Numa-aware, growable physical memory allocator (pmalloc)

Change-Id: Ic4c46bc733afae8bf0d8146623ed15633928de30
Signed-off-by: Damjan Marion <damarion@cisco.com>

5 files changed, 871 insertions, 0 deletions

diff --git a/src/vppinfra/CMakeLists.txt b/src/vppinfra/CMakeLists.txt
index 26368493382..6e1a4fd861a 100644
--- a/src/vppinfra/CMakeLists.txt
+++ b/src/vppinfra/CMakeLists.txt
@@ -56,6 +56,7 @@ set(VPPINFRA_SRCS
   macros.c
   maplog.c
   mhash.c
+  pmalloc.c
   pool.c
   ptclosure.c
   random.c
@@ -135,6 +136,7 @@ set(VPPINFRA_HEADERS
   mheap.h
   os.h
   pipeline.h
+  pmalloc.h
   pool.h
   pmc.h
   ptclosure.h
@@ -227,6 +229,7 @@ if(VPP_BUILD_VPPINFRA_TESTS)
     longjmp
     macros
     maplog
+    pmalloc
     pool_iterate
     ptclosure
     random
diff --git a/src/vppinfra/linux/syscall.h b/src/vppinfra/linux/syscall.h
index 4511b85ac1d..1ae029d5802 100644
--- a/src/vppinfra/linux/syscall.h
+++ b/src/vppinfra/linux/syscall.h
@@ -19,6 +19,12 @@
 #include <unistd.h>
 #include <sys/syscall.h>
 
+static inline int
+getcpu (unsigned *cpu, unsigned *node, void *tcache)
+{
+  return syscall (__NR_getcpu, cpu, node, tcache);
+}
+
 static inline long
 set_mempolicy (int mode, const unsigned long *nodemask, unsigned long maxnode)
 {
diff --git a/src/vppinfra/pmalloc.c b/src/vppinfra/pmalloc.c
new file mode 100644
index 00000000000..d354cce8a7d
--- /dev/null
+++ b/src/vppinfra/pmalloc.c
@@ -0,0 +1,562 @@
+/*
+ * Copyright (c) 2018 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#define _GNU_SOURCE
+#include <stdlib.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <linux/mempolicy.h>
+#include <linux/memfd.h>
+
+#include <vppinfra/format.h>
+#include <vppinfra/linux/syscall.h>
+#include <vppinfra/linux/sysfs.h>
+#include <vppinfra/mem.h>
+#include <vppinfra/hash.h>
+#include <vppinfra/pmalloc.h>
+
+#if __SIZEOF_POINTER__ >= 8
+#define DEFAULT_RESERVED_MB 16384
+#else
+#define DEFAULT_RESERVED_MB 256
+#endif
+
+static inline clib_pmalloc_chunk_t *
+get_chunk (clib_pmalloc_page_t * pp, u32 index)
+{
+  return pool_elt_at_index (pp->chunks, index);
+}
+
+static inline int
+pmalloc_validate_numa_node (u32 * numa_node)
+{
+  if (*numa_node == CLIB_PMALLOC_NUMA_LOCAL)
+    {
+      u32 cpu;
+      if (getcpu (&cpu, numa_node, 0) != 0)
+	return 1;
+    }
+  return 0;
+}
+
+int
+clib_pmalloc_init (clib_pmalloc_main_t * pm, uword size)
+{
+  struct stat st;
+  uword off, pagesize;
+  int fd;
+
+  ASSERT (pm->error == 0);
+
+  pm->log2_page_sz = 21;
+  pm->error = clib_mem_create_hugetlb_fd ("detect_hugepage_size", &fd);
+
+  if (pm->error)
+    return -1;
+
+  if (fd != -1)
+    {
+      if (fstat (fd, &st) == -1)
+	pm->log2_page_sz = min_log2 (st.st_blksize);
+      close (fd);
+    }
+
+  pagesize = 1ULL << pm->log2_page_sz;
+
