forked from ~ljy/RK356X_SDK_RELEASE

blame | history | patch | commit | commitdiff
hc
2023-12-11 d2ccde1c8e90d38cee87a1b0309ad2827f3fd30d 
 kernel/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
....@@ -35,7 +35,14 @@
3535 #include <linux/kfifo.h>
3636 #include <linux/seq_file.h>
3737 #include <linux/kref.h>
38
+#include <linux/sysfs.h>
39
+#include <linux/device_cgroup.h>
40
+#include <drm/drm_file.h>
41
+#include <drm/drm_drv.h>
42
+#include <drm/drm_device.h>
43
+#include <drm/drm_ioctl.h>
3844 #include <kgd_kfd_interface.h>
45
+#include <linux/swap.h>
3946 
4047 #include "amd_shared.h"
4148 
....@@ -54,22 +61,20 @@
5461  * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
5562  *  defines are w.r.t to PAGE_SIZE
5663  */
57
-#define KFD_MMAP_TYPE_SHIFT	(62 - PAGE_SHIFT)
64
+#define KFD_MMAP_TYPE_SHIFT	62
5865 #define KFD_MMAP_TYPE_MASK	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
5966 #define KFD_MMAP_TYPE_DOORBELL	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
6067 #define KFD_MMAP_TYPE_EVENTS	(0x2ULL << KFD_MMAP_TYPE_SHIFT)
6168 #define KFD_MMAP_TYPE_RESERVED_MEM	(0x1ULL << KFD_MMAP_TYPE_SHIFT)
69
+#define KFD_MMAP_TYPE_MMIO	(0x0ULL << KFD_MMAP_TYPE_SHIFT)
6270 
63
-#define KFD_MMAP_GPU_ID_SHIFT (46 - PAGE_SHIFT)
71
+#define KFD_MMAP_GPU_ID_SHIFT 46
6472 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \
6573 				<< KFD_MMAP_GPU_ID_SHIFT)
6674 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\
6775 				& KFD_MMAP_GPU_ID_MASK)
68
-#define KFD_MMAP_GPU_ID_GET(offset)    ((offset & KFD_MMAP_GPU_ID_MASK) \
76
+#define KFD_MMAP_GET_GPU_ID(offset)    ((offset & KFD_MMAP_GPU_ID_MASK) \
6977 				>> KFD_MMAP_GPU_ID_SHIFT)
70
-
71
-#define KFD_MMAP_OFFSET_VALUE_MASK	(0x3FFFFFFFFFFFULL >> PAGE_SHIFT)
72
-#define KFD_MMAP_OFFSET_VALUE_GET(offset) (offset & KFD_MMAP_OFFSET_VALUE_MASK)
7378 
7479 /*
7580  * When working with cp scheduler we should assign the HIQ manually or via
....@@ -92,10 +97,29 @@
9297  * Size of the per-process TBA+TMA buffer: 2 pages
9398  *
9499  * The first page is the TBA used for the CWSR ISA code. The second
95
- * page is used as TMA for daisy changing a user-mode trap handler.
100
+ * page is used as TMA for user-mode trap handler setup in daisy-chain mode.
96101  */
97102 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
98103 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE
104
+
105
+#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE		\
106
+	(KFD_MAX_NUM_OF_PROCESSES *			\
107
+			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
108
+
109
+#define KFD_KERNEL_QUEUE_SIZE 2048
110
+
111
+#define KFD_UNMAP_LATENCY_MS	(4000)
112
+
113
+/*
114
+ * 512 = 0x200
115
+ * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
116
+ * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
117
+ * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
118
+ * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
119
+ * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
120
+ */
121
+#define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
122
+
99123 
100124 /*
101125  * Kernel module parameter to specify maximum number of supported queues per
....@@ -103,12 +127,6 @@
103127  */
104128 extern int max_num_of_queues_per_device;
105129 
106
-#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096
107
-#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE		\
108
-	(KFD_MAX_NUM_OF_PROCESSES *			\
109
-			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
110
-
111
-#define KFD_KERNEL_QUEUE_SIZE 2048
112130 
113131 /* Kernel module parameter to specify the scheduling policy */
114132 extern int sched_policy;
....@@ -139,42 +157,20 @@
139157  */
140158 extern int ignore_crat;
141159 
142
-/*
143
- * Set sh_mem_config.retry_disable on Vega10
144
- */
145
-extern int noretry;
160
+/* Set sh_mem_config.retry_disable on GFX v9 */
161
+extern int amdgpu_noretry;
146162 
147
-/*
148
- * Halt if HWS hang is detected
149
- */
163
+/* Halt if HWS hang is detected */
150164 extern int halt_if_hws_hang;
151165 
152
-/**
153
- * enum kfd_sched_policy
154
- *
155
- * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
156
- * scheduling. In this scheduling mode we're using the firmware code to
157
- * schedule the user mode queues and kernel queues such as HIQ and DIQ.
