From: Jiri Olsa <hidden> Date: 2021-10-23 12:05:00
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
I found code in dedup that seems to handle such situation for arrays,
and added 'some' fix for structs. With that change I can no longer
see vmlinux's structs in kernel module BTF data, but I have no idea
if that breaks anything else.
thoughts? thanks,
jirka
---
Jiri Olsa (2):
kbuild: Unify options for BTF generation for vmlinux and modules
bpf: Add support to detect and dedup instances of same structs
Makefile | 3 +++
scripts/Makefile.modfinal | 2 +-
scripts/link-vmlinux.sh | 11 +----------
scripts/pahole-flags.sh | 20 ++++++++++++++++++++
tools/lib/bpf/btf.c | 12 ++++++++++--
5 files changed, 35 insertions(+), 13 deletions(-)
create mode 100755 scripts/pahole-flags.sh
From: Jiri Olsa <hidden> Date: 2021-10-23 12:05:06
Using new PAHOLE_FLAGS variable to pass extra arguments to
pahole for both vmlinux and modules BTF data generation.
Adding new scripts/pahole-flags.sh script that detect and
prints pahole options.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
Makefile | 3 +++
scripts/Makefile.modfinal | 2 +-
scripts/link-vmlinux.sh | 11 +----------
scripts/pahole-flags.sh | 20 ++++++++++++++++++++
4 files changed, 25 insertions(+), 11 deletions(-)
create mode 100755 scripts/pahole-flags.sh
@@ -40,7 +40,7 @@ quiet_cmd_ld_ko_o = LD [M] $@quiet_cmd_btf_ko=BTF[M]$@cmd_btf_ko=\ if [ -f vmlinux ]; then \-LLVM_OBJCOPY="$(OBJCOPY)"$(PAHOLE)-J--btf_basevmlinux$@;\+LLVM_OBJCOPY="$(OBJCOPY)"$(PAHOLE)-J$(PAHOLE_FLAGS)--btf_basevmlinux$@;\$(RESOLVE_BTFIDS)-bvmlinux$@;\ else \printf"Skipping BTF generation for %s due to unavailability of vmlinux\n"$@1>&2;\
@@ -205,7 +205,6 @@ vmlinux_link() gen_btf(){localpahole_ver-localextra_paholeopt=if![-x"$(command-v${PAHOLE})"];thenecho>&2"BTF: ${1}: pahole (${PAHOLE}) is not available"
@@ -220,16 +219,8 @@ gen_btf()vmlinux_link${1}-if["${pahole_ver}"-ge"118"]&&["${pahole_ver}"-le"121"];then-# pahole 1.18 through 1.21 can't handle zero-sized per-CPU vars-extra_paholeopt="${extra_paholeopt} --skip_encoding_btf_vars"-fi-if["${pahole_ver}"-ge"121"];then-extra_paholeopt="${extra_paholeopt} --btf_gen_floats"-fi-info"BTF"${2}-LLVM_OBJCOPY="${OBJCOPY}"${PAHOLE}-J${extra_paholeopt}${1}+LLVM_OBJCOPY="${OBJCOPY}"${PAHOLE}-J${PAHOLE_FLAGS}${1}# Create ${2} which contains just .BTF section but no symbols. Add# SHF_ALLOC because .BTF will be part of the vmlinux image. --strip-all
From: Jiri Olsa <hidden> Date: 2021-10-23 12:05:16
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I found code in dedup that seems to handle such situation for arrays,
and added btf_dedup_is_equiv call for structs.
With this change I can no longer see vmlinux's structs in kernel
module BTF data, but I have no idea if that breaks anything else.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
tools/lib/bpf/btf.c | 12 ++++++++++--
1 file changed, 10 insertions(+), 2 deletions(-)
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
I found code in dedup that seems to handle such situation for arrays,
and added 'some' fix for structs. With that change I can no longer
see vmlinux's structs in kernel module BTF data, but I have no idea
if that breaks anything else.
thoughts? thanks,
jirka
---
Jiri Olsa (2):
kbuild: Unify options for BTF generation for vmlinux and modules
bpf: Add support to detect and dedup instances of same structs
Makefile | 3 +++
scripts/Makefile.modfinal | 2 +-
scripts/link-vmlinux.sh | 11 +----------
scripts/pahole-flags.sh | 20 ++++++++++++++++++++
tools/lib/bpf/btf.c | 12 ++++++++++--
5 files changed, 35 insertions(+), 13 deletions(-)
create mode 100755 scripts/pahole-flags.sh
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
Using new PAHOLE_FLAGS variable to pass extra arguments to
pahole for both vmlinux and modules BTF data generation.
Adding new scripts/pahole-flags.sh script that detect and
prints pahole options.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
LGTM. I suggest posting it separately from the BTF dedup hack.
Acked-by: Andrii Nakryiko <andrii@kernel.org>
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I found code in dedup that seems to handle such situation for arrays,
and added btf_dedup_is_equiv call for structs.
With this change I can no longer see vmlinux's structs in kernel
module BTF data, but I have no idea if that breaks anything else.
I'm pretty sure this is not the right way to handle this. Let's figure
out what is causing a difference in types (i.e., why they are not
equivalent), before attempting to patch up BTF dedup algorithm.