+  size = size ? size : ((u64) DEFAULT_RESERVED_MB) << 20;
+  size = round_pow2 (size, pagesize);
+
+  pm->max_pages = size >> pm->log2_page_sz;
+
+  /* reserve VA space for future growth */
+  pm->base = mmap (0, size + pagesize, PROT_NONE,
+		   MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+
+  if (pm->base == MAP_FAILED)
+    {
+      pm->error = clib_error_return_unix (0, "failed to reserve %u pages");
+      return -1;
+    }
+
+  off = round_pow2 (pointer_to_uword (pm->base), pagesize) -
+    pointer_to_uword (pm->base);
+
+  /* trim start and end of reservation to be page aligned */
+  if (off)
+    {
+      munmap (pm->base, off);
+      pm->base += off;
+    }
+
+  munmap (pm->base + (pm->max_pages * pagesize), pagesize - off);
+  return 0;
+}
+
+static inline void *
+alloc_chunk_from_page (clib_pmalloc_main_t * pm, clib_pmalloc_page_t * pp,
+		       u32 n_blocks, u32 block_align, u32 numa_node)
+{
+  clib_pmalloc_chunk_t *c;
+  void *va;
+  u32 off;
+  u32 alloc_chunk_index;
+
+  if (pp->chunks == 0)
+    {
+      pool_get (pp->chunks, c);
+      pp->n_free_chunks = 1;
+      pp->first_chunk_index = c - pp->chunks;
+      c->prev = c->next = ~0;
+      c->size = pp->n_free_blocks;
+    }
+
+  alloc_chunk_index = pp->first_chunk_index;
+
+next_chunk:
+  c = pool_elt_at_index (pp->chunks, alloc_chunk_index);
+  off = (block_align - (c->start & (block_align - 1))) & (block_align - 1);
+
+  if (c->used || n_blocks + off > c->size)
+    {
+      if (c->next == ~0)
+	return 0;
+      alloc_chunk_index = c->next;
+      goto next_chunk;
+    }
+
+  /* if alignment is needed create new empty chunk */
+  if (off)
+    {
+      u32 offset_chunk_index;
+      clib_pmalloc_chunk_t *co;
+      pool_get (pp->chunks, c);
+      pp->n_free_chunks++;
+      offset_chunk_index = alloc_chunk_index;
+      alloc_chunk_index = c - pp->chunks;
+
+      co = pool_elt_at_index (pp->chunks, offset_chunk_index);
+      c->size = co->size - off;
+      c->next = co->next;
+      c->start = co->start + off;
+      c->prev = offset_chunk_index;
+      co->size = off;
+      co->next = alloc_chunk_index;
+    }
+
+  c->used = 1;
+  if (c->size > n_blocks)
+    {
+      u32 tail_chunk_index;
+      clib_pmalloc_chunk_t *ct;
+      pool_get (pp->chunks, ct);
+      pp->n_free_chunks++;
+      tail_chunk_index = ct - pp->chunks;
+      c = pool_elt_at_index (pp->chunks, alloc_chunk_index);
+      ct->size = c->size - n_blocks;
+      ct->next = c->next;
+      ct->prev = alloc_chunk_index;
+      ct->start = c->start + n_blocks;
+
+      c->size = n_blocks;
+      c->next = tail_chunk_index;
+      if (ct->next != ~0)
+	pool_elt_at_index (pp->chunks, ct->next)->prev = tail_chunk_index;
+    }
+  else if (c->next != ~0)
+    pool_elt_at_index (pp->chunks, c->next)->prev = alloc_chunk_index;
