Skip to contentSkip to conversation

btf_encoder: fix and complete filtering out zero-sized per-CPU variables

PATCH Applied Andrii Nakryiko 4 participants 10 messages May 24, 2021 – May 28, 2021 BPF

lore
Share & tools
lore

Andrii Nakryiko

author maintainer
#0 lore raw

PATCH 1/1 btf_encoder: fix and complete filtering out zero-sized per-CPU variables

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

See [0] for when this causes big problems.

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

---
 btf_encoder.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)
btf_encoder.c +10 -3
index c711f124b31e..672b9943a4e2 100644
@@ -538,6 +538,7 @@ 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, *dwarf_name;
+		const struct tag *tag;
 		uint64_t addr;
 		int id;
 
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
 		/* addr has to be recorded before we follow spec */
 		addr = var->ip.addr;
+		dwarf_name = variable__name(var, cu);
 
 		/* DWARF takes into account .data..percpu section offset
 		 * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 		 *  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 (var->ip.addr == 0)
 			if (!dwarf_name || strcmp(dwarf_name, name))
 				continue;
-		}
 
 		if (var->spec)
 			var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 			break;
 		}
 
+		tag = cu__type(cu, var->ip.tag.type);
+		if (tag__size(tag, cu) == 0) {
+			if (btf_elf__verbose)
+				fprintf(stderr, "Ignoring zero-sized per-CPU variable '%s'...\n", dwarf_name ?: "<missing name>");
+			continue;
+		}
+
 		type = var->ip.tag.type + type_id_off;
 		linkage = var->external ? BTF_VAR_GLOBAL_ALLOCATED : BTF_VAR_STATIC;
 
Signature · 1 line
2.30.2
#1 lore raw

Em Mon, May 24, 2021 at 04:42:22PM -0700, Andrii Nakryiko escreveu:

Quoted 5 lines

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

I've been following this, just didn't got to process it, will do it soon.

- Arnaldo

Quoted 59 lines

See [0] for when this causes big problems.

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

---
 btf_encoder.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)
btf_encoder.c +10 -3
index c711f124b31e..672b9943a4e2 100644
@@ -538,6 +538,7 @@ 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, *dwarf_name;
+		const struct tag *tag;
 		uint64_t addr;
 		int id;
 
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
 		/* addr has to be recorded before we follow spec */
 		addr = var->ip.addr;
+		dwarf_name = variable__name(var, cu);
 
 		/* DWARF takes into account .data..percpu section offset
 		 * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 		 *  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 (var->ip.addr == 0)
 			if (!dwarf_name || strcmp(dwarf_name, name))
 				continue;
-		}
 
 		if (var->spec)
 			var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 			break;
 		}
 
+		tag = cu__type(cu, var->ip.tag.type);
+		if (tag__size(tag, cu) == 0) {
+			if (btf_elf__verbose)
+				fprintf(stderr, "Ignoring zero-sized per-CPU variable '%s'...\n", dwarf_name ?: "<missing name>");
+			continue;
+		}
+
 		type = var->ip.tag.type + type_id_off;
 		linkage = var->external ? BTF_VAR_GLOBAL_ALLOCATED : BTF_VAR_STATIC;
 
Signature · 1 line
2.30.2
Signature · 1 line
- Arnaldo
#2 lore raw
In reply to Andrii Nakryiko

Em Mon, May 24, 2021 at 04:42:22PM -0700, Andrii Nakryiko escreveu:

btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force,

Quoted 33 lines

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

See [0] for when this causes big problems.

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

+++ b/btf_encoder.c

btf_encoder.c +2 -3
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
 		/* addr has to be recorded before we follow spec */
 		addr = var->ip.addr;
+		dwarf_name = variable__name(var, cu);
 
 		/* DWARF takes into account .data..percpu section offset
 		 * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 		 *  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 (var->ip.addr == 0)
 			if (!dwarf_name || strcmp(dwarf_name, name))
 				continue;
-		}
 
 		if (var->spec)
 			var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,

I just changed the above hunk to be:

@@ -583,7 +585,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,

                 *  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;
                }

Which is shorter and keeps the {} around a multi line if block, ok?

Thanks, applied!

- Arnaldo

Tested-by Michal Suchánek
#3 lore raw

Em Thu, May 27, 2021 at 11:55:10AM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 10 lines

Em Mon, May 24, 2021 at 04:42:22PM -0700, Andrii Nakryiko escreveu:

Quoted 9 lines

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

See [0] for when this causes big problems.