From: Jiri Olsa <hidden> Date: 2021-10-26 12:03:50
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
jirka
quoted
I found code in dedup that seems to handle such situation for arrays,
and added 'some' fix for structs. With that change I can no longer
see vmlinux's structs in kernel module BTF data, but I have no idea
if that breaks anything else.
thoughts? thanks,
jirka
---
Jiri Olsa (2):
kbuild: Unify options for BTF generation for vmlinux and modules
bpf: Add support to detect and dedup instances of same structs
Makefile | 3 +++
scripts/Makefile.modfinal | 2 +-
scripts/link-vmlinux.sh | 11 +----------
scripts/pahole-flags.sh | 20 ++++++++++++++++++++
tools/lib/bpf/btf.c | 12 ++++++++++--
5 files changed, 35 insertions(+), 13 deletions(-)
create mode 100755 scripts/pahole-flags.sh
From: Jiri Olsa <hidden> Date: 2021-10-26 12:04:04
On Mon, Oct 25, 2021 at 09:56:37PM -0700, Andrii Nakryiko wrote:
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
Using new PAHOLE_FLAGS variable to pass extra arguments to
pahole for both vmlinux and modules BTF data generation.
Adding new scripts/pahole-flags.sh script that detect and
prints pahole options.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
LGTM. I suggest posting it separately from the BTF dedup hack.
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
jirka
quoted
quoted
I found code in dedup that seems to handle such situation for arrays,
and added 'some' fix for structs. With that change I can no longer
see vmlinux's structs in kernel module BTF data, but I have no idea
if that breaks anything else.
thoughts? thanks,
jirka
---
Jiri Olsa (2):
kbuild: Unify options for BTF generation for vmlinux and modules
bpf: Add support to detect and dedup instances of same structs
Makefile | 3 +++
scripts/Makefile.modfinal | 2 +-
scripts/link-vmlinux.sh | 11 +----------
scripts/pahole-flags.sh | 20 ++++++++++++++++++++
tools/lib/bpf/btf.c | 12 ++++++++++--
5 files changed, 35 insertions(+), 13 deletions(-)
create mode 100755 scripts/pahole-flags.sh
From: Jiri Olsa <hidden> Date: 2021-10-27 08:53:37
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
thanks,
jirka
quoted
jirka
quoted
quoted
I found code in dedup that seems to handle such situation for arrays,
and added 'some' fix for structs. With that change I can no longer
see vmlinux's structs in kernel module BTF data, but I have no idea
if that breaks anything else.
thoughts? thanks,
jirka
---
Jiri Olsa (2):
kbuild: Unify options for BTF generation for vmlinux and modules
bpf: Add support to detect and dedup instances of same structs
Makefile | 3 +++
scripts/Makefile.modfinal | 2 +-
scripts/link-vmlinux.sh | 11 +----------
scripts/pahole-flags.sh | 20 ++++++++++++++++++++
tools/lib/bpf/btf.c | 12 ++++++++++--
5 files changed, 35 insertions(+), 13 deletions(-)
create mode 100755 scripts/pahole-flags.sh
On Wed, Oct 27, 2021 at 1:53 AM Jiri Olsa [off-list ref] wrote:
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
quoted
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
Ok, now struct module is identical down to the types referenced from
the fields, which means it should have been deduplicated completely.
This will require a more time-consuming debugging, though, so I'll put
it on my TODO list for now. If you get to this earlier, see where the
equivalence check fails in btf_dedup (sprinkle debug outputs around to
see what's going on).
thanks,
jirka
quoted
quoted
jirka
quoted
quoted
I found code in dedup that seems to handle such situation for arrays,
and added 'some' fix for structs. With that change I can no longer
see vmlinux's structs in kernel module BTF data, but I have no idea
if that breaks anything else.
thoughts? thanks,
jirka
---
Jiri Olsa (2):
kbuild: Unify options for BTF generation for vmlinux and modules
bpf: Add support to detect and dedup instances of same structs
Makefile | 3 +++
scripts/Makefile.modfinal | 2 +-
scripts/link-vmlinux.sh | 11 +----------
scripts/pahole-flags.sh | 20 ++++++++++++++++++++
tools/lib/bpf/btf.c | 12 ++++++++++--
5 files changed, 35 insertions(+), 13 deletions(-)
create mode 100755 scripts/pahole-flags.sh
From: Jiri Olsa <hidden> Date: 2021-10-27 18:18:23
On Wed, Oct 27, 2021 at 10:53:55AM -0700, Andrii Nakryiko wrote:
On Wed, Oct 27, 2021 at 1:53 AM Jiri Olsa [off-list ref] wrote:
quoted
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
quoted
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
Ok, now struct module is identical down to the types referenced from
the fields, which means it should have been deduplicated completely.
This will require a more time-consuming debugging, though, so I'll put
it on my TODO list for now. If you get to this earlier, see where the
equivalence check fails in btf_dedup (sprinkle debug outputs around to
see what's going on).
it failed for me on that hypot_type_id check where I did fix,
I thought it's the issue of multiple same struct in the kmod,
but now I see I might have confused cannon_id with cand_id ;-)
I'll check more on this
jirka
On Wed, Oct 27, 2021 at 11:23:55PM IST, Andrii Nakryiko wrote:
On Wed, Oct 27, 2021 at 1:53 AM Jiri Olsa [off-list ref] wrote:
quoted
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
quoted
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
Ok, now struct module is identical down to the types referenced from
the fields, which means it should have been deduplicated completely.
