"Jiaxun Yang" jiaxun.yang@flygoat.com writes:

在2024年8月24日八月 下午3:41，Bjørn Mork写道：

...

Boards based on the same SoC family can use different boot loaders. These may pass numeric arguments which we erroneously interpret as command line or environment pointers. Such errors will cause boot to halt at an early stage since commit 056a68cea01e ("mips: allow firmware to pass RNG seed to kernel").

One known example of this issue is a HPE switch using a BootWare boot loader. It was found to pass these arguments to the kernel:

0x00020000 0x00060000 0xfffdffff 0x0000416c

We can avoid hanging by validating that both passed pointers are in KSEG1 as expected.

Hi Bjorn,

This is actually breaking 64 bit systems passing fw_args in XKPHYS or KSEG0.

Ouch. Thanks for the feedback.

But if so, then aren't those already broken with the current test against CKSEG0? I didn't add that.

IIUC, CKSEGx is the sign extendend version of KSEGx on 64BIT:

#ifdef CONFIG_64BIT

/* * Memory segments (64bit kernel mode addresses) * The compatibility segments use the full 64-bit sign extended value. Note * the R8000 doesn't have them so don't reference these in generic MIPS code. */ #define XKUSEG _CONST64_(0x0000000000000000) #define XKSSEG _CONST64_(0x4000000000000000) #define XKPHYS _CONST64_(0x8000000000000000) #define XKSEG _CONST64_(0xc000000000000000) #define CKSEG0 _CONST64_(0xffffffff80000000) #define CKSEG1 _CONST64_(0xffffffffa0000000) #define CKSSEG _CONST64_(0xffffffffc0000000) #define CKSEG3 _CONST64_(0xffffffffe0000000)

#define CKSEG0ADDR(a) (CPHYSADDR(a) | CKSEG0) #define CKSEG1ADDR(a) (CPHYSADDR(a) | CKSEG1) #define CKSEG2ADDR(a) (CPHYSADDR(a) | CKSEG2) #define CKSEG3ADDR(a) (CPHYSADDR(a) | CKSEG3)

Maybe something like:

static inline bool valid_fw_arg(unsigned long arg) { #ifdef CONFIG_64BIT if (arg >= XKPHYS && arg < XKSEG) return TRUE; #endif return arg >= CKSEG0 && arg < CKSSEG; }

Will be more robust.

Maybe?

But I can't make that match what U-Boot does. AFAICS, u-boot/arch/mips/lib/bootm.c doesn't care about 32 or 64 bit, and simply does:

static void linux_cmdline_init(void) { linux_argc = 1; linux_argv = (char **)CKSEG1ADDR(gd->bd->bi_boot_params); linux_argv[0] = 0; linux_argp = (char *)(linux_argv + LINUX_MAX_ARGS); }

Then it derives the argument and environment pointers from linux_argv. Which means that all these pointers end up somewhere between CKSEG1 and CKSEG2 on both 32 and 64 bit. Or am I missing something?

Tried looking for other MIPS code in U-Boot handling this for 64 bit, but all I found was arch/mips/mach-octeon/bootoctlinux.c which uses a completely different protocol.

Sorry if I am asking stupid questions here. Trying a rather steep learning curve. Maybe too steep :-)

Bjørn

"Maciej W. Rozycki" macro@orcam.me.uk writes:

On Sun, 25 Aug 2024, Bjørn Mork wrote:

...

But I can't make that match what U-Boot does. AFAICS, u-boot/arch/mips/lib/bootm.c doesn't care about 32 or 64 bit, and simply does:

U-Boot isn't the only firmware used with MIPS systems, and in fact it's quite a recent addition; e.g. DEC REX and SGI ARCS firmware goes back at least to 1990 and we continue supporting these systems. There's also CFE, PMON, YAMON, etc.

I know. That's what I'm trying to make us support :-)

We can't blindly assume that all those firmwares implement the exact same command line and environment protocol as U-Boot. Yamon obviously uses something similar enough. But what about the others?

As mentioned, we're currently failing to boot from Bootware because of the relaxed checking here. Something needs to be done about that.

Bjørn

"Maciej W. Rozycki" macro@orcam.me.uk writes:

Even those that do use the function have a choice to override the default handler by setting CONFIG_HAVE_PLAT_FW_INIT_CMDLINE. Perhaps it's what you need to do for your platform too.

Considered that, but this problem isn't directly platform-related, is it? Many of the firmware implementations are multi-platform. And they support many of the same platforms.

My concrete eample showed up on the "realtek" (rtl838x and rtl93xx SoCs) platform in OpenWrt, where a large number of boards from assorted vendors are currently supported. The current implementation works fine for most of them because they using U-Boot. But it failed in what I consider the ugliest way possible (no ooops, no error message, just hanging) on the boards from HPE because they use Bootware.

FWIW, there is also a vendor using BootBase for pretty much the same hardware. But that's still unsupported in OpenWrt for various reasons.

So yes, this could be worked around by having a platform specific fw_init_cmdline(). Or even another out-of-tree OpenWrt specific patch. But whatever the solution is, it can't drop the U-Boot support since most of the boards actually use U-Boot. Which is why I believe it's much better to solve this problem at the root, for the benefit of everyone else.

But of course, if necessary it would be possible to build X kernels with dfferent fw_init_cmdline() implementations to support the same hardware booted from X different boot firmwares.

It's clear to me that this mess has to be cleaned up. Not all kinds of firmware permit the setting of arbitrary environment variables (or ones that survive a reboot) though.

Yes. And I believe it can be solved by improving the pointer validation that's already there. All we need is to reject argument values passed by other firmwares.

But it's probably better to create an inline valid_fw_arg(() or similar function as proposed by Jiaxun, allowing the XKPHYS range too on 64 bit systems.

If necessary we can improve further by adding an alignment requirement, which was a proposal that came up in the OpenWrt discussions.

Another option is to connect the validation of all 4 arguments. There is for example no reason to interpret the Bootware fw_arg2 value as an enviroment pointer after we've already rejected fw_arg1 as cmdline.

It's also possible to validate command line argument pointers and environment variable pointers (except for alignment, I guess?), refusing anything which ends ut pointing outsde KSEG1 or XKPHYS.

How complicated you want this is all a matter of taste IMHO. I tried to make this a simple solution matching the current "forgiving" implementation.

Another improvement would be a pr_info() dumping the fw_argX values. It would make it possible to understand why the code is hanging when the firmware does something unexpected. I don't think pr_debug() helps much in this particular case.

Bjørn