+
+  c = get_chunk (pp, alloc_chunk_index);
+  va = pm->base + ((pp - pm->pages) << pm->log2_page_sz) +
+    (c->start << PMALLOC_LOG2_BLOCK_SZ);
+  hash_set (pm->chunk_index_by_va, pointer_to_uword (va), alloc_chunk_index);
+  pp->n_free_blocks -= n_blocks;
+  pp->n_free_chunks--;
+  return va;
+}
+
+static inline clib_pmalloc_page_t *
+pmalloc_map_pages (clib_pmalloc_main_t * pm, clib_pmalloc_arena_t * a,
+		   u32 numa_node, u32 n_pages)
+{
+  clib_pmalloc_page_t *pp = 0;
+  u64 seek, pa, sys_page_size;
+  int pagemap_fd, status, rv, i, mmap_flags;
+  void *va;
+  int old_mpol = -1;
+  long unsigned int mask[16] = { 0 };
+  long unsigned int old_mask[16] = { 0 };
+
+  clib_error_free (pm->error);
+
+  if (pm->max_pages <= vec_len (pm->pages))
+    {
+      pm->error = clib_error_return (0, "maximum number of pages reached");
+      return 0;
+    }
+
+  pm->error = clib_sysfs_prealloc_hugepages (numa_node, pm->log2_page_sz,
+					     n_pages);
+
+  if (pm->error)
+    return 0;
+
+  rv = get_mempolicy (&old_mpol, old_mask, sizeof (old_mask) * 8 + 1, 0, 0);
+  /* failure to get mempolicy means we can only proceed with numa 0 maps */
+  if (rv == -1 && numa_node != 0)
+    {
+      pm->error = clib_error_return_unix (0, "failed to get mempolicy");
+      return 0;
+    }
+
+  mask[0] = 1 << numa_node;
+  rv = set_mempolicy (MPOL_BIND, mask, sizeof (mask) * 8 + 1);
+  if (rv == -1 && numa_node != 0)
+    {
+      pm->error = clib_error_return_unix (0, "failed to set mempolicy for "
+					  "numa node %u", numa_node);
+      return 0;
+    }
+
+  mmap_flags = MAP_FIXED | MAP_HUGETLB | MAP_LOCKED | MAP_ANONYMOUS;
+  if (a->flags & CLIB_PMALLOC_ARENA_F_SHARED_MEM)
+    {
+      mmap_flags |= MAP_SHARED;
+      pm->error = clib_mem_create_hugetlb_fd ((char *) a->name, &a->fd);
+      if (a->fd == -1)
+	goto error;
+    }
+  else
+    {
+      mmap_flags |= MAP_PRIVATE;
+      a->fd = -1;
+    }
+
+  va = pm->base + (vec_len (pm->pages) << pm->log2_page_sz);
+  if (mmap (va, n_pages << pm->log2_page_sz, PROT_READ | PROT_WRITE,
+	    mmap_flags, a->fd, 0) == MAP_FAILED)
+    {
+      pm->error = clib_error_return_unix (0, "failed to mmap %u pages at %p "
+					  "fd %d numa %d flags 0x%x", n_pages,
+					  va, a->fd, numa_node, mmap_flags);
+      goto error;
+    }
+
+  rv = set_mempolicy (old_mpol, old_mask, sizeof (old_mask) * 8 + 1);
+  if (rv == -1 && numa_node != 0)
+    {
+      pm->error = clib_error_return_unix (0, "failed to restore mempolicy");
+      goto error;
+    }
+
+  /* we tolerate move_pages failure only if request os for numa node 0
+     to support non-numa kernels */
+  rv = move_pages (0, 1, &va, 0, &status, 0);
+  if ((rv == 0 && status != numa_node) || (rv != 0 && numa_node != 0))
+    {
+      pm->error = rv == -1 ?
