Message-Id: <EB6B06BB-C65F-41C2-8D9F-F916F3A706FA@jrtc27.com>
Date:   Thu, 19 Jan 2023 20:33:43 +0000
From:   Jessica Clarke <jrtc27@...c27.com>
To:     Vineet Gupta <vineetg@...osinc.com>
Cc:     linux-riscv <linux-riscv@...ts.infradead.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Palmer Dabbelt <palmer@...osinc.com>,
        Paul Walmsley <paul.walmsley@...ive.com>,
        Albert Ou <aou@...s.berkeley.edu>,
        Eric Biederman <ebiederm@...ssion.com>,
        Kees Cook <keescook@...omium.org>, Guo Ren <guoren@...nel.org>,
        Andy Chiu <andy.chiu@...ive.com>,
        Conor Dooley <conor.dooley@...rochip.com>, linux@...osinc.com
Subject: Re: [PATCH v3] riscv: elf: add .riscv.attributes parsing

On 19 Jan 2023, at 17:43, Vineet Gupta <vineetg@...osinc.com> wrote:
> 
> This implements the elf loader hook to parse RV specific
> .riscv.attributes section. This section is inserted by compilers
> (gcc/llvm) with build related information such as -march organized as
> tag/value attribute pairs.
> 
> It identifies the various attribute tags (and corresponding values) as
> currently specified in the psABI specification.
> 
> This patch only implements the elf parsing mechanics, leaving out the
> recording/usage of the attributes to subsequent patches.
> 
> Signed-off-by: Vineet Gupta <vineetg@...osinc.com>
> ---
> Changes since v2 [2]
>  - Address Jessica's review comments.
>    Mostly robustify code some more, checking for end of buffer etc.
> 
> Changes since v1 [1]
>  - Handling potential oob accesses due to against malformed elf contents
>  - Handling of multiple sub-subsections
> 
> [1]https://lore.kernel.org/linux-riscv/20230110201841.2069353-1-vineetg@rivosinc.com
> [2]https://lore.kernel.org/linux-riscv/20230112210622.2337254-1-vineetg@rivosinc.com
> 
> Given the current state of discussions, the intended Vector extension
> support would likely not need it, still posting the reworked v3 for
> logical conclusion and for posterity in case need comes up in future
> for something like CFI elf annotation.
> Maintainers/reviewers can decide whether to merge it.
> ---
> arch/riscv/Kconfig           |   1 +
> arch/riscv/include/asm/elf.h |  11 ++
> arch/riscv/kernel/Makefile   |   1 +
> arch/riscv/kernel/elf-attr.c | 239 +++++++++++++++++++++++++++++++++++
> 4 files changed, 252 insertions(+)
> create mode 100644 arch/riscv/kernel/elf-attr.c
> 
> diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> index e2b656043abf..f7e0ab05a2d2 100644
> --- a/arch/riscv/Kconfig
> +++ b/arch/riscv/Kconfig
> @@ -12,6 +12,7 @@ config 32BIT
> 
> config RISCV
> 	def_bool y
> +	select ARCH_BINFMT_ELF_STATE
> 	select ARCH_CLOCKSOURCE_INIT
> 	select ARCH_ENABLE_HUGEPAGE_MIGRATION if HUGETLB_PAGE && MIGRATION
> 	select ARCH_ENABLE_SPLIT_PMD_PTLOCK if PGTABLE_LEVELS > 2
> diff --git a/arch/riscv/include/asm/elf.h b/arch/riscv/include/asm/elf.h
> index e7acffdf21d2..7ab8bd0ec330 100644
> --- a/arch/riscv/include/asm/elf.h
> +++ b/arch/riscv/include/asm/elf.h
> @@ -116,6 +116,17 @@ do {							\
> 		*(struct user_regs_struct *)regs;	\
> } while (0);
> 
> +struct arch_elf_state {
> +};
> +
> +#define INIT_ARCH_ELF_STATE {}
> +
> +extern int arch_elf_pt_proc(void *ehdr, void *phdr, struct file *elf,
> +			    bool is_interp, struct arch_elf_state *state);
> +
> +extern int arch_check_elf(void *ehdr, bool has_interpreter, void *interp_ehdr,
> +			  struct arch_elf_state *state);
> +
> #ifdef CONFIG_COMPAT
> 
> #define SET_PERSONALITY(ex)					\
> diff --git a/arch/riscv/kernel/Makefile b/arch/riscv/kernel/Makefile
> index 4cf303a779ab..eff6d845ac9d 100644
> --- a/arch/riscv/kernel/Makefile
> +++ b/arch/riscv/kernel/Makefile
> @@ -50,6 +50,7 @@ obj-y	+= riscv_ksyms.o
> obj-y	+= stacktrace.o
> obj-y	+= cacheinfo.o
> obj-y	+= patch.o
> +obj-y	+= elf-attr.o
> obj-y	+= probes/
> obj-$(CONFIG_MMU) += vdso.o vdso/
> 
> diff --git a/arch/riscv/kernel/elf-attr.c b/arch/riscv/kernel/elf-attr.c
> new file mode 100644
> index 000000000000..8df8f2eea42b
> --- /dev/null
> +++ b/arch/riscv/kernel/elf-attr.c
> @@ -0,0 +1,239 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * Copyright (C) 2023 Rivos Inc.
> + */
> +
> +#include <linux/binfmts.h>
> +#include <linux/elf.h>
> +#include <asm/unaligned.h>
> +
> +#undef pr_fmt
> +#define pr_fmt(fmt) "rv-elf-attr: " fmt
> +
> +#define PT_RISCV_ATTRIBUTES		0x70000003
> +
> +#define RV_ATTR_TAG_file		1
> +
> +#define RV_ATTR_TAG_stack_align		4
> +#define RV_ATTR_TAG_arch		5
> +#define RV_ATTR_TAG_unaligned_access	6
> +
> +#define RV_ATTR_VENDOR_RISCV		"riscv"
> +
> +#define RV_ATTR_SEC_SZ			SZ_1K
> +
> +static void rv_elf_attr_int(u64 tag, u64 val)
> +{
> +	pr_debug("Tag %llx=%llx\n", tag, val);
> +}
> +
> +static void rv_elf_attr_str(u64 tag, const unsigned char *str)
> +{
> +	pr_debug("Tag %llx=[%s]\n", tag, str);
> +}
> +
> +/*
> + * Decode a ule128 encoded value.
> + */
> +static int
> +decode_uleb128_safe(unsigned char **dpp, u64 *val, const unsigned char *p_end)
> +{
> +	unsigned char *bp = *dpp;
> +	unsigned char byte;
> +	unsigned int shift = 0;
> +	u64 result = 0;
> +
> +	while (bp < p_end) {
> +		byte = *bp++;
> +		result |= (byte & 0x7f) << shift;
> +		if ((byte & 0x80) == 0) {
> +			*dpp = bp;
> +			*val = result;
> +			return 0;
> +		}
> +		shift += 7;
> +	}
> +
> +	return -1;
> +}
> +
> +#define ELF_ATTR_TAG_EVEN(t)	(((t) % 2) ? false : true)

