++++
 7 files changed, 98 insertions(+), 79 deletions(-)
 create mode 100644 drivers/gpu/drm/i915/i915_gem_ww.c
 create mode 100644 drivers/gpu/drm/i915/i915_gem_ww.h
drivers/gpu/drm/i915/Makefile +1 -0
index 58d129b5a65a..71503bc26d98 100644
@@ -45,6 +45,7 @@ i915-y += i915_drv.o \
 	  i915_switcheroo.o \
 	  i915_sysfs.o \
 	  i915_utils.o \
+	  i915_gem_ww.o \
 	  intel_device_info.o \
 	  intel_dram.o \
 	  intel_memory_region.o \
drivers/gpu/drm/i915/gem/i915_gem_object.h +1 -0
index f084a25c5121..cd64b1fdf53c 100644
@@ -15,6 +15,7 @@
 #include "i915_gem_object_types.h"
 #include "i915_gem_gtt.h"
 #include "i915_vma_types.h"
+#include "i915_gem_ww.h"
 
 void i915_gem_init__objects(struct drm_i915_private *i915);
 
drivers/gpu/drm/i915/gt/intel_renderstate.h +1 -0
index 713aa1e86c80..d9db833b873b 100644
@@ -26,6 +26,7 @@
 
 #include <linux/types.h>
 #include "i915_gem.h"
+#include "i915_gem_ww.h"
 
 struct i915_request;
 struct intel_context;
drivers/gpu/drm/i915/i915_gem.c +0 -64

70 lines not on this page · raw message

drivers/gpu/drm/i915/i915_gem.h +0 -15
index 4d50afab43f2..db0b2835095d 100644
@@ -27,8 +27,6 @@
 
 #include <linux/bug.h>
 #include <linux/interrupt.h>
-#include <linux/stackdepot.h>
-#include <linux/stacktrace.h>
 #include <drm/drm_drv.h>
 
 #include "i915_utils.h"
@@ -117,17 +115,4 @@ static inline bool __tasklet_is_scheduled(struct tasklet_struct *t)
 	return test_bit(TASKLET_STATE_SCHED, &t->state);
 }
 
-struct i915_gem_ww_ctx {
-	struct ww_acquire_ctx ctx;
-	struct list_head obj_list;
-	bool intr;
-	struct drm_i915_gem_object *contended;
-	depot_stack_handle_t contended_bt;
-};
-
-void i915_gem_ww_ctx_init(struct i915_gem_ww_ctx *ctx, bool intr);
-void i915_gem_ww_ctx_fini(struct i915_gem_ww_ctx *ctx);
-int __must_check i915_gem_ww_ctx_backoff(struct i915_gem_ww_ctx *ctx);
-void i915_gem_ww_unlock_single(struct drm_i915_gem_object *obj);
-
 #endif /* __I915_GEM_H__ */
drivers/gpu/drm/i915/i915_gem_ww.c new file +72 -0

72 lines not on this page · raw message

drivers/gpu/drm/i915/i915_gem_ww.h new file +23 -0
new file mode 100644
index 000000000000..94fdf8c5f89b
@@ -0,0 +1,23 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2020 Intel Corporation
+ */
+#ifndef __I915_GEM_WW_H__
+#define __I915_GEM_WW_H__
+
+#include <linux/stackdepot.h>
+#include <drm/drm_drv.h>
+
+struct i915_gem_ww_ctx {
+	struct ww_acquire_ctx ctx;
+	struct list_head obj_list;
+	struct drm_i915_gem_object *contended;
+	depot_stack_handle_t contended_bt;
+	bool intr;
+};
+
+void i915_gem_ww_ctx_init(struct i915_gem_ww_ctx *ctx, bool intr);
+void i915_gem_ww_ctx_fini(struct i915_gem_ww_ctx *ctx);
+int __must_check i915_gem_ww_ctx_backoff(struct i915_gem_ww_ctx *ctx);
+void i915_gem_ww_unlock_single(struct drm_i915_gem_object *obj);
+#endif
Signature · 6 lines
2.25.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw

RFC PATCH 2/2 drm/i915: Introduce a i915_gem_do_ww(){} utility

In reply to Thomas Hellström (Intel)

From: Thomas Hellström <[email protected]>

With the huge number of sites where multiple-object locking is needed in the driver, it becomes difficult to avoid recursive ww_acquire_ctx initialization, and the function prototypes become bloated passing the ww_acquire_ctx around. Furthermore it's not always easy to get the -EDEADLK handling correct and to follow it.

Introduce a i915_gem_do_ww utility that tries to remedy all these problems by enclosing parts of a ww transaction in the following way:

my_function() {
	struct i915_gem_ww_ctx *ww, template;
	int err;
	bool interruptible = true;

	i915_do_ww(ww, &template, err, interruptible) {
		err = ww_transaction_part(ww);
	}
	return err;
}

The utility will automatically look up an active ww_acquire_ctx if one is initialized previously in the call chain, and if one found will forward the -EDEADLK instead of handling it, which takes care of the recursive initalization. Using the utility also discourages nested ww unlocking / relocking that is both very fragile and hard to follow.