+	clib_error_return_unix (0, "page allocated on wrong node, numa node "
+				"%u status %d", numa_node, status) :
+	clib_error_return (0, "page allocated on wrong node, numa node "
+			   "%u status %d", numa_node, status);
+
+      /* unmap & reesrve */
+      munmap (va, n_pages << pm->log2_page_sz);
+      mmap (va, n_pages << pm->log2_page_sz, PROT_NONE,
+	    MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+      goto error;
+    }
+
+  memset (va, 0, n_pages << pm->log2_page_sz);
+  sys_page_size = sysconf (_SC_PAGESIZE);
+  pagemap_fd = open ((char *) "/proc/self/pagemap", O_RDONLY);
+
+  for (i = 0; i < n_pages; i++)
+    {
+      uword page_va = pointer_to_uword ((u8 *) va + (i << pm->log2_page_sz));
+      vec_add2 (pm->pages, pp, 1);
+      pp->n_free_blocks = 1 << (pm->log2_page_sz - PMALLOC_LOG2_BLOCK_SZ);
+      pp->index = pp - pm->pages;
+      pp->arena_index = a->index;
+
+      vec_add1 (a->page_indices, pp->index);
+      a->n_pages++;
+
+      seek = (page_va / sys_page_size) * sizeof (pa);
+      if (pagemap_fd != -1 &&
+	  lseek (pagemap_fd, seek, SEEK_SET) == seek &&
+	  read (pagemap_fd, &pa, sizeof (pa)) == (sizeof (pa)) &&
+	  pa & (1ULL << 63) /* page present bit */ )
+	{
+	  pp->pa = (pa & pow2_mask (55)) * sys_page_size;
+	}
+      vec_add1_aligned (pm->va_pa_diffs, pp->pa ? page_va - pp->pa : 0,
+			CLIB_CACHE_LINE_BYTES);
+    }
+
+  if (pagemap_fd != -1)
+    close (pagemap_fd);
+
+  /* return pointer to 1st page */
+  return pp - (n_pages - 1);
+
+error:
+  if (a->fd != -1)
+    close (a->fd);
+  return 0;
+}
+
+void *
+clib_pmalloc_create_shared_arena (clib_pmalloc_main_t * pm, char *name,
+				  uword size, u32 numa_node)
+{
+  clib_pmalloc_arena_t *a;
+  clib_pmalloc_page_t *pp;
+  u32 n_pages = round_pow2 (size, 1 << pm->log2_page_sz) >> pm->log2_page_sz;
+
+  if (n_pages + vec_len (pm->pages) > pm->max_pages)
+    return 0;
+
+  if (pmalloc_validate_numa_node (&numa_node))
+    return 0;
+
+  pool_get (pm->arenas, a);
+  a->index = a - pm->arenas;
+  a->name = format (0, "%s%c", name, 0);
+  a->numa_node = numa_node;
+  a->flags = CLIB_PMALLOC_ARENA_F_SHARED_MEM;
+  a->log2_page_sz = pm->log2_page_sz;
+
+  if ((pp = pmalloc_map_pages (pm, a, numa_node, n_pages)) == 0)
+    {
+      vec_free (a->name);
+      memset (a, 0, sizeof (*a));
+      pool_put (pm->arenas, a);
+      return 0;
+    }
+
+  return pm->base + (pp->index << pm->log2_page_sz);
+}
+
+static inline void *
+clib_pmalloc_alloc_inline (clib_pmalloc_main_t * pm, clib_pmalloc_arena_t * a,
+			   uword size, uword align, u32 numa_node)
+{
+  clib_pmalloc_page_t *pp;
+  u32 n_blocks, block_align, *page_index;
+
+  ASSERT (is_pow2 (align));
+
+  if (pmalloc_validate_numa_node (&numa_node))
+    return 0;
+
+  if (a == 0)
+    {
+      vec_validate_init_empty (pm->default_arena_for_numa_node,
+			       numa_node, ~0);
+      if (pm->default_arena_for_numa_node[numa_node] == ~0)
+	{
+	  pool_get (pm->arenas, a);
+	  pm->default_arena_for_numa_node[numa_node] = a - pm->arenas;