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

I also added this:

Michal, so you tested this patch and verified it fixed the problem? If so please let me know so that I also add:

Thanks,

- Arnaldo

Quoted 41 lines

btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force,

Quoted 23 lines

+++ b/btf_encoder.c

btf_encoder.c +2 -3
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
 		/* addr has to be recorded before we follow spec */
 		addr = var->ip.addr;
+		dwarf_name = variable__name(var, cu);
 
 		/* DWARF takes into account .data..percpu section offset
 		 * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 		 *  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 (var->ip.addr == 0)
 			if (!dwarf_name || strcmp(dwarf_name, name))
 				continue;
-		}
 
 		if (var->spec)
 			var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,

I just changed the above hunk to be:

btf_encoder.c +0 -1
@@ -583,7 +585,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
                 *  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;
                }

Which is shorter and keeps the {} around a multi line if block, ok?

Thanks, applied!

- Arnaldo

Signature · 1 line
- Arnaldo
#4 lore raw

Hello,

Quoted 10 lines · Arnaldo Carvalho de Melo

Em Thu, May 27, 2021 at 11:55:10AM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 9 lines

Em Mon, May 24, 2021 at 04:42:22PM -0700, Andrii Nakryiko escreveu:

Quoted 7 lines

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

See [0] for when this causes big problems.

Quoted 8 lines
Quoted 1 line
Quoted 1 line

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

I also added this:

Michal, so you tested this patch and verified it fixed the problem? If so please let me know so that I also add:

This is the first time I see this patch.

Given that linux-next does not build for me at the moment I don't think I will test it soon.

Thanks

Michal

Quoted 52 lines

Thanks,

- Arnaldo

Quoted 41 lines

btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force,

Quoted 23 lines

+++ b/btf_encoder.c

btf_encoder.c +2 -3
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
 		/* addr has to be recorded before we follow spec */
 		addr = var->ip.addr;
+		dwarf_name = variable__name(var, cu);
 
 		/* DWARF takes into account .data..percpu section offset
 		 * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 		 *  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 (var->ip.addr == 0)
 			if (!dwarf_name || strcmp(dwarf_name, name))
 				continue;
-		}
 
 		if (var->spec)
 			var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,

I just changed the above hunk to be:

btf_encoder.c +0 -1
@@ -583,7 +585,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
                 *  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;
                }

Which is shorter and keeps the {} around a multi line if block, ok?

Thanks, applied!

- Arnaldo

Signature · 1 line
- Arnaldo
#5 lore raw

Em Thu, May 27, 2021 at 05:27:58PM +0200, Michal Suchánek escreveu:

Quoted 19 lines

Hello,

Quoted 16 lines · Arnaldo Carvalho de Melo

Em Thu, May 27, 2021 at 11:55:10AM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 10 lines

Em Mon, May 24, 2021 at 04:42:22PM -0700, Andrii Nakryiko escreveu:

Quoted 9 lines

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

See [0] for when this causes big problems.

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

I also added this:

Quoted 7 lines
Quoted 2 lines

Michal, so you tested this patch and verified it fixed the problem? If so please let me know so that I also add:

This is the first time I see this patch.

Given that linux-next does not build for me at the moment I don't think I will test it soon.

Ok, I'm test building with torvalds/master, will try with linux-next afterwards,

Thanks,

- Arnaldo

Quoted 56 lines

Thanks

Michal

Quoted 52 lines

Thanks,

- Arnaldo

Quoted 41 lines

btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force,

Quoted 23 lines

+++ b/btf_encoder.c

btf_encoder.c +2 -3
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
 		/* addr has to be recorded before we follow spec */
 		addr = var->ip.addr;
+		dwarf_name = variable__name(var, cu);
 
 		/* DWARF takes into account .data..percpu section offset
 		 * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 		 *  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 (var->ip.addr == 0)
 			if (!dwarf_name || strcmp(dwarf_name, name))
 				continue;
-		}
 
 		if (var->spec)
 			var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,

I just changed the above hunk to be:

btf_encoder.c +0 -1
@@ -583,7 +585,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
                 *  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;
                }

Which is shorter and keeps the {} around a multi line if block, ok?

Thanks, applied!

- Arnaldo

Signature · 1 line
- Arnaldo
Signature · 1 line
- Arnaldo
#6 lore raw

Em Thu, May 27, 2021 at 12:38:38PM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 23 lines

Em Thu, May 27, 2021 at 05:27:58PM +0200, Michal Suchánek escreveu:

Quoted 22 lines

Hello,

Quoted 18 lines · Arnaldo Carvalho de Melo

Em Thu, May 27, 2021 at 11:55:10AM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 10 lines

Em Mon, May 24, 2021 at 04:42:22PM -0700, Andrii Nakryiko escreveu:

Quoted 9 lines

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

See [0] for when this causes big problems.

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

I also added this:

Michal, so you tested this patch and verified it fixed the problem? If so please let me know so that I also add:

Quoted 7 lines
Quoted 4 lines

This is the first time I see this patch.

Given that linux-next does not build for me at the moment I don't think I will test it soon.