This will require a more time-consuming debugging, though, so I'll put
it on my TODO list for now. If you get to this earlier, see where the
equivalence check fails in btf_dedup (sprinkle debug outputs around to
see what's going on).
Hello Andrii,
I think I'm seeing something similar when working on the conntrack patches [0],
I was looking to match whether the type in a PTR_TO_BTF_ID register is same as
struct nf_conn, but it seems that there can be two BTF IDs for the same struct
type.
When doing bpftool dump, I see:
; bpftool btf dump file /sys/kernel/btf/vmlinux format raw | grep nf_conn
...
[89224] STRUCT 'nf_conn' size=256 vlen=15
...
; bpftool btf dump file /sys/kernel/btf/nf_conntrack format raw | grep nf_conn
...
[103077] STRUCT 'nf_conn' size=256 vlen=15
[104988] STRUCT 'nf_conn' size=256 vlen=15
[106490] STRUCT 'nf_conn' size=256 vlen=15
[108187] STRUCT 'nf_conn' size=256 vlen=15
...
Inside the kernel, when trying to match both, register PTR_TO_BTF_ID refers to
the nf_conntrack BTF ID, while the BTF_ID_LIST resolves to the one in vmlinux,
this ends up making the job of matching the two struct types a bit difficult
(for now, I am thinking of going with btf_struct_ids_match). My original plan
was to compare the result of btf_types_by_id.
[0]: https://github.com/kkdwivedi/linux/commits/conntrack-bpf
quoted
thanks,
jirka
quoted
quoted
jirka
quoted
quoted
I found code in dedup that seems to handle such situation for arrays,
and added 'some' fix for structs. With that change I can no longer
see vmlinux's structs in kernel module BTF data, but I have no idea
if that breaks anything else.
thoughts? thanks,
jirka
---
Jiri Olsa (2):
kbuild: Unify options for BTF generation for vmlinux and modules
bpf: Add support to detect and dedup instances of same structs
Makefile | 3 +++
scripts/Makefile.modfinal | 2 +-
scripts/link-vmlinux.sh | 11 +----------
scripts/pahole-flags.sh | 20 ++++++++++++++++++++
tools/lib/bpf/btf.c | 12 ++++++++++--
5 files changed, 35 insertions(+), 13 deletions(-)
create mode 100755 scripts/pahole-flags.sh
From: Jiri Olsa <hidden> Date: 2021-10-28 19:12:59
On Wed, Oct 27, 2021 at 08:18:11PM +0200, Jiri Olsa wrote:
On Wed, Oct 27, 2021 at 10:53:55AM -0700, Andrii Nakryiko wrote:
quoted
On Wed, Oct 27, 2021 at 1:53 AM Jiri Olsa [off-list ref] wrote:
quoted
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
quoted
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
Ok, now struct module is identical down to the types referenced from
the fields, which means it should have been deduplicated completely.
This will require a more time-consuming debugging, though, so I'll put
it on my TODO list for now. If you get to this earlier, see where the
equivalence check fails in btf_dedup (sprinkle debug outputs around to
see what's going on).
it failed for me on that hypot_type_id check where I did fix,
I thought it's the issue of multiple same struct in the kmod,
but now I see I might have confused cannon_id with cand_id ;-)
I'll check more on this
with more checking I got to the same conclusion as before,
now maybe with little more details ;-)
the problem seems to be that in some cases the module BTF
data stores same structs under new/different IDs, while the
kernel BTF data is already dedup-ed
the dedup algo keeps hypot_map of kernel IDs to kmod IDs,
and in my case it will get to the point that the kernel ID
is already 'known' and points to certain kmod ID 'A', but it
is also equiv to another kmod ID 'B' (so kmod ID 'A' and 'B'
are equiv structs) but the dedup will claim as not equiv
This is where the dedup fails for me on that s390 data:
The pt_regs is defined as:
struct pt_regs
{
union {
user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
...
};
considering just the first union:
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
---------------------------------------------------------------
Comparing the first member 'user_regs':
struct pt_regs
{
union {
---> user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
Which looks like:
typedef struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
} user_pt_regs;
and is also equiv to the next union member struct.. and that's what
kernel knows but not kmod... anyway,
the dedup will compare 'user_pt_regs':
[183] TYPEDEF 'user_pt_regs' type_id=181
[91247] TYPEDEF 'user_pt_regs' type_id=91245
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91245] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
and make them equiv by setting hypot_type_id for 181 to be 91245
---------------------------------------------------------------
Now comparing the second member:
struct pt_regs
{
union {
user_pt_regs user_regs;
---> struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
kernel knows it's same struct as user_pt_regs and uses ID 181
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
but kmod has new ID 91250 (not 91245):
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
and 181 and 91250 are equiv structs:
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91250] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
now hypot_type_id for 181 is 91245, but we have brand new struct
ID 91250, so we fail
what the patch tries to do is at this point to compare ID 91250
with 91245 and if it passes then we are equal and we throw away
ID 91250 because the hypot_type_id for 181 stays 91245
ufff.. thoughts? ;-)
thanks,
jirka
On Thu, Oct 28, 2021 at 12:12 PM Jiri Olsa [off-list ref] wrote:
On Wed, Oct 27, 2021 at 08:18:11PM +0200, Jiri Olsa wrote:
quoted
On Wed, Oct 27, 2021 at 10:53:55AM -0700, Andrii Nakryiko wrote:
quoted
On Wed, Oct 27, 2021 at 1:53 AM Jiri Olsa [off-list ref] wrote:
quoted
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
quoted
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
Ok, now struct module is identical down to the types referenced from
the fields, which means it should have been deduplicated completely.
