lkml.org 
[lkml]   [2018]   [May]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: [External] Re: [PATCH 2/3] include/linux/gfp.h: use unsigned int in gfp_zone
Date

> On Mon, May 07, 2018 at 11:44:10AM -0700, Matthew Wilcox wrote:
> > On Mon, May 07, 2018 at 05:16:50PM +0000, Huaisheng HS1 Ye wrote:
> > > I hope it couldn't cause problem, but based on my analyzation it has the potential
> to go wrong if users still use the flags as usual, which are __GFP_DMA, __GFP_DMA32
> and __GFP_HIGHMEM.
> > > Let me take an example with my testing platform, these logics are much abstract,
> an example will be helpful.
> > >
> > > There is a two sockets X86_64 server, No HIGHMEM and it has 16 + 16GB memories.
> > > Its zone types shall be like this below,
> > >
> > > ZONE_DMA 0 0b0000
> > > ZONE_DMA32 1 0b0001
> > > ZONE_NORMAL 2 0b0010
> > > (OPT_ZONE_HIGHMEM) 2 0b0010
> > > ZONE_MOVABLE 3 0b0011
> > > ZONE_DEVICE 4 0b0100 (virtual zone)
> > > __MAX_NR_ZONES 5
> > >
> > > __GFP_DMA = ZONE_DMA ^ ZONE_NORMAL= 0b0010
> > > __GFP_DMA32 = ZONE_DMA32 ^ ZONE_NORMAL= 0b0011
> > > __GFP_HIGHMEM = OPT_ZONE_HIGHMEM ^ ZONE_NORMAL = 0b0000
> > > __GFP_MOVABLE = ZONE_MOVABLE ^ ZONE_NORMAL | ___GFP_MOVABLE = 0b1001
> > >
> > > Eg.
> > > If a driver uses flags like this below,
> > > Step 1:
> > > gfp_mask | __GFP_DMA32;
> > > (0b 0000 | 0b 0011 = 0b 0011)
> > > gfp_mask's low four bits shall equal to 0011, assuming no __GFP_MOVABLE
> > >
> > > Step 2:
> > > gfp_mask & ~__GFP_DMA;
> > > (0b 0011 & ~0b0010 = 0b0001)
> > > gfp_mask's low four bits shall equal to 0001 now, then when it enter gfp_zone(),
> > >
> > > return ((__force int)flags & ___GFP_ZONE_MASK) ^ ZONE_NORMAL;
> > > (0b0001 ^ 0b0010 = 0b0011)
> > > You know 0011 means that ZONE_MOVABLE will be returned.
> > > In this case, error can be found, because gfp_mask needs to get ZONE_DMA32 originally.
> > > But with existing GFP_ZONE_TABLE/BAD, it is correct. Because the bits are way of
> 0x1, 0x2, 0x4, 0x8
> >
> > Yes, I understand your point here. My point was that this was already a bug;
> > the caller shouldn't simply be clearing __GFP_DMA; they really mean to clear
> > all of the GFP_ZONE bits so that they allocate from ZONE_NORMAL. And for
> > that, they should be using ~GFP_ZONEMASK
> >
> > Unless they already know, of course. For example, this one in
> > arch/x86/mm/pgtable.c is fine:
> >
> > if (strcmp(arg, "nohigh") == 0)
> > __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
> >
> > because it knows that __userpte_alloc_gfp can only have __GFP_HIGHMEM set.
> >
> > But something like btrfs should almost certainly be using ~GFP_ZONEMASK.
>
> Agreed, the direct use of __GFP_DMA32 was added in 3ba7ab220e8918176c6f
> to substitute GFP_NOFS, so the allocation flags are less restrictive but
> still acceptable for allocation from slab.
>
> The requirement from btrfs is to avoid highmem, the 'must be acceptable
> for slab' requirement is more MM internal and should have been hidden
> under some opaque flag mask. There was no strong need for that at the
> time.

Hi Matthew,

Should we add an error detection in gfp_zone? How about this?

@@ -377,6 +377,8 @@ static inline enum zone_type gfp_zone(gfp_t flags)
z = OPT_ZONE_HIGHMEM +
!!((__force unsigned int)flags & ___GFP_MOVABLE);
}
+
+ VM_BUG_ON(z > ZONE_MOVABLE);
return z;
}


Sincerely,
Huaisheng Ye






\
 
 \ /
  Last update: 2018-05-09 16:58    [W:0.073 / U:1.456 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site