Ok, I'm test building with torvalds/master, will try with linux-next afterwards,

I build and booted torvalds/master, all seems to work, now moving to linux-next.

Quoted 64 lines

Thanks,

- Arnaldo

Quoted 56 lines

Thanks

Michal

Quoted 52 lines

Thanks,

- Arnaldo

Quoted 41 lines

btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force,

Quoted 23 lines

+++ b/btf_encoder.c

btf_encoder.c +2 -3
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
 		/* addr has to be recorded before we follow spec */
 		addr = var->ip.addr;
+		dwarf_name = variable__name(var, cu);
 
 		/* DWARF takes into account .data..percpu section offset
 		 * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 		 *  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 (var->ip.addr == 0)
 			if (!dwarf_name || strcmp(dwarf_name, name))
 				continue;
-		}
 
 		if (var->spec)
 			var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,

I just changed the above hunk to be:

btf_encoder.c +0 -1
@@ -583,7 +585,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
                 *  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;
                }

Which is shorter and keeps the {} around a multi line if block, ok?

Thanks, applied!

- Arnaldo

Signature · 1 line
- Arnaldo
Signature · 1 line
- Arnaldo
Signature · 1 line
- Arnaldo
#7 lore raw
Quoted 26 lines · Arnaldo Carvalho de Melo

Em Thu, May 27, 2021 at 12:38:38PM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 25 lines

Em Thu, May 27, 2021 at 05:27:58PM +0200, Michal Suchánek escreveu:

Quoted 24 lines

Hello,

Quoted 10 lines · Arnaldo Carvalho de Melo

Em Thu, May 27, 2021 at 11:55:10AM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 9 lines

Em Mon, May 24, 2021 at 04:42:22PM -0700, Andrii Nakryiko escreveu:

Quoted 7 lines

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

See [0] for when this causes big problems.

Quoted 8 lines
Quoted 1 line
Quoted 1 line

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

I also added this:

Michal, so you tested this patch and verified it fixed the problem? If so please let me know so that I also add:

This is the first time I see this patch.

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

Given that linux-next does not build for me at the moment I don't think I will test it soon.

Ok, I'm test building with torvalds/master, will try with linux-next afterwards,

I build and booted torvalds/master, all seems to work, now moving to linux-next.

[ 9s] /usr/bin/qemu-kvm -nodefaults -no-reboot -nographic -vga none -cpu host -object rng-random,filename=/dev/random,id=rng0 -device virtio-rng-pci,rng=rng0 -runas qemu -net none -kernel /var/cache/obs/worker/root_2/.mount/boot/kernel -initrd /var/cache/obs/worker/root_2/.mount/boot/initrd -append root=/dev/disk/by-id/virtio-0 rootfstype=ext4 rootflags=noatime ext4.allow_unsupported=1 mitigations=off panic=1 quiet no-kvmclock elevator=noop nmi_watchdog=0 rw rd.driver.pre=binfmt_misc console=ttyS0 init=/.build/build -m 8192 -drive file=/var/cache/obs/worker/root_2/root,format=raw,if=none,id=disk,cache=unsafe -device virtio-blk-pci,drive=disk,serial=0 -drive file=/var/cache/obs/worker/root_2/swap,format=raw,if=none,id=swap,cache=unsafe -device virtio-blk-pci,drive=swap,serial=1 -serial stdio -chardev socket,id=monitor,server,nowait,path=/var/cache/obs/worker/root_2/root.qemu/monitor -mon chardev=monitor,mode=readline -smp 8
[ 9s] c[?7l[2J[0mSeaBIOS (version rel-1.14.0-0-g155821a-rebuilt.opensuse.org)

[   15s] Booting from ROM..c[?7l[2J### VM INTERACTION END ###
[   15s] 2nd stage started in virtual machine
[   15s] machine type: x86_64
[   15s] Linux version: 5.13.0-rc3-next-2[    5.210478] sysrq: Changing Loglevel
[   15s] 0210526-2.g3bd43[    5.211566] sysrq: Loglevel set to 4
[   15s] 9f-vanilla #1 SMP Thu May 27 18:27:17 UTC 2021 (3bd439f)
[   15s] Increasing log level from now on...
[   15s] Enable sysrq operations
[   15s] Setting up swapspace version 1, size = 2 GiB (2097147904 bytes)
[   15s] no label, UUID=6e22796c-6df5-4859-a291-24acde2e92e5
[   15s] swapon: /dev/vdb: found signature [pagesize=4096, signature=swap]
[   15s] swapon: /dev/vdb: pagesize=4096, swapsize=2097152000, devsize=2097152000
[   15s] swapon /dev/vdb
[   15s] WARNING: udev not running, creating extra device nodes
[   15s] logging output to //.build.log...

next-20210527 does not build becuase of missing symbol but next-20210526 built with the patched pahole boots.

