[Qemu-devel] [PATCH] mark host pages as reserved

[Qemu-devel] [PATCH] mark host pages as reserved

[Magnus Damm]

[-- Attachment #1: Type: text/plain, Size: 692 bytes --]

Hi all,

This patch teaches the user space emulator about host pages. It marks
present host page mappings with PAGE_RESERVED so mmap_find_vma()
properly can detect that pages at mmap_next_start should be skipped
over instead of being overwritten using mmap(). Without this patch I
experience crashes with the arm or sh4 user space emulator together
with sbox2. This combination sometimes result in that host libraries
are mapped into the qemu process space starting from 0x4xxxxxxx
instead of 0xb7xxxxxx. This together with the initial value of
mmap_next_start results in a segfault when the elf loader overwrites
the mapped host libraries with the target binary.

Comments anyone?

/ magnus

[-- Attachment #2: qemu-cvs-20071205b-reserve-host-pages.patch --]
[-- Type: application/octet-stream, Size: 1735 bytes --]

--- 0001/cpu-all.h
+++ work/cpu-all.h	2007-12-05 17:05:36.000000000 +0900
@@ -716,6 +716,7 @@ extern unsigned long qemu_host_page_mask
 /* original state of the write flag (used when tracking self-modifying
    code */
 #define PAGE_WRITE_ORG 0x0010
+#define PAGE_RESERVED  0x0020
 
 void page_dump(FILE *f);
 int page_get_flags(target_ulong address);
--- 0001/exec.c
+++ work/exec.c	2007-12-05 17:06:07.000000000 +0900
@@ -209,6 +209,27 @@ static void page_init(void)
     qemu_host_page_mask = ~(qemu_host_page_size - 1);
     l1_phys_map = qemu_vmalloc(L1_SIZE * sizeof(void *));
     memset(l1_phys_map, 0, L1_SIZE * sizeof(void *));
+
+#if !defined(_WIN32) && defined(CONFIG_USER_ONLY)
+    {
+        long long startaddr, endaddr;
+        FILE *f;
+        int n;
+
+        f = fopen("/proc/self/maps", "r");
+        if (f) {
+            do {
+                n = fscanf (f, "%llx-%llx %*[^\n]\n", &startaddr, &endaddr);
+                if (n == 2) {
+                    page_set_flags(TARGET_PAGE_ALIGN(startaddr),
+				   TARGET_PAGE_ALIGN(endaddr),
+				   PAGE_RESERVED); 
+                }
+            } while (!feof(f));
+            fclose(f);
+        }
+    }
+#endif
 }
 
 static inline PageDesc *page_find_alloc(unsigned int index)
--- 0001/linux-user/mmap.c
+++ work/linux-user/mmap.c	2007-12-05 17:05:36.000000000 +0900
@@ -162,7 +162,7 @@ static abi_ulong mmap_next_start = 0x400
    'start'. If 'start' == 0, then a default start address is used.
    Return -1 if error.
 */
-/* XXX: should mark pages used by the host as reserved to be sure not
+/* page_init() marks pages used by the host as reserved to be sure not
    to use them. */
 static abi_ulong mmap_find_vma(abi_ulong start, abi_ulong size)
 {

[Qemu-devel] Re: [PATCH] mark host pages as reserved

[Magnus Damm]

Hi everyone,

On Dec 5, 2007 5:45 PM, Magnus Damm <magnus.damm@gmail.com> wrote:
> Hi all,
>
> This patch teaches the user space emulator about host pages. It marks
> present host page mappings with PAGE_RESERVED so mmap_find_vma()
> properly can detect that pages at mmap_next_start should be skipped
> over instead of being overwritten using mmap(). Without this patch I
> experience crashes with the arm or sh4 user space emulator together
> with sbox2. This combination sometimes result in that host libraries
> are mapped into the qemu process space starting from 0x4xxxxxxx
> instead of 0xb7xxxxxx. This together with the initial value of
> mmap_next_start results in a segfault when the elf loader overwrites
> the mapped host libraries with the target binary.
>
> Comments anyone?

Any reason for not applying this patch? This fixes user space
emulation for all targets.

Thanks,

/ magnus

Re: [Qemu-devel] Re: [PATCH] mark host pages as reserved

[Fabrice Bellard]

Magnus Damm wrote:
> Hi everyone,
> 
> On Dec 5, 2007 5:45 PM, Magnus Damm <magnus.damm@gmail.com> wrote:
>> Hi all,
>>
>> This patch teaches the user space emulator about host pages. It marks
>> present host page mappings with PAGE_RESERVED so mmap_find_vma()
>> properly can detect that pages at mmap_next_start should be skipped
>> over instead of being overwritten using mmap(). Without this patch I
>> experience crashes with the arm or sh4 user space emulator together
>> with sbox2. This combination sometimes result in that host libraries
>> are mapped into the qemu process space starting from 0x4xxxxxxx
>> instead of 0xb7xxxxxx. This together with the initial value of
>> mmap_next_start results in a segfault when the elf loader overwrites
>> the mapped host libraries with the target binary.
>>
>> Comments anyone?
> 
> Any reason for not applying this patch? This fixes user space
> emulation for all targets.

It seems OK.

Fabrice.