This will require a more time-consuming debugging, though, so I'll put
it on my TODO list for now. If you get to this earlier, see where the
equivalence check fails in btf_dedup (sprinkle debug outputs around to
see what's going on).
it failed for me on that hypot_type_id check where I did fix,
I thought it's the issue of multiple same struct in the kmod,
but now I see I might have confused cannon_id with cand_id ;-)
I'll check more on this
with more checking I got to the same conclusion as before,
now maybe with little more details ;-)
the problem seems to be that in some cases the module BTF
data stores same structs under new/different IDs, while the
kernel BTF data is already dedup-ed
the dedup algo keeps hypot_map of kernel IDs to kmod IDs,
and in my case it will get to the point that the kernel ID
is already 'known' and points to certain kmod ID 'A', but it
is also equiv to another kmod ID 'B' (so kmod ID 'A' and 'B'
are equiv structs) but the dedup will claim as not equiv
This is where the dedup fails for me on that s390 data:
The pt_regs is defined as:
struct pt_regs
{
union {
user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
...
};
considering just the first union:
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
---------------------------------------------------------------
Comparing the first member 'user_regs':
struct pt_regs
{
union {
---> user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
Which looks like:
typedef struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
} user_pt_regs;
and is also equiv to the next union member struct.. and that's what
kernel knows but not kmod... anyway,
the dedup will compare 'user_pt_regs':
[183] TYPEDEF 'user_pt_regs' type_id=181
[91247] TYPEDEF 'user_pt_regs' type_id=91245
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91245] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
and make them equiv by setting hypot_type_id for 181 to be 91245
---------------------------------------------------------------
Now comparing the second member:
struct pt_regs
{
union {
user_pt_regs user_regs;
---> struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
kernel knows it's same struct as user_pt_regs and uses ID 181
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
but kmod has new ID 91250 (not 91245):
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
and 181 and 91250 are equiv structs:
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91250] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
now hypot_type_id for 181 is 91245, but we have brand new struct
ID 91250, so we fail
what the patch tries to do is at this point to compare ID 91250
with 91245 and if it passes then we are equal and we throw away
ID 91250 because the hypot_type_id for 181 stays 91245
ufff.. thoughts? ;-)
Oh, this is a really great analysis, thanks a lot! It makes everything
clear. Basically, BTF dedup algo does too good job deduping vmlinux
BTF. :)
What's not clear is what to do about that, because a (current)
fundamental assumption of is_equiv() check is that any type within CU
(or in this case deduped vmlinux BTF) has exactly one unique mapping.
Clearly that's not the case now. That array fix you mentioned worked
around GCC bug where this assumption broke. In this case it's not a
bug of a compiler (neither of algo, really), we just need to make algo
smarter.
Let me think about this a bit, we'll need to make the equivalence
check be aware that there could be multiple equivalent mappings and be
ok with that as long as all candidates are equivalent between
themselves. Lots of equivalence and recursion to think about.
It would be great to have a simplified test case to play with that. Do
you mind distilling the chain of types above into a selftests and
posting it to the mailing list so that I can play with it? It
shouldn't be hard to write given BTF writing APIs. And we'll need a
selftests anyway once we improve the algo, so it's definitely not a
wasted work.
And thanks again for analysis and writing this down, it would take me
ages to get to this otherwise.
P.S. If the improved BTF dedup algo will be able to handle this, we
should also remove the array workaround, because that one should work
automatically. I don't know if we have a test for duplicate array
scenario, but it's probably good to have that as well.
From: Jiri Olsa <hidden> Date: 2021-11-02 14:14:56
On Mon, Nov 01, 2021 at 04:14:29PM -0700, Andrii Nakryiko wrote:
On Thu, Oct 28, 2021 at 12:12 PM Jiri Olsa [off-list ref] wrote:
quoted
On Wed, Oct 27, 2021 at 08:18:11PM +0200, Jiri Olsa wrote:
quoted
On Wed, Oct 27, 2021 at 10:53:55AM -0700, Andrii Nakryiko wrote:
quoted
On Wed, Oct 27, 2021 at 1:53 AM Jiri Olsa [off-list ref] wrote:
quoted
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
quoted
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
Ok, now struct module is identical down to the types referenced from
the fields, which means it should have been deduplicated completely.
This will require a more time-consuming debugging, though, so I'll put
it on my TODO list for now. If you get to this earlier, see where the
equivalence check fails in btf_dedup (sprinkle debug outputs around to
see what's going on).
it failed for me on that hypot_type_id check where I did fix,
I thought it's the issue of multiple same struct in the kmod,
but now I see I might have confused cannon_id with cand_id ;-)
I'll check more on this
with more checking I got to the same conclusion as before,
now maybe with little more details ;-)
the problem seems to be that in some cases the module BTF
data stores same structs under new/different IDs, while the
kernel BTF data is already dedup-ed
the dedup algo keeps hypot_map of kernel IDs to kmod IDs,
and in my case it will get to the point that the kernel ID
is already 'known' and points to certain kmod ID 'A', but it
is also equiv to another kmod ID 'B' (so kmod ID 'A' and 'B'
are equiv structs) but the dedup will claim as not equiv
This is where the dedup fails for me on that s390 data:
The pt_regs is defined as:
struct pt_regs
{
union {
user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
...