+	  a->name = format (0, "default-numa-%u%c", numa_node, 0);
+	  a->numa_node = numa_node;
+	}
+      else
+	a = pool_elt_at_index (pm->arenas,
+			       pm->default_arena_for_numa_node[numa_node]);
+    }
+
+  n_blocks = round_pow2 (size, PMALLOC_BLOCK_SZ) / PMALLOC_BLOCK_SZ;
+  block_align = align >> PMALLOC_LOG2_BLOCK_SZ;
+
+  vec_foreach (page_index, a->page_indices)
+  {
+    pp = vec_elt_at_index (pm->pages, *page_index);
+    void *rv = alloc_chunk_from_page (pm, pp, n_blocks, block_align,
+				      numa_node);
+
+    if (rv)
+      return rv;
+  }
+
+  if ((a->flags & CLIB_PMALLOC_ARENA_F_SHARED_MEM) == 0 &&
+      (pp = pmalloc_map_pages (pm, a, numa_node, 1)))
+    return alloc_chunk_from_page (pm, pp, n_blocks, block_align, numa_node);
+
+  return 0;
+}
+
+void *
+clib_pmalloc_alloc_aligned_on_numa (clib_pmalloc_main_t * pm, uword size,
+				    uword align, u32 numa_node)
+{
+  return clib_pmalloc_alloc_inline (pm, 0, size, align, numa_node);
+}
+
+void *
+clib_pmalloc_alloc_aligned (clib_pmalloc_main_t * pm, uword size, uword align)
+{
+  return clib_pmalloc_alloc_inline (pm, 0, size, align,
+				    CLIB_PMALLOC_NUMA_LOCAL);
+}
+
+void *
+clib_pmalloc_alloc_from_arena (clib_pmalloc_main_t * pm, void *arena_va,
+			       uword size, uword align)
+{
+  clib_pmalloc_arena_t *a = clib_pmalloc_get_arena (pm, arena_va);
+  return clib_pmalloc_alloc_inline (pm, a, size, align, 0);
+}
+
+void
+clib_pmalloc_free (clib_pmalloc_main_t * pm, void *va)
+{
+  clib_pmalloc_page_t *pp;
+  clib_pmalloc_chunk_t *c;
+  uword *p;
+  u32 chunk_index, page_index;
+
+  p = hash_get (pm->chunk_index_by_va, pointer_to_uword (va));
+
+  if (p == 0)
+    os_panic ();
+
+  chunk_index = p[0];
+  page_index = clib_pmalloc_get_page_index (pm, va);
+  hash_unset (pm->chunk_index_by_va, pointer_to_uword (va));
+
+  pp = vec_elt_at_index (pm->pages, page_index);
+  c = pool_elt_at_index (pp->chunks, chunk_index);
+  c->used = 0;
+  pp->n_free_blocks += c->size;
+  pp->n_free_chunks++;
+
+  /* merge with next if free */
+  if (c->next != ~0 && get_chunk (pp, c->next)->used == 0)
+    {
+      clib_pmalloc_chunk_t *next = get_chunk (pp, c->next);
+      c->size += next->size;
+      c->next = next->next;
+      if (next->next != ~0)
+	get_chunk (pp, next->next)->prev = chunk_index;
+      memset (next, 0, sizeof (*next));
+      pool_put (pp->chunks, next);
+      pp->n_free_chunks--;
+    }
+
+  /* merge with prev if free */
+  if (c->prev != ~0 && get_chunk (pp, c->prev)->used == 0)
+    {
+      clib_pmalloc_chunk_t *prev = get_chunk (pp, c->prev);
+      prev->size += c->size;
+      prev->next = c->next;
+      if (c->next != ~0)
+	get_chunk (pp, c->next)->prev = c->prev;
+      memset (c, 0, sizeof (*c));
+      pool_put (pp->chunks, c);
+      pp->n_free_chunks--;
+    }
+}
+
+static u8 *
+format_pmalloc_page (u8 * s, va_list * va)
+{
+  clib_pmalloc_page_t *pp = va_arg (*va, clib_pmalloc_page_t *);
+  int verbose = va_arg (*va, int);
+  u32 indent = format_get_indent (s);
+