158
- * the HIQ queue is used as a special queue that dispatches the configuration
159
- * to the cp and the user mode queues list that are currently running.
160
- * the DIQ queue is a debugging queue that dispatches debugging commands to the
161
- * firmware.
162
- * in this scheduling mode user mode queues over subscription feature is
163
- * enabled.
164
- *
165
- * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
166
- * subscription feature disabled.
167
- *
168
- * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
169
- * set the command processor registers and sets the queues "manually". This
170
- * mode is used *ONLY* for debugging proposes.
171
- *
172
- */
173
-enum kfd_sched_policy {
174
-	KFD_SCHED_POLICY_HWS = 0,
175
-	KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION,
176
-	KFD_SCHED_POLICY_NO_HWS
177
-};
166
+/* Whether MEC FW support GWS barriers */
167
+extern bool hws_gws_support;
168
+
169
+/* Queue preemption timeout in ms */
170
+extern int queue_preemption_timeout_ms;
171
+
172
+/* Enable eviction debug messages */
173
+extern bool debug_evictions;
178174 
179175 enum cache_policy {
180176 	cache_policy_coherent,
....@@ -193,6 +189,7 @@
193189 
194190 struct kfd_device_info {
195191 	enum amd_asic_type asic_family;
192
+	const char *asic_name;
196193 	const struct kfd_event_interrupt_class *event_interrupt_class;
197194 	unsigned int max_pasid_bits;
198195 	unsigned int max_no_of_hqd;
....@@ -204,6 +201,8 @@
204201 	bool needs_iommu_device;
205202 	bool needs_pci_atomics;
206203 	unsigned int num_sdma_engines;
204
+	unsigned int num_xgmi_sdma_engines;
205
+	unsigned int num_sdma_queues_per_engine;
207206 };
208207 
209208 struct kfd_mem_obj {
....@@ -225,6 +224,7 @@
225224 
226225 	const struct kfd_device_info *device_info;
227226 	struct pci_dev *pdev;
227
+	struct drm_device *ddev;
228228 
229229 	unsigned int id;		/* topology stub index */
230230 
....@@ -232,9 +232,10 @@
232232 					 * KFD. It is aligned for mapping
233233 					 * into user mode
234234 					 */
235
-	size_t doorbell_id_offset;	/* Doorbell offset (from KFD doorbell
236
-					 * to HW doorbell, GFX reserved some
237
-					 * at the start)
235
+	size_t doorbell_base_dw_offset;	/* Offset from the start of the PCI
236
+					 * doorbell BAR to the first KFD
237
+					 * doorbell in dwords. GFX reserves
238
+					 * the segment before this offset.
