On Tue, Sep 3, 2024 at 12:33 AM Tony Ambardar <tony.ambardar@xxxxxxxxx> wrote:
>
> Support for handling BTF data of either endianness was added in [1], but
> did not include BTF.ext data for lack of use cases. Later, support for
> static linking [2] provided a use case, but this feature and later ones
> were restricted to native-endian usage.
>
> Add support for BTF.ext handling in either endianness. Convert BTF.ext data
> to native endianness when read into memory for further processing, and
> support raw data access that restores the original byte-order for output.
> Add internal header functions for byte-swapping func, line, and core info
> records.
>
> Add new API functions btf_ext__endianness() and btf_ext__set_endianness()
> for query and setting byte-order, as already exist for BTF data.
>
> [1] 3289959b97ca ("libbpf: Support BTF loading and raw data output in both endianness")
> [2] 8fd27bf69b86 ("libbpf: Add BPF static linker BTF and BTF.ext support")
>
> Signed-off-by: Tony Ambardar <tony.ambardar@xxxxxxxxx>
> ---
> tools/lib/bpf/btf.c | 238 +++++++++++++++++++++++++++++---
> tools/lib/bpf/btf.h | 3 +
> tools/lib/bpf/libbpf.map | 2 +
> tools/lib/bpf/libbpf_internal.h | 28 ++++
> 4 files changed, 255 insertions(+), 16 deletions(-)
>
> diff --git a/tools/lib/bpf/btf.c b/tools/lib/bpf/btf.c
> index 5f094c1f4388..c11dfc81d007 100644
> --- a/tools/lib/bpf/btf.c
> +++ b/tools/lib/bpf/btf.c
> @@ -3023,25 +3023,140 @@ static int btf_ext_setup_core_relos(struct btf_ext *btf_ext)
> return btf_ext_setup_info(btf_ext, ¶m);
> }
>
> -static int btf_ext_parse_hdr(__u8 *data, __u32 data_size)
> +/* Swap byte-order of BTF.ext header with any endianness */
> +static void btf_ext_bswap_hdr(struct btf_ext *btf_ext, __u32 hdr_len)
> {
> - const struct btf_ext_header *hdr = (struct btf_ext_header *)data;
> + struct btf_ext_header *h = btf_ext->hdr;
>
> - if (data_size < offsetofend(struct btf_ext_header, hdr_len) ||
> - data_size < hdr->hdr_len) {
> - pr_debug("BTF.ext header not found\n");
> + h->magic = bswap_16(h->magic);
> + h->hdr_len = bswap_32(h->hdr_len);
> + h->func_info_off = bswap_32(h->func_info_off);
> + h->func_info_len = bswap_32(h->func_info_len);
> + h->line_info_off = bswap_32(h->line_info_off);
> + h->line_info_len = bswap_32(h->line_info_len);
> +
> + if (hdr_len < offsetofend(struct btf_ext_header, core_relo_len))
> + return;
> +
> + h->core_relo_off = bswap_32(h->core_relo_off);
> + h->core_relo_len = bswap_32(h->core_relo_len);
> +}
> +
> +/* Swap metadata byte-order of generic info subsection */
> +static int info_subsec_bswap_metadata(const struct btf_ext *btf_ext, struct btf_ext_info *ext_info)
> +{
> + const bool is_native = btf_ext->swapped_endian;
> + __u32 left, *rs, rec_size, num_info;
> + struct btf_ext_info_sec *si;
> +
> + if (ext_info->len == 0)
> + return 0;
> +
> + rs = ext_info->info - sizeof(__u32); /* back up to record size */
> + rec_size = is_native ? *rs : bswap_32(*rs);
> + if (rec_size != ext_info->rec_size)
> return -EINVAL;
> + *rs = bswap_32(*rs);
> +
> + si = ext_info->info; /* sec info #1 */
> + left = ext_info->len;
> + while (left > 0) {
> + num_info = is_native ? si->num_info : bswap_32(si->num_info);
> + si->sec_name_off = bswap_32(si->sec_name_off);
> + si->num_info = bswap_32(si->num_info);
> + si = (void *)si->data + rec_size * num_info;
> + left -= offsetof(struct btf_ext_info_sec, data) +
> + rec_size * num_info;
> }
>
> + return 0;
> +}
> +
> +/* Swap byte order of info subsection metadata and records in the correct
> + * order depending on whether or not data is in native endianness.
