forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2024-01-31 f70575805708cabdedea7498aaa3f710fde4d920 
 kernel/drivers/gpu/arm/bifrost/jm/mali_kbase_jm_defs.h
....@@ -1,7 +1,7 @@
11 /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
22 /*
33  *
4
- * (C) COPYRIGHT 2019-2021 ARM Limited. All rights reserved.
4
+ * (C) COPYRIGHT 2019-2023 ARM Limited. All rights reserved.
55  *
66  * This program is free software and is provided to you under the terms of the
77  * GNU General Public License version 2 as published by the Free Software
....@@ -87,8 +87,6 @@
8787 #define KBASE_KATOM_FLAG_FAIL_BLOCKER (1<<8)
8888 /* Atom is currently in the list of atoms blocked on cross-slot dependencies */
8989 #define KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST (1<<9)
90
-/* Atom is currently holding a context reference */
91
-#define KBASE_KATOM_FLAG_HOLDING_CTX_REF (1<<10)
9290 /* Atom requires GPU to be in protected mode */
9391 #define KBASE_KATOM_FLAG_PROTECTED (1<<11)
9492 /* Atom has been stored in runnable_tree */
....@@ -125,6 +123,25 @@
125123 #define KBASE_SERIALIZE_INTER_SLOT (1 << 1)
126124 /* Reset the GPU after each atom completion */
127125 #define KBASE_SERIALIZE_RESET (1 << 2)
126
+
127
+/**
128
+ * enum kbase_timeout_selector - The choice of which timeout to get scaled
129
+ *                               using the lowest GPU frequency.
130
+ * @MMU_AS_INACTIVE_WAIT_TIMEOUT: Maximum waiting time in ms for the completion
131
+ *                                of a MMU operation
132
+ * @JM_DEFAULT_JS_FREE_TIMEOUT: Maximum timeout to wait for JS_COMMAND_NEXT
133
+ *                              to be updated on HW side so a Job Slot is
134
+ *                              considered free.
135
+ * @KBASE_TIMEOUT_SELECTOR_COUNT: Number of timeout selectors. Must be last in
136
+ *                                the enum.
137
+ */
138
+enum kbase_timeout_selector {
139
+	MMU_AS_INACTIVE_WAIT_TIMEOUT,
140
+	JM_DEFAULT_JS_FREE_TIMEOUT,
141
+
142
+	/* Must be the last in the enum */
143
+	KBASE_TIMEOUT_SELECTOR_COUNT
144
+};
128145 
129146 #if IS_ENABLED(CONFIG_DEBUG_FS)
130147 /**
....@@ -176,8 +193,6 @@
176193 static inline const struct kbase_jd_atom *
177194 kbase_jd_katom_dep_atom(const struct kbase_jd_atom_dependency *dep)
178195 {
179
-	LOCAL_ASSERT(dep != NULL);
180
-
181196 	return (const struct kbase_jd_atom *)(dep->atom);
182197 }
183198 
....@@ -191,8 +206,6 @@
191206 static inline u8 kbase_jd_katom_dep_type(
192207 		const struct kbase_jd_atom_dependency *dep)
193208 {
194
-	LOCAL_ASSERT(dep != NULL);
195
-
196209 	return dep->dep_type;
197210 }
198211 
....@@ -209,8 +222,6 @@
209222 {
210223 	struct kbase_jd_atom_dependency *dep;
211224 
212
-	LOCAL_ASSERT(const_dep != NULL);
213
-
214225 	dep = (struct kbase_jd_atom_dependency *)const_dep;
215226 
216227 	dep->atom = a;
....@@ -226,8 +237,6 @@
226237 		const struct kbase_jd_atom_dependency *const_dep)
227238 {
228239 	struct kbase_jd_atom_dependency *dep;
229
-
230
-	LOCAL_ASSERT(const_dep != NULL);
231240 
232241 	dep = (struct kbase_jd_atom_dependency *)const_dep;
233242 
....@@ -343,19 +352,6 @@
343352 };
344353 
345354 /**
346
- * struct kbase_ext_res - Contains the info for external resources referred
347
- *                        by an atom, which have been mapped on GPU side.
