[Qemu-devel] [PATCH] linux-user/mmap.c: Support shared memory mapping in mmap_frag()

[Qemu-devel] [PATCH] linux-user/mmap.c: Support shared memory mapping in mmap_frag()

[chengang]

From: Chen Gang <chengang@emindsoft.com.cn>

It is a temporary fix for i386 target system running Windows.

Also remove useless variable 'p'.

Signed-off-by: Chen Gang <gang.chen.5i5j@gmail.com>
---
 linux-user/mmap.c |   22 +++++++++++++++++++---
 1 files changed, 19 insertions(+), 3 deletions(-)

diff --git a/linux-user/mmap.c b/linux-user/mmap.c
index 445e8c6..07758d4 100644
--- a/linux-user/mmap.c
+++ b/linux-user/mmap.c
@@ -156,12 +156,28 @@ static int mmap_frag(abi_ulong real_start,
             prot1 |= page_get_flags(addr);
     }
 
+    /*
+     * It is a temporary fix. Normally, target system will let shared memory
+     * aligned with 64KB, and allocate them near with each other, no any other
+     * kinds of mapping regions nearby. e.g. Windows on i386.
+     */
+    if ((start == real_start) && (flags & MAP_SHARED)) {
+        if (prot1) {
+            munmap(host_start, qemu_host_page_size);
+        }
+        if (mmap(host_start, qemu_host_page_size, prot, flags, fd, offset)
+            == MAP_FAILED) {
+            return -1;
+        }
+        return 0;
+    }
+
     if (prot1 == 0) {
         /* no page was there, so we allocate one */
-        void *p = mmap(host_start, qemu_host_page_size, prot,
-                       flags | MAP_ANONYMOUS, -1, 0);
-        if (p == MAP_FAILED)
+        if (mmap(host_start, qemu_host_page_size, prot, flags | MAP_ANONYMOUS,
+                 -1, 0) == MAP_FAILED) {
             return -1;
+        }
         prot1 = prot;
     }
     prot1 &= PAGE_BITS;
-- 
1.7.3.4

Re: [Qemu-devel] [PATCH] linux-user/mmap.c: Support shared memory mapping in mmap_frag()

[Chen Gang]

Hello all:

This patch is only for a discussion, I guess this patch is valuable for
i386 wine running Windows.

Theoretically, this patch is incorrect, we have to implement softmmu to
support different host and target pages (e.g. host 8KB, target 4KB):

 - If host 8KB is mapped as PRIVATE | FIXED, and runs a while (trigger
   copy on write).

 - Then map its low 4KB as SHARED | FIXED.

 - It has to fail.

So this patch is only a temporary fix:

 - It looks Windows maps shared memory with 64KB alignments (at least,
   at present, it is true).

 - OS will manage the process address area: let most shared memory area
   nearby.

 - After this patch, we can run Windows Notepad.exe and ACDSee5.0 setup
   program successfully. Next, I will run wine with the true windows
   system dlls for a test (e.g. windows own kernel32.dll).

I guess, we can use a switch macro or a new input parameter to enable or
disable the related code (if they are really valuable enough for some
using cases).

Thanks.

On 2015年12月30日 09:11, chengang@emindsoft.com.cn wrote:
> From: Chen Gang <chengang@emindsoft.com.cn>
> 
> It is a temporary fix for i386 target system running Windows.
> 
> Also remove useless variable 'p'.
> 
> Signed-off-by: Chen Gang <gang.chen.5i5j@gmail.com>
> ---
>  linux-user/mmap.c |   22 +++++++++++++++++++---
>  1 files changed, 19 insertions(+), 3 deletions(-)
> 
> diff --git a/linux-user/mmap.c b/linux-user/mmap.c
> index 445e8c6..07758d4 100644
> --- a/linux-user/mmap.c
> +++ b/linux-user/mmap.c
> @@ -156,12 +156,28 @@ static int mmap_frag(abi_ulong real_start,
>              prot1 |= page_get_flags(addr);
>      }
>  
> +    /*
> +     * It is a temporary fix. Normally, target system will let shared memory
> +     * aligned with 64KB, and allocate them near with each other, no any other
> +     * kinds of mapping regions nearby. e.g. Windows on i386.
> +     */
> +    if ((start == real_start) && (flags & MAP_SHARED)) {
> +        if (prot1) {
> +            munmap(host_start, qemu_host_page_size);
> +        }
> +        if (mmap(host_start, qemu_host_page_size, prot, flags, fd, offset)
> +            == MAP_FAILED) {
> +            return -1;
> +        }
> +        return 0;
> +    }
> +
>      if (prot1 == 0) {
>          /* no page was there, so we allocate one */
> -        void *p = mmap(host_start, qemu_host_page_size, prot,
> -                       flags | MAP_ANONYMOUS, -1, 0);
> -        if (p == MAP_FAILED)
> +        if (mmap(host_start, qemu_host_page_size, prot, flags | MAP_ANONYMOUS,
> +                 -1, 0) == MAP_FAILED) {
>              return -1;
> +        }
>          prot1 = prot;
>      }
>      prot1 &= PAGE_BITS;
> 

