Pull sem2mutex-ioc4 into release branch
[linux-2.6] / Documentation / io_ordering.txt
1 On some platforms, so-called memory-mapped I/O is weakly ordered.  On such
2 platforms, driver writers are responsible for ensuring that I/O writes to
3 memory-mapped addresses on their device arrive in the order intended.  This is
4 typically done by reading a 'safe' device or bridge register, causing the I/O
5 chipset to flush pending writes to the device before any reads are posted.  A
6 driver would usually use this technique immediately prior to the exit of a
7 critical section of code protected by spinlocks.  This would ensure that
8 subsequent writes to I/O space arrived only after all prior writes (much like a
9 memory barrier op, mb(), only with respect to I/O).
10
11 A more concrete example from a hypothetical device driver:
12
13         ...
14 CPU A:  spin_lock_irqsave(&dev_lock, flags)
15 CPU A:  val = readl(my_status);
16 CPU A:  ...
17 CPU A:  writel(newval, ring_ptr);
18 CPU A:  spin_unlock_irqrestore(&dev_lock, flags)
19         ...
20 CPU B:  spin_lock_irqsave(&dev_lock, flags)
21 CPU B:  val = readl(my_status);
22 CPU B:  ...
23 CPU B:  writel(newval2, ring_ptr);
24 CPU B:  spin_unlock_irqrestore(&dev_lock, flags)
25         ...
26
27 In the case above, the device may receive newval2 before it receives newval,
28 which could cause problems.  Fixing it is easy enough though:
29
30         ...
31 CPU A:  spin_lock_irqsave(&dev_lock, flags)
32 CPU A:  val = readl(my_status);
33 CPU A:  ...
34 CPU A:  writel(newval, ring_ptr);
35 CPU A:  (void)readl(safe_register); /* maybe a config register? */
36 CPU A:  spin_unlock_irqrestore(&dev_lock, flags)
37         ...
38 CPU B:  spin_lock_irqsave(&dev_lock, flags)
39 CPU B:  val = readl(my_status);
40 CPU B:  ...
41 CPU B:  writel(newval2, ring_ptr);
42 CPU B:  (void)readl(safe_register); /* maybe a config register? */
43 CPU B:  spin_unlock_irqrestore(&dev_lock, flags)
44
45 Here, the reads from safe_register will cause the I/O chipset to flush any
46 pending writes before actually posting the read to the chipset, preventing
47 possible data corruption.