1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 /*
25  * KFD Interrupts.
26  *
27  * AMD GPUs deliver interrupts by pushing an interrupt description onto the
28  * interrupt ring and then sending an interrupt. KGD receives the interrupt
29  * in ISR and sends us a pointer to each new entry on the interrupt ring.
30  *
31  * We generally can't process interrupt-signaled events from ISR, so we call
32  * out to each interrupt client module (currently only the scheduler) to ask if
33  * each interrupt is interesting. If they return true, then it requires further
34  * processing so we copy it to an internal interrupt ring and call each
35  * interrupt client again from a work-queue.
36  *
37  * There's no acknowledgment for the interrupts we use. The hardware simply
38  * queues a new interrupt each time without waiting.
39  *
40  * The fixed-size internal queue means that it's possible for us to lose
41  * interrupts because we have no back-pressure to the hardware.
42  */
43 
44 #include <linux/slab.h>
45 #include <linux/device.h>
46 #include <linux/kfifo.h>
47 #include "kfd_priv.h"
48 
49 #define KFD_IH_NUM_ENTRIES 8192
50 
51 static void interrupt_wq(struct work_struct *);
52 
53 int kfd_interrupt_init(struct kfd_dev *kfd)
54 {
55 	int r;
56 
57 	r = kfifo_alloc(&kfd->ih_fifo,
58 		KFD_IH_NUM_ENTRIES * kfd->device_info.ih_ring_entry_size,
59 		GFP_KERNEL);
60 	if (r) {
61 		dev_err(kfd->adev->dev, "Failed to allocate IH fifo\n");
62 		return r;
63 	}
64 
65 	kfd->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI, 1);
66 	if (unlikely(!kfd->ih_wq)) {
67 		kfifo_free(&kfd->ih_fifo);
68 		dev_err(kfd->adev->dev, "Failed to allocate KFD IH workqueue\n");
69 		return -ENOMEM;
70 	}
71 	spin_lock_init(&kfd->interrupt_lock);
72 
73 	INIT_WORK(&kfd->interrupt_work, interrupt_wq);
74 
75 	kfd->interrupts_active = true;
76 
77 	/*
78 	 * After this function returns, the interrupt will be enabled. This
79 	 * barrier ensures that the interrupt running on a different processor
80 	 * sees all the above writes.
81 	 */
82 	smp_wmb();
83 
84 	return 0;
85 }
86 
87 void kfd_interrupt_exit(struct kfd_dev *kfd)
88 {
89 	/*
90 	 * Stop the interrupt handler from writing to the ring and scheduling
91 	 * workqueue items. The spinlock ensures that any interrupt running
92 	 * after we have unlocked sees interrupts_active = false.
93 	 */
94 	unsigned long flags;
95 
96 	spin_lock_irqsave(&kfd->interrupt_lock, flags);
97 	kfd->interrupts_active = false;
98 	spin_unlock_irqrestore(&kfd->interrupt_lock, flags);
99 
100 	/*
101 	 * flush_work ensures that there are no outstanding
102 	 * work-queue items that will access interrupt_ring. New work items
103 	 * can't be created because we stopped interrupt handling above.
104 	 */
105 	flush_workqueue(kfd->ih_wq);
106 
107 	kfifo_free(&kfd->ih_fifo);
108 }
109 
110 /*
111  * Assumption: single reader/writer. This function is not re-entrant
112  */
113 bool enqueue_ih_ring_entry(struct kfd_dev *kfd,	const void *ih_ring_entry)
114 {
115 	int count;
116 
117 	count = kfifo_in(&kfd->ih_fifo, ih_ring_entry,
118 				kfd->device_info.ih_ring_entry_size);
119 	if (count != kfd->device_info.ih_ring_entry_size) {
120 		dev_dbg_ratelimited(kfd->adev->dev,
121 			"Interrupt ring overflow, dropping interrupt %d\n",
122 			count);
123 		return false;
124 	}
125 
126 	return true;
127 }
128 
129 /*
130  * Assumption: single reader/writer. This function is not re-entrant
131  */
132 static bool dequeue_ih_ring_entry(struct kfd_dev *kfd, void *ih_ring_entry)
133 {
134 	int count;
135 
136 	count = kfifo_out(&kfd->ih_fifo, ih_ring_entry,
137 				kfd->device_info.ih_ring_entry_size);
138 
139 	WARN_ON(count && count != kfd->device_info.ih_ring_entry_size);
140 
141 	return count == kfd->device_info.ih_ring_entry_size;
142 }
143 
144 static void interrupt_wq(struct work_struct *work)
145 {
146 	struct kfd_dev *dev = container_of(work, struct kfd_dev,
147 						interrupt_work);
148 	uint32_t ih_ring_entry[KFD_MAX_RING_ENTRY_SIZE];
149 	unsigned long start_jiffies = jiffies;
150 
151 	if (dev->device_info.ih_ring_entry_size > sizeof(ih_ring_entry)) {
152 		dev_err_once(dev->adev->dev, "Ring entry too small\n");
153 		return;
154 	}
155 
156 	while (dequeue_ih_ring_entry(dev, ih_ring_entry)) {
157 		dev->device_info.event_interrupt_class->interrupt_wq(dev,
158 								ih_ring_entry);
159 		if (time_is_before_jiffies(start_jiffies + HZ)) {
160 			/* If we spent more than a second processing signals,
161 			 * reschedule the worker to avoid soft-lockup warnings
162 			 */
163 			queue_work(dev->ih_wq, &dev->interrupt_work);
164 			break;
165 		}
166 	}
167 }
168 
169 bool interrupt_is_wanted(struct kfd_dev *dev,
170 			const uint32_t *ih_ring_entry,
171 			uint32_t *patched_ihre, bool *flag)
172 {
173 	/* integer and bitwise OR so there is no boolean short-circuiting */
174 	unsigned int wanted = 0;
175 
176 	wanted |= dev->device_info.event_interrupt_class->interrupt_isr(dev,
177 					 ih_ring_entry, patched_ihre, flag);
178 
179 	return wanted != 0;
180 }
181