I don’t think you need a macro for "unsigned integer is even".

> +
> +/*
> + * Parse a single elf attribute.
> + */
> +static int rv_parse_elf_attr_safe(unsigned char **dpp, unsigned char *p_end)
> +{
> +	unsigned char *p = *dpp;
> +	unsigned char *str;
> +	u64 tag, val;
> +	u32 s_len;
> +
> +	if (decode_uleb128_safe(&p, &tag, p_end))
> +		goto bad_attr;
> +
> +	switch (tag) {
> +	case RV_ATTR_TAG_stack_align:
> +		if (decode_uleb128_safe(&p, &val, p_end))
> +			goto bad_attr;
> +		if (!ELF_ATTR_TAG_EVEN(tag))
> +			goto bad_attr;

Huh? You just checked it against a constant so you know exactly what it
is. This is just Static_assert(RV_ATTR_TAG_stack_align % 2 == 0) but at
run time. And you know that’s going to be the case, because the spec is
self-consistent by design (wouldn’t be a worthwhile spec otherwise).

> +		rv_elf_attr_int(tag, val);
> +		break;
> +
> +	case RV_ATTR_TAG_unaligned_access:
> +		if (decode_uleb128_safe(&p, &val, p_end))
> +			goto bad_attr;
> +		if (!ELF_ATTR_TAG_EVEN(tag))
> +			goto bad_attr;
> +		rv_elf_attr_int(tag, val);
> +		break;
> +
> +	case RV_ATTR_TAG_arch:
> +		if (ELF_ATTR_TAG_EVEN(tag))
> +			goto bad_attr;
> +		str = p;
> +		s_len = strnlen(p, p_end - p) + 1;
> +		if ((p + s_len) > p_end)
> +			goto bad_attr;
> +		p += s_len;
> +		rv_elf_attr_str(tag, str);
> +		break;
> +
> +	default:

The whole point of the even/odd split is so that when you *don’t* know
what the tag means you can still decode its value and thus know how to
skip past it. That is, *here* is where you need to be checking
even/odd, and deciding whether to treat it as a string or a ULEB128,
which is why I annotated *here* not one of the other case labels before.

