Merge branch 'overlayfs-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mszer...
[sfrench/cifs-2.6.git] / Documentation / io_ordering.txt
1 ==============================================
2 Ordering I/O writes to memory-mapped addresses
3 ==============================================
4
5 On some platforms, so-called memory-mapped I/O is weakly ordered.  On such
6 platforms, driver writers are responsible for ensuring that I/O writes to
7 memory-mapped addresses on their device arrive in the order intended.  This is
8 typically done by reading a 'safe' device or bridge register, causing the I/O
9 chipset to flush pending writes to the device before any reads are posted.  A
10 driver would usually use this technique immediately prior to the exit of a
11 critical section of code protected by spinlocks.  This would ensure that
12 subsequent writes to I/O space arrived only after all prior writes (much like a
13 memory barrier op, mb(), only with respect to I/O).
14
15 A more concrete example from a hypothetical device driver::
16
17                 ...
18         CPU A:  spin_lock_irqsave(&dev_lock, flags)
19         CPU A:  val = readl(my_status);
20         CPU A:  ...
21         CPU A:  writel(newval, ring_ptr);
22         CPU A:  spin_unlock_irqrestore(&dev_lock, flags)
23                 ...
24         CPU B:  spin_lock_irqsave(&dev_lock, flags)
25         CPU B:  val = readl(my_status);
26         CPU B:  ...
27         CPU B:  writel(newval2, ring_ptr);
28         CPU B:  spin_unlock_irqrestore(&dev_lock, flags)
29                 ...
30
31 In the case above, the device may receive newval2 before it receives newval,
32 which could cause problems.  Fixing it is easy enough though::
33
34                 ...
35         CPU A:  spin_lock_irqsave(&dev_lock, flags)
36         CPU A:  val = readl(my_status);
37         CPU A:  ...
38         CPU A:  writel(newval, ring_ptr);
39         CPU A:  (void)readl(safe_register); /* maybe a config register? */
40         CPU A:  spin_unlock_irqrestore(&dev_lock, flags)
41                 ...
42         CPU B:  spin_lock_irqsave(&dev_lock, flags)
43         CPU B:  val = readl(my_status);
44         CPU B:  ...
45         CPU B:  writel(newval2, ring_ptr);
46         CPU B:  (void)readl(safe_register); /* maybe a config register? */
47         CPU B:  spin_unlock_irqrestore(&dev_lock, flags)
48
49 Here, the reads from safe_register will cause the I/O chipset to flush any
50 pending writes before actually posting the read to the chipset, preventing
51 possible data corruption.