To look up and register an active ww_acquire_ctx, use a driver-wide hash table for now. But noting that a task could only have a single active ww_acqurie_ctx per ww_class, the active CTX is really task state and a generic version of this utility in the ww_mutex code could thus probably use a quick lookup from a list in the struct task_struct.

---
 drivers/gpu/drm/i915/i915_gem_ww.c | 73 +++++++++++++++++++++++++++++-
 drivers/gpu/drm/i915/i915_gem_ww.h | 55 +++++++++++++++++++++-
 2 files changed, 126 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem_ww.c +72 -1
index 3490b72cf613..9b51cb535b65 100644
@@ -1,10 +1,12 @@
 // SPDX-License-Identifier: MIT
 /*
- * Copyright © 2020 Intel Corporation
+ * Copyright © 2019 Intel Corporation
  */
+#include <linux/rhashtable.h>
 #include <linux/dma-resv.h>
 #include <linux/stacktrace.h>
 #include "i915_gem_ww.h"
+#include "i915_globals.h"
 #include "gem/i915_gem_object.h"
 
 void i915_gem_ww_ctx_init(struct i915_gem_ww_ctx *ww, bool intr)
@@ -70,3 +72,72 @@ int __must_check i915_gem_ww_ctx_backoff(struct i915_gem_ww_ctx *ww)
 
 	return ret;
 }
+
+static struct rhashtable ww_ht;
+static const struct rhashtable_params ww_params = {
+	.key_len = sizeof(struct task_struct *),
+	.key_offset = offsetof(struct i915_gem_ww_ctx, ctx.task),
+	.head_offset = offsetof(struct i915_gem_ww_ctx, head),
+	.min_size = 128,
+};
+
+static void i915_ww_item_free(void *ptr, void *arg)
+{
+	WARN_ON_ONCE(1);
+}
+
+static void i915_global_ww_exit(void)
+{
+	rhashtable_free_and_destroy(&ww_ht, i915_ww_item_free, NULL);
+}
+
+static void i915_global_ww_shrink(void)
+{
+}
+
+static struct i915_global global = {
+	.shrink = i915_global_ww_shrink,
+	.exit = i915_global_ww_exit,
+};
+
+int __init i915_global_ww_init(void)
+{
+	int ret = rhashtable_init(&ww_ht, &ww_params);
+
+	if (ret)
+		return ret;
+
+	i915_global_register(&global);
+
+	return 0;
+}
+
+void __i915_gem_ww_mark_unused(struct i915_gem_ww_ctx *ww)
+{
+	GEM_WARN_ON(rhashtable_remove_fast(&ww_ht, &ww->head, ww_params));
+}
+
+/**
+ * __i915_gem_ww_locate_or_find - return the task's i915_gem_ww_ctx context to use.
+ *
+ * @template: The context to use if there was none initialized previously
+ * in the call chain.
+ *
+ * RETURN: The task's i915_gem_ww_ctx context.
+ */
+struct i915_gem_ww_ctx *
+__i915_gem_ww_locate_or_use(struct i915_gem_ww_ctx *template)
+{
+	struct i915_gem_ww_ctx *tmp;
+
+	/* ctx.task is the hash key, so set it first. */
+	template->ctx.task = current;
+
+	/*
+	 * Ideally we'd just hook the active context to the
+	 * struct task_struct. But for now use a hash table.
+	 */
+	tmp = rhashtable_lookup_get_insert_fast(&ww_ht, &template->head,
+						ww_params);
+	return tmp;
+}
drivers/gpu/drm/i915/i915_gem_ww.h +54 -1

72 lines not on this page · raw message

Signature · 6 lines
2.25.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#3 lore raw

[Intel-gfx] ✗ Fi.CI.BUILD: failure for Introduce a ww transaction utility

In reply to Thomas Hellström (Intel)
== Series Details ==

== Series Details ==

Series: Introduce a ww transaction utility
URL : https://patchwork.freedesktop.org/series/81787/
State : failure

== Summary ==

Applying: drm/i915: Break out dma_resv ww locking utilities to separate files
error: sha1 information is lacking or useless (drivers/gpu/drm/i915/gem/i915_gem_object.h).
error: could not build fake ancestor
hint: Use 'git am --show-current-patch=diff' to see the failed patch
Patch failed at 0001 drm/i915: Break out dma_resv ww locking utilities to separate files When you have resolved this problem, run "git am --continue".
If you prefer to skip this patch, run "git am --skip" instead.
To restore the original branch and stop patching, run "git am --abort".

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx