1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
|
From 27a044ad049618cd14de6d5701e51ff2c17dac13 Mon Sep 17 00:00:00 2001
From: Felix Kuehling <Felix.Kuehling@amd.com>
Date: Thu, 30 Jun 2016 16:42:36 -0400
Subject: [PATCH 1466/4131] drm/amdkfd: Add support for resuming SDMA queues
w/o HWS
Add support for resuming SDMA queues w/o HWS in update_queue_nocpsch
and process_restory_queues.
Remote process memory access in hqd_sdma_load requires the mmap_sem
to be locked. mmap_sem locking is done before locking dqm->lock to
avoid a circular lock dependency. There is still a circular lock
dependency warning between the process lock and mmap_sem. However,
the process locks involved are all read-locks, so there is no
possibility of a deadlock.
Change-Id: Iae2c1d05de7c7ff732cf000de91535ab73656c3c
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
---
.../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 59 +++++++++++++---------
1 file changed, 34 insertions(+), 25 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
index 9fbbeee..aacc4dc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -389,18 +389,20 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q)
BUG_ON(!dqm || !q || !q->mqd);
+ if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
+ down_read(&q->process->mm->mmap_sem);
mutex_lock(&dqm->lock);
pdd = kfd_get_process_device_data(q->device, q->process);
if (!pdd) {
- mutex_unlock(&dqm->lock);
- return -ENODEV;
+ retval = -ENODEV;
+ goto out_unlock;
}
mqd = dqm->ops.get_mqd_manager(dqm,
get_mqd_type_from_queue_type(q->properties.type));
if (mqd == NULL) {
- mutex_unlock(&dqm->lock);
- return -ENOMEM;
+ retval = -ENOMEM;
+ goto out_unlock;
}
/*
* Eviction state logic: we only mark active queues as evicted
@@ -418,8 +420,8 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q)
retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
- q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
- /* FIXME: Handle SDMA queues as well */
+ (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
+ q->properties.type == KFD_QUEUE_TYPE_SDMA)) {
if (q->properties.is_active)
retval = mqd->load_mqd(mqd, q->mqd, q->pipe,
q->queue,
@@ -442,7 +444,11 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q)
if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS)
retval = execute_queues_cpsch(dqm, false);
+out_unlock:
mutex_unlock(&dqm->lock);
+ if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
+ up_read(&q->process->mm->mmap_sem);
+
return retval;
}
@@ -494,8 +500,9 @@ int process_evict_queues(struct device_queue_manager *dqm,
retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
- q->properties.type == KFD_QUEUE_TYPE_COMPUTE &&
- q->properties.is_evicted)
+ (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
+ q->properties.type == KFD_QUEUE_TYPE_SDMA) &&
+ q->properties.is_evicted)
retval = mqd->destroy_mqd(mqd, q->mqd,
KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
@@ -515,21 +522,21 @@ int process_restore_queues(struct device_queue_manager *dqm,
{
struct queue *q, *next;
struct mqd_manager *mqd;
+ struct kfd_process_device *pdd =
+ container_of(qpd, struct kfd_process_device, qpd);
int retval = 0;
-
BUG_ON(!dqm || !qpd);
+ if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
+ down_read(&pdd->process->mm->mmap_sem);
mutex_lock(&dqm->lock);
- if (qpd->evicted == 0) { /* already restored, do nothing */
- mutex_unlock(&dqm->lock);
- return 0;
- }
+ if (qpd->evicted == 0) /* already restored, do nothing */
+ goto out_unlock;
if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
qpd->evicted--;
- mutex_unlock(&dqm->lock);
- return 0;
+ goto out_unlock;
}
/* activate all active queues on the qpd */
@@ -544,15 +551,13 @@ int process_restore_queues(struct device_queue_manager *dqm,
q->properties.is_evicted = false;
retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
- q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
- retval =
- mqd->load_mqd(
- mqd,
- q->mqd,
- q->pipe,
- q->queue,
- (uint32_t __user *)q->properties.write_ptr,
- q->process->mm);
+ (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
+ q->properties.type == KFD_QUEUE_TYPE_SDMA))
+ retval = mqd->load_mqd(
+ mqd, q->mqd, q->pipe, q->queue,
+ (uint32_t __user *)
+ q->properties.write_ptr,
+ q->process->mm);
dqm->queue_count++;
}
}
@@ -561,9 +566,13 @@ int process_restore_queues(struct device_queue_manager *dqm,
if (retval == 0)
qpd->evicted = 0;
+
+out_unlock:
mutex_unlock(&dqm->lock);
- return retval;
+ if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
+ up_read(&pdd->process->mm->mmap_sem);
+ return retval;
}
static int register_process_nocpsch(struct device_queue_manager *dqm,
--
2.7.4
|
