Re: Lockless file reading

[Jamie Lokier <jamie@shareable.org>]

Richard B. Johnson wrote:
> Let's see if it is possible for the middle byte of
> a 3-byte sequence to not be written when both
> other bytes are written:

> Even in machines that do load/store operations where individual
> components of those stores happen in groups, access to/from
> a buffer of such data is controlled (by hardware) so a write
> will complete before a read occurs.

I don't understand what you mean by this.

Do you mean that the writes are forced to appear on a different CPU in
the same order that they were written?  That isn't true on x86, for
two reasons: 1. writes aren't always in processor order (see
CONFIG_X86_OOSTORE and CONFIG_X86_PPRO_FENCE); 2. reads on the other
processor are out of order anyway.

> With hardware that can perform byte-access (ix86), the only
> byte-access that is going to happen is at the end(s) of buffers.

Not true.  Take a look at __copy_user() in arch/i386/lib/usercopy.c.
The first few bytes are copied using "rep;movsb", which is not
guaranteed to use a word write for the aligned pair, nor is it
guaranteed a particular timing (there could be an interrupt between
each byte).

Other architectures are similar.

-- Jamie