+  s = format (s, "page %u: phys-addr %p ", pp->index, pp->pa);
+
+  if (pp->chunks == 0)
+    return s;
+
+  s = format (s, "free %u chunks %u free-chunks %d ",
+	      (pp->n_free_blocks) << PMALLOC_LOG2_BLOCK_SZ,
+	      pool_elts (pp->chunks), pp->n_free_chunks);
+
+  if (verbose >= 2)
+    {
+      clib_pmalloc_chunk_t *c;
+      c = pool_elt_at_index (pp->chunks, pp->first_chunk_index);
+      s = format (s, "\n%U%12s%12s%8s%8s%8s%8s",
+		  format_white_space, indent + 2,
+		  "chunk offset", "size", "used", "index", "prev", "next");
+      while (1)
+	{
+	  s = format (s, "\n%U%12u%12u%8s%8d%8d%8d",
+		      format_white_space, indent + 2,
+		      c->start << PMALLOC_LOG2_BLOCK_SZ,
+		      c->size << PMALLOC_LOG2_BLOCK_SZ,
+		      c->used ? "yes" : "no",
+		      c - pp->chunks, c->prev, c->next);
+	  if (c->next == ~0)
+	    break;
+	  c = pool_elt_at_index (pp->chunks, c->next);
+	}
+    }
+  return s;
+}
+
+u8 *
+format_pmalloc (u8 * s, va_list * va)
+{
+  clib_pmalloc_main_t *pm = va_arg (*va, clib_pmalloc_main_t *);
+  int verbose = va_arg (*va, int);
+  u32 indent = format_get_indent (s);
+
+  clib_pmalloc_page_t *pp;
+  clib_pmalloc_arena_t *a;
+
+  s = format (s, "used-pages %u reserved-pages %u pagesize %uKB",
+	      vec_len (pm->pages), pm->max_pages,
+	      1 << (pm->log2_page_sz - 10));
+
+  if (verbose >= 2)
+    s = format (s, " va-start %p", pm->base);
+
+  if (pm->error)
+    s = format (s, "\n%Ulast-error: %U", format_white_space, indent + 2,
+		format_clib_error, pm->error);
+
+
+  /* *INDENT-OFF* */
+  pool_foreach (a, pm->arenas,
+    {
+      u32 *page_index;
+      s = format (s, "\n%Uarena '%s' pages %u numa-node %u",
+		  format_white_space, indent + 2,
+		  a->name, vec_len (a->page_indices), a->numa_node);
+      if (a->fd != -1)
+        s = format (s, " shared fd %d", a->fd);
+      if (verbose >= 1)
+	vec_foreach (page_index, a->page_indices)
+	  {
+	    pp = vec_elt_at_index (pm->pages, *page_index);
+	    s = format (s, "\n%U%U", format_white_space, indent + 4,
+			format_pmalloc_page, pp, verbose);
+	  }
+    });
+  /* *INDENT-ON* */
+
+  return s;
+}
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/vppinfra/pmalloc.h b/src/vppinfra/pmalloc.h
new file mode 100644
index 00000000000..4d9906ab6e7
--- /dev/null
+++ b/src/vppinfra/pmalloc.h
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2018 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef included_palloc_h
+#define included_palloc_h
+#include <vppinfra/format.h>
+#include <vppinfra/pool.h>
+
+#define PMALLOC_LOG2_BLOCK_SZ          CLIB_LOG2_CACHE_LINE_BYTES
+#define PMALLOC_BLOCK_SZ               (1 << 6)
+
+#define CLIB_PMALLOC_NUMA_LOCAL 0xffffffff
+
+typedef struct
+{
+  u32 start, prev, next;
+  u32 size:31;
+  u32 used:1;
+} clib_pmalloc_chunk_t;
+
+STATIC_ASSERT_SIZEOF (clib_pmalloc_chunk_t, 16);
+
+typedef struct
+{
+  u32 index;
+  u32 arena_index;
+  uword pa;
+  clib_pmalloc_chunk_t *chunks;
+  u32 first_chunk_index;
+  u32 n_free_chunks;
+  u32 n_free_blocks;
+} clib_pmalloc_page_t;
+
+typedef struct
