1 <?xml version="1.0" encoding="UTF-8"?>
2 <!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook XML V4.1.2//EN"
3 "http://www.oasis-open.org/docbook/xml/4.1.2/docbookx.dtd" []>
7 <title>Bus-Independent Device Accesses</title>
11 <firstname>Matthew</firstname>
12 <surname>Wilcox</surname>
15 <email>matthew@wil.cx</email>
23 <firstname>Alan</firstname>
24 <surname>Cox</surname>
27 <email>alan@lxorguk.ukuu.org.uk</email>
35 <holder>Matthew Wilcox</holder>
40 This documentation is free software; you can redistribute
41 it and/or modify it under the terms of the GNU General Public
42 License as published by the Free Software Foundation; either
43 version 2 of the License, or (at your option) any later
48 This program is distributed in the hope that it will be
49 useful, but WITHOUT ANY WARRANTY; without even the implied
50 warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
51 See the GNU General Public License for more details.
55 You should have received a copy of the GNU General Public
56 License along with this program; if not, write to the Free
57 Software Foundation, Inc., 59 Temple Place, Suite 330, Boston,
62 For more details see the file COPYING in the source
63 distribution of Linux.
71 <title>Introduction</title>
73 Linux provides an API which abstracts performing IO across all busses
74 and devices, allowing device drivers to be written independently of
80 <title>Known Bugs And Assumptions</title>
87 <title>Memory Mapped IO</title>
88 <sect1 id="getting_access_to_the_device">
89 <title>Getting Access to the Device</title>
91 The most widely supported form of IO is memory mapped IO.
92 That is, a part of the CPU's address space is interpreted
93 not as accesses to memory, but as accesses to a device. Some
94 architectures define devices to be at a fixed address, but most
95 have some method of discovering devices. The PCI bus walk is a
96 good example of such a scheme. This document does not cover how
97 to receive such an address, but assumes you are starting with one.
98 Physical addresses are of type unsigned long.
102 This address should not be used directly. Instead, to get an
103 address suitable for passing to the accessor functions described
104 below, you should call <function>ioremap</function>.
105 An address suitable for accessing the device will be returned to you.
109 After you've finished using the device (say, in your module's
110 exit routine), call <function>iounmap</function> in order to return
111 the address space to the kernel. Most architectures allocate new
112 address space each time you call <function>ioremap</function>, and
113 they can run out unless you call <function>iounmap</function>.
117 <sect1 id="accessing_the_device">
118 <title>Accessing the device</title>
120 The part of the interface most used by drivers is reading and
121 writing memory-mapped registers on the device. Linux provides
122 interfaces to read and write 8-bit, 16-bit, 32-bit and 64-bit
123 quantities. Due to a historical accident, these are named byte,
124 word, long and quad accesses. Both read and write accesses are
125 supported; there is no prefetch support at this time.
129 The functions are named <function>readb</function>,
130 <function>readw</function>, <function>readl</function>,
131 <function>readq</function>, <function>readb_relaxed</function>,
132 <function>readw_relaxed</function>, <function>readl_relaxed</function>,
133 <function>readq_relaxed</function>, <function>writeb</function>,
134 <function>writew</function>, <function>writel</function> and
135 <function>writeq</function>.
139 Some devices (such as framebuffers) would like to use larger
140 transfers than 8 bytes at a time. For these devices, the
141 <function>memcpy_toio</function>, <function>memcpy_fromio</function>
142 and <function>memset_io</function> functions are provided.
143 Do not use memset or memcpy on IO addresses; they
144 are not guaranteed to copy data in order.
148 The read and write functions are defined to be ordered. That is the
149 compiler is not permitted to reorder the I/O sequence. When the
150 ordering can be compiler optimised, you can use <function>
151 __readb</function> and friends to indicate the relaxed ordering. Use
156 While the basic functions are defined to be synchronous with respect
157 to each other and ordered with respect to each other the busses the
158 devices sit on may themselves have asynchronicity. In particular many
159 authors are burned by the fact that PCI bus writes are posted
160 asynchronously. A driver author must issue a read from the same
161 device to ensure that writes have occurred in the specific cases the
162 author cares. This kind of property cannot be hidden from driver
163 writers in the API. In some cases, the read used to flush the device
164 may be expected to fail (if the card is resetting, for example). In
165 that case, the read should be done from config space, which is
166 guaranteed to soft-fail if the card doesn't respond.
170 The following is an example of flushing a write to a device when
171 the driver would like to ensure the write's effects are visible prior
172 to continuing execution.
