diff mbox series

[v6,5/8] linux-user: Do not align brk with host page size

Message ID 20230801232745.4125-6-deller@gmx.de
State New
Headers show
Series linux-user: brk fixes | expand

Commit Message

Helge Deller Aug. 1, 2023, 11:27 p.m. UTC 
From: Akihiko Odaki <akihiko.odaki@daynix.com>

do_brk() minimizes calls into target_mmap() by aligning the address
with host page size, which is potentially larger than the target page
size. However, the current implementation of this optimization has two
bugs:

- The start of brk is rounded up with the host page size while brk
  advertises an address aligned with the target page size as the
  beginning of brk. This makes the beginning of brk unmapped.
- Content clearing after mapping is flawed. The size to clear is
  specified as HOST_PAGE_ALIGN(brk_page) - brk_page, but brk_page is
  aligned with the host page size so it is always zero.

This optimization actually has no practical benefit. It makes difference
when brk() is called multiple times with values in a range of the host
page size. However, sophisticated memory allocators try to avoid to
make such frequent brk() calls. For example, glibc 2.37 calls brk() to
shrink the heap only when there is a room more than 128 KiB. It is
rare to have a page size larger than 128 KiB if it happens.

Let's remove the optimization to fix the bugs and make the code simpler.

Fixes: 86f04735ac ("linux-user: Fix brk() to release pages")
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1616
Signed-off-by: Akihiko Odaki <akihiko.odaki@daynix.com>
Reviewed-by: Helge Deller <deller@gmx.de>
Signed-off-by: Helge Deller <deller@gmx.de>
---
 linux-user/elfload.c |  4 ++--
 linux-user/syscall.c | 54 ++++++++++----------------------------------
 2 files changed, 14 insertions(+), 44 deletions(-)

--
2.41.0
diff mbox series

Patch

diff --git a/linux-user/elfload.c b/linux-user/elfload.c
index 861ec07abc..2aee2298ec 100644
--- a/linux-user/elfload.c
+++ b/linux-user/elfload.c
@@ -3678,8 +3678,8 @@  int load_elf_binary(struct linux_binprm *bprm, struct image_info *info)
      * to mmap pages in this space.
      */
     if (info->reserve_brk) {
-        abi_ulong start_brk = HOST_PAGE_ALIGN(info->brk);
-        abi_ulong end_brk = HOST_PAGE_ALIGN(info->brk + info->reserve_brk);
+        abi_ulong start_brk = TARGET_PAGE_ALIGN(info->brk);
+        abi_ulong end_brk = TARGET_PAGE_ALIGN(info->brk + info->reserve_brk);
         target_munmap(start_brk, end_brk - start_brk);
     }

diff --git a/linux-user/syscall.c b/linux-user/syscall.c
index ebdc8c144c..475260b7ce 100644
--- a/linux-user/syscall.c
+++ b/linux-user/syscall.c
@@ -802,81 +802,51 @@  static inline int host_to_target_sock_type(int host_type)
 }

 static abi_ulong target_brk, initial_target_brk;
-static abi_ulong brk_page;

 void target_set_brk(abi_ulong new_brk)
 {
     target_brk = TARGET_PAGE_ALIGN(new_brk);
     initial_target_brk = target_brk;
-    brk_page = HOST_PAGE_ALIGN(target_brk);
 }

 /* do_brk() must return target values and target errnos. */
 abi_long do_brk(abi_ulong brk_val)
 {
     abi_long mapped_addr;
-    abi_ulong new_alloc_size;
-    abi_ulong new_brk, new_host_brk_page;
+    abi_ulong new_brk;
+    abi_ulong old_brk;

     /* brk pointers are always untagged */

-    /* return old brk value if brk_val unchanged */
-    if (brk_val == target_brk) {
-        return target_brk;
-    }
-
     /* do not allow to shrink below initial brk value */
     if (brk_val < initial_target_brk) {
         return target_brk;
     }

     new_brk = TARGET_PAGE_ALIGN(brk_val);
-    new_host_brk_page = HOST_PAGE_ALIGN(brk_val);
+    old_brk = TARGET_PAGE_ALIGN(target_brk);

-    /* brk_val and old target_brk might be on the same page */
-    if (new_brk == TARGET_PAGE_ALIGN(target_brk)) {
-        /* empty remaining bytes in (possibly larger) host page */
-        memset(g2h_untagged(new_brk), 0, new_host_brk_page - new_brk);
+    /* new and old target_brk might be on the same page */
+    if (new_brk == old_brk) {
         target_brk = brk_val;
         return target_brk;
     }

     /* Release heap if necesary */
-    if (new_brk < target_brk) {
-        /* empty remaining bytes in (possibly larger) host page */
-        memset(g2h_untagged(new_brk), 0, new_host_brk_page - new_brk);
-
-        /* free unused host pages and set new brk_page */
-        target_munmap(new_host_brk_page, brk_page - new_host_brk_page);
-        brk_page = new_host_brk_page;
+    if (new_brk < old_brk) {
+        target_munmap(new_brk, old_brk - new_brk);

         target_brk = brk_val;
         return target_brk;
     }

-    if (new_host_brk_page > brk_page) {
-        new_alloc_size = new_host_brk_page - brk_page;
-        mapped_addr = target_mmap(brk_page, new_alloc_size,
-                                  PROT_READ | PROT_WRITE,
-                                  MAP_FIXED_NOREPLACE | MAP_ANON | MAP_PRIVATE,
-                                  0, 0);
-    } else {
-        new_alloc_size = 0;
-        mapped_addr = brk_page;
-    }
-
-    if (mapped_addr == brk_page) {
-        /* Heap contents are initialized to zero, as for anonymous
-         * mapped pages.  Technically the new pages are already
-         * initialized to zero since they *are* anonymous mapped
-         * pages, however we have to take care with the contents that
-         * come from the remaining part of the previous page: it may
-         * contains garbage data due to a previous heap usage (grown
-         * then shrunken).  */
-        memset(g2h_untagged(brk_page), 0, HOST_PAGE_ALIGN(brk_page) - brk_page);
+    mapped_addr = target_mmap(old_brk, new_brk - old_brk,
+                              PROT_READ | PROT_WRITE,
+                              MAP_FIXED_NOREPLACE | MAP_ANON | MAP_PRIVATE,
+                              0, 0);

+    if (mapped_addr == old_brk) {
         target_brk = brk_val;
-        brk_page = new_host_brk_page;
         return target_brk;
     }