[Valgrind-developers] valgrind: r13351: Don't assume that page size is 4K when loading PIEs. Fixes #263034.

[<sv...@va...>]

sewardj	2013-03-29 09:57:24 +0000 (Fri, 29 Mar 2013)

  New Revision: 13351

  Log:
    Don't assume that page size is 4K when loading PIEs.  Fixes #263034.
    (Dodji Seketeli, do...@re...)

  Modified files:
    trunk/coregrind/m_ume/elf.c


  Modified: trunk/coregrind/m_ume/elf.c (+12 -9)
===================================================================
--- trunk/coregrind/m_ume/elf.c	2013-03-29 09:40:48 +00:00 (rev 13350)
+++ trunk/coregrind/m_ume/elf.c	2013-03-29 09:57:24 +00:00 (rev 13351)
@@ -334,18 +334,21 @@
          become legit, which is really bad) and causes problems for
          exp-ptrcheck, which assumes all numbers below 1MB are
          nonpointers.  So, hackily, move it above 1MB. */
-      /* Later .. is appears ppc32-linux tries to put [vdso] at 1MB,
+      /* Later .. it appears ppc32-linux tries to put [vdso] at 1MB,
          which totally screws things up, because nothing else can go
-         there.  So bump the hacky load addess along by 0x8000, to
-         0x108000. */
-      /* Later .. on mips64 we can't use 0x108000, because mapelf will fail. */
-#if defined(VGP_mips64_linux)
+         there.  The size of [vdso] is around 2 or 3 pages, so bump
+         the hacky load addess along by 8 * VKI_PAGE_SIZE to be safe. */
+      /* Later .. on mips64 we can't use 0x108000, because mapelf will
+         fail. */
+#     if defined(VGP_mips64_linux)
       if (ebase < 0x100000)
          ebase = 0x100000;
-#else
-      if (ebase < 0x108000)
-         ebase = 0x108000;
-#endif
+#     else
+      vg_assert(VKI_PAGE_SIZE >= 4096); /* stay sane */
+      ESZ(Addr) hacky_load_address = 0x100000 + 8 * VKI_PAGE_SIZE;
+      if (ebase < hacky_load_address)
+         ebase = hacky_load_address;
+#     endif
    }
 
    info->phnum = e->e.e_phnum;