1 <!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook V3.1//EN"[]>
5 <title>Bus-Independent Device Accesses</title>
9 <firstname>Matthew</firstname>
10 <surname>Wilcox</surname>
13 <email>matthew@wil.cx</email>
21 <firstname>Alan</firstname>
22 <surname>Cox</surname>
25 <email>alan@redhat.com</email>
33 <holder>Matthew Wilcox</holder>
38 This documentation is free software; you can redistribute
39 it and/or modify it under the terms of the GNU General Public
40 License as published by the Free Software Foundation; either
41 version 2 of the License, or (at your option) any later
46 This program is distributed in the hope that it will be
47 useful, but WITHOUT ANY WARRANTY; without even the implied
48 warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
49 See the GNU General Public License for more details.
53 You should have received a copy of the GNU General Public
54 License along with this program; if not, write to the Free
55 Software Foundation, Inc., 59 Temple Place, Suite 330, Boston,
60 For more details see the file COPYING in the source
61 distribution of Linux.
69 <title>Introduction</title>
71 Linux provides an API which abstracts performing IO across all busses
72 and devices, allowing device drivers to be written independently of
78 <title>Known Bugs And Assumptions</title>
85 <title>Memory Mapped IO</title>
87 <title>Getting Access to the Device</title>
89 The most widely supported form of IO is memory mapped IO.
90 That is, a part of the CPU's address space is interpreted
91 not as accesses to memory, but as accesses to a device. Some
92 architectures define devices to be at a fixed address, but most
93 have some method of discovering devices. The PCI bus walk is a
94 good example of such a scheme. This document does not cover how
95 to receive such an address, but assumes you are starting with one.
96 Physical addresses are of type unsigned long.
100 This address should not be used directly. Instead, to get an
101 address suitable for passing to the accessor functions described
102 below, you should call <function>ioremap</function>.
103 An address suitable for accessing the device will be returned to you.
107 After you've finished using the device (say, in your module's
108 exit routine), call <function>iounmap</function> in order to return
109 the address space to the kernel. Most architectures allocate new
110 address space each time you call <function>ioremap</function>, and
111 they can run out unless you call <function>iounmap</function>.
116 <title>Accessing the device</title>
118 The part of the interface most used by drivers is reading and
119 writing memory-mapped registers on the device. Linux provides
120 interfaces to read and write 8-bit, 16-bit, 32-bit and 64-bit
121 quantities. Due to a historical accident, these are named byte,
122 word, long and quad accesses. Both read and write accesses are
123 supported; there is no prefetch support at this time.
127 The functions are named <function>readb</function>,
128 <function>readw</function>, <function>readl</function>,
129 <function>readq</function>, <function>readb_relaxed</function>,
130 <function>readw_relaxed</function>, <function>readl_relaxed</function>,
131 <function>readq_relaxed</function>, <function>writeb</function>,
132 <function>writew</function>, <function>writel</function> and
133 <function>writeq</function>.
137 Some devices (such as framebuffers) would like to use larger
138 transfers than 8 bytes at a time. For these devices, the
139 <function>memcpy_toio</function>, <function>memcpy_fromio</function>
140 and <function>memset_io</function> functions are provided.
141 Do not use memset or memcpy on IO addresses; they
142 are not guaranteed to copy data in order.
146 The read and write functions are defined to be ordered. That is the
147 compiler is not permitted to reorder the I/O sequence. When the
148 ordering can be compiler optimised, you can use <function>
149 __readb</function> and friends to indicate the relaxed ordering. Use
154 While the basic functions are defined to be synchronous with respect
155 to each other and ordered with respect to each other the busses the
156 devices sit on may themselves have asynchronicity. In particular many
157 authors are burned by the fact that PCI bus writes are posted
158 asynchronously. A driver author must issue a read from the same
159 device to ensure that writes have occurred in the specific cases the
160 author cares. This kind of property cannot be hidden from driver
161 writers in the API. In some cases, the read used to flush the device
162 may be expected to fail (if the card is resetting, for example). In
163 that case, the read should be done from config space, which is
164 guaranteed to soft-fail if the card doesn't respond.
168 The following is an example of flushing a write to a device when
169 the driver would like to ensure the write's effects are visible prior
170 to continuing execution.
