Andrii, sorry for the late reply. This change looks good. It makes much more sense than before. Thanks, Hao On Thu, Dec 10, 2020 at 8:11 PM Andrii Nakryiko <andrii@xxxxxxxxxx> wrote: > > Adjust pahole logic of skipping any per-CPU symbol with offset 0, which is > especially bad for kernel modules, because it most certainly skips the very > first per-CPU variable. > > Instead, do collect per-CPU ELF symbol with 0 offset, but do extra check for > non-kernel module case by verifying that ELF symbol name and DWARF variable > name match. Due to the bug of DWARF name of variable sometimes being NULL, > this is necessarily too pessimistic check (e.g., on my vmlinux image, > fixed_percpu_data variable is still not emitted due to missing DWARF variable > name), it allows to emit data for all module per-CPU variables. > > Fixes: f3d9054ba8ff ("btf_encoder: Teach pahole to store percpu variables in vmlinux BTF.") > Signed-off-by: Andrii Nakryiko <andrii@xxxxxxxxxx> > --- > btf_encoder.c | 40 ++++++++++++++++++++++++---------------- > 1 file changed, 24 insertions(+), 16 deletions(-) > > diff --git a/btf_encoder.c b/btf_encoder.c > index a7d484765ce2..1d7817078f89 100644 > --- a/btf_encoder.c > +++ b/btf_encoder.c > @@ -412,21 +412,6 @@ static int collect_percpu_var(struct btf_elf *btfe, GElf_Sym *sym) > return 0; > > addr = elf_sym__value(sym); > - /* > - * Store only those symbols that have allocated space in the percpu section. > - * This excludes the following three types of symbols: > - * > - * 1. __ADDRESSABLE(sym), which are forcely emitted as symbols. > - * 2. __UNIQUE_ID(prefix), which are introduced to generate unique ids. > - * 3. __exitcall(fn), functions which are labeled as exit calls. > - * > - * In addition, the variables defined using DEFINE_PERCPU_FIRST are > - * also not included, which currently includes: > - * > - * 1. fixed_percpu_data > - */ > - if (!addr) > - return 0; > > size = elf_sym__size(sym); > if (!size) > @@ -652,7 +637,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force, > > cu__for_each_variable(cu, core_id, pos) { > uint32_t size, type, linkage; > - const char *name; > + const char *name, *dwarf_name; > uint64_t addr; > int id; > > @@ -680,6 +665,29 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force, > if (!percpu_var_exists(addr, &size, &name)) > continue; /* not a per-CPU variable */ > > + /* A lot of "special" DWARF variables (e.g, __UNIQUE_ID___xxx) > + * have addr == 0, which is the same as, say, valid > + * fixed_percpu_data per-CPU variable. To distinguish between > + * them, additionally compare DWARF and ELF symbol names. If > + * DWARF doesn't provide proper name, pessimistically assume > + * bad variable. > + * > + * Examples of such special variables are: > + * > + * 1. __ADDRESSABLE(sym), which are forcely emitted as symbols. > + * 2. __UNIQUE_ID(prefix), which are introduced to generate unique ids. > + * 3. __exitcall(fn), functions which are labeled as exit calls. > + * > + * This is relevant only for vmlinux image, as for kernel > + * modules per-CPU data section has non-zero offset so all > + * per-CPU symbols have non-zero values. > + */ > + if (var->ip.addr == 0) { > + dwarf_name = variable__name(var, cu); > + if (!dwarf_name || strcmp(dwarf_name, name)) > + continue; > + } > + > if (var->spec) > var = var->spec; > > -- > 2.24.1 >