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
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
|
==================
User Mode Queues
==================
Introduction
============
Similar to the KFD, GPU engine queues move into userspace. The idea is to let
user processes manage their submissions to the GPU engines directly, bypassing
IOCTL calls to the driver to submit work. This reduces overhead and also allows
the GPU to submit work to itself. Applications can set up work graphs of jobs
across multiple GPU engines without needing trips through the CPU.
UMDs directly interface with firmware via per application shared memory areas.
The main vehicle for this is queue. A queue is a ring buffer with a read
pointer (rptr) and a write pointer (wptr). The UMD writes IP specific packets
into the queue and the firmware processes those packets, kicking off work on the
GPU engines. The CPU in the application (or another queue or device) updates
the wptr to tell the firmware how far into the ring buffer to process packets
and the rtpr provides feedback to the UMD on how far the firmware has progressed
in executing those packets. When the wptr and the rptr are equal, the queue is
idle.
Theory of Operation
===================
The various engines on modern AMD GPUs support multiple queues per engine with a
scheduling firmware which handles dynamically scheduling user queues on the
available hardware queue slots. When the number of user queues outnumbers the
available hardware queue slots, the scheduling firmware dynamically maps and
unmaps queues based on priority and time quanta. The state of each user queue
is managed in the kernel driver in an MQD (Memory Queue Descriptor). This is a
buffer in GPU accessible memory that stores the state of a user queue. The
scheduling firmware uses the MQD to load the queue state into an HQD (Hardware
Queue Descriptor) when a user queue is mapped. Each user queue requires a
number of additional buffers which represent the ring buffer and any metadata
needed by the engine for runtime operation. On most engines this consists of
the ring buffer itself, a rptr buffer (where the firmware will shadow the rptr
to userspace), a wptr buffer (where the application will write the wptr for the
firmware to fetch it), and a doorbell. A doorbell is a piece of one of the
device's MMIO BARs which can be mapped to specific user queues. When the
application writes to the doorbell, it will signal the firmware to take some
action. Writing to the doorbell wakes the firmware and causes it to fetch the
wptr and start processing the packets in the queue. Each 4K page of the doorbell
BAR supports specific offset ranges for specific engines. The doorbell of a
queue must be mapped into the aperture aligned to the IP used by the queue
(e.g., GFX, VCN, SDMA, etc.). These doorbell apertures are set up via NBIO
registers. Doorbells are 32 bit or 64 bit (depending on the engine) chunks of
the doorbell BAR. A 4K doorbell page provides 512 64-bit doorbells for up to
512 user queues. A subset of each page is reserved for each IP type supported
on the device. The user can query the doorbell ranges for each IP via the INFO
IOCTL. See the IOCTL Interfaces section for more information.
When an application wants to create a user queue, it allocates the necessary
buffers for the queue (ring buffer, wptr and rptr, context save areas, etc.).
These can be separate buffers or all part of one larger buffer. The application
would map the buffer(s) into its GPUVM and use the GPU virtual addresses of for
the areas of memory they want to use for the user queue. They would also
allocate a doorbell page for the doorbells used by the user queues. The
application would then populate the MQD in the USERQ IOCTL structure with the
GPU virtual addresses and doorbell index they want to use. The user can also
specify the attributes for the user queue (priority, whether the queue is secure
for protected content, etc.). The application would then call the USERQ
CREATE IOCTL to create the queue using the specified MQD details in the IOCTL.
The kernel driver then validates the MQD provided by the application and
translates the MQD into the engine specific MQD format for the IP. The IP
specific MQD would be allocated and the queue would be added to the run list
maintained by the scheduling firmware. Once the queue has been created, the
application can write packets directly into the queue, update the wptr, and
write to the doorbell offset to kick off work in the user queue.
When the application is done with the user queue, it would call the USERQ
FREE IOCTL to destroy it. The kernel driver would preempt the queue and
remove it from the scheduling firmware's run list. Then the IP specific MQD
would be freed and the user queue state would be cleaned up.
Some engines may require the aggregated doorbell too if the engine does not
support doorbells from unmapped queues. The aggregated doorbell is a special
page of doorbell space which wakes the scheduler. In cases where the engine may
be oversubscribed, some queues may not be mapped. If the doorbell is rung when
the queue is not mapped, the engine firmware may miss the request. Some
scheduling firmware may work around this by polling wptr shadows when the
hardware is oversubscribed, other engines may support doorbell updates from
unmapped queues. In the event that one of these options is not available, the
kernel driver will map a page of aggregated doorbell space into each GPUVM
space. The UMD will then update the doorbell and wptr as normal and then write
to the aggregated doorbell as well.