238239 					 */
239240 	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
240241 					   * page used by kernel queue
....@@ -273,7 +274,12 @@
273274 	bool interrupts_active;
274275 
275276 	/* Debug manager */
276
-	struct kfd_dbgmgr           *dbgmgr;
277
+	struct kfd_dbgmgr *dbgmgr;
278
+
279
+	/* Firmware versions */
280
+	uint16_t mec_fw_version;
281
+	uint16_t mec2_fw_version;
282
+	uint16_t sdma_fw_version;
277283 
278284 	/* Maximum process number mapped to HW scheduler */
279285 	unsigned int max_proc_per_quantum;
....@@ -282,15 +288,38 @@
282288 	bool cwsr_enabled;
283289 	const void *cwsr_isa;
284290 	unsigned int cwsr_isa_size;
285
-};
286291 
287
-/* KGD2KFD callbacks */
288
-void kgd2kfd_exit(void);
289
-struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
290
-			struct pci_dev *pdev, const struct kfd2kgd_calls *f2g);
291
-bool kgd2kfd_device_init(struct kfd_dev *kfd,
292
-			const struct kgd2kfd_shared_resources *gpu_resources);
293
-void kgd2kfd_device_exit(struct kfd_dev *kfd);
292
+	/* xGMI */
293
+	uint64_t hive_id;
294
+
295
+	/* UUID */
296
+	uint64_t unique_id;
297
+
298
+	bool pci_atomic_requested;
299
+
300
+	/* Use IOMMU v2 flag */
301
+	bool use_iommu_v2;
302
+
303
+	/* SRAM ECC flag */
304
+	atomic_t sram_ecc_flag;
305
+
306
+	/* Compute Profile ref. count */
307
+	atomic_t compute_profile;
308
+
309
+	/* Global GWS resource shared between processes */
310
+	void *gws;
311
+
312
+	/* Clients watching SMI events */
313
+	struct list_head smi_clients;
314
+	spinlock_t smi_lock;
315
+
316
+	uint32_t reset_seq_num;
317
+
318
+	struct ida doorbell_ida;
319
+	unsigned int max_doorbell_slices;
320
+
321
+	int noretry;
322
+};
294323 
295324 enum kfd_mempool {
296325 	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
....@@ -304,7 +333,7 @@
304333 struct device *kfd_chardev(void);
305334 
306335 /**
307
- * enum kfd_unmap_queues_filter
336
+ * enum kfd_unmap_queues_filter - Enum for queue filters.
308337  *
309338  * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
310339  *
....@@ -323,26 +352,34 @@
323352 };
324353 
325354 /**
326
- * enum kfd_queue_type
355
+ * enum kfd_queue_type - Enum for various queue types.
327356  *
328357  * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
329358  *
330
- * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
359
+ * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
331360  *
332361  * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
333362  *
334363  * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
364
+ *
365
+ * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
335366  */
336367 enum kfd_queue_type  {
337368 	KFD_QUEUE_TYPE_COMPUTE,
338369 	KFD_QUEUE_TYPE_SDMA,
339370 	KFD_QUEUE_TYPE_HIQ,
340
-	KFD_QUEUE_TYPE_DIQ
371
+	KFD_QUEUE_TYPE_DIQ,
372
+	KFD_QUEUE_TYPE_SDMA_XGMI
341373 };
342374 
343375 enum kfd_queue_format {
344376 	KFD_QUEUE_FORMAT_PM4,
345377 	KFD_QUEUE_FORMAT_AQL
378
+};
379
+
380
+enum KFD_QUEUE_PRIORITY {
381
+	KFD_QUEUE_PRIORITY_MINIMUM = 0,
382
+	KFD_QUEUE_PRIORITY_MAXIMUM = 15
346383 };
347384 
348385 /**
....@@ -371,9 +408,9 @@
371408  *
372409  * @write_ptr: Defines the number of dwords written to the ring buffer.
373410  *
374
- * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
375
- * the queue ring buffer. This field should be similar to write_ptr and the
376
- * user should update this field after he updated the write_ptr.
411
+ * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
412
+ * buffer. This field should be similar to write_ptr and the user should
413
+ * update this field after updating the write_ptr.
377414  *
378415  * @doorbell_off: The doorbell offset in the doorbell pci-bar.
379416  *
....@@ -385,6 +422,10 @@
385422  *
386423  * @is_active: Defines if the queue is active or not. @is_active and
387424  * @is_evicted are protected by the DQM lock.
425
+ *
426
+ * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
427
+ * @is_gws should be protected by the DQM lock, since changing it can yield the
428
+ * possibility of updating DQM state on number of GWS queues.
388429  *
389430  * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
390431  * of the queue.
....@@ -408,6 +449,7 @@
408449 	bool is_interop;
409450 	bool is_evicted;
410451 	bool is_active;
452
+	bool is_gws;
411453 	/* Not relevant for user mode queues in cp scheduling */
412454 	unsigned int vmid;
413455 	/* Relevant only for sdma queues*/
....@@ -427,12 +469,17 @@
427469 	uint32_t *cu_mask;
428470 };
429471 
472
+#define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\
473
+			    (q).queue_address != 0 &&	\
474
+			    (q).queue_percent > 0 &&	\
475
+			    !(q).is_evicted)
476
+
430477 /**
431478  * struct queue
432479  *
433480  * @list: Queue linked list.