177 qla1280_disable_intrs(struct scsi_qla_host *ha)
179 struct device_reg *reg;
182 /* disable risc and host interrupts */
183 WRT_REG_WORD(&reg->ictrl, 0);
185 * The following read will ensure that the above write
186 * has been received by the device before we return from this
189 RD_REG_WORD(&reg->ictrl);
190 ha->flags.ints_enabled = 0;
195 In addition to write posting, on some large multiprocessing systems
196 (e.g. SGI Challenge, Origin and Altix machines) posted writes won't
197 be strongly ordered coming from different CPUs. Thus it's important
198 to properly protect parts of your driver that do memory-mapped writes
199 with locks and use the <function>mmiowb</function> to make sure they
200 arrive in the order intended. Issuing a regular <function>readX
201 </function> will also ensure write ordering, but should only be used
202 when the driver has to be sure that the write has actually arrived
203 at the device (not that it's simply ordered with respect to other
204 writes), since a full <function>readX</function> is a relatively
209 Generally, one should use <function>mmiowb</function> prior to
210 releasing a spinlock that protects regions using <function>writeb
211 </function> or similar functions that aren't surrounded by <function>
212 readb</function> calls, which will ensure ordering and flushing. The
213 following pseudocode illustrates what might occur if write ordering
214 isn't guaranteed via <function>mmiowb</function> or one of the
215 <function>readX</function> functions.
219 CPU A: spin_lock_irqsave(&dev_lock, flags)
221 CPU A: writel(newval, ring_ptr);
222 CPU A: spin_unlock_irqrestore(&dev_lock, flags)
224 CPU B: spin_lock_irqsave(&dev_lock, flags)
225 CPU B: writel(newval2, ring_ptr);
227 CPU B: spin_unlock_irqrestore(&dev_lock, flags)
231 In the case above, newval2 could be written to ring_ptr before
232 newval. Fixing it is easy though:
236 CPU A: spin_lock_irqsave(&dev_lock, flags)
238 CPU A: writel(newval, ring_ptr);
239 CPU A: mmiowb(); /* ensure no other writes beat us to the device */
240 CPU A: spin_unlock_irqrestore(&dev_lock, flags)
242 CPU B: spin_lock_irqsave(&dev_lock, flags)
243 CPU B: writel(newval2, ring_ptr);
246 CPU B: spin_unlock_irqrestore(&dev_lock, flags)
250 See tg3.c for a real world example of how to use <function>mmiowb
255 PCI ordering rules also guarantee that PIO read responses arrive
256 after any outstanding DMA writes from that bus, since for some devices
257 the result of a <function>readb</function> call may signal to the
258 driver that a DMA transaction is complete. In many cases, however,
259 the driver may want to indicate that the next
260 <function>readb</function> call has no relation to any previous DMA
261 writes performed by the device. The driver can use
262 <function>readb_relaxed</function> for these cases, although only
263 some platforms will honor the relaxed semantics. Using the relaxed
264 read functions will provide significant performance benefits on
265 platforms that support it. The qla2xxx driver provides examples
266 of how to use <function>readX_relaxed</function>. In many cases,
267 a majority of the driver's <function>readX</function> calls can
268 safely be converted to <function>readX_relaxed</function> calls, since
269 only a few will indicate or depend on DMA completion.
275 <chapter id="port_space_accesses">
276 <title>Port Space Accesses</title>
277 <sect1 id="port_space_explained">
278 <title>Port Space Explained</title>
281 Another form of IO commonly supported is Port Space. This is a
282 range of addresses separate to the normal memory address space.
283 Access to these addresses is generally not as fast as accesses
284 to the memory mapped addresses, and it also has a potentially
285 smaller address space.
289 Unlike memory mapped IO, no preparation is required
290 to access port space.
294 <sect1 id="accessing_port_space">
295 <title>Accessing Port Space</title>
297 Accesses to this space are provided through a set of functions
298 which allow 8-bit, 16-bit and 32-bit accesses; also
299 known as byte, word and long. These functions are
300 <function>inb</function>, <function>inw</function>,
301 <function>inl</function>, <function>outb</function>,
302 <function>outw</function> and <function>outl</function>.
306 Some variants are provided for these functions. Some devices
307 require that accesses to their ports are slowed down. This
308 functionality is provided by appending a <function>_p</function>
309 to the end of the function. There are also equivalents to memcpy.
310 The <function>ins</function> and <function>outs</function>
311 functions copy bytes, words or longs to the given port.
317 <chapter id="pubfunctions">
318 <title>Public Functions Provided</title>
319 !Iarch/x86/include/asm/io_32.h