Thanks

Michal

Quoted 69 lines
Quoted 65 lines

Thanks,

- Arnaldo

Quoted 56 lines

Thanks

Michal

Quoted 52 lines

Thanks,

- Arnaldo

Quoted 41 lines

btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force,

Quoted 23 lines

+++ b/btf_encoder.c

btf_encoder.c +2 -3
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
 		/* addr has to be recorded before we follow spec */
 		addr = var->ip.addr;
+		dwarf_name = variable__name(var, cu);
 
 		/* DWARF takes into account .data..percpu section offset
 		 * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 		 *  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 (var->ip.addr == 0)
 			if (!dwarf_name || strcmp(dwarf_name, name))
 				continue;
-		}
 
 		if (var->spec)
 			var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,

I just changed the above hunk to be:

btf_encoder.c +0 -1
@@ -583,7 +585,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
                 *  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;
                }

Which is shorter and keeps the {} around a multi line if block, ok?

Thanks, applied!

- Arnaldo

Signature · 1 line
- Arnaldo
Signature · 1 line
- Arnaldo
Signature · 1 line
- Arnaldo
#8 lore raw
In reply to Michal Suchánek
Quoted 16 lines · Michal Suchánek

Hello,

Quoted 12 lines · Arnaldo Carvalho de Melo

Em Thu, May 27, 2021 at 11:55:10AM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 10 lines

Em Mon, May 24, 2021 at 04:42:22PM -0700, Andrii Nakryiko escreveu:

Quoted 9 lines

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

See [0] for when this causes big problems.

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

Quoted 9 lines
Quoted 6 lines

I also added this:

Michal, so you tested this patch and verified it fixed the problem? If so please let me know so that I also add:

This is the first time I see this patch.

I've posted a link to it in your thread [0].

[0] https://lore.kernel.org/bpf/CAEf4BzZ9=aLVD7ytgCcSxcbOLqFNK-p1mj14Rv_TGnOyL3aO_g@mail.gmail.com/

Quoted 2 lines

Given that linux-next does not build for me at the moment I don't think I will test it soon.

This patch applied to pahole master will fix linux-next build.

Quoted 57 lines

Thanks

Michal

Quoted 52 lines

Thanks,

- Arnaldo

Quoted 36 lines

btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force,

Quoted 23 lines

+++ b/btf_encoder.c

btf_encoder.c +2 -3
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
          /* addr has to be recorded before we follow spec */
          addr = var->ip.addr;
+         dwarf_name = variable__name(var, cu);
 
          /* DWARF takes into account .data..percpu section offset
           * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
           *  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 (var->ip.addr == 0)
                  if (!dwarf_name || strcmp(dwarf_name, name))
                          continue;
-         }
 
          if (var->spec)
                  var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,

I just changed the above hunk to be:

btf_encoder.c +0 -1
@@ -583,7 +585,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
                 *  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;
                }
Quoted 5 lines

Which is shorter and keeps the {} around a multi line if block, ok?

Thanks, applied!

- Arnaldo

--

- Arnaldo

#9 lore raw
In reply to Arnaldo Carvalho de Melo
Quoted 48 lines · Arnaldo Carvalho de Melo

Em Mon, May 24, 2021 at 04:42:22PM -0700, Andrii Nakryiko escreveu:

btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force,

Quoted 33 lines

btf_encoder is ignoring zero-sized per-CPU ELF symbols, but the same has to be done for DWARF variables when matching them with ELF symbols. This is due to zero-sized DWARF variables matching unrelated (non-zero-sized) variable that happens to be allocated at the exact same address, leading to a lot of confusion in BTF.

See [0] for when this causes big problems.

[0] https://lore.kernel.org/bpf/CAEf4BzZ0-sihSL-UAm21JcaCCY92CqfNxycHRZYXcoj8OYb=wA@mail.gmail.com/

+++ b/btf_encoder.c

btf_encoder.c +2 -3
@@ -550,6 +551,7 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
 
              /* addr has to be recorded before we follow spec */
              addr = var->ip.addr;
+             dwarf_name = variable__name(var, cu);
 
              /* DWARF takes into account .data..percpu section offset
               * within its segment, which for vmlinux is 0, but for kernel
@@ -582,11 +584,9 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
               *  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 (var->ip.addr == 0)
                      if (!dwarf_name || strcmp(dwarf_name, name))
                              continue;
-             }
 
              if (var->spec)
                      var = var->spec;
@@ -600,6 +600,13 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,

I just changed the above hunk to be:

btf_encoder.c +0 -1
@@ -583,7 +585,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,
                 *  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;
                }
Quoted 1 line

Which is shorter and keeps the {} around a multi line if block, ok?

yeah, no problem

Quoted 2 lines

Thanks, applied!

Thanks!

Quoted 2 lines

- Arnaldo