434481  *
435
- * @mqd: The queue MQD.
482
+ * @mqd: The queue MQD (memory queue descriptor).
436483  *
437484  * @mqd_mem_obj: The MQD local gpu memory object.
438485  *
....@@ -441,7 +488,7 @@
441488  * @properties: The queue properties.
442489  *
443490  * @mec: Used only in no cp scheduling mode and identifies to micro engine id
444
- *	 that the queue should be execute on.
491
+ *	 that the queue should be executed on.
445492  *
446493  * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
447494  *	  id.
....@@ -451,6 +498,9 @@
451498  * @process: The kfd process that created this queue.
452499  *
453500  * @device: The kfd device that created this queue.
501
+ *
502
+ * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
503
+ * otherwise.
454504  *
455505  * This structure represents user mode compute queues.
456506  * It contains all the necessary data to handle such queues.
....@@ -473,17 +523,24 @@
473523 
474524 	struct kfd_process	*process;
475525 	struct kfd_dev		*device;
526
+	void *gws;
527
+
528
+	/* procfs */
529
+	struct kobject kobj;
476530 };
477531 
478
-/*
479
- * Please read the kfd_mqd_manager.h description.
480
- */
481532 enum KFD_MQD_TYPE {
482
-	KFD_MQD_TYPE_COMPUTE = 0,	/* for no cp scheduling */
483
-	KFD_MQD_TYPE_HIQ,		/* for hiq */
533
+	KFD_MQD_TYPE_HIQ = 0,		/* for hiq */
484534 	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
485535 	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
536
+	KFD_MQD_TYPE_DIQ,		/* for diq */
486537 	KFD_MQD_TYPE_MAX
538
+};
539
+
540
+enum KFD_PIPE_PRIORITY {
541
+	KFD_PIPE_PRIORITY_CS_LOW = 0,
542
+	KFD_PIPE_PRIORITY_CS_MEDIUM,
543
+	KFD_PIPE_PRIORITY_CS_HIGH
487544 };
488545 
489546 struct scheduling_resources {
....@@ -521,15 +578,22 @@
521578 	 */
522579 	bool reset_wavefronts;
523580 
524
-	/*
525
-	 * All the memory management data should be here too
581
+	/* This flag tells us if this process has a GWS-capable
582
+	 * queue that will be mapped into the runlist. It's
583
+	 * possible to request a GWS BO, but not have the queue
584
+	 * currently mapped, and this changes how the MAP_PROCESS
585
+	 * PM4 packet is configured.
526586 	 */
587
+	bool mapped_gws_queue;
588
+
589
+	/* All the memory management data should be here too */
527590 	uint64_t gds_context_area;
591
+	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
592
+	uint64_t page_table_base;
528593 	uint32_t sh_mem_config;
529594 	uint32_t sh_mem_bases;
530595 	uint32_t sh_mem_ape1_base;
531596 	uint32_t sh_mem_ape1_limit;
532
-	uint32_t page_table_base;
533597 	uint32_t gds_size;
534598 	uint32_t num_gws;
535599 	uint32_t num_oac;
....@@ -558,11 +622,6 @@
558622 /* Approx. time before evicting the process again */
559623 #define PROCESS_ACTIVE_TIME_MS 10
560624 
561
-int kgd2kfd_quiesce_mm(struct mm_struct *mm);
562
-int kgd2kfd_resume_mm(struct mm_struct *mm);
563
-int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
564
-					       struct dma_fence *fence);
565
-
566625 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
567626  * idr_handle in the least significant 4 bytes
568627  */
....@@ -576,6 +635,15 @@
576635 	PDD_BOUND,
577636 	PDD_BOUND_SUSPENDED,
578637 };
638
+
639
+#define MAX_SYSFS_FILENAME_LEN 15
640
+
641
+/*
642
+ * SDMA counter runs at 100MHz frequency.
643
+ * We display SDMA activity in microsecond granularity in sysfs.
644
+ * As a result, the divisor is 100.
645
+ */
646
+#define SDMA_ACTIVITY_DIVISOR  100
579647 
580648 /* Data that is per-process-per device. */
581649 struct kfd_process_device {
....@@ -615,9 +683,53 @@
615683 	 * function.