> + */
> +#define ORDER_INFO_BSWAP(btf_ext, ext_info, info_t, swap_fn) \
> +{ \
> + const bool is_native = btf_ext->swapped_endian; \
> + struct btf_ext_info_sec *si; \
> + int c, err; \
> + info_t *i; \
> + if (is_native) \
> + for_each_btf_ext_sec(ext_info, si) \
> + for_each_btf_ext_rec(ext_info, si, c, i) \
> + swap_fn(i); \
> + err = info_subsec_bswap_metadata(btf_ext, ext_info); \
> + if (err) { \
> + pr_warn(#info_t " record size mismatch!\n"); \
> + return err; \
> + } \
> + if (!is_native) \
> + for_each_btf_ext_sec(ext_info, si) \
> + for_each_btf_ext_rec(ext_info, si, c, i) \
> + swap_fn(i); \
> +}
> +
Nope-nope-nope, I'm never landing something like this :)
Ok, if there is no clean swapping and setup separation we can come up
with, let's still do a decent job at trying to keep all this coherent
and simple.
How about this? btf_ext_bswap_hdr() stays, it seems fine.
btf_ext_setup_func_info(), btf_ext_setup_line_info(), and
btf_ext_setup_core_relos() all call into generic btf_ext_setup_info,
right? We teach btf_ext_setup_info() to work with both endianness (but
not swap data at all). The only two fields that we need to byte-swap
on the fly are record_size and sinfo->num_info, so it's minimal
changes to btf_ext_setup_info().
Oh, and while we are at it, if data is in non-native endianness,
btf_ext_setup_info() should return failure if any of the record sizes
are not *exactly* matching our expectations. This should be a trivial
addition as we have never extended those records, I believe.
This will take care of correctness checking regardless of endianness.
Now, forget about for_each_btf_ext_{sec,rec}(), it will be too fragile
to make them work for inverted endianness. But we know now that the
data is correct, right? So, fine, let's have a bit of duplication (but
without any error checking) to go over raw .BTF.ext data and swap all
the records and all metadata fields. This logic now will be used after
btf_ext_setup_*() steps, and in btf_ext_raw_data(). For
btf_ext_raw_data() you'll also btf_ext_bswap_hdr() at the very end,
right?
How does that sound? Am I missing something big again?
And fine, let's use callbacks for different record types to keep this
simple. I dislike callbacks, in principle, but sometimes they are the
simplest way forward, unfortunately (a proper iterator would be
better, but that's another story and I don't want to get into that
implementation task just yet).
> +/*
> + * Swap endianness of the whole info segment in a BTF.ext data section:
> + * - requires BTF.ext header data in native byte order
> + * - only support info structs from BTF version 1
> + */
> +static int btf_ext_bswap_info(struct btf_ext *btf_ext)
> +{
> + const struct btf_ext_header *h = btf_ext->hdr;
> + struct btf_ext_info ext = {};
> +
> + /* Swap func_info subsection byte-order */
> + ext.info = (void *)h + h->hdr_len + h->func_info_off + sizeof(__u32);
> + ext.len = h->func_info_len - (h->func_info_len ? sizeof(__u32) : 0);
> + ext.rec_size = sizeof(struct bpf_func_info);
> +
> + ORDER_INFO_BSWAP(btf_ext, &ext, struct bpf_func_info, bpf_func_info_bswap);
You shouldn't have bent over backwards just to use
for_each_btf_ext_{sec,rec}() macros. Just because I proposed it
(initially, without actually coding anything) doesn't mean it's the
best and final solution :)
> +
> + /* Swap line_info subsection byte-order */
> + ext.info = (void *)h + h->hdr_len + h->line_info_off + sizeof(__u32);
> + ext.len = h->line_info_len - (h->line_info_len ? sizeof(__u32) : 0);
> + ext.rec_size = sizeof(struct bpf_line_info);
> +
> + ORDER_INFO_BSWAP(btf_ext, &ext, struct bpf_line_info, bpf_line_info_bswap);
> +
> + /* Swap core_relo subsection byte-order (if present) */
> + if (h->hdr_len < offsetofend(struct btf_ext_header, core_relo_len))
> + return 0;
> +
> + ext.info = (void *)h + h->hdr_len + h->core_relo_off + sizeof(__u32);
> + ext.len = h->core_relo_len - (h->core_relo_len ? sizeof(__u32) : 0);
> + ext.rec_size = sizeof(struct bpf_core_relo);
> +
> + ORDER_INFO_BSWAP(btf_ext, &ext, struct bpf_core_relo, bpf_core_relo_bswap);
> +
> + return 0;
> +}
> +#undef ORDER_INFO_BSWAP
[...]