};
considering just the first union:
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
---------------------------------------------------------------
Comparing the first member 'user_regs':
struct pt_regs
{
union {
---> user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
Which looks like:
typedef struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
} user_pt_regs;
and is also equiv to the next union member struct.. and that's what
kernel knows but not kmod... anyway,
the dedup will compare 'user_pt_regs':
[183] TYPEDEF 'user_pt_regs' type_id=181
[91247] TYPEDEF 'user_pt_regs' type_id=91245
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91245] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
and make them equiv by setting hypot_type_id for 181 to be 91245
---------------------------------------------------------------
Now comparing the second member:
struct pt_regs
{
union {
user_pt_regs user_regs;
---> struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
kernel knows it's same struct as user_pt_regs and uses ID 181
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
but kmod has new ID 91250 (not 91245):
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
and 181 and 91250 are equiv structs:
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91250] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
now hypot_type_id for 181 is 91245, but we have brand new struct
ID 91250, so we fail
what the patch tries to do is at this point to compare ID 91250
with 91245 and if it passes then we are equal and we throw away
ID 91250 because the hypot_type_id for 181 stays 91245
ufff.. thoughts? ;-)
Oh, this is a really great analysis, thanks a lot! It makes everything
clear. Basically, BTF dedup algo does too good job deduping vmlinux
BTF. :)
What's not clear is what to do about that, because a (current)
fundamental assumption of is_equiv() check is that any type within CU
(or in this case deduped vmlinux BTF) has exactly one unique mapping.
Clearly that's not the case now. That array fix you mentioned worked
around GCC bug where this assumption broke. In this case it's not a
bug of a compiler (neither of algo, really), we just need to make algo
smarter.
Let me think about this a bit, we'll need to make the equivalence
check be aware that there could be multiple equivalent mappings and be
ok with that as long as all candidates are equivalent between
themselves. Lots of equivalence and recursion to think about.
It would be great to have a simplified test case to play with that. Do
you mind distilling the chain of types above into a selftests and
posting it to the mailing list so that I can play with it? It
shouldn't be hard to write given BTF writing APIs. And we'll need a
selftests anyway once we improve the algo, so it's definitely not a
wasted work.
And thanks again for analysis and writing this down, it would take me
ages to get to this otherwise.
P.S. If the improved BTF dedup algo will be able to handle this, we
should also remove the array workaround, because that one should work
automatically. I don't know if we have a test for duplicate array
scenario, but it's probably good to have that as well.
right, I'll try to add test for both
thanks for checking on this
jirka
From: Jiri Olsa <hidden> Date: 2021-11-07 14:58:00
On Tue, Nov 02, 2021 at 03:14:52PM +0100, Jiri Olsa wrote:
On Mon, Nov 01, 2021 at 04:14:29PM -0700, Andrii Nakryiko wrote:
quoted
On Thu, Oct 28, 2021 at 12:12 PM Jiri Olsa [off-list ref] wrote:
quoted
On Wed, Oct 27, 2021 at 08:18:11PM +0200, Jiri Olsa wrote:
quoted
On Wed, Oct 27, 2021 at 10:53:55AM -0700, Andrii Nakryiko wrote:
quoted
On Wed, Oct 27, 2021 at 1:53 AM Jiri Olsa [off-list ref] wrote:
quoted
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
quoted
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
Ok, now struct module is identical down to the types referenced from
the fields, which means it should have been deduplicated completely.
This will require a more time-consuming debugging, though, so I'll put
it on my TODO list for now. If you get to this earlier, see where the
equivalence check fails in btf_dedup (sprinkle debug outputs around to
see what's going on).
it failed for me on that hypot_type_id check where I did fix,
I thought it's the issue of multiple same struct in the kmod,
but now I see I might have confused cannon_id with cand_id ;-)
I'll check more on this
with more checking I got to the same conclusion as before,
now maybe with little more details ;-)
the problem seems to be that in some cases the module BTF
data stores same structs under new/different IDs, while the
kernel BTF data is already dedup-ed
the dedup algo keeps hypot_map of kernel IDs to kmod IDs,
and in my case it will get to the point that the kernel ID
is already 'known' and points to certain kmod ID 'A', but it
is also equiv to another kmod ID 'B' (so kmod ID 'A' and 'B'
are equiv structs) but the dedup will claim as not equiv
This is where the dedup fails for me on that s390 data:
The pt_regs is defined as:
struct pt_regs
{
union {
user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
...
