Re: [PATCH] hw/elf_ops: clear uninitialized segment space - Philippe Mathieu-Daudé

From: "Philippe Mathieu-Daudé" <philmd@redhat.com>
To: Laurent Vivier <laurent@vivier.eu>, qemu-devel@nongnu.org
Cc: Peter Maydell <peter.maydell@linaro.org>,
	Stefano Garzarella <sgarzare@redhat.com>
Subject: Re: [PATCH] hw/elf_ops: clear uninitialized segment space
Date: Thu, 15 Apr 2021 11:57:44 +0200	[thread overview]
Message-ID: <934a3f68-8800-4774-0326-6881edff94e4@redhat.com> (raw)
In-Reply-To: <97b4d188-9fb6-3ba3-183a-82f926ab016c@amsat.org>

On 4/15/21 10:44 AM, Philippe Mathieu-Daudé wrote:
> On 4/14/21 2:41 PM, Laurent Vivier wrote:
>> Le 14/04/2021 à 14:16, Philippe Mathieu-Daudé a écrit :
>>> On 4/14/21 12:58 PM, Laurent Vivier wrote:
>>>> When the mem_size of the segment is bigger than the file_size,
>>>> and if this space doesn't overlap another segment, it needs
>>>> to be cleared.
>>>>
>>>> This bug is very similar to the one we had for linux-user,
>>>> 22d113b52f41 ("linux-user: Fix loading of BSS segments"),
>>>> where .bss section is encoded as an extension of the the data
>>>> one by setting the segment p_memsz > p_filesz.
>>>>
>>>> Signed-off-by: Laurent Vivier <laurent@vivier.eu>
>>>> ---
>>>>  include/hw/elf_ops.h | 17 +++++++++++++++++
>>>>  1 file changed, 17 insertions(+)
>>>>
>>>> diff --git a/include/hw/elf_ops.h b/include/hw/elf_ops.h
>>>> index 6ee458e7bc3c..e3dcee3ee349 100644
>>>> --- a/include/hw/elf_ops.h
>>>> +++ b/include/hw/elf_ops.h
>>>> @@ -562,6 +562,23 @@ static int glue(load_elf, SZ)(const char *name, int fd,
>>>>                      if (res != MEMTX_OK) {
>>>>                          goto fail;
>>>>                      }
>>>> +                    /*
>>>> +                     * We need to zero'ify the space that is not copied
>>>> +                     * from file
>>>> +                     */
>>>> +                    if (file_size < mem_size) {
>>>> +                        static uint8_t zero[4096];
>>>
>>> Given it is unlikely, maybe better use:
>>>
>>>               g_autofree uint8_t *zero = g_new0(uint8_t, 4096);
>>
>> I don't know what is the best solution but this seems to introduce a lot of complexity only to have
>> a page of 0s.
> 
> Less complex alternative is to use dma_memory_set():
> 
>    dma_memory_set(as, file_size, 0, mem_size - file_size);
> 
> Actually we should extract address_space_set() from it, keeping
> the dma_barrier() call in dma_memory_set(), and use address_space_set()
> here.
> 
> What do you think?

I'll post a patch.