Re: [PATCH 1/6] mm: introduce kv[mz]alloc helpers

[John Hubbard]

On 01/18/2017 12:21 AM, Michal Hocko wrote:
> On Tue 17-01-17 21:59:13, John Hubbard wrote:
> > On 01/16/2017 11:51 PM, Michal Hocko wrote:
> > > On Mon 16-01-17 13:57:43, John Hubbard wrote:
> > > > On 01/16/2017 01:48 PM, Michal Hocko wrote:
> > > > > On Mon 16-01-17 13:15:08, John Hubbard wrote:
> > > > > > On 01/16/2017 11:40 AM, Michal Hocko wrote:
> > > > > > > On Mon 16-01-17 11:09:37, John Hubbard wrote:
> > > > > > > > On 01/16/2017 12:47 AM, Michal Hocko wrote:
> > > > > > > > > On Sun 15-01-17 20:34:13, John Hubbard wrote:
> > > > > > > [...]
> > > > > > > > > > Is that "Reclaim modifiers" line still true, or is it a leftover from an
> > > > > > > > > > earlier approach? I am having trouble reconciling it with rest of the
> > > > > > > > > > patchset, because:
> > > > > > > > > >
> > > > > > > > > > a) the flags argument below is effectively passed on to either kmalloc_node
> > > > > > > > > > (possibly adding, but not removing flags), or to __vmalloc_node_flags.
> > > > > > > > >
> > > > > > > > > The above only says thos are _unsupported_ - in other words the behavior
> > > > > > > > > is not defined. Even if flags are passed down to kmalloc resp. vmalloc
> > > > > > > > > it doesn't mean they are used that way.  Remember that vmalloc uses
> > > > > > > > > some hardcoded GFP_KERNEL allocations.  So while I could be really
> > > > > > > > > strict about this and mask away these flags I doubt this is worth the
> > > > > > > > > additional code.
> > > > > > > >
> > > > > > > > I do wonder about passing those flags through to kmalloc. Maybe it is worth
> > > > > > > > stripping out __GFP_NORETRY and __GFP_NOFAIL, after all. It provides some
> > > > > > > > insulation from any future changes to the implementation of kmalloc, and it
> > > > > > > > also makes the documentation more believable.
> > > > > > >
> > > > > > > I am not really convinced that we should take an extra steps for these
> > > > > > > flags. There are no existing users for those flags and new users should
> > > > > > > follow the documentation.
> > > > > >
> > > > > > OK, let's just fortify the documentation ever so slightly, then, so that
> > > > > > users are more likely to do the right thing. How's this sound:
> > > > > >
> > > > > > * Reclaim modifiers - __GFP_NORETRY and __GFP_NOFAIL are not supported. (Even
> > > > > > * though the current implementation passes the flags on through to kmalloc and
> > > > > > * vmalloc, that is done for efficiency and to avoid unnecessary code. The caller
> > > > > > * should not pass in these flags.)
> > > > > > *
> > > > > > * __GFP_REPEAT is supported, but only for large (>64kB) allocations.
> > > > > >
> > > > > >
> > > > > > ? Or is that documentation overkill?
> > > > >
> > > > > Dunno, it sounds like an overkill to me. It is telling more than
> > > > > necessary. If we want to be so vocal about gfp flags then we would have
> > > > > to say much more I suspect. E.g. what about __GFP_HIGHMEM? This flag is
> > > > > supported for vmalloc while unsupported for kmalloc. I am pretty sure
> > > > > there would be other gfp flags to consider and then this would grow
> > > > > borringly large and uninteresting to the point when people simply stop
> > > > > reading it. Let's just be as simple as possible.
> > > >
> > > > Agreed, on the simplicity point: simple and clear is ideal. But here, it's
> > > > merely short, and not quite simple. :)  People will look at that short bit
> > > > of documentation, and then notice that the flags are, in fact, all passed
> > > > right on through down to both kmalloc_node and __vmalloc_node_flags.
> > > >
> > > > If you don't want too much documentation, then I'd be inclined to say
> > > > something higher-level, about the intent, rather than mentioning those two
> > > > flags directly. Because as it stands, the documentation contradicts what the
> > > > code does.
> > >
> > > Feel free to suggest a better wording. I am, of course, open to any
> > > changes.
> >
> > OK, here's the best I've got, I tried to keep it concise, but (as you
> > suspected) I'm not sure it's actually any better than the original:
> >
> >  * Reclaim modifiers - __GFP_NORETRY and __GFP_NOFAIL should not be passed in.
> >  * Passing in __GFP_REPEAT is supported, but note that it is ignored for small
> >  * (<=64KB) allocations, during the kmalloc attempt.
>
> > __GFP_REPEAT is fully
> >  * honored for  all allocation sizes during the second part: the vmalloc attempt.
>
> this is not true to be really precise because vmalloc doesn't respect
> the given gfp mask all the way down (look at the pte initialization).

I'm having some difficulty in locating that pte initialization part, am I on the 
wrong code path? Here's what I checked, before making the claim about __GFP_REPEAT 
being honored:

kvmalloc_node
  __vmalloc_node_flags
    __vmalloc_node
      __vmalloc_node_range
        __vmalloc_area_node
            alloc_pages_node
              __alloc_pages_node
                __alloc_pages
                  __alloc_pages_nodemask
                    __alloc_pages_slowpath


...and __alloc_pages_slowpath does the __GFP_REPEAT handling:

    /*
     * Do not retry costly high order allocations unless they are
     * __GFP_REPEAT
     */
    if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
        goto nopage;

thanks,
john h