[Bug 217138] head (e.g.) -r314638 for arm64: sh vs. jemalloc asserts: include/jemalloc/internal/tsd.h:687: Failed assertion: "tsd_booted" once swapped in after being swapped out (comment 10)

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=217138

--- Comment #21 from Mark Millard <markmi at dsl-only.net> ---
I my test program I reported:

#define region_size (4u*1024u+1u)
                        // Bad dyn_region pattern, parent and child
                        // processes:
                        //     256u,  2u*1024u,  4u*1024u, 8u*1024u,
                        // 9u*1024u, 12u*1024u, 14u*1024u
                        // (but see the partial_test_check() call
                        //  notes above).

                        // Works:
                        // 14u*1024u+1u, 15u*1024u, 16u*1024u,
                        // 32u*1024u, 256u*1024u*1024u

It appears to turn out that the
14u*1024u vs. 14u*1024+1u boundary matches
up with jemalloc:

contrib/jemalloc/include/jemalloc/internal/size_classes.h :

. . . size == 1<<lg_grp + ndelta<<lg_delta
. . .\
 *     bin: 'yes' if a small bin size class, 'no' otherwise.
. . .
#if (LG_SIZEOF_PTR == 3 && LG_TINY_MIN == 3 && LG_QUANTUM == 3 && LG_PAGE ==
12)
#define SIZE_CLASSES \
  /* index, lg_grp, lg_delta, ndelta, psz, bin, lg_delta_lookup */ \
. . .
    SC( 38,     13,       11,      3,  no, yes, no) \
    SC( 39,     13,       11,      4, yes,  no, no) \
. . .
#define SMALL_MAXCLASS          ((((size_t)1) << 13) + (((size_t)3) << 11))

So SMALL_MAXCLASS==(14u*1024u).

The problem in the example program is tied to small bin size
handling in jemalloc.

-- 
You are receiving this mail because:
You are on the CC list for the bug.