xref: /freebsd/contrib/ofed/include/udma_barrier.h (revision af23369a6deaaeb612ab266eb88b8bb8d560c322)
1 /*
2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #ifndef __UTIL_UDMA_BARRIER_H
34 #define __UTIL_UDMA_BARRIER_H
35 
36 #include <pthread.h>
37 
38 /* Barriers for DMA.
39 
40    These barriers are expliclty only for use with user DMA operations. If you
41    are looking for barriers to use with cache-coherent multi-threaded
42    consitency then look in stdatomic.h. If you need both kinds of synchronicity
43    for the same address then use an atomic operation followed by one
44    of these barriers.
45 
46    When reasoning about these barriers there are two objects:
47      - CPU attached address space (the CPU memory could be a range of things:
48        cached/uncached/non-temporal CPU DRAM, uncached MMIO space in another
49        device, pMEM). Generally speaking the ordering is only relative
50        to the local CPU's view of the system. Eg if the local CPU
51        is not guaranteed to see a write from another CPU then it is also
52        OK for the DMA device to also not see the write after the barrier.
53      - A DMA initiator on a bus. For instance a PCI-E device issuing
54        MemRd/MemWr TLPs.
55 
56    The ordering guarantee is always stated between those two streams. Eg what
57    happens if a MemRd TLP is sent in via PCI-E relative to a CPU WRITE to the
58    same memory location.
59 
60    The providers have a very regular and predictable use of these barriers,
61    to make things very clear each narrow use is given a name and the proper
62    name should be used in the provider as a form of documentation.
63 */
64 
65 /* Ensure that the device's view of memory matches the CPU's view of memory.
66    This should be placed before any MMIO store that could trigger the device
67    to begin doing DMA, such as a device doorbell ring.
68 
69    eg
70     *dma_buf = 1;
71     udma_to_device_barrier();
72     mmio_write(DO_DMA_REG, dma_buf);
73    Must ensure that the device sees the '1'.
74 
75    This is required to fence writes created by the libibverbs user. Those
76    writes could be to any CPU mapped memory object with any cachability mode.
77 
78    NOTE: x86 has historically used a weaker semantic for this barrier, and
79    only fenced normal stores to normal memory. libibverbs users using other
80    memory types or non-temporal stores are required to use SFENCE in their own
81    code prior to calling verbs to start a DMA.
82 */
83 #if defined(__i386__)
84 #define udma_to_device_barrier() asm volatile("" ::: "memory")
85 #elif defined(__x86_64__)
86 #define udma_to_device_barrier() asm volatile("" ::: "memory")
87 #elif defined(__PPC64__)
88 #define udma_to_device_barrier() asm volatile("sync" ::: "memory")
89 #elif defined(__PPC__)
90 #define udma_to_device_barrier() asm volatile("sync" ::: "memory")
91 #elif defined(__ia64__)
92 #define udma_to_device_barrier() asm volatile("mf" ::: "memory")
93 #elif defined(__sparc_v9__)
94 #define udma_to_device_barrier() asm volatile("membar #StoreStore" ::: "memory")
95 #elif defined(__aarch64__)
96 #define udma_to_device_barrier() asm volatile("dsb st" ::: "memory");
97 #elif defined(__sparc__) || defined(__s390x__)
98 #define udma_to_device_barrier() asm volatile("" ::: "memory")
99 #elif defined(__mips__)
100 #include <sys/types.h>
101 #include <machine/atomic.h>
102 #define udma_to_device_barrier() mips_sync()
103 #elif defined(__arm__)
104 #include <sys/types.h>
105 #include <machine/atomic.h>
106 #define udma_to_device_barrier() dmb()
107 #elif defined(__riscv)
108 #include <sys/types.h>
109 #include <machine/atomic.h>
110 #define udma_to_device_barrier() fence()
111 #else
112 #error No architecture specific memory barrier defines found!
113 #endif
114 
115 /* Ensure that all ordered stores from the device are observable from the
116    CPU. This only makes sense after something that observes an ordered store
117    from the device - eg by reading a MMIO register or seeing that CPU memory is
118    updated.
119 
120    This guarantees that all reads that follow the barrier see the ordered
121    stores that preceded the observation.
122 
123    For instance, this would be used after testing a valid bit in a memory
124    that is a DMA target, to ensure that the following reads see the
125    data written before the MemWr TLP that set the valid bit.
126 */
127 #if defined(__i386__)
128 #define udma_from_device_barrier() asm volatile("lock; addl $0,0(%%esp) " ::: "memory")
129 #elif defined(__x86_64__)
130 #define udma_from_device_barrier() asm volatile("lfence" ::: "memory")
131 #elif defined(__PPC64__)
132 #define udma_from_device_barrier() asm volatile("lwsync" ::: "memory")
133 #elif defined(__PPC__)
134 #define udma_from_device_barrier() asm volatile("sync" ::: "memory")
135 #elif defined(__ia64__)
136 #define udma_from_device_barrier() asm volatile("mf" ::: "memory")
137 #elif defined(__sparc_v9__)
138 #define udma_from_device_barrier() asm volatile("membar #LoadLoad" ::: "memory")
139 #elif defined(__aarch64__)
140 #define udma_from_device_barrier() asm volatile("dsb ld" ::: "memory");
141 #elif defined(__sparc__) || defined(__s390x__)
142 #define udma_from_device_barrier() asm volatile("" ::: "memory")
143 #elif defined(__mips__)
144 #define udma_from_device_barrier() mips_sync()
145 #elif defined(__arm__)
146 #define udma_from_device_barrier() dmb()
147 #elif defined(__riscv)
148 #define udma_from_device_barrier() fence()
149 #else
150 #error No architecture specific memory barrier defines found!
