Loading...
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
51static void interrupt_wq(struct work_struct *);
52
53int kfd_interrupt_init(struct kfd_node *node)
54{
55 int r;
56
57 r = kfifo_alloc(&node->ih_fifo,
58 KFD_IH_NUM_ENTRIES * node->kfd->device_info.ih_ring_entry_size,
59 GFP_KERNEL);
60 if (r) {
61 dev_err(node->adev->dev, "Failed to allocate IH fifo\n");
62 return r;
63 }
64
65 node->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI, 1);
66 if (unlikely(!node->ih_wq)) {
67 kfifo_free(&node->ih_fifo);
68 dev_err(node->adev->dev, "Failed to allocate KFD IH workqueue\n");
69 return -ENOMEM;
70 }
71 spin_lock_init(&node->interrupt_lock);
72
73 INIT_WORK(&node->interrupt_work, interrupt_wq);
74
75 node->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
87void kfd_interrupt_exit(struct kfd_node *node)
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(&node->interrupt_lock, flags);
97 node->interrupts_active = false;
98 spin_unlock_irqrestore(&node->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(node->ih_wq);
106
107 kfifo_free(&node->ih_fifo);
108}
109
110/*
111 * Assumption: single reader/writer. This function is not re-entrant
112 */
113bool enqueue_ih_ring_entry(struct kfd_node *node, const void *ih_ring_entry)
114{
115 int count;
116
117 count = kfifo_in(&node->ih_fifo, ih_ring_entry,
118 node->kfd->device_info.ih_ring_entry_size);
119 if (count != node->kfd->device_info.ih_ring_entry_size) {
120 dev_dbg_ratelimited(node->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 */
132static bool dequeue_ih_ring_entry(struct kfd_node *node, void *ih_ring_entry)
133{
134 int count;
135
136 count = kfifo_out(&node->ih_fifo, ih_ring_entry,
137 node->kfd->device_info.ih_ring_entry_size);
138
139 WARN_ON(count && count != node->kfd->device_info.ih_ring_entry_size);
140
141 return count == node->kfd->device_info.ih_ring_entry_size;
142}
143
144static void interrupt_wq(struct work_struct *work)
145{
146 struct kfd_node *dev = container_of(work, struct kfd_node,
147 interrupt_work);
148 uint32_t ih_ring_entry[KFD_MAX_RING_ENTRY_SIZE];
149 unsigned long start_jiffies = jiffies;
150
151 if (dev->kfd->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->kfd->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
169bool interrupt_is_wanted(struct kfd_node *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->kfd->device_info.event_interrupt_class->interrupt_isr(dev,
177 ih_ring_entry, patched_ihre, flag);
178
179 return wanted != 0;
180}
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
51static void interrupt_wq(struct work_struct *);
52
53int kfd_interrupt_init(struct kfd_node *node)
54{
55 int r;
56
57 r = kfifo_alloc(&node->ih_fifo,
58 KFD_IH_NUM_ENTRIES * node->kfd->device_info.ih_ring_entry_size,
59 GFP_KERNEL);
60 if (r) {
61 dev_err(node->adev->dev, "Failed to allocate IH fifo\n");
62 return r;
63 }
64
65 if (!node->kfd->ih_wq) {
66 node->kfd->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI | WQ_UNBOUND,
67 node->kfd->num_nodes);
68 if (unlikely(!node->kfd->ih_wq)) {
69 kfifo_free(&node->ih_fifo);
70 dev_err(node->adev->dev, "Failed to allocate KFD IH workqueue\n");
71 return -ENOMEM;
72 }
73 }
74 spin_lock_init(&node->interrupt_lock);
75
76 INIT_WORK(&node->interrupt_work, interrupt_wq);
77
78 node->interrupts_active = true;
79
80 /*
81 * After this function returns, the interrupt will be enabled. This
82 * barrier ensures that the interrupt running on a different processor
83 * sees all the above writes.
84 */
85 smp_wmb();
86
87 return 0;
88}
89
90void kfd_interrupt_exit(struct kfd_node *node)
91{
92 /*
93 * Stop the interrupt handler from writing to the ring and scheduling
94 * workqueue items. The spinlock ensures that any interrupt running
95 * after we have unlocked sees interrupts_active = false.
96 */
97 unsigned long flags;
98
99 spin_lock_irqsave(&node->interrupt_lock, flags);
100 node->interrupts_active = false;
101 spin_unlock_irqrestore(&node->interrupt_lock, flags);
102 kfifo_free(&node->ih_fifo);
103}
104
105/*
106 * Assumption: single reader/writer. This function is not re-entrant
107 */
108bool enqueue_ih_ring_entry(struct kfd_node *node, const void *ih_ring_entry)
109{
110 int count;
111
112 count = kfifo_in(&node->ih_fifo, ih_ring_entry,
113 node->kfd->device_info.ih_ring_entry_size);
114 if (count != node->kfd->device_info.ih_ring_entry_size) {
115 dev_dbg_ratelimited(node->adev->dev,
116 "Interrupt ring overflow, dropping interrupt %d\n",
117 count);
118 return false;
119 }
120
121 return true;
122}
123
124/*
125 * Assumption: single reader/writer. This function is not re-entrant
126 */
127static bool dequeue_ih_ring_entry(struct kfd_node *node, void *ih_ring_entry)
128{
129 int count;
130
131 count = kfifo_out(&node->ih_fifo, ih_ring_entry,
132 node->kfd->device_info.ih_ring_entry_size);
133
134 WARN_ON(count && count != node->kfd->device_info.ih_ring_entry_size);
135
136 return count == node->kfd->device_info.ih_ring_entry_size;
137}
138
139static void interrupt_wq(struct work_struct *work)
140{
141 struct kfd_node *dev = container_of(work, struct kfd_node,
142 interrupt_work);
143 uint32_t ih_ring_entry[KFD_MAX_RING_ENTRY_SIZE];
144 unsigned long start_jiffies = jiffies;
145
146 if (dev->kfd->device_info.ih_ring_entry_size > sizeof(ih_ring_entry)) {
147 dev_err_once(dev->adev->dev, "Ring entry too small\n");
148 return;
149 }
150
151 while (dequeue_ih_ring_entry(dev, ih_ring_entry)) {
152 dev->kfd->device_info.event_interrupt_class->interrupt_wq(dev,
153 ih_ring_entry);
154 if (time_is_before_jiffies(start_jiffies + HZ)) {
155 /* If we spent more than a second processing signals,
156 * reschedule the worker to avoid soft-lockup warnings
157 */
158 queue_work(dev->kfd->ih_wq, &dev->interrupt_work);
159 break;
160 }
161 }
162}
163
164bool interrupt_is_wanted(struct kfd_node *dev,
165 const uint32_t *ih_ring_entry,
166 uint32_t *patched_ihre, bool *flag)
167{
168 /* integer and bitwise OR so there is no boolean short-circuiting */
169 unsigned int wanted = 0;
170
171 wanted |= dev->kfd->device_info.event_interrupt_class->interrupt_isr(dev,
172 ih_ring_entry, patched_ihre, flag);
173
174 return wanted != 0;
175}