[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Libguestfs] [PATCH libnbd] copy: Set POSIX_FADV_SEQUENTIAL and POSIX_FADV_DONTNEED.



On 2/19/21 1:41 PM, Richard W.M. Jones wrote:

>>> +#if defined (HAVE_POSIX_FADVISE) && defined (POSIX_FADV_DONTNEED)
>>> +    /* On Linux this will evict the pages we just read from the page cache. */
>>> +    posix_fadvise (fd, offset, r, POSIX_FADV_DONTNEED);
>>
>> I don't think this is a good idea, since this affects the current page
>> cache, for
>> the entire host.
>>
>> So if the host is having an image in cache for good reason, running nbdcopy
>> will drop the cache since nbdcopy does not need it, but maybe the host will
>> need that cache after running nbdcopy.
>>
>> The right way to avoid polluting the page cache is to bypass the cache using
>> O_DIRECT, so nbdcopy is not using or affecting the page cache.
>>
>> This can be useful if the user can enable this with a flag.
> 
> The trouble with O_DIRECT is it's a pain to use correctly, and I guess
> doesn't use readahead or the existing cache (kind of the opposite
> problem).
> 
> However I take your point that we probably ought not to remove files
> from the cache that are already there.  I somehow thought that
> POSIX_FADV_DONTNEED only affected the current 'fd', but looking at the
> Linux impl I see it affects the whole system.
> 
> This argues for having a flag to enable this.

If I understand correctly, unconditionally attempting FADV_SEQUENTIAL
should always should be fine both for reads (from existing file) and
writes (when copying to a new file), so its only FADV_DONTNEED for
clearing out the cache that needs a flag?

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]