};
considering just the first union:
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
---------------------------------------------------------------
Comparing the first member 'user_regs':
struct pt_regs
{
union {
---> user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
Which looks like:
typedef struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
} user_pt_regs;
and is also equiv to the next union member struct.. and that's what
kernel knows but not kmod... anyway,
the dedup will compare 'user_pt_regs':
[183] TYPEDEF 'user_pt_regs' type_id=181
[91247] TYPEDEF 'user_pt_regs' type_id=91245
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91245] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
and make them equiv by setting hypot_type_id for 181 to be 91245
---------------------------------------------------------------
Now comparing the second member:
struct pt_regs
{
union {
user_pt_regs user_regs;
---> struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
kernel knows it's same struct as user_pt_regs and uses ID 181
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
but kmod has new ID 91250 (not 91245):
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
and 181 and 91250 are equiv structs:
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91250] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
now hypot_type_id for 181 is 91245, but we have brand new struct
ID 91250, so we fail
what the patch tries to do is at this point to compare ID 91250
with 91245 and if it passes then we are equal and we throw away
ID 91250 because the hypot_type_id for 181 stays 91245
ufff.. thoughts? ;-)
Oh, this is a really great analysis, thanks a lot! It makes everything
clear. Basically, BTF dedup algo does too good job deduping vmlinux
BTF. :)
What's not clear is what to do about that, because a (current)
fundamental assumption of is_equiv() check is that any type within CU
(or in this case deduped vmlinux BTF) has exactly one unique mapping.
Clearly that's not the case now. That array fix you mentioned worked
around GCC bug where this assumption broke. In this case it's not a
bug of a compiler (neither of algo, really), we just need to make algo
smarter.
Let me think about this a bit, we'll need to make the equivalence
check be aware that there could be multiple equivalent mappings and be
ok with that as long as all candidates are equivalent between
themselves. Lots of equivalence and recursion to think about.
It would be great to have a simplified test case to play with that. Do
you mind distilling the chain of types above into a selftests and
posting it to the mailing list so that I can play with it? It
shouldn't be hard to write given BTF writing APIs. And we'll need a
selftests anyway once we improve the algo, so it's definitely not a
wasted work.
I ended up with simply test, where the idea is to use
type id which is defined after currently processed type
the last VALIDATE_RAW_BTF fails
I'm not sending full atch, because I assume this is not
to merge yet also I assume you might want to change that
anyway ;-)
I'll check later on that special array case
thanks,
jirka
---
.../bpf/prog_tests/btf_dedup_split.c | 113 ++++++++++++++++++
1 file changed, 113 insertions(+)
On Sun, Nov 7, 2021 at 6:57 AM Jiri Olsa [off-list ref] wrote:
On Tue, Nov 02, 2021 at 03:14:52PM +0100, Jiri Olsa wrote:
quoted
On Mon, Nov 01, 2021 at 04:14:29PM -0700, Andrii Nakryiko wrote:
quoted
On Thu, Oct 28, 2021 at 12:12 PM Jiri Olsa [off-list ref] wrote:
quoted
On Wed, Oct 27, 2021 at 08:18:11PM +0200, Jiri Olsa wrote:
quoted
On Wed, Oct 27, 2021 at 10:53:55AM -0700, Andrii Nakryiko wrote:
quoted
On Wed, Oct 27, 2021 at 1:53 AM Jiri Olsa [off-list ref] wrote:
quoted
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
quoted
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
Ok, now struct module is identical down to the types referenced from
the fields, which means it should have been deduplicated completely.
This will require a more time-consuming debugging, though, so I'll put
it on my TODO list for now. If you get to this earlier, see where the
equivalence check fails in btf_dedup (sprinkle debug outputs around to
see what's going on).
it failed for me on that hypot_type_id check where I did fix,
I thought it's the issue of multiple same struct in the kmod,
but now I see I might have confused cannon_id with cand_id ;-)
I'll check more on this
with more checking I got to the same conclusion as before,
now maybe with little more details ;-)
the problem seems to be that in some cases the module BTF
data stores same structs under new/different IDs, while the
kernel BTF data is already dedup-ed
the dedup algo keeps hypot_map of kernel IDs to kmod IDs,
and in my case it will get to the point that the kernel ID
is already 'known' and points to certain kmod ID 'A', but it
is also equiv to another kmod ID 'B' (so kmod ID 'A' and 'B'
are equiv structs) but the dedup will claim as not equiv
This is where the dedup fails for me on that s390 data:
The pt_regs is defined as:
struct pt_regs
{
union {
user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
...
};
considering just the first union:
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
---------------------------------------------------------------
Comparing the first member 'user_regs':
struct pt_regs
{
union {
---> user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
Which looks like:
typedef struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
} user_pt_regs;
and is also equiv to the next union member struct.. and that's what
kernel knows but not kmod... anyway,
the dedup will compare 'user_pt_regs':
[183] TYPEDEF 'user_pt_regs' type_id=181
[91247] TYPEDEF 'user_pt_regs' type_id=91245
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91245] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
and make them equiv by setting hypot_type_id for 181 to be 91245
---------------------------------------------------------------
Now comparing the second member:
struct pt_regs
{
union {
user_pt_regs user_regs;
---> struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
kernel knows it's same struct as user_pt_regs and uses ID 181
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
but kmod has new ID 91250 (not 91245):
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
and 181 and 91250 are equiv structs:
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91250] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
now hypot_type_id for 181 is 91245, but we have brand new struct
ID 91250, so we fail
what the patch tries to do is at this point to compare ID 91250
with 91245 and if it passes then we are equal and we throw away
ID 91250 because the hypot_type_id for 181 stays 91245
ufff.. thoughts? ;-)
Oh, this is a really great analysis, thanks a lot! It makes everything
clear. Basically, BTF dedup algo does too good job deduping vmlinux
BTF. :)
What's not clear is what to do about that, because a (current)
fundamental assumption of is_equiv() check is that any type within CU
(or in this case deduped vmlinux BTF) has exactly one unique mapping.
