1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  */
22 
23 /*
24  * KFD Interrupts.
25  *
26  * AMD GPUs deliver interrupts by pushing an interrupt description onto the
27  * interrupt ring and then sending an interrupt. KGD receives the interrupt
28  * in ISR and sends us a pointer to each new entry on the interrupt ring.
29  *
30  * We generally can't process interrupt-signaled events from ISR, so we call
31  * out to each interrupt client module (currently only the scheduler) to ask if
32  * each interrupt is interesting. If they return true, then it requires further
33  * processing so we copy it to an internal interrupt ring and call each
34  * interrupt client again from a work-queue.
35  *
36  * There's no acknowledgment for the interrupts we use. The hardware simply
37  * queues a new interrupt each time without waiting.
38  *
39  * The fixed-size internal queue means that it's possible for us to lose
40  * interrupts because we have no back-pressure to the hardware.
41  */
42 
43 #include <linux/slab.h>
44 #include <linux/device.h>
45 #include <linux/kfifo.h>
46 #include "kfd_priv.h"
47 
48 #define KFD_IH_NUM_ENTRIES 8192
49 
50 static void interrupt_wq(struct work_struct *);
51 
52 int kfd_interrupt_init(struct kfd_dev *kfd)
53 {
54 	int r;
55 
56 	r = kfifo_alloc(&kfd->ih_fifo,
57 		KFD_IH_NUM_ENTRIES * kfd->device_info.ih_ring_entry_size,
58 		GFP_KERNEL);
59 	if (r) {
60 		dev_err(kfd_chardev(), "Failed to allocate IH fifo\n");
61 		return r;
62 	}
63 
64 	kfd->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI, 1);
65 	if (unlikely(!kfd->ih_wq)) {
66 		kfifo_free(&kfd->ih_fifo);
67 		dev_err(kfd_chardev(), "Failed to allocate KFD IH workqueue\n");
68 		return -ENOMEM;
69 	}
70 	spin_lock_init(&kfd->interrupt_lock);
71 
72 	INIT_WORK(&kfd->interrupt_work, interrupt_wq);
73 
74 	kfd->interrupts_active = true;
75 
76 	/*
77 	 * After this function returns, the interrupt will be enabled. This
78 	 * barrier ensures that the interrupt running on a different processor
79 	 * sees all the above writes.
80 	 */
81 	smp_wmb();
82 
83 	return 0;
84 }
85 
86 void kfd_interrupt_exit(struct kfd_dev *kfd)
87 {
88 	/*
89 	 * Stop the interrupt handler from writing to the ring and scheduling
90 	 * workqueue items. The spinlock ensures that any interrupt running
91 	 * after we have unlocked sees interrupts_active = false.
92 	 */
93 	unsigned long flags;
94 
95 	spin_lock_irqsave(&kfd->interrupt_lock, flags);
96 	kfd->interrupts_active = false;
97 	spin_unlock_irqrestore(&kfd->interrupt_lock, flags);
98 
99 	/*
100 	 * flush_work ensures that there are no outstanding
101 	 * work-queue items that will access interrupt_ring. New work items
102 	 * can't be created because we stopped interrupt handling above.
103 	 */
104 	flush_workqueue(kfd->ih_wq);
105 
106 	kfifo_free(&kfd->ih_fifo);
107 }
108 
109 /*
110  * Assumption: single reader/writer. This function is not re-entrant
111  */
112 bool enqueue_ih_ring_entry(struct kfd_dev *kfd,	const void *ih_ring_entry)
113 {
114 	int count;
115 
116 	count = kfifo_in(&kfd->ih_fifo, ih_ring_entry,
117 				kfd->device_info.ih_ring_entry_size);
118 	if (count != kfd->device_info.ih_ring_entry_size) {
119 		dev_err_ratelimited(kfd_chardev(),
120 			"Interrupt ring overflow, dropping interrupt %d\n",
121 			count);
122 		return false;
123 	}
124 
125 	return true;
126 }
127 
128 /*
129  * Assumption: single reader/writer. This function is not re-entrant
130  */
131 static bool dequeue_ih_ring_entry(struct kfd_dev *kfd, void *ih_ring_entry)
132 {
133 	int count;
134 
135 	count = kfifo_out(&kfd->ih_fifo, ih_ring_entry,
136 				kfd->device_info.ih_ring_entry_size);
137 
138 	WARN_ON(count && count != kfd->device_info.ih_ring_entry_size);
139 
140 	return count == kfd->device_info.ih_ring_entry_size;
141 }
142 
143 static void interrupt_wq(struct work_struct *work)
144 {
145 	struct kfd_dev *dev = container_of(work, struct kfd_dev,
146 						interrupt_work);
147 	uint32_t ih_ring_entry[KFD_MAX_RING_ENTRY_SIZE];
148 
149 	if (dev->device_info.ih_ring_entry_size > sizeof(ih_ring_entry)) {
150 		dev_err_once(kfd_chardev(), "Ring entry too small\n");
151 		return;
152 	}
153 
154 	while (dequeue_ih_ring_entry(dev, ih_ring_entry))
155 		dev->device_info.event_interrupt_class->interrupt_wq(dev,
156 								ih_ring_entry);
157 }
158 
159 bool interrupt_is_wanted(struct kfd_dev *dev,
160 			const uint32_t *ih_ring_entry,
161 			uint32_t *patched_ihre, bool *flag)
162 {
163 	/* integer and bitwise OR so there is no boolean short-circuiting */
164 	unsigned int wanted = 0;
165 
166 	wanted |= dev->device_info.event_interrupt_class->interrupt_isr(dev,
167 					 ih_ring_entry, patched_ihre, flag);
168 
169 	return wanted != 0;
170 }
171