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/*
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 "kfd_priv.h"
46
47#define KFD_INTERRUPT_RING_SIZE 1024
48
49static void interrupt_wq(struct work_struct *);
50
51int kfd_interrupt_init(struct kfd_dev *kfd)
52{
53 void *interrupt_ring = kmalloc_array(KFD_INTERRUPT_RING_SIZE,
54 kfd->device_info->ih_ring_entry_size,
55 GFP_KERNEL);
56 if (!interrupt_ring)
57 return -ENOMEM;
58
59 kfd->interrupt_ring = interrupt_ring;
60 kfd->interrupt_ring_size =
61 KFD_INTERRUPT_RING_SIZE * kfd->device_info->ih_ring_entry_size;
62 atomic_set(&kfd->interrupt_ring_wptr, 0);
63 atomic_set(&kfd->interrupt_ring_rptr, 0);
64
65 spin_lock_init(&kfd->interrupt_lock);
66
67 INIT_WORK(&kfd->interrupt_work, interrupt_wq);
68
69 kfd->interrupts_active = true;
70
71 /*
72 * After this function returns, the interrupt will be enabled. This
73 * barrier ensures that the interrupt running on a different processor
74 * sees all the above writes.
75 */
76 smp_wmb();
77
78 return 0;
79}
80
81void kfd_interrupt_exit(struct kfd_dev *kfd)
82{
83 /*
84 * Stop the interrupt handler from writing to the ring and scheduling
85 * workqueue items. The spinlock ensures that any interrupt running
86 * after we have unlocked sees interrupts_active = false.
87 */
88 unsigned long flags;
89
90 spin_lock_irqsave(&kfd->interrupt_lock, flags);
91 kfd->interrupts_active = false;
92 spin_unlock_irqrestore(&kfd->interrupt_lock, flags);
93
94 /*
95 * Flush_scheduled_work ensures that there are no outstanding
96 * work-queue items that will access interrupt_ring. New work items
97 * can't be created because we stopped interrupt handling above.
98 */
99 flush_scheduled_work();
100
101 kfree(kfd->interrupt_ring);
102}
103
104/*
105 * This assumes that it can't be called concurrently with itself
106 * but only with dequeue_ih_ring_entry.
107 */
108bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry)
109{
110 unsigned int rptr = atomic_read(&kfd->interrupt_ring_rptr);
111 unsigned int wptr = atomic_read(&kfd->interrupt_ring_wptr);
112
113 if ((rptr - wptr) % kfd->interrupt_ring_size ==
114 kfd->device_info->ih_ring_entry_size) {
115 /* This is very bad, the system is likely to hang. */
116 dev_err_ratelimited(kfd_chardev(),
117 "Interrupt ring overflow, dropping interrupt.\n");
118 return false;
119 }
120
121 memcpy(kfd->interrupt_ring + wptr, ih_ring_entry,
122 kfd->device_info->ih_ring_entry_size);
123
124 wptr = (wptr + kfd->device_info->ih_ring_entry_size) %
125 kfd->interrupt_ring_size;
126 smp_wmb(); /* Ensure memcpy'd data is visible before wptr update. */
127 atomic_set(&kfd->interrupt_ring_wptr, wptr);
128
129 return true;
130}
131
132/*
133 * This assumes that it can't be called concurrently with itself
134 * but only with enqueue_ih_ring_entry.
135 */
136static bool dequeue_ih_ring_entry(struct kfd_dev *kfd, void *ih_ring_entry)
137{
138 /*
139 * Assume that wait queues have an implicit barrier, i.e. anything that
140 * happened in the ISR before it queued work is visible.
141 */
142
143 unsigned int wptr = atomic_read(&kfd->interrupt_ring_wptr);
144 unsigned int rptr = atomic_read(&kfd->interrupt_ring_rptr);
145
146 if (rptr == wptr)
147 return false;
148
149 memcpy(ih_ring_entry, kfd->interrupt_ring + rptr,
150 kfd->device_info->ih_ring_entry_size);
151
152 rptr = (rptr + kfd->device_info->ih_ring_entry_size) %
153 kfd->interrupt_ring_size;
154
155 /*
156 * Ensure the rptr write update is not visible until
157 * memcpy has finished reading.
158 */
159 smp_mb();
160 atomic_set(&kfd->interrupt_ring_rptr, rptr);
161
162 return true;
163}
164
165static void interrupt_wq(struct work_struct *work)
166{
167 struct kfd_dev *dev = container_of(work, struct kfd_dev,
168 interrupt_work);
169
170 uint32_t ih_ring_entry[DIV_ROUND_UP(
171 dev->device_info->ih_ring_entry_size,
172 sizeof(uint32_t))];
173
174 while (dequeue_ih_ring_entry(dev, ih_ring_entry))
175 dev->device_info->event_interrupt_class->interrupt_wq(dev,
176 ih_ring_entry);
177}
178
179bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry)
180{
181 /* integer and bitwise OR so there is no boolean short-circuiting */
182 unsigned wanted = 0;
183
184 wanted |= dev->device_info->event_interrupt_class->interrupt_isr(dev,
185 ih_ring_entry);
186
187 return wanted != 0;
188}