151 #endif
152 
153 /* Order writes to CPU memory so that a DMA device cannot view writes after
154    the barrier without also seeing all writes before the barrier. This does
155    not guarantee any writes are visible to DMA.
156 
157    This would be used in cases where a DMA buffer might have a valid bit and
158    data, this barrier is placed after writing the data but before writing the
159    valid bit to ensure the DMA device cannot observe a set valid bit with
160    unwritten data.
161 
162    Compared to udma_to_device_barrier() this barrier is not required to fence
163    anything but normal stores to normal malloc memory. Usage should be:
164 
165    write_wqe
166       udma_to_device_barrier();    // Get user memory ready for DMA
167       wqe->addr = ...;
168       wqe->flags = ...;
169       udma_ordering_write_barrier();  // Guarantee WQE written in order
170       wqe->valid = 1;
171 */
172 #define udma_ordering_write_barrier() udma_to_device_barrier()
173 
174 /* Promptly flush writes to MMIO Write Cominbing memory.
175    This should be used after a write to WC memory. This is both a barrier
176    and a hint to the CPU to flush any buffers to reduce latency to TLP
177    generation.
178 
179    This is not required to have any effect on CPU memory.
180 
181    If done while holding a lock then the ordering of MMIO writes across CPUs
182    must be guaranteed to follow the natural ordering implied by the lock.
183 
184    This must also act as a barrier that prevents write combining, eg
185      *wc_mem = 1;
186      mmio_flush_writes();
187      *wc_mem = 2;
188    Must always produce two MemWr TLPs, '1' and '2'. Without the barrier
189    the CPU is allowed to produce a single TLP '2'.
190 
191    Note that there is no order guarantee for writes to WC memory without
192    barriers.
193 
194    This is intended to be used in conjunction with WC memory to generate large
195    PCI-E MemWr TLPs from the CPU.
196 */
197 #if defined(__i386__)
198 #define mmio_flush_writes() asm volatile("lock; addl $0,0(%%esp) " ::: "memory")
199 #elif defined(__x86_64__)
200 #define mmio_flush_writes() asm volatile("sfence" ::: "memory")
201 #elif defined(__PPC64__)
202 #define mmio_flush_writes() asm volatile("sync" ::: "memory")
203 #elif defined(__PPC__)
204 #define mmio_flush_writes() asm volatile("sync" ::: "memory")
205 #elif defined(__ia64__)
206 #define mmio_flush_writes() asm volatile("fwb" ::: "memory")
207 #elif defined(__sparc_v9__)
208 #define mmio_flush_writes() asm volatile("membar #StoreStore" ::: "memory")
209 #elif defined(__aarch64__)
210 #define mmio_flush_writes() asm volatile("dsb st" ::: "memory");
211 #elif defined(__sparc__) || defined(__s390x__)
212 #define mmio_flush_writes() asm volatile("" ::: "memory")
213 #elif defined(__mips__)
214 #define mmio_flush_writes() mips_sync()
215 #elif defined(__arm__)
216 #define mmio_flush_writes() dmb()
217 #elif defined(__riscv)
218 #define mmio_flush_writes() fence()
219 #else
220 #error No architecture specific memory barrier defines found!
221 #endif
222 
223 /* Prevent WC writes from being re-ordered relative to other MMIO
224    writes. This should be used before a write to WC memory.
225 
226    This must act as a barrier to prevent write re-ordering from different
227    memory types:
228      *mmio_mem = 1;
229      mmio_flush_writes();
230      *wc_mem = 2;
231    Must always produce a TLP '1' followed by '2'.
232 
233    This barrier implies udma_to_device_barrier()
234 
235    This is intended to be used in conjunction with WC memory to generate large
236    PCI-E MemWr TLPs from the CPU.
237 */
238 #define mmio_wc_start() mmio_flush_writes()
239 
240 /* Keep MMIO writes in order.
241    Currently we lack writel macros that universally guarantee MMIO
242    writes happen in order, like the kernel does. Even worse many
243    providers haphazardly open code writes to MMIO memory omitting even
244    volatile.
245 
246    Until this can be fixed with a proper writel macro, this barrier
247    is a stand in to indicate places where MMIO writes should be switched
248    to some future writel.
249 */
250 #define mmio_ordered_writes_hack() mmio_flush_writes()
251 
252 /* Write Combining Spinlock primitive
253 
254    Any access to a multi-value WC region must ensure that multiple cpus do not
255    write to the same values concurrently, these macros make that
256    straightforward and efficient if the choosen exclusion is a spinlock.
257 
258    The spinlock guarantees that the WC writes issued within the critical
259    section are made visible as TLP to the device. The TLP must be seen by the
260    device strictly in the order that the spinlocks are acquired, and combining
261    WC writes between different sections is not permitted.
262 
263    Use of these macros allow the fencing inside the spinlock to be combined
264    with the fencing required for DMA.
265  */
266 static inline void mmio_wc_spinlock(pthread_spinlock_t *lock)
267 {
268 	pthread_spin_lock(lock);
269 #if !defined(__i386__) && !defined(__x86_64__)
270 	/* For x86 the serialization within the spin lock is enough to
271 	 * strongly order WC and other memory types. */
272 	mmio_wc_start();
273 #endif
274 }
275 
276 static inline void mmio_wc_spinunlock(pthread_spinlock_t *lock)
277 {
278 	/* It is possible that on x86 the atomic in the lock is strong enough
279 	 * to force-flush the WC buffers quickly, and this SFENCE can be
280 	 * omitted too. */
281 	mmio_flush_writes();
282 	pthread_spin_unlock(lock);
283 }
284 
285 #endif
286