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
|
Memory Addressing
=================
There is 3 types of addresses: virtual, physical, and bus. For DMA a bus
address is used. However, on x86 physical and bus addresses are the same (on
other architectures it is not guaranteed). Anyway, this assumption is still
used by xdma driver, it uses phiscal address for DMA access. I have ported
in the same way. Now, we need to provide additionaly bus-addresses in kmem
abstraction and use it in NWL DMA implementation.
DMA Access Synchronization
==========================
- At driver level, few types of buffers are supported:
* SIMPLE - non-reusable buffers, the use infomation can be used for cleanup
after crashed applications.
* EXCLUSIVE - reusable buffers which can be mmaped by a single appliction
only. There is two modes of these buffers:
+ Buffers in a STANDARD mode are created for a single DMA operation and
if such buffer is detected while trying to reuse, the last operation
has failed and reset is needed.
+ Buffers in a PERSISTENT mode are preserved between invocations of
control application and cleaned up only after the PERSISTENT flag is
removed
* SHARED - reusable buffers shared by multiple processes. Not really
needed at the moment.
KMEM_FLAG_HW - indicates that buffer can be used by hardware, acually this
means that DMA will be enabled afterwards. The driver is not able to check
if it really was enable and therefore will block any attempt to release
buffer until KMEM_HW_FLAG is passed to kmem_free routine as well. The later
should only called with KMEM_HW_FLAG after the DMA engine is stopped. Then,
the driver can be realesd by kmem_free if ref count reaches 0.
KMEM_FLAG_EXCLUSIVE - prevents multiple processes mmaping the buffer
simultaneously. This is used to prevent multiple processes use the same
DMA engine at the same time. When passed to kmem_free, allows to clean
buffers with lost clients even for shared buffers.
KMEM_FLAG_REUSE - requires reuse of existing buffer. If reusable buffer is
found (non-reusable buffers, i.e. allocated without KMEM_FLAG_REUSE are
ignored), it is returned instead of allocation. Three types of usage
counters are used. At moment of allocation, the HW reference is set if
neccessary. The usage counter is increased by kmem_alloc function and
decreased by kmem_free. Finally, the reference is obtained at returned
during mmap/munmap. So, on kmem_free, we do not clean
a) buffers with reference count above zero or hardware reference set.
REUSE flag should be supplied, overwise the error is returned
b) PERSISTENT buffer. REUSE flash should be supplied, overwise the
error is returned
c) non-exclusive buffers with usage counter above zero (For exclusive
buffer the value of usage counter above zero just means that application
have failed without cleaning buffers first. There is no easy way to
detect that for shared buffers, so it is left as manual operation in
this case)
d) any buffer if KMEM_FLAG_REUSE was provided to function
During module unload, only buffers with references can prevent cleanup. In
this case the only possiblity to free the driver is to call kmem_free
passing FORCE flags.
KMEM_FLAG_PERSISTENT - if passed to allocation routine, changes mode of
buffer to PERSISTENT, if passed to free routine, vice-versa changes mode
of buffer to NORMAL. Basically, if we call 'pci --dma-start' this flag
should be passed to alloc and if we call 'pci --dma-stop' it should be
passed to free. In other case, the flag should not be present.
If application crashed, the munmap while be still called cleaning software
references. However, the hardware reference will stay since it is not clear
if hardware channel was closed or not. To lift hardware reference, the
application can be re-executed (or dma_stop called, for instance).
* If there is no hardware reference, the buffers will be reused by next
call to application and for EXCLUSIVE buffer cleaned at the end. For SHARED
buffers they will be cleaned during module cleanup only (no active
references).
* The buffer will be reused by next call which can result in wrong behaviour
if buffer left in incoherent stage. This should be handled on upper level.
- At pcilib/kmem level synchronization of multiple buffers is performed
* The HW reference and following modes should be consistent between member
parts: REUSABLE, PERSISTENT, EXCLUSIVE (only HW reference and PERSISTENT
mode should be checked, others are handled on dirver level)
* It is fine if only part of buffers are reused and others are newly
allocated. However, on higher level this can be checked and resulting
in failure.