348
- * @gpu_address:          Start address of the memory region allocated for
349
- *                        the resource from GPU virtual address space.
350
- * @alloc:                pointer to physical pages tracking object, set on
351
- *                        mapping the external resource on GPU side.
352
- */
353
-struct kbase_ext_res {
354
-	u64 gpu_address;
355
-	struct kbase_mem_phy_alloc *alloc;
356
-};
357
-
358
-/**
359355  * struct kbase_jd_atom  - object representing the atom, containing the complete
360356  *                         state and attributes of an atom.
361357  * @work:                  work item for the bottom half processing of the atom,
....@@ -388,7 +384,8 @@
388384  *                         each allocation is read in order to enforce an
389385  *                         overall physical memory usage limit.
390386  * @nr_extres:             number of external resources referenced by the atom.
391
- * @extres:                pointer to the location containing info about
387
+ * @extres:                Pointer to @nr_extres VA regions containing the external
388
+ *                         resource allocation and other information.
392389  *                         @nr_extres external resources referenced by the atom.
393390  * @device_nr:             indicates the coregroup with which the atom is
394391  *                         associated, when
....@@ -406,16 +403,21 @@
406403  *                         sync through soft jobs and for the implicit
407404  *                         synchronization required on access to external
408405  *                         resources.
409
- * @dma_fence.fence_in:    Input fence
406
+ * @dma_fence.fence_in:    Points to the dma-buf input fence for this atom.
407
+ *                         The atom would complete only after the fence is
408
+ *                         signaled.
410409  * @dma_fence.fence:       Points to the dma-buf output fence for this atom.
410
+ * @dma_fence.fence_cb:    The object that is passed at the time of adding the
411
+ *                         callback that gets invoked when @dma_fence.fence_in
412
+ *                         is signaled.
413
+ * @dma_fence.fence_cb_added: Flag to keep a track if the callback was successfully
414
+ *                            added for @dma_fence.fence_in, which is supposed to be
415
+ *                            invoked on the signaling of fence.
411416  * @dma_fence.context:     The dma-buf fence context number for this atom. A
412417  *                         unique context number is allocated to each katom in
413418  *                         the context on context creation.
414419  * @dma_fence.seqno:       The dma-buf fence sequence number for this atom. This
415420  *                         is increased every time this katom uses dma-buf fence
416
- * @dma_fence.callbacks:   List of all callbacks set up to wait on other fences
417
- * @dma_fence.dep_count:   Atomic counter of number of outstandind dma-buf fence
418
- *                         dependencies for this atom.
419421  * @event_code:            Event code for the job chain represented by the atom,
420422  *                         both HW and low-level SW events are represented by
421423  *                         event codes.
....@@ -498,7 +500,6 @@
498500  *                 BASE_JD_REQ_START_RENDERPASS set in its core requirements
499501  *                 with an atom that has BASE_JD_REQ_END_RENDERPASS set.
500502  * @jc_fragment:          Set of GPU fragment job chains
501
- * @retry_count:          TODO: Not used,to be removed
502503  */
503504 struct kbase_jd_atom {
504505 	struct work_struct work;
....@@ -518,21 +519,17 @@
518519 #endif /* MALI_JIT_PRESSURE_LIMIT_BASE */
519520 
520521 	u16 nr_extres;
521
-	struct kbase_ext_res *extres;
522
+	struct kbase_va_region **extres;
522523 
523524 	u32 device_nr;
524525 	u64 jc;
525526 	void *softjob_data;
526
-#if defined(CONFIG_SYNC)
527
-	struct sync_fence *fence;
528
-	struct sync_fence_waiter sync_waiter;
529
-#endif				/* CONFIG_SYNC */
530
-#if defined(CONFIG_MALI_BIFROST_DMA_FENCE) || defined(CONFIG_SYNC_FILE)
527
+#if IS_ENABLED(CONFIG_SYNC_FILE)
531528 	struct {
532529 		/* Use the functions/API defined in mali_kbase_fence.h to
533530 		 * when working with this sub struct
534531 		 */
535
-#if defined(CONFIG_SYNC_FILE)
532
+#if IS_ENABLED(CONFIG_SYNC_FILE)
536533 #if (KERNEL_VERSION(4, 10, 0) > LINUX_VERSION_CODE)
537534 		struct fence *fence_in;
538535 #else
....@@ -555,38 +552,21 @@
555552 #else
556553 		struct dma_fence *fence;
557554 #endif
555
+
556
+		/* This is the callback object that is registered for the fence_in.