616684 	 */
617685 	bool already_dequeued;
686
+	bool runtime_inuse;
618687 
619688 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
620689 	enum kfd_pdd_bound bound;
690
+
691
+	/* VRAM usage */
692
+	uint64_t vram_usage;
693
+	struct attribute attr_vram;
694
+	char vram_filename[MAX_SYSFS_FILENAME_LEN];
695
+
696
+	/* SDMA activity tracking */
697
+	uint64_t sdma_past_activity_counter;
698
+	struct attribute attr_sdma;
699
+	char sdma_filename[MAX_SYSFS_FILENAME_LEN];
700
+
701
+	/* Eviction activity tracking */
702
+	uint64_t last_evict_timestamp;
703
+	atomic64_t evict_duration_counter;
704
+	struct attribute attr_evict;
705
+
706
+	struct kobject *kobj_stats;
707
+	unsigned int doorbell_index;
708
+
709
+	/*
710
+	 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
711
+	 * that is associated with device encoded by "this" struct instance. The
712
+	 * value reflects CU usage by all of the waves launched by this process
713
+	 * on this device. A very important property of occupancy parameter is
714
+	 * that its value is a snapshot of current use.
715
+	 *
716
+	 * Following is to be noted regarding how this parameter is reported:
717
+	 *
718
+	 *  The number of waves that a CU can launch is limited by couple of
719
+	 *  parameters. These are encoded by struct amdgpu_cu_info instance
720
+	 *  that is part of every device definition. For GFX9 devices this
721
+	 *  translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
722
+	 *  do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
723
+	 *  when they do use scratch memory. This could change for future
724
+	 *  devices and therefore this example should be considered as a guide.
725
+	 *
726
+	 *  All CU's of a device are available for the process. This may not be true
727
+	 *  under certain conditions - e.g. CU masking.
728
+	 *
729
+	 *  Finally number of CU's that are occupied by a process is affected by both
730
+	 *  number of CU's a device has along with number of other competing processes
731
+	 */
732
+	struct attribute attr_cu_occupancy;
621733 };
622734 
623735 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
....@@ -654,11 +766,7 @@
654766 	/* We want to receive a notification when the mm_struct is destroyed */
655767 	struct mmu_notifier mmu_notifier;
656768 
657
-	/* Use for delayed freeing of kfd_process structure */
658
-	struct rcu_head	rcu;
659
-
660
-	unsigned int pasid;
661
-	unsigned int doorbell_index;
769
+	u32 pasid;
662770 
663771 	/*
664772 	 * List of kfd_process_device structures,
....@@ -698,6 +806,11 @@
698806 	 * restored after an eviction
699807 	 */
700808 	unsigned long last_restore_timestamp;
809
+
810
+	/* Kobj for our procfs */
811
+	struct kobject *kobj;
812
+	struct kobject *kobj_queues;
813
+	struct attribute attr_pasid;
701814 };
702815 
703816 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
....@@ -705,11 +818,13 @@
705818 extern struct srcu_struct kfd_processes_srcu;
706819 
707820 /**
708
- * Ioctl function type.
821
+ * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
709822  *
710
- * \param filep pointer to file structure.
711
- * \param p amdkfd process pointer.
712
- * \param data pointer to arg that was copied from user.
823
+ * @filep: pointer to file structure.
824
+ * @p: amdkfd process pointer.
825
+ * @data: pointer to arg that was copied from user.
826
+ *
827
+ * Return: returns ioctl completion code.