Treatment of inconsistencies:
* Buffers are in PRESISTENT mode, but newly allocated, OK
* Buffers are reused, but are not in PERSISTENT mode (for EXCLUSIVE buffers
this means that application has crashed during the last execution), OK
* Some of buffers are reused (not just REUSABLE, but actually reused),
others - not, OK until
a) either PERSISTENT flag is set or reused buffers are non-PERSISTENT
b) either HW flag is set or reused buffers does not hold HW reference
* PERSISTENT mode inconsistency, FAIL (even if we are going to set
PERSISTENT mode anyway)
* HW reference inconsistency, FAIL (even if we are going to set
HW flag anyway)
On allocation error at some of the buffer, call clean routine and
* Preserve PERSISTENT mode and HW reference if buffers held them before
unsuccessful kmem initialization. Until the last failed block, the blocks
of kmem should be consistent. The HW/PERSISTENT flags should be removed
if all reused blocks were in HW/PERSISTENT mode. The last block needs
special treatment. The flags may be removed for the block if it was
HW/PERSISTENT state (and others not).
* Remove REUSE flag, we want to clean if allowed by current buffer status
* EXCLUSIVE flag is not important for kmem_free routine.
- At DMA level
There is 4 components of DMA access:
* DMA engine enabled/disabled
* DMA engine IRQs enabled/disabled - always enabled at startup
* Memory buffers
* Ring start/stop pointers
To prevent multiple processes accessing DMA engine in parallel, the first
action is buffer initialization which will fail if buffers already used
* Always with REUSE, EXCLUSIVE, and HW flags
* Optionally with PERSISTENT flag (if DMA_PERSISTENT flag is set)
If another DMA app is running, the buffer allocation will fail (no dma_stop
is executed in this case)
Depending on PRESERVE flag, kmem_free will be called with REUSE flag
keeping buffer in memory (this is redundant since HW flag is enough) or HW
flag indicating that DMA engine is stopped and buffer could be cleaned.
PERSISTENT flag is defined by DMA_PERSISTENT flag passed to stop routine.
PRESERVE flag is enforced if DMA_PERSISTENT is not passed to dma_stop
routine and either it:
a) Explicitely set by DMA_PERMANENT flag passed to dma_start
function
b) Implicitely set if DMA engine is already enabled during dma_start,
all buffers are reused, and are in persistent mode.
If PRESERVE flag is on, the engine will not be stopped at the end of
execution (and buffers will stay because of HW flag).
If buffers are reused and are already in PERSISTENT mode, DMA engine was on
before dma_start (PRESERVE flag is ignored, because it can be enforced),
ring pointers are calculated from LAST_BD and states of ring elements.
If previous application crashed (i.e. buffers may be corrupted). Two
cases are possible:
* If during the call buffers were in non-PERSISTENT mode, it can be
easily detected - buffers are reused, but are not in PERSISTENT mode
(or at least was not before we set them to). In this case we just
reinitialize all buffers.
* If during the call buffers were in PERSISTENT mode, it is up to
user to check their consistency and restart DMA engine.]
IRQs are enabled and disabled at each call
Register Access Synchronization
===============================
We need to serialize access to the registers by the different running
applications and handle case when registers are accessed indirectly by
writting PCI BARs (DMA implementations, for instance).
- Module-assisted locking:
* During initialization the locking context is created (which is basicaly
a kmem_handle of type LOCK_PAGE.
* This locking context is passed to the kernel module along with lock type
(LOCK_BANK) and lock item (BANK ADDRESS). If lock context is already owns
lock on the specified bank, just reference number is increased, otherwise
we are trying to obtain new lock.
* Kernel module just iterates over all registered lock pages and checks if
any holds the specified lock. if not, the lock is obtained and registered
in the our lock page.
* This allows to share access between multiple threads of single application
(by using the same lock page) or protect (by using own lock pages by each of
the threads)
* Either on application cleanup or if application crashed, the memory mapping
of lock page is removed and, hence, locks are freed.
- Multiple-ways of accessing registers
Because of reference counting, we can successfully obtain locks multiple
times if necessary. The following locks are protecting register access:
a) Global register_read/write lock bank before executing implementation
b) DMA bank is locked by global DMA functions. So we can access the
registers using plain PCI bar read/write.
c) Sequence of register operations can be protected with pcilib_lock_bank
function
Reading raw register space or PCI bank is not locked.
* Ok. We can detect banks which will be affected by PCI read/write and
lock them. But shall we do it?
Register/DMA Configuration
==========================
- XML description of registers
- Formal XML-based (or non XML-based) language for DMA implementation.
a) Writting/Reading register values
b) Wait until <register1>=<value> on <register2>=<value> report error
c) ... ?
|
