ssvm: reuse clib mem infra for memfd segments

Change-Id: I67648dbed3c7ed291b3e1ce617d83a776d3623bb
Signed-off-by: Florin Coras <fcoras@cisco.com>

5 files changed, 119 insertions, 79 deletions

diff --git a/src/svm/ssvm.c b/src/svm/ssvm.c
index 9b935e4a3bd..36abc208636 100644
--- a/src/svm/ssvm.c
+++ b/src/svm/ssvm.c
@@ -205,62 +205,38 @@ ssvm_delete (ssvm_private_t * ssvm)
 int
 ssvm_master_init_memfd (ssvm_private_t * memfd, u32 master_index)
 {
-  int flags;
+  uword page_size, flags = MHEAP_FLAG_DISABLE_VM | MHEAP_FLAG_THREAD_SAFE;
   ssvm_shared_header_t *sh;
-  u64 ticks = clib_cpu_time_now ();
-  u64 randomize_baseva;
   void *oldheap;
+  clib_mem_vm_alloc_t alloc = { 0 };
+  clib_error_t *err;
 
   if (memfd->ssvm_size == 0)
     return SSVM_API_ERROR_NO_SIZE;
 
   ASSERT (vec_c_string_is_terminated (memfd->name));
-  memfd->name = format (0, "memfd svm region %d", master_index);
+  memfd->name = format (0, "memfd svm region %d%c", master_index, 0);
 
-  memfd->fd = memfd_create ((char *) memfd->name, MFD_ALLOW_SEALING);
-  if (memfd->fd < 0)
+  alloc.name = (char *) memfd->name;
+  alloc.size = memfd->ssvm_size;
+  alloc.flags = CLIB_MEM_VM_F_SHARED;
+  if ((err = clib_mem_vm_ext_alloc (&alloc)))
     {
-      clib_unix_warning ("create segment '%s'", memfd->name);
+      clib_error_report (err);
       return SSVM_API_ERROR_CREATE_FAILURE;
     }
 
-  if ((ftruncate (memfd->fd, memfd->ssvm_size)) == -1)
-    {
-      clib_unix_warning ("set memfd size");
-      return SSVM_API_ERROR_SET_SIZE;
-    }
-
-  if ((fcntl (memfd->fd, F_ADD_SEALS, F_SEAL_SHRINK)) == -1)
-    clib_unix_warning ("fcntl (F_ADD_SEALS, F_SEAL_SHRINK)");
-
-  flags = MAP_SHARED;
-  if (memfd->requested_va)
-    flags |= MAP_FIXED;
-
-  randomize_baseva = (ticks & 15) * MMAP_PAGESIZE;
-
-  if (memfd->requested_va)
-    memfd->requested_va += randomize_baseva;
-
-  sh = memfd->sh =
-    (ssvm_shared_header_t *) mmap ((void *) memfd->requested_va,
-				   memfd->ssvm_size, PROT_READ | PROT_WRITE,
-				   flags, memfd->fd, 0);
-
-  if (memfd->sh == MAP_FAILED)
-    {
-      clib_unix_warning ("mmap");
-      close (memfd->fd);
-      return SSVM_API_ERROR_MMAP;
-    }
-
+  memfd->fd = alloc.fd;
+  memfd->sh = alloc.addr;
   memfd->my_pid = getpid ();
+  memfd->i_am_master = 1;
+
+  page_size = 1 << alloc.log2_page_size;
+  sh = (ssvm_shared_header_t *) memfd->sh;
   sh->master_pid = memfd->my_pid;
   sh->ssvm_size = memfd->ssvm_size;
-  sh->heap = mheap_alloc_with_flags
-    (((u8 *) sh) + MMAP_PAGESIZE, memfd->ssvm_size - MMAP_PAGESIZE,
-     MHEAP_FLAG_DISABLE_VM | MHEAP_FLAG_THREAD_SAFE);
-
+  sh->heap = mheap_alloc_with_flags (((u8 *) sh) + page_size,
+				     memfd->ssvm_size - page_size, flags);
   sh->ssvm_va = pointer_to_uword (sh);
   sh->master_index = master_index;
 
@@ -268,8 +244,6 @@ ssvm_master_init_memfd (ssvm_private_t * memfd, u32 master_index)
   sh->name = format (0, "%s%c", memfd->name, 0);
   ssvm_pop_heap (oldheap);
 
-  memfd->i_am_master = 1;
-
   /* The application has to set set sh->ready... */
   return 0;
 }
@@ -283,39 +257,63 @@ ssvm_master_init_memfd (ssvm_private_t * memfd, u32 master_index)
 int
 ssvm_slave_init_memfd (ssvm_private_t * memfd)
 {
+  clib_mem_vm_map_t mapa = { 0 };
   ssvm_shared_header_t *sh;
+  uword page_size;
 
   memfd->i_am_master = 0;
 
-  /* Map the segment once, to look at the shared header */
-  sh = (void *) mmap (0, MMAP_PAGESIZE, PROT_READ | PROT_WRITE, MAP_SHARED,
-		      memfd->fd, 0);
-  if (sh == MAP_FAILED)
+  page_size = clib_mem_vm_get_page_size (memfd->fd);
+  if (!page_size)
+    {
+      clib_unix_warning ("page size unknown");
+      return SSVM_API_ERROR_MMAP;
+    }
+
+  /*
+   * Map the segment once, to look at the shared header
+   */
+  mapa.fd = memfd->fd;
+  mapa.size = page_size;
+
+  if (clib_mem_vm_ext_map (&mapa))
     {
       clib_unix_warning ("slave research mmap");
       close (memfd->fd);
       return SSVM_API_ERROR_MMAP;
     }
 
-  memfd->requested_va = (u64) sh->ssvm_va;
+  sh = mapa.addr;
+  memfd->requested_va = sh->ssvm_va;
   memfd->ssvm_size = sh->ssvm_size;
-  munmap (sh, MMAP_PAGESIZE);
-
-  sh = (void *) mmap ((void *) memfd->requested_va, memfd->ssvm_size,
-		      PROT_READ | PROT_WRITE, MAP_SHARED | MAP_FIXED,
-		      memfd->fd, 0);
-
-  if (sh == MAP_FAILED)
+  munmap (sh, page_size);
+
+  /*
+   * Remap the segment at the 'right' address
+   */
+  mapa.requested_va = memfd->requested_va;
+  mapa.size = memfd->ssvm_size;
+  if (clib_mem_vm_ext_map (&mapa))
     {
       clib_unix_warning ("slave final mmap");
       close (memfd->fd);
       return SSVM_API_ERROR_MMAP;
     }
+
+  sh = mapa.addr;
   sh->slave_pid = getpid ();
   memfd->sh = sh;
   return 0;
 }
 
+void
+ssvm_delete_memfd (ssvm_private_t * memfd)
+{
+  vec_free (memfd->name);
+  clib_mem_vm_free (memfd->sh, memfd->ssvm_size);
+  close (memfd->fd);
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/svm/ssvm.h b/src/svm/ssvm.h
index 0295e302e72..a181f2ef864 100644
--- a/src/svm/ssvm.h
+++ b/src/svm/ssvm.h
@@ -168,21 +168,7 @@ void ssvm_delete (ssvm_private_t * ssvm);
 
 int ssvm_master_init_memfd (ssvm_private_t * memfd, u32 master_index);
 int ssvm_slave_init_memfd (ssvm_private_t * memfd);
-void memfd_delete (ssvm_private_t * memfd);
-
-/* These do not belong here, but the original keeps running around... */
-/* $$$$ work w/ Damjan to fix properly */
-#ifndef F_LINUX_SPECIFIC_BASE
-#define F_LINUX_SPECIFIC_BASE 1024
-#endif
-#define MFD_ALLOW_SEALING       0x0002U
-#define F_ADD_SEALS (F_LINUX_SPECIFIC_BASE + 9)
-#define F_GET_SEALS (F_LINUX_SPECIFIC_BASE + 10)
-
-#define F_SEAL_SEAL     0x0001	/* prevent further seals from being set */
-#define F_SEAL_SHRINK   0x0002	/* prevent file from shrinking */
-#define F_SEAL_GROW     0x0004	/* prevent file from growing */
-#define F_SEAL_WRITE    0x0008	/* prevent writes */
+void ssvm_delete_memfd (ssvm_private_t * memfd);
 
 #endif /* __included_ssvm_h__ */
 
diff --git a/src/svm/svm_fifo_segment.h b/src/svm/svm_fifo_segment.h
index 7a30efabdc4..405df934578 100644
--- a/src/svm/svm_fifo_segment.h
+++ b/src/svm/svm_fifo_segment.h
@@ -15,8 +15,8 @@
 #ifndef __included_ssvm_fifo_segment_h__
 #define __included_ssvm_fifo_segment_h__
 
-#include <svm/svm_fifo.h>
 #include <svm/ssvm.h>
+#include <svm/svm_fifo.h>
 #include <vppinfra/lock.h>
 
 typedef enum
diff --git a/src/vppinfra/linux/mem.c b/src/vppinfra/linux/mem.c
index e0e22b0a775..3c449226273 100644
--- a/src/vppinfra/linux/mem.c
+++ b/src/vppinfra/linux/mem.c
@@ -26,6 +26,7 @@
 
 #include <vppinfra/clib.h>
 #include <vppinfra/mem.h>
+#include <vppinfra/time.h>
 #include <vppinfra/format.h>
 #include <vppinfra/clib_error.h>
 #include <vppinfra/linux/syscall.h>
@@ -45,13 +46,34 @@
 #define F_SEAL_WRITE    0x0008	/* prevent writes */
 #endif
 
-int
-clib_mem_vm_get_log2_page_size (int fd)
+uword
+clib_mem_vm_get_page_size (int fd)
 {
   struct stat st = { 0 };
   if (fstat (fd, &st) == -1)
     return 0;
-  return min_log2 (st.st_blksize);
+  return st.st_blksize;
+}
+
+int
+clib_mem_vm_get_log2_page_size (int fd)
+{
+  return min_log2 (clib_mem_vm_get_page_size (fd));
+}
+
+static void
+clib_mem_vm_randomize_va (uword * requested_va, u32 log2_page_size)
+{
+  u8 bit_mask = 15;
+
+  if (log2_page_size <= 12)
+    bit_mask = 15;
+  else if (log2_page_size > 12 && log2_page_size <= 16)
+    bit_mask = 3;
+  else
+    bit_mask = 0;
+
+  *requested_va += (clib_cpu_time_now () & bit_mask) * (1 << log2_page_size);
 }
 
 clib_error_t *
@@ -131,13 +153,19 @@ clib_mem_vm_ext_alloc (clib_mem_vm_alloc_t * a)
 	      goto error;
 	    }
 	}
-      log2_page_size = clib_mem_vm_get_log2_page_size (fd);
 
+      log2_page_size = clib_mem_vm_get_log2_page_size (fd);
       if (log2_page_size == 0)
 	{
 	  err = clib_error_return_unix (0, "cannot determine page size");
 	  goto error;
 	}
+
+      if (a->requested_va)
+	{
+	  clib_mem_vm_randomize_va (&a->requested_va, log2_page_size);
+	  mmap_flags |= MAP_FIXED;
+	}
     }
   else				/* not CLIB_MEM_VM_F_SHARED */
     {
@@ -155,7 +183,6 @@ clib_mem_vm_ext_alloc (clib_mem_vm_alloc_t * a)
 
   n_pages = ((a->size - 1) >> log2_page_size) + 1;
 