Clearly that's not the case now. That array fix you mentioned worked
around GCC bug where this assumption broke. In this case it's not a
bug of a compiler (neither of algo, really), we just need to make algo
smarter.
Let me think about this a bit, we'll need to make the equivalence
check be aware that there could be multiple equivalent mappings and be
ok with that as long as all candidates are equivalent between
themselves. Lots of equivalence and recursion to think about.
It would be great to have a simplified test case to play with that. Do
you mind distilling the chain of types above into a selftests and
posting it to the mailing list so that I can play with it? It
shouldn't be hard to write given BTF writing APIs. And we'll need a
selftests anyway once we improve the algo, so it's definitely not a
wasted work.
I ended up with simply test, where the idea is to use
type id which is defined after currently processed type
the last VALIDATE_RAW_BTF fails
I'm not sending full atch, because I assume this is not
to merge yet also I assume you might want to change that
anyway ;-)
Thanks, Jiri! I'll get to playing with this some time this week,
hopefully. I hope this is not a huge blocker for you?
quoted hunk
I'll check later on that special array case
thanks,
jirka
---
.../bpf/prog_tests/btf_dedup_split.c | 113 ++++++++++++++++++
1 file changed, 113 insertions(+)
From: Jiri Olsa <hidden> Date: 2021-11-12 07:23:48
On Tue, Nov 09, 2021 at 03:23:25PM -0800, Andrii Nakryiko wrote:
On Sun, Nov 7, 2021 at 6:57 AM Jiri Olsa [off-list ref] wrote:
quoted
On Tue, Nov 02, 2021 at 03:14:52PM +0100, Jiri Olsa wrote:
quoted
On Mon, Nov 01, 2021 at 04:14:29PM -0700, Andrii Nakryiko wrote:
quoted
On Thu, Oct 28, 2021 at 12:12 PM Jiri Olsa [off-list ref] wrote:
quoted
On Wed, Oct 27, 2021 at 08:18:11PM +0200, Jiri Olsa wrote:
quoted
On Wed, Oct 27, 2021 at 10:53:55AM -0700, Andrii Nakryiko wrote:
quoted
On Wed, Oct 27, 2021 at 1:53 AM Jiri Olsa [off-list ref] wrote:
quoted
On Tue, Oct 26, 2021 at 09:12:31PM -0700, Andrii Nakryiko wrote:
quoted
On Tue, Oct 26, 2021 at 5:03 AM Jiri Olsa [off-list ref] wrote:
quoted
On Mon, Oct 25, 2021 at 09:54:48PM -0700, Andrii Nakryiko wrote:
quoted
On Sat, Oct 23, 2021 at 5:05 AM Jiri Olsa [off-list ref] wrote:
quoted
hi,
I'm trying to enable BTF for kernel module in fedora,
and I'm getting big increase on modules sizes on s390x arch.
Size of modules in total - kernel dir under /lib/modules/VER/
from kernel-core and kernel-module packages:
current new
aarch64 60M 76M
ppc64le 53M 66M
s390x 21M 41M
x86_64 64M 79M
The reason for higher increase on s390x was that dedup algorithm
did not detect some of the big kernel structs like 'struct module',
so they are duplicated in the kernel module BTF data. The s390x
has many small modules that increased significantly in size because
of that even after compression.
First issues was that the '--btf_gen_floats' option is not passed
to pahole for kernel module BTF generation.
The other problem is more tricky and is the reason why this patchset
is RFC ;-)
The s390x compiler generates multiple definitions of the same struct
and dedup algorithm does not seem to handle this at the moment.
I put the debuginfo and btf dump of the s390x pnet.ko module in here:
http://people.redhat.com/~jolsa/kmodbtf/
Please let me know if you'd like to see other info/files.
Hard to tell what's going on without vmlinux itself. Can you upload a
corresponding kernel image with BTF in it?
sure, uploaded
vmlinux.btfdump:
[174] FLOAT 'float' size=4
[175] FLOAT 'double' size=8
VS
pnet.btfdump:
[89318] INT 'float' size=4 bits_offset=0 nr_bits=32 encoding=(none)
[89319] INT 'double' size=8 bits_offset=0 nr_bits=64 encoding=(none)
ugh, that's with no fix applied, sry
I applied the first patch and uploaded new files
now when I compare the 'module' struct from vmlinux:
[885] STRUCT 'module' size=1280 vlen=70
and same one from pnet.ko:
[89323] STRUCT 'module' size=1280 vlen=70
they seem to completely match, all the fields
and yet it still appears in the kmod's BTF
Ok, now struct module is identical down to the types referenced from
the fields, which means it should have been deduplicated completely.
This will require a more time-consuming debugging, though, so I'll put
it on my TODO list for now. If you get to this earlier, see where the
equivalence check fails in btf_dedup (sprinkle debug outputs around to
see what's going on).
it failed for me on that hypot_type_id check where I did fix,
I thought it's the issue of multiple same struct in the kmod,
but now I see I might have confused cannon_id with cand_id ;-)
I'll check more on this
with more checking I got to the same conclusion as before,
now maybe with little more details ;-)
the problem seems to be that in some cases the module BTF
data stores same structs under new/different IDs, while the
kernel BTF data is already dedup-ed
the dedup algo keeps hypot_map of kernel IDs to kmod IDs,
and in my case it will get to the point that the kernel ID
is already 'known' and points to certain kmod ID 'A', but it
is also equiv to another kmod ID 'B' (so kmod ID 'A' and 'B'
are equiv structs) but the dedup will claim as not equiv
This is where the dedup fails for me on that s390 data:
The pt_regs is defined as:
struct pt_regs
{
union {
user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
...
