lkml.org 
[lkml]   [2022]   [May]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Date
SubjectRe: [PATCH -next 1/2] bpf: Unify data extension operation of jited_ksyms and jited_linfo
On Thu, Apr 28, 2022 at 2:47 AM Pu Lehui <pulehui@huawei.com> wrote:
>
> Hi Andrii,
>
> On 2022/4/28 6:33, Andrii Nakryiko wrote:
> > On Tue, Apr 26, 2022 at 6:40 AM Pu Lehui <pulehui@huawei.com> wrote:
> >>
> >> We found that 32-bit environment can not print bpf line info due
> >> to data inconsistency between jited_ksyms[0] and jited_linfo[0].
> >>
> >> For example:
> >> jited_kyms[0] = 0xb800067c, jited_linfo[0] = 0xffffffffb800067c
> >>
> >> We know that both of them store bpf func address, but due to the
> >> different data extension operations when extended to u64, they may
> >> not be the same. We need to unify the data extension operations of
> >> them.
> >>
> >> Signed-off-by: Pu Lehui <pulehui@huawei.com>
> >> ---
> >> kernel/bpf/syscall.c | 5 ++++-
> >> tools/lib/bpf/bpf_prog_linfo.c | 8 ++++----
> >> tools/testing/selftests/bpf/prog_tests/btf.c | 18 +++++++++---------
> >
> > please split kernel changes, libbpf changes, and selftests/bpf changes
> > into separate patches
> Thanks for your review. Alright, I will split it next time.
>
> >
> >> 3 files changed, 17 insertions(+), 14 deletions(-)
> >>
> >> diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> >> index e9621cfa09f2..4c417c806d92 100644
> >> --- a/kernel/bpf/syscall.c
> >> +++ b/kernel/bpf/syscall.c
> >> @@ -3868,13 +3868,16 @@ static int bpf_prog_get_info_by_fd(struct file *file,
> >> info.nr_jited_line_info = 0;
> >> if (info.nr_jited_line_info && ulen) {
> >> if (bpf_dump_raw_ok(file->f_cred)) {
> >> + unsigned long jited_linfo_addr;
> >> __u64 __user *user_linfo;
> >> u32 i;
> >>
> >> user_linfo = u64_to_user_ptr(info.jited_line_info);
> >> ulen = min_t(u32, info.nr_jited_line_info, ulen);
> >> for (i = 0; i < ulen; i++) {
> >> - if (put_user((__u64)(long)prog->aux->jited_linfo[i],
> >> + jited_linfo_addr = (unsigned long)
> >> + prog->aux->jited_linfo[i];
> >> + if (put_user((__u64) jited_linfo_addr,
> >> &user_linfo[i]))
> >> return -EFAULT;
> >> }
> Please let me to explain more detail, sorry if I'm wordy.
> The main reason that 32-bit env does not print bpf line info is here:
>
> kernel/bpf/syscall.c:
> bpf_prog_get_info_by_fd {
> ...
> user_ksyms = u64_to_user_ptr(info.jited_ksyms);
> ksym_addr = (unsigned long)prog->aux->func[i]->bpf_func;
> if (put_user((u64) ksym_addr, &user_ksyms[i]))
> ...
>
> user_linfo = u64_to_user_ptr(info.jited_line_info);
> if (put_user((__u64)(long)prog->aux->jited_linfo[i],
> &user_linfo[i]))
> ...
> }
>
> In 32-bit env, ksym_addr and prog->aux->jited_linfo[0] both store the
> 32-bit address of bpf_func, but the first one is zero-extension to u64,
> while the other is sign-extension to u64.
> For example:
> prog->aux->func[0]->bpf_func = 0xb800067c
> user_ksyms[0] = 0xb800067c, user_linfo[0] = 0xffffffffb800067c
>
> Both zero-extension and sign-extension are fine, but if operating
> directly between them without casting in 32-bit env, there will have
> some potential problems. Such as:
>
> tools/lib/bpf/bpf_prog_linfo.c:
> dissect_jited_func {
> ...
> if (ksym_func[0] != *jited_linfo) //always missmatch in 32 env
> goto errout;
> ...
> if (ksym_func[f] == *jited_linfo) {
> ...
> last_jited_linfo = *jited_linfo;
> if (last_jited_linfo - ksym_func[f - 1] + 1 >
> ksym_len[f - 1])
> ...
> }
>
> We could cast them to 32-bit data type, but I think unify data extension
> operation will be better.
>
> >> diff --git a/tools/lib/bpf/bpf_prog_linfo.c b/tools/lib/bpf/bpf_prog_linfo.c
> >> index 5c503096ef43..5cf41a563ef5 100644
> >> --- a/tools/lib/bpf/bpf_prog_linfo.c
> >> +++ b/tools/lib/bpf/bpf_prog_linfo.c
> >> @@ -127,7 +127,7 @@ struct bpf_prog_linfo *bpf_prog_linfo__new(const struct bpf_prog_info *info)
> >> prog_linfo->raw_linfo = malloc(data_sz);
> >> if (!prog_linfo->raw_linfo)
> >> goto err_free;
> >> - memcpy(prog_linfo->raw_linfo, (void *)(long)info->line_info, data_sz);
> >> + memcpy(prog_linfo->raw_linfo, (void *)(unsigned long)info->line_info, data_sz);
> >>
> >> nr_jited_func = info->nr_jited_ksyms;
> >> if (!nr_jited_func ||
> >> @@ -148,7 +148,7 @@ struct bpf_prog_linfo *bpf_prog_linfo__new(const struct bpf_prog_info *info)
> >> if (!prog_linfo->raw_jited_linfo)
> >> goto err_free;
> >> memcpy(prog_linfo->raw_jited_linfo,
> >> - (void *)(long)info->jited_line_info, data_sz);
> >> + (void *)(unsigned long)info->jited_line_info, data_sz);
> >>
> >> /* Number of jited_line_info per jited func */
> >> prog_linfo->nr_jited_linfo_per_func = malloc(nr_jited_func *
> >> @@ -166,8 +166,8 @@ struct bpf_prog_linfo *bpf_prog_linfo__new(const struct bpf_prog_info *info)
> >> goto err_free;
> >>
> >> if (dissect_jited_func(prog_linfo,
> >> - (__u64 *)(long)info->jited_ksyms,
> >> - (__u32 *)(long)info->jited_func_lens))
> >> + (__u64 *)(unsigned long)info->jited_ksyms,
> >> + (__u32 *)(unsigned long)info->jited_func_lens))
> >
> > so I'm trying to understand how this is changing anything for 32-bit
> > architecture and I must be missing something, sorry if I'm being
> > dense. The example you used below
> >
> > jited_kyms[0] = 0xb800067c, jited_linfo[0] = 0xffffffffb800067c
> >
> > Wouldn't (unsigned long)0xffffffffb800067c == (long)0xffffffffb800067c
> > == 0xb800067c ?
> If I understand correctly, info->jited_ksyms or info->jited_func_lens is
> just a u64 address that point to the corresponding space. The bpf_func
> address is stored in the item of info->jited_ksyms but not
> info->jited_ksyms.
>
> And here, I may have misled you. Both (__u64 *)(long)info->jited_ksyms
> and (__u64 *)(unsigned long)info->jited_ksyms are the same, I just want
> to unify the style. I will remove them in v2.
>
> Please let me know if there is any problem with my understanding.
>