175 qla1280_disable_intrs(struct scsi_qla_host *ha)
177 struct device_reg *reg;
180 /* disable risc and host interrupts */
181 WRT_REG_WORD(&reg->ictrl, 0);
183 * The following read will ensure that the above write
184 * has been received by the device before we return from this
187 RD_REG_WORD(&reg->ictrl);
188 ha->flags.ints_enabled = 0;
193 In addition to write posting, on some large multiprocessing systems
194 (e.g. SGI Challenge, Origin and Altix machines) posted writes won't
195 be strongly ordered coming from different CPUs. Thus it's important
196 to properly protect parts of your driver that do memory-mapped writes
197 with locks and use the <function>mmiowb</function> to make sure they
198 arrive in the order intended. Issuing a regular <function>readX
199 </function> will also ensure write ordering, but should only be used
200 when the driver has to be sure that the write has actually arrived
201 at the device (not that it's simply ordered with respect to other
202 writes), since a full <function>readX</function> is a relatively
207 Generally, one should use <function>mmiowb</function> prior to
208 releasing a spinlock that protects regions using <function>writeb
209 </function> or similar functions that aren't surrounded by <function>
210 readb</function> calls, which will ensure ordering and flushing. The
211 following pseudocode illustrates what might occur if write ordering
212 isn't guaranteed via <function>mmiowb</function> or one of the
213 <function>readX</function> functions.
217 CPU A: spin_lock_irqsave(&dev_lock, flags)
219 CPU A: writel(newval, ring_ptr);
220 CPU A: spin_unlock_irqrestore(&dev_lock, flags)
222 CPU B: spin_lock_irqsave(&dev_lock, flags)
223 CPU B: writel(newval2, ring_ptr);
225 CPU B: spin_unlock_irqrestore(&dev_lock, flags)
229 In the case above, newval2 could be written to ring_ptr before
230 newval. Fixing it is easy though:
234 CPU A: spin_lock_irqsave(&dev_lock, flags)
236 CPU A: writel(newval, ring_ptr);
237 CPU A: mmiowb(); /* ensure no other writes beat us to the device */
238 CPU A: spin_unlock_irqrestore(&dev_lock, flags)
240 CPU B: spin_lock_irqsave(&dev_lock, flags)
241 CPU B: writel(newval2, ring_ptr);
244 CPU B: spin_unlock_irqrestore(&dev_lock, flags)
248 See tg3.c for a real world example of how to use <function>mmiowb
253 PCI ordering rules also guarantee that PIO read responses arrive
254 after any outstanding DMA writes from that bus, since for some devices
255 the result of a <function>readb</function> call may signal to the
256 driver that a DMA transaction is complete. In many cases, however,
257 the driver may want to indicate that the next
258 <function>readb</function> call has no relation to any previous DMA
259 writes performed by the device. The driver can use
260 <function>readb_relaxed</function> for these cases, although only
261 some platforms will honor the relaxed semantics. Using the relaxed
262 read functions will provide significant performance benefits on
263 platforms that support it. The qla2xxx driver provides examples
264 of how to use <function>readX_relaxed</function>. In many cases,
265 a majority of the driver's <function>readX</function> calls can
266 safely be converted to <function>readX_relaxed</function> calls, since
267 only a few will indicate or depend on DMA completion.
272 <title>ISA legacy functions</title>
274 On older kernels (2.2 and earlier) the ISA bus could be read or
275 written with these functions and without ioremap being used. This is
276 no longer true in Linux 2.4. A set of equivalent functions exist for
277 easy legacy driver porting. The functions available are prefixed
278 with 'isa_' and are <function>isa_readb</function>,
279 <function>isa_writeb</function>, <function>isa_readw</function>,
280 <function>isa_writew</function>, <function>isa_readl</function>,
281 <function>isa_writel</function>, <function>isa_memcpy_fromio</function>
282 and <function>isa_memcpy_toio</function>
285 These functions should not be used in new drivers, and will
286 eventually be going away.
293 <title>Port Space Accesses</title>
295 <title>Port Space Explained</title>
298 Another form of IO commonly supported is Port Space. This is a
299 range of addresses separate to the normal memory address space.
300 Access to these addresses is generally not as fast as accesses
301 to the memory mapped addresses, and it also has a potentially
302 smaller address space.
306 Unlike memory mapped IO, no preparation is required
307 to access port space.
312 <title>Accessing Port Space</title>
314 Accesses to this space are provided through a set of functions
315 which allow 8-bit, 16-bit and 32-bit accesses; also
316 known as byte, word and long. These functions are
317 <function>inb</function>, <function>inw</function>,
318 <function>inl</function>, <function>outb</function>,
319 <function>outw</function> and <function>outl</function>.
323 Some variants are provided for these functions. Some devices
324 require that accesses to their ports are slowed down. This
325 functionality is provided by appending a <function>_p</function>
326 to the end of the function. There are also equivalents to memcpy.
327 The <function>ins</function> and <function>outs</function>
328 functions copy bytes, words or longs to the given port.
334 <chapter id="pubfunctions">
335 <title>Public Functions Provided</title>
336 !Einclude/asm-i386/io.h