+{
+  u32 index;
+  u32 flags;
+#define CLIB_PMALLOC_ARENA_F_SHARED_MEM (1 << 0)
+  int fd;
+  u32 numa_node;
+  u32 first_page_index;
+  u32 log2_page_sz;
+  u32 n_pages;
+  u8 *name;
+  u32 *page_indices;
+} clib_pmalloc_arena_t;
+
+typedef struct
+{
+  u8 *base;
+  uword log2_page_sz;
+  uword *va_pa_diffs;
+  u32 max_pages;
+  clib_pmalloc_page_t *pages;
+  uword *chunk_index_by_va;
+  clib_pmalloc_arena_t *arenas;
+  u32 *default_arena_for_numa_node;
+
+  clib_error_t *error;
+} clib_pmalloc_main_t;
+
+
+int clib_pmalloc_init (clib_pmalloc_main_t * pm, uword size);
+void *clib_pmalloc_alloc_aligned_on_numa (clib_pmalloc_main_t * pm,
+					  uword size, uword align,
+					  u32 numa_node);
+void *clib_pmalloc_alloc_aligned (clib_pmalloc_main_t * pm, uword size,
+				  uword align);
+void clib_pmalloc_free (clib_pmalloc_main_t * pm, void *va);
+
+void *clib_pmalloc_create_shared_arena (clib_pmalloc_main_t * pm, char *name,
+					uword size, u32 numa_node);
+
+void *clib_pmalloc_alloc_from_arena (clib_pmalloc_main_t * pm, void *arena_va,
+				     uword size, uword align);
+
+format_function_t format_pmalloc;
+
+always_inline clib_error_t *
+clib_pmalloc_last_error (clib_pmalloc_main_t * pm)
+{
+  return pm->error;
+}
+
+always_inline u32
+clib_pmalloc_get_page_index (clib_pmalloc_main_t * pm, void *va)
+{
+  uword index = (pointer_to_uword (va) - pointer_to_uword (pm->base)) >>
+    pm->log2_page_sz;
+
+  ASSERT (index < vec_len (pm->pages));
+
+  return index;
+}
+
+always_inline clib_pmalloc_arena_t *
+clib_pmalloc_get_arena (clib_pmalloc_main_t * pm, void *va)
+{
+  u32 index = clib_pmalloc_get_page_index (pm, va);
+  return pm->arenas + pm->pages[index].arena_index;
+}
+
+always_inline uword
+clib_pmalloc_get_pa (clib_pmalloc_main_t * pm, void *va)
+{
+  u32 index = clib_pmalloc_get_page_index (pm, va);
+  return pointer_to_uword (va) - pm->va_pa_diffs[index];
+}
+
+
+#endif /* included_palloc_h */
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/vppinfra/test_pmalloc.c b/src/vppinfra/test_pmalloc.c
new file mode 100644
index 00000000000..c9ae01fc936
--- /dev/null
+++ b/src/vppinfra/test_pmalloc.c
@@ -0,0 +1,169 @@
+/*
+ * Copyright (c) 2018 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <vppinfra/format.h>
+#include <vppinfra/pmalloc.h>
+#include <vppinfra/random.h>
+
+typedef struct
+{
+  uword baseva;
+  uword size;
+  uword *vas;
+  u32 nitems;
+  u32 item_size;
+  u32 align;
+  int max_numa;
+  u32 arena_pages;
+  u32 arena_numa;
+  u32 arena_items;
+  int verbose;
+  clib_pmalloc_main_t pmalloc_main;
+} test_main_t;
+
+test_main_t test_main;
+
+clib_error_t *
+test_palloc (test_main_t * tm)
+{
+  clib_pmalloc_main_t *pm = &tm->pmalloc_main;
+  void *arena;
+  int i;
+  uword *va;
+
+  if (clib_pmalloc_init (pm, 0) != 0)
+    return clib_error_return (0, "pmalloc init failure");
+
+  fformat (stdout, "Allocate %d items...\n", tm->nitems);
+
+  for (i = 0; i < tm->nitems; i++)
+    {
+      u32 size = tm->item_size ? tm->item_size : 64 + 64 * (i % 8);
