Re: [PATCH v15 5/5] kunit: Add tests for csum_ipv6_magic and ip_fast_csum
From: Guenter Roeck
Date: Mon Jan 22 2024 - 21:07:28 EST
On 1/22/24 15:55, Charlie Jenkins wrote:
On Mon, Jan 22, 2024 at 03:39:16PM -0800, Palmer Dabbelt wrote:
On Mon, 22 Jan 2024 13:41:48 PST (-0800), David.Laight@xxxxxxxxxx wrote:
From: Guenter Roeck
Sent: 22 January 2024 17:16
On 1/22/24 08:52, David Laight wrote:
From: Guenter Roeck
Sent: 22 January 2024 16:40
Hi,
On Mon, Jan 08, 2024 at 03:57:06PM -0800, Charlie Jenkins wrote:
Supplement existing checksum tests with tests for csum_ipv6_magic and
ip_fast_csum.
Signed-off-by: Charlie Jenkins <charlie@xxxxxxxxxxxx>
---
With this patch in the tree, the arm:mps2-an385 qemu emulation gets a bad hiccup.
[ 1.839556] Unhandled exception: IPSR = 00000006 LR = fffffff1
[ 1.839804] CPU: 0 PID: 164 Comm: kunit_try_catch Tainted: G N 6.8.0-rc1 #1
[ 1.839948] Hardware name: Generic DT based system
[ 1.840062] PC is at __csum_ipv6_magic+0x8/0xb4
[ 1.840408] LR is at test_csum_ipv6_magic+0x3d/0xa4
[ 1.840493] pc : [<21212f34>] lr : [<21117fd5>] psr: 0100020b
[ 1.840586] sp : 2180bebc ip : 46c7f0d2 fp : 21275b38
[ 1.840664] r10: 21276b60 r9 : 21275b28 r8 : 21465cfc
[ 1.840751] r7 : 00003085 r6 : 21275b4e r5 : 2138702c r4 : 00000001
[ 1.840847] r3 : 2c000000 r2 : 1ac7f0d2 r1 : 21275b39 r0 : 21275b29
[ 1.840942] xPSR: 0100020b
This translates to:
PC is at __csum_ipv6_magic (arch/arm/lib/csumipv6.S:15)
LR is at test_csum_ipv6_magic (./arch/arm/include/asm/checksum.h:60
./arch/arm/include/asm/checksum.h:163 lib/checksum_kunit.c:617)
Obviously I can not say if this is a problem with qemu or a problem with
the Linux kernel. Given that, and the presumably low interest in
running mps2-an385 with Linux, I'll simply disable that test. Just take
it as a heads up that there _may_ be a problem with this on arm
nommu systems.
Can you drop in a disassembly of __csum_ipv6_magic ?
Actually I think it is:
It is, as per the PC pointer above. I don't know anything about arm assembler,
much less about its behavior with THUMB code.
Doesn't look like thumb to me (offset 8 is two 4-byte instructions) and
the code I found looks like arm to me.
(I haven't written any arm asm since before they invented thumb!)
ENTRY(__csum_ipv6_magic)
str lr, [sp, #-4]!
adds ip, r2, r3
ldmia r1, {r1 - r3, lr}
So the fault is (probably) a misaligned ldmia ?
Are they ever supported?
Good question. My primary guess is that this never worked. As I said,
this was just intended to be informational, (probably) no reason to bother.
Of course one might ask if it makes sense to even keep the arm nommu code
in the kernel, but that is of course a different question. I do wonder though
if anyone but me is running it.
If it is an alignment fault it isn't a 'nommu' bug.
And traditionally arm didn't support misaligned transfers (well not
in anyway any other cpu did!).
It might be that the kernel assumes that all ethernet packets are
aligned, but the test suite isn't aligning the buffer.
Which would make it a test suite bug.
From talking to Evan and Vineet, I think you're right and this is a test
suite bug: specifically the tests weren't respecting NET_IP_ALIGN. That
didn't crop up for ip_fast_csum() as it just uses ldr which supports
misaligned accesses on the M3 (at least as far as I can tell).