557
+		 * The callback is invoked when the fence_in is signaled.
558
+		 */
559
+#if (KERNEL_VERSION(4, 10, 0) > LINUX_VERSION_CODE)
560
+		struct fence_cb fence_cb;
561
+#else
562
+		struct dma_fence_cb fence_cb;
563
+#endif
564
+		bool fence_cb_added;
565
+
558566 		unsigned int context;
559567 		atomic_t seqno;
560
-		/* This contains a list of all callbacks set up to wait on
561
-		 * other fences.  This atom must be held back from JS until all
562
-		 * these callbacks have been called and dep_count have reached
563
-		 * 0. The initial value of dep_count must be equal to the
564
-		 * number of callbacks on this list.
565
-		 *
566
-		 * This list is protected by jctx.lock. Callbacks are added to
567
-		 * this list when the atom is built and the wait are set up.
568
-		 * All the callbacks then stay on the list until all callbacks
569
-		 * have been called and the atom is queued, or cancelled, and
570
-		 * then all callbacks are taken off the list and freed.
571
-		 */
572
-		struct list_head callbacks;
573
-		/* Atomic counter of number of outstandind dma-buf fence
574
-		 * dependencies for this atom. When dep_count reaches 0 the
575
-		 * atom may be queued.
576
-		 *
577
-		 * The special value "-1" may only be set after the count
578
-		 * reaches 0, while holding jctx.lock. This indicates that the
579
-		 * atom has been handled, either queued in JS or cancelled.
580
-		 *
581
-		 * If anyone but the dma-fence worker sets this to -1 they must
582
-		 * ensure that any potentially queued worker must have
583
-		 * completed before allowing the atom to be marked as unused.
584
-		 * This can be done by flushing the fence work queue:
585
-		 * kctx->dma_fence.wq.
586
-		 */
587
-		atomic_t dep_count;
588568 	} dma_fence;
589
-#endif /* CONFIG_MALI_BIFROST_DMA_FENCE || CONFIG_SYNC_FILE */
569
+#endif /* CONFIG_SYNC_FILE */
590570 
591571 	/* Note: refer to kbasep_js_atom_retained_state, which will take a copy
592572 	 * of some of the following members
....@@ -605,11 +585,9 @@
605585 #if IS_ENABLED(CONFIG_GPU_TRACEPOINTS)
606586 	int work_id;
607587 #endif
608
-	int slot_nr;
588
+	unsigned int slot_nr;
609589 
610590 	u32 atom_flags;
611
-
612
-	int retry_count;
613591 
614592 	enum kbase_atom_gpu_rb_state gpu_rb_state;
615593 
....@@ -653,6 +631,51 @@
653631 	return (bool)(katom->atom_flags & KBASE_KATOM_FLAG_PROTECTED);
654632 }
655633 
634
+/**
635
+ * kbase_jd_atom_is_younger - query if one atom is younger by age than another
636
+ *
637
+ * @katom_a: the first atom
638
+ * @katom_b: the second atom
639
+ *
640
+ * Return: true if the first atom is strictly younger than the second,
641
+ *         false otherwise.
642
+ */
643
+static inline bool kbase_jd_atom_is_younger(const struct kbase_jd_atom *katom_a,
644
+					    const struct kbase_jd_atom *katom_b)
645
+{
646
+	return ((s32)(katom_a->age - katom_b->age) < 0);
647
+}
648
+
649
+/**
650
+ * kbase_jd_atom_is_earlier - Check whether the first atom has been submitted
651
+ *                            earlier than the second one
652
+ *
653
+ * @katom_a: the first atom
654
+ * @katom_b: the second atom
655
+ *
656
+ * Return: true if the first atom has been submitted earlier than the
657
+ * second atom. It is used to understand if an atom that is ready has been
658
+ * submitted earlier than the currently running atom, so that the currently
659
+ * running atom should be preempted to allow the ready atom to run.