Special Packets
---------------
In order to support legacy implicit synchronization, as well as mixed user and
kernel queues, we need a synchronization mechanism that is secure. Because
kernel queues or memory management tasks depend on kernel fences, we need a way
for user queues to update memory that the kernel can use for a fence, that can't
be messed with by a bad actor. To support this, we've added a protected fence
packet. This packet works by writing a monotonically increasing value to
a memory location that only privileged clients have write access to. User
queues only have read access. When this packet is executed, the memory location
is updated and other queues (kernel or user) can see the results. The
user application would submit this packet in their command stream. The actual
packet format varies from IP to IP (GFX/Compute, SDMA, VCN, etc.), but the
behavior is the same. The packet submission is handled in userspace. The
kernel driver sets up the privileged memory used for each user queue when it
sets the queues up when the application creates them.
Memory Management
=================
It is assumed that all buffers mapped into the GPUVM space for the process are
valid when engines on the GPU are running. The kernel driver will only allow
user queues to run when all buffers are mapped. If there is a memory event that
requires buffer migration, the kernel driver will preempt the user queues,
migrate buffers to where they need to be, update the GPUVM page tables and
invaldidate the TLB, and then resume the user queues.
Interaction with Kernel Queues
==============================
Depending on the IP and the scheduling firmware, you can enable kernel queues
and user queues at the same time, however, you are limited by the HQD slots.
Kernel queues are always mapped so any work that goes into kernel queues will
take priority. This limits the available HQD slots for user queues.
Not all IPs will support user queues on all GPUs. As such, UMDs will need to
support both user queues and kernel queues depending on the IP. For example, a
GPU may support user queues for GFX, compute, and SDMA, but not for VCN, JPEG,
and VPE. UMDs need to support both. The kernel driver provides a way to
determine if user queues and kernel queues are supported on a per IP basis.
UMDs can query this information via the INFO IOCTL and determine whether to use
kernel queues or user queues for each IP.
Queue Resets
============
For most engines, queues can be reset individually. GFX, compute, and SDMA
queues can be reset individually. When a hung queue is detected, it can be
reset either via the scheduling firmware or MMIO. Since there are no kernel
fences for most user queues, they will usually only be detected when some other
event happens; e.g., a memory event which requires migration of buffers. When
the queues are preempted, if the queue is hung, the preemption will fail.
Driver will then look up the queues that failed to preempt and reset them and
record which queues are hung.
On the UMD side, we will add a USERQ QUERY_STATUS IOCTL to query the queue
status. UMD will provide the queue id in the IOCTL and the kernel driver
will check if it has already recorded the queue as hung (e.g., due to failed
peemption) and report back the status.
IOCTL Interfaces
================
GPU virtual addresses used for queues and related data (rptrs, wptrs, context
save areas, etc.) should be validated by the kernel mode driver to prevent the
user from specifying invalid GPU virtual addresses. If the user provides
invalid GPU virtual addresses or doorbell indicies, the IOCTL should return an
error message. These buffers should also be tracked in the kernel driver so
that if the user attempts to unmap the buffer(s) from the GPUVM, the umap call
would return an error.
INFO
----
There are several new INFO queries related to user queues in order to query the
size of user queue meta data needed for a user queue (e.g., context save areas
or shadow buffers), whether kernel or user queues or both are supported
for each IP type, and the offsets for each IP type in each doorbell page.
USERQ
-----
The USERQ IOCTL is used for creating, freeing, and querying the status of user
queues. It supports 3 opcodes:
1. CREATE - Create a user queue. The application provides an MQD-like structure
that defines the type of queue and associated metadata and flags for that
queue type. Returns the queue id.
2. FREE - Free a user queue.
3. QUERY_STATUS - Query that status of a queue. Used to check if the queue is
healthy or not. E.g., if the queue has been reset. (WIP)
USERQ_SIGNAL
------------
The USERQ_SIGNAL IOCTL is used to provide a list of sync objects to be signaled.
USERQ_WAIT
----------
The USERQ_WAIT IOCTL is used to provide a list of sync object to be waited on.
Kernel and User Queues
======================
In order to properly validate and test performance, we have a driver option to
select what type of queues are enabled (kernel queues, user queues or both).
The user_queue driver parameter allows you to enable kernel queues only (0),
user queues and kernel queues (1), and user queues only (2). Enabling user
queues only will free up static queue assignments that would otherwise be used
by kernel queues for use by the scheduling firmware. Some kernel queues are
required for kernel driver operation and they will always be created. When the
kernel queues are not enabled, they are not registered with the drm scheduler
and the CS IOCTL will reject any incoming command submissions which target those
queue types. Kernel queues only mirrors the behavior on all existing GPUs.
Enabling both queues allows for backwards compatibility with old userspace while
still supporting user queues.
|
