Date:   Fri, 05 Nov 2021 16:07:08 +0100
Message-ID: <s5h5yt6fxpf.wl-tiwai@suse.de>
From:   Takashi Iwai <tiwai@suse.de>
To:     Pkshih <pkshih@realtek.com>
Cc:     "kvalo@codeaurora.org" <kvalo@codeaurora.org>,
        "linux-wireless@vger.kernel.org" <linux-wireless@vger.kernel.org>,
        "Larry.Finger@gmail.com" <Larry.Finger@gmail.com>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] rtw89: Fix crash by loading compressed firmware file
In-Reply-To: <bd80d3b6cdc42d7818d7d5c6a5036d8188eb4a67.camel@realtek.com>
References: <20211105071725.31539-1-tiwai@suse.de>
        <s5hpmrfgj93.wl-tiwai@suse.de>
        <87zgqjqaae.fsf@codeaurora.org>
        <s5hh7crgflg.wl-tiwai@suse.de>
        <87v917q8hw.fsf@codeaurora.org>
        <bd80d3b6cdc42d7818d7d5c6a5036d8188eb4a67.camel@realtek.com>
User-Agent: Wanderlust/2.15.9 (Almost Unreal) SEMI/1.14.6 (Maruoka)
 FLIM/1.14.9 (=?UTF-8?B?R29qxY0=?=) APEL/10.8 Emacs/25.3
 (x86_64-suse-linux-gnu) MULE/6.0 (HANACHIRUSATO)
MIME-Version: 1.0 (generated by SEMI 1.14.6 - "Maruoka")
Content-Type: text/plain; charset=US-ASCII
Precedence: bulk

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@suse.de> writes:
> > 
> > > On Fri, 05 Nov 2021 09:25:13 +0100,
> > > Kalle Valo wrote:
> > > > Takashi Iwai <tiwai@suse.de> 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