+      u32 align = tm->align ? tm->align : 64 << (i % 5);
+      u32 numa = i % (tm->max_numa + 1);
+      va = clib_pmalloc_alloc_aligned_on_numa (pm, size, align, numa);
+
+      if (va == 0)
+	clib_error ("Failed to alloc %u byte chunk with align %u on numa %u,"
+		    "\nerror: %U", size, align, numa, format_clib_error,
+		    clib_pmalloc_last_error (pm));
+
+      if ((pointer_to_uword (va) & (align - 1)) != 0)
+	clib_error (0, "Alignment error: %p not aligned with %u", va, align);
+
+      vec_add1 (tm->vas, pointer_to_uword (va));
+    }
+  fformat (stdout, "%U\n", format_pmalloc, pm, tm->verbose);
+
+  /* alloc from arena */
+  if (tm->arena_items)
+    {
+      fformat (stdout, "Allocate %d items from arena ...\n", tm->arena_items);
+      arena = clib_pmalloc_create_shared_arena (pm, "test arena",
+						tm->arena_pages << 21,
+						tm->arena_numa);
+      if (arena == 0)
+	clib_error ("Failed to alloc shared arena: %U", format_clib_error,
+		    clib_pmalloc_last_error (pm));
+
+      for (i = 0; i < tm->arena_items; i++)
+	{
+	  u32 size = tm->item_size ? tm->item_size : 64 + 64 * (i % 8);
+	  u32 align = tm->align ? tm->align : 64 << (i % 5);
+	  va = clib_pmalloc_alloc_from_arena (pm, arena, size, align);
+	  vec_add1 (tm->vas, pointer_to_uword (va));
+	}
+      fformat (stdout, "\n%U\n", format_pmalloc, pm, tm->verbose);
+    }
+
+
+  fformat (stdout, "Freeing %d items ...\n", vec_len (tm->vas));
+  for (i = 0; i < vec_len (tm->vas); i++)
+    clib_pmalloc_free (pm, (void *) tm->vas[i]);
+
+  fformat (stdout, "\n%U\n", format_pmalloc, pm, tm->verbose);
+  return 0;
+}
+
+clib_error_t *
+test_palloc_main (unformat_input_t * i)
+{
+  test_main_t *tm = &test_main;
+  clib_error_t *error;
+
+  tm->nitems = 5;
+  tm->arena_pages = 2;
+  tm->arena_numa = CLIB_PMALLOC_NUMA_LOCAL;
+
+  while (unformat_check_input (i) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (i, "nitems %u", &tm->nitems))
+	;
+      else if (unformat (i, "max-numa %u", &tm->max_numa))
+	;
+      else if (unformat (i, "item-size %u", &tm->item_size))
+	;
+      else if (unformat (i, "align %u", &tm->align))
+	;
+      else if (unformat (i, "verbose %d", &tm->verbose))
+	;
+      else if (unformat (i, "arena-pages %u", &tm->arena_pages))
+	;
+      else if (unformat (i, "arena-numa %u", &tm->arena_numa))
+	;
+      else if (unformat (i, "arena-items %u", &tm->arena_items))
+	;
+      else if (unformat (i, "verbose"))
+	tm->verbose = 1;
+      else
+	return clib_error_return (0, "unknown input '%U'",
+				  format_unformat_error, i);
+    }
+
+  error = test_palloc (tm);
+
+  return error;
+}
+
+#ifdef CLIB_UNIX
+int
+main (int argc, char *argv[])
+{
+  unformat_input_t i;
+  int rv = 0;
+  clib_error_t *error;
+
+  clib_mem_init (0, 3ULL << 30);
+
+  unformat_init_command_line (&i, argv);
+  error = test_palloc_main (&i);
+  if (error)
+    {
+      clib_error_report (error);
+      rv = 1;
+    }
+  unformat_free (&i);
+
+  return rv;
+}
+#endif /* CLIB_UNIX */
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */