[Qemu-devel] [PATCH for-2.5] util/mmap-alloc: fix hugetlb support on ppc64

["Michael S. Tsirkin" <mst@redhat.com>]

Since commit 8561c9244ddf1122d "exec: allocate PROT_NONE pages on top of
RAM", it is no longer possible to back guest RAM with hugepages on ppc64
hosts:

mmap(NULL, 285212672, PROT_NONE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) =
0x3fff57000000
mmap(0x3fff57000000, 268435456, PROT_READ|PROT_WRITE,
MAP_PRIVATE|MAP_FIXED, 19, 0) = -1 EBUSY (Device or resource busy)

This is because on ppc64, Linux fixes a page size for a virtual address
at mmap size, so we can't switch a range of memory from anonymous
small pages to hugetlbs with MAP_FIXED.

See commit d0f13e3c20b6fb73ccb467bdca97fa7cf5a574cd
("[POWERPC] Introduce address space "slices"") in Linux
history for the details.

Detect this and create the PROT_NONE mapping using the same fd.

Natually, this makes the guard page bigger with hugetlbfs.

Based on patch by Greg Kurz.

Cc: Rik van Riel <riel@redhat.com>
CC: Greg Kurz <gkurz@linux.vnet.ibm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---

To me, this seems way safer.
Greg, could you send a Tested-by tag pls?

 include/qemu/mmap-alloc.h |  2 ++
 util/mmap-alloc.c         | 39 +++++++++++++++++++++++++++++++++++++++
 util/oslib-posix.c        | 24 +-----------------------
 3 files changed, 42 insertions(+), 23 deletions(-)

diff --git a/include/qemu/mmap-alloc.h b/include/qemu/mmap-alloc.h
index 56388e6..0899b2f 100644
--- a/include/qemu/mmap-alloc.h
+++ b/include/qemu/mmap-alloc.h
@@ -3,6 +3,8 @@
 
 #include "qemu-common.h"
 
+size_t qemu_fd_getpagesize(int fd);
+
 void *qemu_ram_mmap(int fd, size_t size, size_t align, bool shared);
 
 void qemu_ram_munmap(void *ptr, size_t size);
diff --git a/util/mmap-alloc.c b/util/mmap-alloc.c
index c37acbe..0b139cd 100644
--- a/util/mmap-alloc.c
+++ b/util/mmap-alloc.c
@@ -14,6 +14,32 @@
 #include <sys/mman.h>
 #include <assert.h>
 
+#define HUGETLBFS_MAGIC       0x958458f6
+
+#ifdef CONFIG_LINUX
+#include <sys/vfs.h>
+#endif
+
+size_t qemu_fd_getpagesize(int fd)
+{
+#ifdef CONFIG_LINUX
+    struct statfs fs;
+    int ret;
+
+    if (fd != -1) {
+        do {
+            ret = fstatfs(fd, &fs);
+        } while (ret != 0 && errno == EINTR);
+
+        if (ret == 0 && fs.f_type == HUGETLBFS_MAGIC) {
+            return fs.f_bsize;
+        }
+    }
+#endif
+
+    return getpagesize();
+}
+
 void *qemu_ram_mmap(int fd, size_t size, size_t align, bool shared)
 {
     /*
@@ -21,7 +47,20 @@ void *qemu_ram_mmap(int fd, size_t size, size_t align, bool shared)
      * space, even if size is already aligned.
      */
     size_t total = size + align;
+#if defined(__powerpc64__) && defined(__linux__)
+    /* On ppc64 mappings in the same segment (aka slice) must share the same
+     * page size. Since we will be re-allocating part of this segment
+     * from the supplied fd, we should make sure to use the same page size,
+     * unless we are using the system page size, in which case anonymous memory
+     * is OK. Use align as a hint for the page size.
+     * In this case, set MAP_NORESERVE to avoid allocating backing store memory.
+     */
+    int anonfd = (fd == -1 || fd_getpagesize(fd) == getpagesize()) ? -1 : fd;
+    int flags = anonfd == -1 ? MAP_ANONYMOUS : MAP_NORESERVE;
+    void *ptr = mmap(0, total, PROT_NONE, flags | MAP_PRIVATE, anonfd, 0);
+#else
     void *ptr = mmap(0, total, PROT_NONE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+#endif
     size_t offset = QEMU_ALIGN_UP((uintptr_t)ptr, align) - (uintptr_t)ptr;
     void *ptr1;
 
diff --git a/util/oslib-posix.c b/util/oslib-posix.c
index 914cef5..d25f671 100644
--- a/util/oslib-posix.c
+++ b/util/oslib-posix.c
@@ -46,7 +46,6 @@ extern int daemon(int, int);
 #else
 #  define QEMU_VMALLOC_ALIGN getpagesize()
 #endif
-#define HUGETLBFS_MAGIC       0x958458f6
 
 #include <termios.h>
 #include <unistd.h>
@@ -65,7 +64,6 @@ extern int daemon(int, int);
 
 #ifdef CONFIG_LINUX
 #include <sys/syscall.h>
-#include <sys/vfs.h>
 #endif
 
 #ifdef __FreeBSD__
@@ -340,26 +338,6 @@ static void sigbus_handler(int signal)
     siglongjmp(sigjump, 1);
 }
 
-static size_t fd_getpagesize(int fd)
-{
-#ifdef CONFIG_LINUX
-    struct statfs fs;
-    int ret;
-
-    if (fd != -1) {
-        do {
-            ret = fstatfs(fd, &fs);
-        } while (ret != 0 && errno == EINTR);
-
-        if (ret == 0 && fs.f_type == HUGETLBFS_MAGIC) {
-            return fs.f_bsize;
-        }
-    }
-#endif
-
-    return getpagesize();
-}
-
 void os_mem_prealloc(int fd, char *area, size_t memory)
 {
     int ret;
@@ -387,7 +365,7 @@ void os_mem_prealloc(int fd, char *area, size_t memory)
         exit(1);
     } else {
         int i;
-        size_t hpagesize = fd_getpagesize(fd);
+        size_t hpagesize = qemu_fd_getpagesize(fd);
         size_t numpages = DIV_ROUND_UP(memory, hpagesize);
 
         /* MAP_POPULATE silently ignores failures */
-- 
MST