Handle huge pages in ibv_fork_init() and madvise tracking

When fork support is enabled in libibverbs, madvise() is called for
every memory page that is registered as a memory region.  Memory
ranges that are passed to madvise() must be page aligned and the size
must be a multiple of the page size.

libibverbs uses sysconf(_SC_PAGESIZE) to find out the system page size
and rounds all ranges passed to reg_mr() according to this page size.
When memory from libhugetlbfs is passed to reg_mr(), this does not
work as the page size for this memory range might be different
(e.g. 16MB).  So libibverbs would have to use the huge page size to
calculate a page aligned range for madvise.

As huge pages are provided to the application "under the hood" when
preloading libhugetlbfs, the application does not have any knowledge
about when it registers a huge page or a usual page.

To work around this issue, detect the use of huge pages in libibverbs
and align memory ranges passed to madvise according to the huge page
size.  Determining the page size of a given memory range by watching
madvise() fail has proven to be unreliable.  So we introduce the
RDMAV_HUGEPAGES_SAFE environment variable to let the user decide if
the page size should be checked on every reg_mr() call or not. This
requires the user to be aware if huge pages are used by the running
application or not.

I did not add an aditional API call to enable this, as applications
can use setenv() + ibv_fork_init() to enable checking for huge pages
in the code.

Signed-off-by: Alexander Schmidt <alexs@linux.vnet.ibm.com>
[ Updated ibv_fork_init() manpage for RDMAV_HUGEPAGES_SAFE.  - Roland ]

Signed-off-by: Roland Dreier <roland@purestorage.com>

diff --git a/man/ibv_fork_init.3 b/man/ibv_fork_init.3
index 6f2a2878390de654943b2cedbceaa1dcb0d434e9..acffe3cb7c67cc3419e699d0c26460b14b285387 100644 (file)
--- a/man/ibv_fork_init.3
+++ b/man/ibv_fork_init.3
@@ -41,12 +41,22 @@ or
 has the same effect as calling
 .B ibv_fork_init()\fR.
 .PP
+Setting the environment variable
+.BR RDMAV_HUGEPAGES_SAFE
+tells the library to check the underlying page size used by the kernel
+for memory regions.  This is required if an application uses huge
+pages either directly or indirectly via a library such as libhugetlbfs.
+.PP
 Calling
 .B ibv_fork_init()
 will reduce performance due to an extra system call for every memory
 registration, and the additional memory allocated to track memory
 regions.  The precise performance impact depends on the workload and
 usually will not be significant.
+.PP
+Setting
+.BR RDMAV_HUGEPAGES_SAFE
+adds further overhead to all memory registrations.
 .SH "SEE ALSO"
 .BR fork (2),
 .BR wait (2),

diff --git a/src/memory.c b/src/memory.c
index ce58ae8c6f435434d19e8f9d6c3ab4061f948bd1..faa43f37c79948793aea5804b0e6f48faedf73c4 100644 (file)
--- a/src/memory.c
+++ b/src/memory.c
@@ -40,6 +40,10 @@
 #include <unistd.h>
 #include <stdlib.h>
 #include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <dirent.h>
+#include <limits.h>
 
 #include "ibverbs.h"
 
@@ -68,12 +72,72 @@ struct ibv_mem_node {
 static struct ibv_mem_node *mm_root;
 static pthread_mutex_t mm_mutex = PTHREAD_MUTEX_INITIALIZER;
 static int page_size;
+static int huge_page_enabled;
 static int too_late;
 
+static unsigned long smaps_page_size(FILE *file)
+{
+       int n;
+       unsigned long size = page_size;
+       char buf[1024];
+
+       while (fgets(buf, sizeof(buf), file) != NULL) {
+               if (!strstr(buf, "KernelPageSize:"))
+                       continue;
+
+               n = sscanf(buf, "%*s %lu", &size);
+               if (n < 1)
+                       continue;
+
+               /* page size is printed in Kb */
+               size = size * 1024;
+
+               break;
+       }
+
+       return size;
+}
+
+static unsigned long get_page_size(void *base)
+{
+       unsigned long ret = page_size;
+       pid_t pid;
+       FILE *file;
+       char buf[1024];
+
+       pid = getpid();
+       snprintf(buf, sizeof(buf), "/proc/%d/smaps", pid);
+
+       file = fopen(buf, "r");
+       if (!file)
+               goto out;
+
+       while (fgets(buf, sizeof(buf), file) != NULL) {
+               int n;
+               uintptr_t range_start, range_end;
+
+               n = sscanf(buf, "%lx-%lx", &range_start, &range_end);
+
+               if (n < 2)
+                       continue;
+
+               if ((uintptr_t) base >= range_start && (uintptr_t) base < range_end) {
+                       ret = smaps_page_size(file);
+                       break;
+               }
+       }
+
+       fclose(file);
+
+out:
+       return ret;
+}
+
 int ibv_fork_init(void)
 {
-       void *tmp;
+       void *tmp, *tmp_aligned;
        int ret;
+       unsigned long size;
 
        if (mm_root)
                return 0;
@@ -88,8 +152,21 @@ int ibv_fork_init(void)
        if (posix_memalign(&tmp, page_size, page_size))
                return ENOMEM;
 
-       ret = madvise(tmp, page_size, MADV_DONTFORK) ||
-             madvise(tmp, page_size, MADV_DOFORK);
+       if (getenv("RDMAV_HUGEPAGES_SAFE"))
+               huge_page_enabled = 1;
+       else
+               huge_page_enabled = 0;
+
+       if (huge_page_enabled) {
+               size = get_page_size(tmp);
+               tmp_aligned = (void *) ((uintptr_t) tmp & ~(size - 1));
+       } else {
+               size = page_size;
+               tmp_aligned = tmp;
+       }
+
+       ret = madvise(tmp_aligned, size, MADV_DONTFORK) ||
+             madvise(tmp_aligned, size, MADV_DOFORK);
 
        free(tmp);
 
@@ -529,13 +606,19 @@ static int ibv_madvise_range(void *base, size_t size, int advice)
        int inc;
        int rolling_back = 0;
        int ret = 0;
+       unsigned long range_page_size;
 
        if (!size)
                return 0;
 
-       start = (uintptr_t) base & ~(page_size - 1);
-       end   = ((uintptr_t) (base + size + page_size - 1) &
-                ~(page_size - 1)) - 1;
+       if (huge_page_enabled)
+               range_page_size = get_page_size(base);
+       else
+               range_page_size = page_size;
+
+       start = (uintptr_t) base & ~(range_page_size - 1);
+       end   = ((uintptr_t) (base + size + range_page_size - 1) &
+                ~(range_page_size - 1)) - 1;
 
        pthread_mutex_lock(&mm_mutex);
 again: