On Tue, Jun 23, 2020 at 12:40:02PM -0700, Andrii Nakryiko wrote:
> On Tue, Jun 23, 2020 at 12:29 PM Jakub Bogusz <qboosh@xxxxxxxxxxxxx> wrote:
> >
> > On Mon, Jun 22, 2020 at 10:44:56PM -0700, Andrii Nakryiko wrote:
> > > On Sun, Jun 21, 2020 at 7:34 AM Jakub Bogusz <qboosh@xxxxxxxxxxxxx> wrote:
> > > >
> > > > Hello,
> > > >
> > > > I noticed that _bpftool crashes when building kernel tools (5.7.x) for
> > > > 32-bit targets because in libbpf hashmap implementation hash_bits()
> > > > function returning numbers exceeding hashmap buckets capacity.
> > > >
> > > > Attached patch fixes this problem.
> > > >
> > >
> > > Thanks! But this was already fixed by Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
> > > in 8ca8d4a84173 ("libbpf: Define __WORDSIZE if not available").
> >
> > No, it's not:
> > This change worked around __WORDSIZE not always being available.
> >
> > But the issue on (I)LP32 platforms is that 64-bit value is shifted by
> > (32-bits) instead of (64-bits).
> >
> > (__SIZEOF_LONG__ * 8) is 32 on such architectures (i686, arm).
> > I used __SIZEOF_LONG_LONG__ to get proper bit shift both on (I)LP32 and
> > LP64 architectures.
> >
>
> Ah, I see. I actually mentioned __SIZEOF_ constants on the original
> fix patch. But I think in this case it has to use __SIZEOF_SIZE_T,
> which on 32-bit should be 4, right?
After changing constant to 32-bit, yes (to be precise, it should use maximum
of __SIZEOF_SIZE_T__ and __SIZEOF_LONG__ if constant is specified with
UL suffix; there is no constant suffix available for size_t).
> > Should I provide an updated patch to apply on top of acme change?
>
> Yes, that would be good. But I think there is no need to penalize
> 32-bit arches with use of 64-bit long longs, and instead it's better
> to use #ifdef for 32-bit case vs 64-bit case. The multiplication
> constant will change, of course, should be 2654435769. I'd appreciate
> it if you can do the patch, thanks!
OK, so now the patch provides two variants:
- "long long" case for LP64 architectures
- "long" case for (I)LP32 architectures
(selected basing of __SIZEOF_ constants)
matter)
Regards,
--
Jakub Bogusz http://qboosh.pl/
Fix libbpf hashmap on (I)LP32 architectures
On ILP32, 64-bit result was shifted by value calculated for 32-bit long type
and returned value was much outside hashmap capacity.
As advised by Andrii Nakryiko, this patch uses different hashing variant for
architectures with size_t shorter than long long.
Signed-off-by: Jakub Bogusz <qboosh@xxxxxxxxxxxxx>
--- linux/tools/lib/bpf/hashmap.h.orig 2020-06-01 01:49:15.000000000 +0200
+++ linux/tools/lib/bpf/hashmap.h 2020-06-21 15:22:07.298466419 +0200
@@ -11,14 +11,18 @@
#include <stdbool.h>
#include <stddef.h>
#include <limits.h>
-#ifndef __WORDSIZE
-#define __WORDSIZE (__SIZEOF_LONG__ * 8)
-#endif
static inline size_t hash_bits(size_t h, int bits)
{
/* shuffle bits and return requested number of upper bits */
- return (h * 11400714819323198485llu) >> (__WORDSIZE - bits);
+#if (__SIZEOF_SIZE_T__ == __SIZEOF_LONG_LONG__)
+ /* LP64 case */
+ return (h * 11400714819323198485llu) >> (__SIZEOF_LONG_LONG__ * 8 - bits);
+#elif (__SIZEOF_SIZE_T__ <= __SIZEOF_LONG__)
+ return (h * 2654435769lu) >> (__SIZEOF_LONG__ * 8 - bits);
+#else
+# error "Unsupported size_t size"
+#endif
}
typedef size_t (*hashmap_hash_fn)(const void *key, void *ctx);