713828  */
714829 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
715830 				void *data);
....@@ -721,12 +836,13 @@
721836 	unsigned int cmd_drv;
722837 	const char *name;
723838 };
839
+bool kfd_dev_is_large_bar(struct kfd_dev *dev);
724840 
725841 int kfd_process_create_wq(void);
726842 void kfd_process_destroy_wq(void);
727843 struct kfd_process *kfd_create_process(struct file *filep);
728844 struct kfd_process *kfd_get_process(const struct task_struct *);
729
-struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
845
+struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid);
730846 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
731847 void kfd_unref_process(struct kfd_process *p);
732848 int kfd_process_evict_queues(struct kfd_process *p);
....@@ -767,8 +883,8 @@
767883 void kfd_pasid_exit(void);
768884 bool kfd_set_pasid_limit(unsigned int new_limit);
769885 unsigned int kfd_get_pasid_limit(void);
770
-unsigned int kfd_pasid_alloc(void);
771
-void kfd_pasid_free(unsigned int pasid);
886
+u32 kfd_pasid_alloc(void);
887
+void kfd_pasid_free(u32 pasid);
772888 
773889 /* Doorbells */
774890 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
....@@ -782,14 +898,14 @@
782898 u32 read_kernel_doorbell(u32 __iomem *db);
783899 void write_kernel_doorbell(void __iomem *db, u32 value);
784900 void write_kernel_doorbell64(void __iomem *db, u64 value);
785
-unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
786
-					struct kfd_process *process,
901
+unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
902
+					struct kfd_process_device *pdd,
787903 					unsigned int doorbell_id);
788
-phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
789
-					struct kfd_process *process);
790
-int kfd_alloc_process_doorbells(struct kfd_process *process);
791
-void kfd_free_process_doorbells(struct kfd_process *process);
792
-
904
+phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
905
+int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
906
+				unsigned int *doorbell_index);
907
+void kfd_free_process_doorbells(struct kfd_dev *kfd,
908
+				unsigned int doorbell_index);
793909 /* GTT Sub-Allocator */
794910 
795911 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
....@@ -798,6 +914,12 @@
798914 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
799915 
800916 extern struct device *kfd_device;
917
+
918
+/* KFD's procfs */
919
+void kfd_procfs_init(void);
920
+void kfd_procfs_shutdown(void);
921
+int kfd_procfs_add_queue(struct queue *q);
922
+void kfd_procfs_del_queue(struct queue *q);
801923 
802924 /* Topology */
803925 int kfd_topology_init(void);
....@@ -809,25 +931,18 @@
809931 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
810932 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
811933 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
934
+struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd);
812935 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
813936 int kfd_numa_node_to_apic_id(int numa_node_id);
937
+void kfd_double_confirm_iommu_support(struct kfd_dev *gpu);
814938 
815939 /* Interrupts */
816940 int kfd_interrupt_init(struct kfd_dev *dev);
817941 void kfd_interrupt_exit(struct kfd_dev *dev);
818
-void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
819942 bool enqueue_ih_ring_entry(struct kfd_dev *kfd,	const void *ih_ring_entry);
820943 bool interrupt_is_wanted(struct kfd_dev *dev,
821944 				const uint32_t *ih_ring_entry,
822945 				uint32_t *patched_ihre, bool *flag);
823
-
824
-/* Power Management */
825
-void kgd2kfd_suspend(struct kfd_dev *kfd);
826
-int kgd2kfd_resume(struct kfd_dev *kfd);
827
-
828
-/* GPU reset */
829
-int kgd2kfd_pre_reset(struct kfd_dev *kfd);
830
-int kgd2kfd_post_reset(struct kfd_dev *kfd);
831946 
832947 /* amdkfd Apertures */
833948 int kfd_init_apertures(struct kfd_process *process);
....@@ -838,8 +953,6 @@
838953 void print_queue_properties(struct queue_properties *q);
839954 void print_queue(struct queue *q);
840955 
841
-struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
842
-					struct kfd_dev *dev);
843956 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
844957 		struct kfd_dev *dev);
845958 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
....@@ -850,12 +963,14 @@
850963 		struct kfd_dev *dev);
851964 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
852965 		struct kfd_dev *dev);
966
+struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
967
+		struct kfd_dev *dev);
853968 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
854969 void device_queue_manager_uninit(struct device_queue_manager *dqm);
855970 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
856971 					enum kfd_queue_type type);
857
-void kernel_queue_uninit(struct kernel_queue *kq);
858
-int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid);
972
+void kernel_queue_uninit(struct kernel_queue *kq, bool hanging);
973
+int kfd_process_vm_fault(struct device_queue_manager *dqm, u32 pasid);
859974 
860975 /* Process Queue Manager */
861976 struct process_queue_node {
....@@ -872,18 +987,28 @@