> +		if (decode_uleb128_safe(&p, &val, p_end))
> +			goto bad_attr;
> +		pr_debug("skipping Unrecognized Tag [%llx]=%llx\n", tag, val);
> +		break;
> +	}
> +
> +	*dpp = p;
> +	return 0;
> +bad_attr:
> +	return -ENOEXEC;
> +}
> +
> +/*
> + * Parse .riscv.attributes elf section.
> + */
> +static int rv_parse_elf_attributes(struct file *f, const struct elf_phdr *phdr,
> +				   struct arch_elf_state *state)
> +{
> +	unsigned char buf[RV_ATTR_SEC_SZ];
> +	unsigned char *p, *p_end;
> +	ssize_t n;
> +	int ret = 0;
> +	loff_t pos;
> +
> +	pr_debug("Section .riscv.attributes found\n");
> +
> +	/* Assume a reasonable size for now */
> +	if (phdr->p_filesz > sizeof(buf))
> +		goto bad_elf;
> +
> +	memset(buf, 0, RV_ATTR_SEC_SZ);
> +	pos = phdr->p_offset;
> +	n = kernel_read(f, &buf, phdr->p_filesz, &pos);
> +
> +	if (n <= 0)
> +		return -EIO;

0 should be ENOEXEC not EIO? And surely in the < 0 case you want to be
forwarding on the exact error from kernel_read, not squashing it into
EIO?

> +
> +	p = buf;
> +	p_end = p + n;
> +
> +	/* sanity check format-version */
> +	if (*p++ != 'A')
> +		goto bad_elf;
> +
> +	/*
> +	 * elf attribute section organized as Vendor sub-sections(s)
> +	 *   {sub-section length, vendor name, vendor data}
> +	 * Vendor data organized as sub-subsection(s)
> +	 *   {tag, sub-subsection length, attributes contents}
> +	 * Attribute contents organized as
> +	 *   {tag, value} pair(s).
> +	 */
> +	while ((p_end - p) >= 4) {
> +		u32 sub_len, vname_len;
> +
> +		sub_len = get_unaligned_le32(p);
> +		if (sub_len <= 4 || sub_len > (p_end - p))
> +			goto bad_elf;
> +
> +		p += 4;
> +		sub_len -= 4;
> +
> +		/* Vendor name string */
> +		vname_len = strnlen(p, sub_len) + 1;
> +		if (vname_len > sub_len)
> +			goto bad_elf;
> +
> +		/* skip non-mandatory sub-section for now */
> +		if (strncmp(p, RV_ATTR_VENDOR_RISCV, sub_len)) {
> +			p += sub_len;
> +			continue;
> +		}
> +
> +		p += vname_len;
> +		sub_len -= vname_len;
> +
> +		/* Vendor data: sub-subsections(s) */
> +		while (sub_len > 0) {
> +			unsigned char *p_ss_end, *p_ss_start = p;
> +			u32 ss_len;
> +			u64 tag;
> +
> +			if (decode_uleb128_safe(&p, &tag, p_end))
> +				goto bad_elf;
> +
> +			if ((p_end - p) < 4)
> +				goto bad_elf;
> +
> +			ss_len = get_unaligned_le32(p);
> +			if (ss_len > sub_len)
> +				goto bad_elf;
> +
> +			p += 4;
> +			sub_len -= ss_len;
> +			p_ss_end = p_ss_start + ss_len;
> +
> +			/* For now handle attributes relating to whole file */
> +			if (tag != RV_ATTR_TAG_file) {
> +				p = p_ss_end;
> +				continue;
> +			}
> +
> +			/* Attribute(s): tag:value pairs */
> +			while (p < p_ss_end) {
> +				ret = rv_parse_elf_attr_safe(&p, p_end);
> +				if (ret)
> +					goto bad_elf;
> +			}
> +		}
> +	}
> +
> +	return ret;
> +bad_elf:
> +	return -ENOEXEC;
> +}
> +
> +/*
> + * Hook invoked by generic elf loader to parse riscv specific elf segments.
> + */
> +int arch_elf_pt_proc(void *_ehdr, void *_phdr, struct file *elf,
> +		     bool is_interp, struct arch_elf_state *state)
> +{
> +	struct elf_phdr *phdr = _phdr;
> +	int ret = 0;
> +
> +	if (phdr->p_type == PT_RISCV_ATTRIBUTES && !is_interp)

Both the executable and its interpreter matter.

Jess

> +		ret = rv_parse_elf_attributes(elf, phdr, state);
> +
> +	return ret;
> +}
> +
> +int arch_check_elf(void *_ehdr, bool has_interpreter, void *_interp_ehdr,
> +		   struct arch_elf_state *state)
> +{
> +	return 0;
> +}
> -- 
> 2.34.1
> 

Powered by blists - more mailing lists