summaryrefslogtreecommitdiffstats
path: root/src/vppinfra
diff options
context:
space:
mode:
Diffstat (limited to 'src/vppinfra')
-rw-r--r--src/vppinfra/dlmalloc.h23
1 files changed, 22 insertions, 1 deletions
diff --git a/src/vppinfra/dlmalloc.h b/src/vppinfra/dlmalloc.h
index e710b0c56ca..b7a8eea80a0 100644
--- a/src/vppinfra/dlmalloc.h
+++ b/src/vppinfra/dlmalloc.h
@@ -699,7 +699,28 @@ MAX_RELEASE_CHECK_RATE default: 4095 unless not HAVE_MMAP
#endif /* DEFAULT_TRIM_THRESHOLD */
#ifndef DEFAULT_MMAP_THRESHOLD
#if HAVE_MMAP
-#define DEFAULT_MMAP_THRESHOLD ((size_t)256U * (size_t)1024U)
+/*
+ * The default value in the dlmalloc was set as follows:
+ *
+ * #define DEFAULT_MMAP_THRESHOLD ((size_t)256U * (size_t)1024U)
+ *
+ * Above this threshold the sys_alloc() calls mmap_alloc() to directly mmap the memory.
+ * However, the interaction of this path with the rest of the vpp infra results
+ * in vpp infra considering directly mmap-allocated pieces to not
+ * be part of the heap, with predictable consequence.
+ *
+ * A simple unit-test to show that behavior is to make a small private
+ * heap and repeatedly perform vec_add1() within that heap.
+ *
+ * The issue is because there is no tracking which mmap-allocated chunk
+ * belongs to which heap.
+ *
+ * The temporary approach is to dial up the threshold so that the problematic
+ * code path never gets called. The full fix needs to introduce
+ * introduce the vector of mmap-allocated chunks to mspace, and in general
+ * do some more thorough testing.
+ */
+#define DEFAULT_MMAP_THRESHOLD ((size_t)~0ULL)
#else /* HAVE_MMAP */
#define DEFAULT_MMAP_THRESHOLD MAX_SIZE_T
#endif /* HAVE_MMAP */