On Fri, Jan 22, 2021 at 12:16:42PM -0800, Andrii Nakryiko wrote:
> On Fri, Jan 22, 2021 at 8:46 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> >
> > For very large ELF objects (with many sections), we could
> > get special value SHN_XINDEX (65535) for symbol's st_shndx.
> >
> > This patch is adding code to detect the optional extended
> > section index table and use it to resolve symbol's section
> > index.
> >
> > Adding elf_symtab__for_each_symbol_index macro that returns
> > symbol's section index and usign it in collect functions.
> >
> > Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> > ---
> > btf_encoder.c | 59 +++++++++++++++++++++++++++++++++++++--------------
> > elf_symtab.c | 39 +++++++++++++++++++++++++++++++++-
> > elf_symtab.h | 2 ++
> > 3 files changed, 83 insertions(+), 17 deletions(-)
> >
> > diff --git a/btf_encoder.c b/btf_encoder.c
> > index 5557c9efd365..56ee55965093 100644
> > --- a/btf_encoder.c
> > +++ b/btf_encoder.c
> > @@ -63,13 +63,13 @@ static void delete_functions(void)
> > #define max(x, y) ((x) < (y) ? (y) : (x))
> > #endif
> >
> > -static int collect_function(struct btf_elf *btfe, GElf_Sym *sym)
> > +static int collect_function(struct btf_elf *btfe, GElf_Sym *sym,
> > + Elf32_Word sym_sec_idx)
>
> nit: we use size_t or int for this, no need for libelf types here, imo
ok
>
>
>
> > {
> > struct elf_function *new;
> > static GElf_Shdr sh;
> > - static int last_idx;
> > + static Elf32_Word last_idx;
> > const char *name;
> > - int idx;
> >
> > if (elf_sym__type(sym) != STT_FUNC)
> > return 0;
> > @@ -90,12 +90,10 @@ static int collect_function(struct btf_elf *btfe, GElf_Sym *sym)
> > functions = new;
> > }
> >
> > - idx = elf_sym__section(sym);
> > -
> > - if (idx != last_idx) {
> > - if (!elf_section_by_idx(btfe->elf, &sh, idx))
> > + if (sym_sec_idx != last_idx) {
> > + if (!elf_section_by_idx(btfe->elf, &sh, sym_sec_idx))
> > return 0;
> > - last_idx = idx;
> > + last_idx = sym_sec_idx;
> > }
> >
> > functions[functions_cnt].name = name;
> > @@ -542,14 +540,15 @@ static bool percpu_var_exists(uint64_t addr, uint32_t *sz, const char **name)
> > return true;
> > }
> >
> > -static int collect_percpu_var(struct btf_elf *btfe, GElf_Sym *sym)
> > +static int collect_percpu_var(struct btf_elf *btfe, GElf_Sym *sym,
> > + Elf32_Word sym_sec_idx)
>
> nit: same, size_t or just int would be fine
>
> > {
> > const char *sym_name;
> > uint64_t addr;
> > uint32_t size;
> >
> > /* compare a symbol's shndx to determine if it's a percpu variable */
> > - if (elf_sym__section(sym) != btfe->percpu_shndx)
> > + if (sym_sec_idx != btfe->percpu_shndx)
> > return 0;
> > if (elf_sym__type(sym) != STT_OBJECT)
> > return 0;
> > @@ -585,12 +584,13 @@ static int collect_percpu_var(struct btf_elf *btfe, GElf_Sym *sym)
> > return 0;
> > }
> >
> > -static void collect_symbol(GElf_Sym *sym, struct funcs_layout *fl)
> > +static void collect_symbol(GElf_Sym *sym, struct funcs_layout *fl,
> > + Elf32_Word sym_sec_idx)
> > {
> > if (!fl->mcount_start &&
> > !strcmp("__start_mcount_loc", elf_sym__name(sym, btfe->symtab))) {
> > fl->mcount_start = sym->st_value;
> > - fl->mcount_sec_idx = sym->st_shndx;
> > + fl->mcount_sec_idx = sym_sec_idx;
> > }
> >
> > if (!fl->mcount_stop &&
> > @@ -598,9 +598,36 @@ static void collect_symbol(GElf_Sym *sym, struct funcs_layout *fl)
> > fl->mcount_stop = sym->st_value;
> > }
> >
> > +static bool elf_sym__get(Elf_Data *syms, Elf_Data *syms_sec_idx_table,
> > + int id, GElf_Sym *sym, Elf32_Word *sym_sec_idx)
>
> This is a generic function, why don't you want to move it into elf_symtab.h?
>
> > +{
> > + if (!gelf_getsym(syms, id, sym))
> > + return false;
> > +
> > + *sym_sec_idx = sym->st_shndx;
> > +
> > + if (sym->st_shndx == SHN_XINDEX) {
> > + if (!syms_sec_idx_table)
> > + return false;
> > + if (!gelf_getsymshndx(syms, syms_sec_idx_table,
> > + id, sym, sym_sec_idx))
> > + return false;
>
> You also ignored my feedback about not fetching symbol twice. Why?
ugh, I did not read your last 2 comments.. let me check that, sry
>
> > + }
> > +
> > + return true;
> > +}
> > +
> > +#define elf_symtab__for_each_symbol_index(symtab, id, sym, sym_sec_idx) \
> > + for (id = 0; \
> > + id < symtab->nr_syms && \
> > + elf_sym__get(symtab->syms, symtab->syms_sec_idx_table, \
> > + id, &sym, &sym_sec_idx); \
> > + id++)
>
> This should be in elf_symtab.h next to elf_symtab__for_each_symbol.
>
> And thinking a bit more, the variant with just ignoring symbols that
> we failed to get is probably a safer alternative. I.e., currently
> there is no way to communicate that we terminated iteration with
> error, so it's probably better to skip failed symbols and still get
> the rest, no? I was hoping to discuss stuff like this on the previous
> version of the patch...
I was thinking of that, but then I thought it's better to fail,
than have possibly wrong data in BTF, because the ELF data is
possibly damaged? no idea.. so I took the safer way
jirka