660
+ */
661
+static inline bool kbase_jd_atom_is_earlier(const struct kbase_jd_atom *katom_a,
662
+					    const struct kbase_jd_atom *katom_b)
663
+{
664
+	/* No seq_nr set? */
665
+	if (!katom_a->seq_nr || !katom_b->seq_nr)
666
+		return false;
667
+
668
+	/* Efficiently handle the unlikely case of wrapping.
669
+	 * The following code assumes that the delta between the sequence number
670
+	 * of the two atoms is less than INT64_MAX.
671
+	 * In the extremely unlikely case where the delta is higher, the comparison
672
+	 * defaults for no preemption.
673
+	 * The code also assumes that the conversion from unsigned to signed types
674
+	 * works because the signed integers are 2's complement.
675
+	 */
676
+	return (s64)(katom_a->seq_nr - katom_b->seq_nr) < 0;
677
+}
678
+
656679 /*
657680  * Theory of operations:
658681  *
....@@ -690,17 +713,13 @@
690713  * A state machine is used to control incremental rendering.
691714  */
692715 enum kbase_jd_renderpass_state {
693
-	KBASE_JD_RP_COMPLETE,       /* COMPLETE => START */
694
-	KBASE_JD_RP_START,          /* START => PEND_OOM or COMPLETE */
695
-	KBASE_JD_RP_PEND_OOM,       /* PEND_OOM => OOM or COMPLETE */
696
-	KBASE_JD_RP_OOM,            /* OOM => RETRY */
697
-	KBASE_JD_RP_RETRY,          /* RETRY => RETRY_PEND_OOM or
698
-				     *          COMPLETE
699
-				     */
700
-	KBASE_JD_RP_RETRY_PEND_OOM, /* RETRY_PEND_OOM => RETRY_OOM or
701
-				     *                   COMPLETE
702
-				     */
703
-	KBASE_JD_RP_RETRY_OOM,      /* RETRY_OOM => RETRY */
716
+	KBASE_JD_RP_COMPLETE, /* COMPLETE => START */
717
+	KBASE_JD_RP_START, /* START => PEND_OOM or COMPLETE */
718
+	KBASE_JD_RP_PEND_OOM, /* PEND_OOM => OOM or COMPLETE */
719
+	KBASE_JD_RP_OOM, /* OOM => RETRY */
720
+	KBASE_JD_RP_RETRY, /* RETRY => RETRY_PEND_OOM or COMPLETE */
721
+	KBASE_JD_RP_RETRY_PEND_OOM, /* RETRY_PEND_OOM => RETRY_OOM or COMPLETE */
722
+	KBASE_JD_RP_RETRY_OOM /* RETRY_OOM => RETRY */
704723 };
705724 
706725 /**
....@@ -773,7 +792,7 @@
773792  *                            atom completes
774793  *                            execution on GPU or the input fence get signaled.
775794  * @tb_lock:                  Lock to serialize the write access made to @tb to
776
- *                            to store the register access trace messages.
795
+ *                            store the register access trace messages.
777796  * @tb:                       Pointer to the Userspace accessible buffer storing
778797  *                            the trace messages for register read/write
779798  *                            accesses made by the Kbase. The buffer is filled
....@@ -840,6 +859,10 @@
840859  * @pf_data:           Data relating to Page fault.
841860  * @bf_data:           Data relating to Bus fault.
842861  * @current_setup:     Stores the MMU configuration for this address space.
862
+ * @is_unresponsive:   Flag to indicate MMU is not responding.
863
+ *                     Set if a MMU command isn't completed within
864
+ *                     &kbase_device:mmu_as_inactive_wait_time_ms.
865
+ *                     Clear by kbase_ctx_sched_restore_all_as() after GPU reset completes.
843866  */
844867 struct kbase_as {
845868 	int number;
....@@ -849,6 +872,7 @@
849872 	struct kbase_fault pf_data;
850873 	struct kbase_fault bf_data;
851874 	struct kbase_mmu_setup current_setup;
875
+	bool is_unresponsive;
852876 };
853877 
854878 #endif /* _KBASE_JM_DEFS_H_ */