So I think the right fix is something like
diff --git a/lib/checksum_kunit.c b/lib/checksum_kunit.c
index 225bb7701460..2dd282e27dd4 100644
--- a/lib/checksum_kunit.c
+++ b/lib/checksum_kunit.c
@@ -5,6 +5,7 @@
#include <kunit/test.h>
#include <asm/checksum.h>
+#include <asm/checksum.h>
#include <net/ip6_checksum.h>
#define MAX_LEN 512
@@ -15,6 +16,7 @@
#define IPv4_MAX_WORDS 15
#define NUM_IPv6_TESTS 200
#define NUM_IP_FAST_CSUM_TESTS 181
+#define SUPPORTED_ALIGNMENT (1 << NET_IP_ALIGN)
/* Values for a little endian CPU. Byte swap each half on big endian CPU. */
static const u32 random_init_sum = 0x2847aab;
@@ -486,7 +488,7 @@ static void test_csum_fixed_random_inputs(struct kunit *test)
__sum16 result, expec;
assert_setup_correct(test);
- for (align = 0; align < TEST_BUFLEN; ++align) {
+ for (align = 0; align < TEST_BUFLEN; align += SUPPORTED_ALIGNMENT) {
memcpy(&tmp_buf[align], random_buf,
min(MAX_LEN, TEST_BUFLEN - align));
for (len = 0; len < MAX_LEN && (align + len) < TEST_BUFLEN;
@@ -513,7 +515,7 @@ static void test_csum_all_carry_inputs(struct kunit *test)
assert_setup_correct(test);
memset(tmp_buf, 0xff, TEST_BUFLEN);
- for (align = 0; align < TEST_BUFLEN; ++align) {
+ for (align = 0; align < TEST_BUFLEN; align += SUPPORTED_ALIGNMENT) {
for (len = 0; len < MAX_LEN && (align + len) < TEST_BUFLEN;
++len) {
/*
@@ -553,7 +555,7 @@ static void test_csum_no_carry_inputs(struct kunit *test)
assert_setup_correct(test);
memset(tmp_buf, 0x4, TEST_BUFLEN);
- for (align = 0; align < TEST_BUFLEN; ++align) {
+ for (align = 0; align < TEST_BUFLEN; align += SUPPORTED_ALIGNMENT) {
for (len = 0; len < MAX_LEN && (align + len) < TEST_BUFLEN;
++len) {
/*
but I haven't even build tested it...
This doesn't fix the test_csum_ipv6_magic test case that was causing the
initial problem, but the same trick can be done in that test.
The above didn't (and still doesn't) fail for me. The following fixes the problem.
So, yes, I guess the problem has to do with alignment. I don't know if NET_IP_ALIGN
would do the trick, though - it works, but it seems to me that the definition of
NET_IP_ALIGN is supposed to address potential performance issues, not mandatory
IP header alignment.
Thanks,
Guenter
---
diff --git a/lib/checksum_kunit.c b/lib/checksum_kunit.c
index 225bb7701460..c8730af2a474 100644
--- a/lib/checksum_kunit.c
+++ b/lib/checksum_kunit.c
@@ -591,6 +591,8 @@ static void test_ip_fast_csum(struct kunit *test)
}
}
+#define SUPPORTED_ALIGNMENT (1 << NET_IP_ALIGN)
+
static void test_csum_ipv6_magic(struct kunit *test)
{
#if defined(CONFIG_NET)
@@ -607,7 +609,7 @@ static void test_csum_ipv6_magic(struct kunit *test)
const int csum_offset = sizeof(struct in6_addr) + sizeof(struct in6_addr) +
sizeof(int) + sizeof(char);
- for (int i = 0; i < NUM_IPv6_TESTS; i++) {
+ for (int i = 0; i < NUM_IPv6_TESTS; i+=SUPPORTED_ALIGNMENT) {
saddr = (const struct in6_addr *)(random_buf + i);
daddr = (const struct in6_addr *)(random_buf + i +
daddr_offset);