Re: [PATCH] rtw89: Fix crash by loading compressed firmware file

From: Takashi Iwai
Date: Fri Nov 05 2021 - 11:07:12 EST


On Fri, 05 Nov 2021 15:28:39 +0100,
Pkshih wrote:
>
> On Fri, 2021-11-05 at 11:03 +0200, Kalle Valo wrote:
> > Takashi Iwai <tiwai@xxxxxxx> writes:
> >
> > > On Fri, 05 Nov 2021 09:25:13 +0100,
> > > Kalle Valo wrote:
> > > > Takashi Iwai <tiwai@xxxxxxx> writes:
> > > >
> > > > > On Fri, 05 Nov 2021 08:17:25 +0100,
> > > > > Takashi Iwai wrote:
> > > > > > When a firmware is loaded in the compressed format or via user-mode
> > > > > > helper, it's mapped in read-only, and the rtw89 driver crashes at
> > > > > > rtw89_fw_download() when it tries to modify some data.
> > > > > >
> > > > > > This patch is an attemp to avoid the crash by re-allocating the data
> > > > > > via vmalloc() for the data modification.
> > > > >
> > > > > Alternatively, we may drop the code that modifies the loaded firmware
> > > > > data? At least SET_FW_HDR_PART_SIZE() in rtw89_fw_hdr_parser() looks
> > > > > writing it, and I have no idea why this overwrite is needed.
> > > >
> > > > Strange, isn't the firmware data marked as const just to avoid this kind
> > > > of problem? Does rtw89 have wrong casts somewhere which removes the
> > > > const?
> > >
> > > Yes. SET_FW_HDR_PART_SIZE() does the cast, dropping the const.
> >
> > Oh man, all of GET and SET macros in fw.h have those casts:
> >
> > #define GET_FW_HDR_MAJOR_VERSION(fwhdr) \
> > le32_get_bits(*((__le32 *)(fwhdr) + 1), GENMASK(7, 0))
> > #define GET_FW_HDR_MINOR_VERSION(fwhdr) \
> > le32_get_bits(*((__le32 *)(fwhdr) + 1), GENMASK(15, 8))
> > #define GET_FW_HDR_SUBVERSION(fwhdr) \
> > le32_get_bits(*((__le32 *)(fwhdr) + 1), GENMASK(23, 16))
> >
> > I don't know how I missed those during my review :( But this is exactly
> > why I prefer having a proper struct for commands and events, instead of
> > u8 buf used with these macros.
> >
>
>
> I can use a struct to access firmware header, becuase their fields
> are multiple of 8 bits.
>
> But, the "firmware section header" that is additional header followed
> by firmware header, and it contains bit fields, likes:
>
> #define GET_FWSECTION_HDR_SEC_SIZE(fwhdr) \
> le32_get_bits(*((__le32 *)(fwhdr) + 1), GENMASK(23, 0))
> #define GET_FWSECTION_HDR_CHECKSUM(fwhdr) \
> le32_get_bits(*((__le32 *)(fwhdr) + 1), BIT(28))
> #define GET_FWSECTION_HDR_REDL(fwhdr) \
> le32_get_bits(*((__le32 *)(fwhdr) + 1), BIT(29))
> #define GET_FWSECTION_HDR_DL_ADDR(fwhdr) \
> le32_get_bits(*((__le32 *)(fwhdr)), GENMASK(31, 0))
>
> If we use a struct, it needs big-/little- endians parts.
>
> Then, we will access firmware header with two methods; is
> it reasonable?

You should put const in the cast in le32_get_bits() invocations, at
least.

For the le32_replace_bits(), ideally it should be rewritten in some
better way the compiler can catch. e.g. use an inline function to
take a void * argument without const,

static inline void RTW89_SET_FWCMD_CXRFK_TYPE(void *cmd, unsigned int val)
{
le32p_replace_bits((__le32 *)((u8 *)(cmd) + 2), val, GENMASK(17, 10));
}

Then the compiler will warn when you pass a const pointer there.


BTW, while reading your reply, I noticed that it's an unaligned access
to a 32bit value, which is another potential breakage on some
architectures. So the whole stuff has to be rewritten in anyway...


> The macro SET_FW_HDR_PART_SIZE() is used to set the firmware
> partition size we are going to download, and it is only used
> by rtw89_fw_download_hdr(). So, I will set the partition size
> after copying constant firmware header into skb->data.

Sounds good.


thanks,

Takashi