-- 
Chen Gang (陈刚)

Open, share, and attitude like air, water, and life which God blessed

Re: [Qemu-devel] [PATCH] linux-user/mmap.c: Support shared memory mapping in mmap_frag()

[Chen Gang]

Oh, sorry, please skip this patch (or discussion), I have send a new fix
patch for the related issue, so let the old ACDSee.exe and the latest
WinRAR.exe setup programs run successfully under real WinXP system dlls.

The related root cause is not related with shared memory, so at least now,
we can skip this patch (or discussion).

Thanks.

On 2015年12月30日 09:35, Chen Gang wrote:
> Hello all:
> 
> This patch is only for a discussion, I guess this patch is valuable for
> i386 wine running Windows.
> 
> Theoretically, this patch is incorrect, we have to implement softmmu to
> support different host and target pages (e.g. host 8KB, target 4KB):
> 
>  - If host 8KB is mapped as PRIVATE | FIXED, and runs a while (trigger
>    copy on write).
> 
>  - Then map its low 4KB as SHARED | FIXED.
> 
>  - It has to fail.
> 
> So this patch is only a temporary fix:
> 
>  - It looks Windows maps shared memory with 64KB alignments (at least,
>    at present, it is true).
> 
>  - OS will manage the process address area: let most shared memory area
>    nearby.
> 
>  - After this patch, we can run Windows Notepad.exe and ACDSee5.0 setup
>    program successfully. Next, I will run wine with the true windows
>    system dlls for a test (e.g. windows own kernel32.dll).
> 
> I guess, we can use a switch macro or a new input parameter to enable or
> disable the related code (if they are really valuable enough for some
> using cases).
> 
> Thanks.
> 
> On 2015年12月30日 09:11, chengang@emindsoft.com.cn wrote:
>> From: Chen Gang <chengang@emindsoft.com.cn>
>>
>> It is a temporary fix for i386 target system running Windows.
>>
>> Also remove useless variable 'p'.
>>
>> Signed-off-by: Chen Gang <gang.chen.5i5j@gmail.com>
>> ---
>>  linux-user/mmap.c |   22 +++++++++++++++++++---
>>  1 files changed, 19 insertions(+), 3 deletions(-)
>>
>> diff --git a/linux-user/mmap.c b/linux-user/mmap.c
>> index 445e8c6..07758d4 100644
>> --- a/linux-user/mmap.c
>> +++ b/linux-user/mmap.c
>> @@ -156,12 +156,28 @@ static int mmap_frag(abi_ulong real_start,
>>              prot1 |= page_get_flags(addr);
>>      }
>>  
>> +    /*
>> +     * It is a temporary fix. Normally, target system will let shared memory
>> +     * aligned with 64KB, and allocate them near with each other, no any other
>> +     * kinds of mapping regions nearby. e.g. Windows on i386.
>> +     */
>> +    if ((start == real_start) && (flags & MAP_SHARED)) {
>> +        if (prot1) {
>> +            munmap(host_start, qemu_host_page_size);
>> +        }
>> +        if (mmap(host_start, qemu_host_page_size, prot, flags, fd, offset)
>> +            == MAP_FAILED) {
>> +            return -1;
>> +        }
>> +        return 0;
>> +    }
>> +
>>      if (prot1 == 0) {
>>          /* no page was there, so we allocate one */
>> -        void *p = mmap(host_start, qemu_host_page_size, prot,
>> -                       flags | MAP_ANONYMOUS, -1, 0);
>> -        if (p == MAP_FAILED)
>> +        if (mmap(host_start, qemu_host_page_size, prot, flags | MAP_ANONYMOUS,
>> +                 -1, 0) == MAP_FAILED) {
>>              return -1;
>> +        }
>>          prot1 = prot;
>>      }
>>      prot1 &= PAGE_BITS;
>>
> 

-- 
Chen Gang (陈刚)

Open, share, and attitude like air, water, and life which God blessed