-
   if (a->flags & CLIB_MEM_VM_F_HUGETLB_PREALLOC)
     {
       err = clib_sysfs_prealloc_hugepages (a->numa_node,
@@ -186,7 +213,8 @@ clib_mem_vm_ext_alloc (clib_mem_vm_alloc_t * a)
 	}
     }
 
-  addr = mmap (0, a->size, (PROT_READ | PROT_WRITE), mmap_flags, fd, 0);
+  addr = mmap (uword_to_pointer (a->requested_va, void *), a->size,
+	       (PROT_READ | PROT_WRITE), mmap_flags, fd, 0);
   if (addr == MAP_FAILED)
     {
       err = clib_error_return_unix (0, "mmap");
@@ -255,7 +283,24 @@ done:
   return r;
 }
 
+clib_error_t *
+clib_mem_vm_ext_map (clib_mem_vm_map_t * a)
+{
+  int mmap_flags = MAP_SHARED;
+  void *addr;
 
+  if (a->requested_va)
+    mmap_flags |= MAP_FIXED;
+
+  addr = (void *) mmap (uword_to_pointer (a->requested_va, void *), a->size,
+			PROT_READ | PROT_WRITE, mmap_flags, a->fd, 0);
+
+  if (addr == MAP_FAILED)
+    return clib_error_return_unix (0, "mmap");
+
+  a->addr = addr;
+  return 0;
+}
 
 /*
  * fd.io coding-style-patch-verification: ON
diff --git a/src/vppinfra/mem.h b/src/vppinfra/mem.h
index 69ab880317d..b566e8d8895 100644
--- a/src/vppinfra/mem.h
+++ b/src/vppinfra/mem.h
@@ -330,7 +330,7 @@ typedef struct
 #define CLIB_MEM_VM_F_HUGETLB_PREALLOC (1 << 4)
   u32 flags; /**< vm allocation flags:
                 <br> CLIB_MEM_VM_F_SHARED: request shared memory, file
-		destiptor will be provided on successful allocation.
+		descriptor will be provided on successful allocation.
                 <br> CLIB_MEM_VM_F_HUGETLB: request hugepages.
 		<br> CLIB_MEM_VM_F_NUMA_PREFER: numa_node field contains valid
 		numa node preference.
@@ -342,15 +342,26 @@ typedef struct
   uword size; /**< Allocation size, set by caller. */
   int numa_node; /**< numa node preference. Valid if CLIB_MEM_VM_F_NUMA_PREFER set. */
   void *addr; /**< Pointer to allocated memory, set on successful allocation. */
-  int fd; /**< File desriptor, set on successful allocation if CLIB_MEM_VM_F_SHARED is set. */
+  int fd; /**< File descriptor, set on successful allocation if CLIB_MEM_VM_F_SHARED is set. */
   int log2_page_size;		/* Page size in log2 format, set on successful allocation. */
   int n_pages;			/* Number of pages. */
+  uword requested_va;		/**< Request fixed position mapping */
 } clib_mem_vm_alloc_t;
 
 clib_error_t *clib_mem_vm_ext_alloc (clib_mem_vm_alloc_t * a);
+u64 clib_mem_vm_get_page_size (int fd);
 int clib_mem_vm_get_log2_page_size (int fd);
 u64 *clib_mem_vm_get_paddr (void *mem, int log2_page_size, int n_pages);
 
+typedef struct
+{
+  uword size;		/**< Map size */
+  int fd;		/**< File descriptor to be mapped */
+  uword requested_va;	/**< Request fixed position mapping */
+  void *addr;		/**< Pointer to mapped memory, if successful */
+} clib_mem_vm_map_t;
+clib_error_t *clib_mem_vm_ext_map (clib_mem_vm_map_t * a);
+
 
 #include <vppinfra/error.h>	/* clib_panic */