};
considering just the first union:
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
---------------------------------------------------------------
Comparing the first member 'user_regs':
struct pt_regs
{
union {
---> user_pt_regs user_regs;
struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
Which looks like:
typedef struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
} user_pt_regs;
and is also equiv to the next union member struct.. and that's what
kernel knows but not kmod... anyway,
the dedup will compare 'user_pt_regs':
[183] TYPEDEF 'user_pt_regs' type_id=181
[91247] TYPEDEF 'user_pt_regs' type_id=91245
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91245] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
and make them equiv by setting hypot_type_id for 181 to be 91245
---------------------------------------------------------------
Now comparing the second member:
struct pt_regs
{
union {
user_pt_regs user_regs;
---> struct {
unsigned long args[1];
psw_t psw;
unsigned long gprs[NUM_GPRS];
};
};
kernel knows it's same struct as user_pt_regs and uses ID 181
[186] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=183 bits_offset=0
'(anon)' type_id=181 bits_offset=0
but kmod has new ID 91250 (not 91245):
[91251] UNION '(anon)' size=152 vlen=2
'user_regs' type_id=91247 bits_offset=0
'(anon)' type_id=91250 bits_offset=0
and 181 and 91250 are equiv structs:
[181] STRUCT '(anon)' size=152 vlen=3
'args' type_id=182 bits_offset=0
'psw' type_id=179 bits_offset=64
'gprs' type_id=48 bits_offset=192
[91250] STRUCT '(anon)' size=152 vlen=3
'args' type_id=91246 bits_offset=0
'psw' type_id=91243 bits_offset=64
'gprs' type_id=91132 bits_offset=192
now hypot_type_id for 181 is 91245, but we have brand new struct
ID 91250, so we fail
what the patch tries to do is at this point to compare ID 91250
with 91245 and if it passes then we are equal and we throw away
ID 91250 because the hypot_type_id for 181 stays 91245
ufff.. thoughts? ;-)
Oh, this is a really great analysis, thanks a lot! It makes everything
clear. Basically, BTF dedup algo does too good job deduping vmlinux
BTF. :)
What's not clear is what to do about that, because a (current)
fundamental assumption of is_equiv() check is that any type within CU
(or in this case deduped vmlinux BTF) has exactly one unique mapping.
Clearly that's not the case now. That array fix you mentioned worked
around GCC bug where this assumption broke. In this case it's not a
bug of a compiler (neither of algo, really), we just need to make algo
smarter.
Let me think about this a bit, we'll need to make the equivalence
check be aware that there could be multiple equivalent mappings and be
ok with that as long as all candidates are equivalent between
themselves. Lots of equivalence and recursion to think about.
It would be great to have a simplified test case to play with that. Do
you mind distilling the chain of types above into a selftests and
posting it to the mailing list so that I can play with it? It
shouldn't be hard to write given BTF writing APIs. And we'll need a
selftests anyway once we improve the algo, so it's definitely not a
wasted work.
I ended up with simply test, where the idea is to use
type id which is defined after currently processed type
the last VALIDATE_RAW_BTF fails
I'm not sending full atch, because I assume this is not
to merge yet also I assume you might want to change that
anyway ;-)
Thanks, Jiri! I'll get to playing with this some time this week,
hopefully. I hope this is not a huge blocker for you?
I switched off s390x for the moment, so we can move on
I don't think it's huge blocker, nobody is screaming yet ;-)
thanks,
jirka
quoted
I'll check later on that special array case
thanks,
jirka
---
.../bpf/prog_tests/btf_dedup_split.c | 113 ++++++++++++++++++
1 file changed, 113 insertions(+)
From: Alexander Egorenkov <hidden> Date: 2022-02-25 13:19:35
Hi Jiri and Andrii,
we also have discovered this problem recently on Fedora 35 and linux-next.
Is there any status update here ?
@Jiri
Is the increase of total kernel modules size by 20MB really a big deal
on s390x ? We would like to have it enabled on our architecture
again ;-) And 20MB seems okay or am i missing something maybe ?
Another question i have wrt to BTF is why is it necessary to have e.g.
_struct module_ be present within kernel module BTF if it is already
present within vmlinux's one ? Can't the one from vmlinux be reused for
kernel modules as well, they should be identical, right ?
Thanks
Regards
Alex
@Jiri
Is the increase of total kernel modules size by 20MB really a big deal
on s390x ? We would like to have it enabled on our architecture
again ;-) And 20MB seems okay or am i missing something maybe ?
20M is not that much, the problem was that it's double the size
of all the modules and it was problem on rhel, where the impact
was much bigger for some reason
the fix seems to be already in fedora kernel so we could enable
BTF for s390x, I'll check on that
Another question i have wrt to BTF is why is it necessary to have e.g.
_struct module_ be present within kernel module BTF if it is already
present within vmlinux's one ? Can't the one from vmlinux be reused for
kernel modules as well, they should be identical, right ?
that's basically the issue.. the dedup algo did not cover all the
cases so BTF kept both module structs because they were 'different'
jirka