Thanks for explanation. I guess in my mind I was always sign extending
32-bit to 64-bit, but I think memory addresses are conceptually
unsigned, so (unsigned long) casting makes more sense, and u64
representation of 0xb800067c should be 0x00000000b800067c and not
0xffffffffb800067c. So your changes make sense, and I agree that
libbpf-side changes for conceptual uniformity are also good.


> Thanks,
> Lehui
> >
> > isn't sizeof(long) == sizeof(void*) == 4?
> >
> > It would be nice if you could elaborate a bit more on what problems
> > did you see in practice?
> >
> >> goto err_free;
> >>
> >> return prog_linfo;
> >> diff --git a/tools/testing/selftests/bpf/prog_tests/btf.c b/tools/testing/selftests/bpf/prog_tests/btf.c
> >> index 84aae639ddb5..d9ba1ec1d5b3 100644
> >> --- a/tools/testing/selftests/bpf/prog_tests/btf.c
> >> +++ b/tools/testing/selftests/bpf/prog_tests/btf.c
> >> @@ -6451,8 +6451,8 @@ static int test_get_linfo(const struct prog_info_raw_test *test,
> >> info.nr_jited_line_info, jited_cnt,
> >> info.line_info_rec_size, rec_size,
> >> info.jited_line_info_rec_size, jited_rec_size,
> >> - (void *)(long)info.line_info,
> >> - (void *)(long)info.jited_line_info)) {
> >> + (void *)(unsigned long)info.line_info,
> >> + (void *)(unsigned long)info.jited_line_info)) {
> >> err = -1;
> >> goto done;
> >> }
> >> @@ -6500,8 +6500,8 @@ static int test_get_linfo(const struct prog_info_raw_test *test,
> >> }
> >>
> >> if (CHECK(jited_linfo[0] != jited_ksyms[0],
> >> - "jited_linfo[0]:%lx != jited_ksyms[0]:%lx",
> >> - (long)(jited_linfo[0]), (long)(jited_ksyms[0]))) {
> >> + "jited_linfo[0]:%llx != jited_ksyms[0]:%llx",
> >> + jited_linfo[0], jited_ksyms[0])) {
> >> err = -1;
> >> goto done;
> >> }
> >> @@ -6519,16 +6519,16 @@ static int test_get_linfo(const struct prog_info_raw_test *test,
> >> }
> >>
> >> if (CHECK(jited_linfo[i] <= jited_linfo[i - 1],
> >> - "jited_linfo[%u]:%lx <= jited_linfo[%u]:%lx",
> >> - i, (long)jited_linfo[i],
> >> - i - 1, (long)(jited_linfo[i - 1]))) {
> >> + "jited_linfo[%u]:%llx <= jited_linfo[%u]:%llx",
> >> + i, jited_linfo[i],
> >> + i - 1, (jited_linfo[i - 1]))) {
> >> err = -1;
> >> goto done;
> >> }
> >>
> >> if (CHECK(jited_linfo[i] - cur_func_ksyms > cur_func_len,
> >> - "jited_linfo[%u]:%lx - %lx > %u",
> >> - i, (long)jited_linfo[i], (long)cur_func_ksyms,
> >> + "jited_linfo[%u]:%llx - %llx > %u",
> >> + i, jited_linfo[i], cur_func_ksyms,
> >> cur_func_len)) {
> >> err = -1;
> >> goto done;
> >> --
> >> 2.25.1
> >>
> > .
> >

\
 
 \ /
  Last update: 2022-05-06 23:02    [W:0.094 / U:2.016 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site