872987 			    struct kfd_dev *dev,
873988 			    struct file *f,
874989 			    struct queue_properties *properties,
875
-			    unsigned int *qid);
990
+			    unsigned int *qid,
991
+			    uint32_t *p_doorbell_offset_in_process);
876992 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
877993 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
878994 			struct queue_properties *p);
879995 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
880996 			struct queue_properties *p);
997
+int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
998
+			void *gws);
881999 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
8821000 						unsigned int qid);
1001
+struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1002
+						unsigned int qid);
1003
+int pqm_get_wave_state(struct process_queue_manager *pqm,
1004
+		       unsigned int qid,
1005
+		       void __user *ctl_stack,
1006
+		       u32 *ctl_stack_used_size,
1007
+		       u32 *save_area_used_size);
8831008 
884
-int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
885
-				unsigned int fence_value,
886
-				unsigned int timeout_ms);
1009
+int amdkfd_fence_wait_timeout(uint64_t *fence_addr,
1010
+			      uint64_t fence_value,
1011
+			      unsigned int timeout_ms);
8871012 
8881013 /* Packet Manager */
8891014 
....@@ -897,6 +1022,7 @@
8971022 	bool allocated;
8981023 	struct kfd_mem_obj *ib_buffer_obj;
8991024 	unsigned int ib_size_bytes;
1025
+	bool is_over_subscription;
9001026 
9011027 	const struct packet_manager_funcs *pmf;
9021028 };
....@@ -917,7 +1043,7 @@
9171043 			uint32_t filter_param, bool reset,
9181044 			unsigned int sdma_engine);
9191045 	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
920
-			uint64_t fence_address,	uint32_t fence_value);
1046
+			uint64_t fence_address,	uint64_t fence_value);
9211047 	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
9221048 
9231049 	/* Packet sizes */
....@@ -934,12 +1060,12 @@
9341060 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
9351061 
9361062 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
937
-void pm_uninit(struct packet_manager *pm);
1063
+void pm_uninit(struct packet_manager *pm, bool hanging);
9381064 int pm_send_set_resources(struct packet_manager *pm,
9391065 				struct scheduling_resources *res);
9401066 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
9411067 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
942
-				uint32_t fence_value);
1068
+				uint64_t fence_value);
9431069 
9441070 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
9451071 			enum kfd_unmap_queues_filter mode,
....@@ -950,8 +1076,6 @@
9501076 
9511077 /* Following PM funcs can be shared among VI and AI */
9521078 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
953
-int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer,
954
-				struct scheduling_resources *res);
9551079 
9561080 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
9571081 
....@@ -968,12 +1092,12 @@
9681092 		       uint32_t num_events, void __user *data,
9691093 		       bool all, uint32_t user_timeout_ms,
9701094 		       uint32_t *wait_result);
971
-void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
1095
+void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
9721096 				uint32_t valid_id_bits);
9731097 void kfd_signal_iommu_event(struct kfd_dev *dev,
974
-		unsigned int pasid, unsigned long address,
975
-		bool is_write_requested, bool is_execute_requested);
976
-void kfd_signal_hw_exception_event(unsigned int pasid);
1098
+			    u32 pasid, unsigned long address,
1099
+			    bool is_write_requested, bool is_execute_requested);
1100
+void kfd_signal_hw_exception_event(u32 pasid);
9771101 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
9781102 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
9791103 int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
....@@ -984,7 +1108,7 @@
9841108 		     uint64_t *event_page_offset, uint32_t *event_slot_index);
9851109 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
9861110 
987
-void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
1111
+void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
9881112 				struct kfd_vm_fault_info *info);
9891113 
9901114 void kfd_signal_reset_event(struct kfd_dev *dev);
....@@ -995,6 +1119,25 @@
9951119 
9961120 bool kfd_is_locked(void);
9971121 
1122
+/* Compute profile */
1123
+void kfd_inc_compute_active(struct kfd_dev *dev);
1124
+void kfd_dec_compute_active(struct kfd_dev *dev);
1125
+
1126
+/* Cgroup Support */
1127
+/* Check with device cgroup if @kfd device is accessible */
1128
+static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)
1129
+{
1130
+#if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1131
+	struct drm_device *ddev = kfd->ddev;
1132
+
1133
+	return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1134
+					  ddev->render->index,
1135
+					  DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1136
+#else
1137
+	return 0;
1138
+#endif
1139
+}
1140
+
9981141 /* Debugfs */
9991142 #if defined(CONFIG_DEBUG_FS)
10001143