* [PATCH v2] cg: fix offset for > 8 bit bitfields in dt_cg_ctf_offsetof()
@ 2025-09-04 9:08 Alan Maguire
2025-09-05 1:39 ` Eugene Loh
0 siblings, 1 reply; 2+ messages in thread
From: Alan Maguire @ 2025-09-04 9:08 UTC (permalink / raw)
To: dtrace; +Cc: dtrace-devel, Alan Maguire
The tcp provider uses dt_cg_tramp_get_member() to retrieve the
offset of the sk_protocol field in struct sock. However it
returns the wrong value on UEK6 since it is an 8-bit bitfield.
From pahole we see:
unsigned int __sk_flags_offset[0]; /* 560 0 */
unsigned int sk_padding:1; /* 560: 0 4 */
unsigned int sk_kern_sock:1; /* 560: 1 4 */
unsigned int sk_no_check_tx:1; /* 560: 2 4 */
unsigned int sk_no_check_rx:1; /* 560: 3 4 */
unsigned int sk_userlocks:4; /* 560: 4 4 */
unsigned int sk_protocol:8; /* 560: 8 4 */
In other words it is really at offset 561 but because we just
lookup the member offset and not the member type offset we get the
wrong value for the sk_protoocol.
This in turn causes tcp state-change probes (and in-progress UDP
probes) to not fire since we verify that sk_protocol == IPPROTO_TCP.
The fix is to look up the member _type_ offset and add it to the
bit offset we get for the member itself. With this in place the
state-change probes fire, but the local tcp tests still fail due
to separate issues with the tcp:::accept-established probe.
This issue is not seen on more recent kernels because sk_protocol
becomes a __u16 as the number of protocols exceeds 256.
Signed-off-by: Alan Maguire <alan.maguire@oracle.com>
Changes since v1:
- added comment reflecting need to handle more complex bitfield values
in future (Nick)
---
libdtrace/dt_cg.c | 15 ++++++++++++++-
1 file changed, 14 insertions(+), 1 deletion(-)
diff --git a/libdtrace/dt_cg.c b/libdtrace/dt_cg.c
index cd9e7f4e..e378cf96 100644
--- a/libdtrace/dt_cg.c
+++ b/libdtrace/dt_cg.c
@@ -1959,6 +1959,8 @@ dt_cg_ctf_offsetof(const char *structname, const char *membername,
dtrace_typeinfo_t sym;
ctf_file_t *ctfp;
ctf_membinfo_t ctm;
+ ctf_encoding_t cte;
+ int offset;
if (dtrace_lookup_by_type(yypcb->pcb_hdl, DTRACE_OBJ_EVERY, structname,
&sym))
@@ -1973,6 +1975,17 @@ dt_cg_ctf_offsetof(const char *structname, const char *membername,
longjmp(yypcb->pcb_jmpbuf, EDT_NOCTF);
}
+ offset = ctm.ctm_offset;
+
+ /* a bitfield may have an additional bit offset which means we need
+ * to adjust the reported byte offset. Currently there is no need
+ * to bitshift values as they always both fall on 8-bit boundaries
+ * and are sizes which are multiples of 8 bits, but in future if
+ * more complex bitfield retrieval is required, changes such as those
+ * in dt_cg_field_get() will be needed to support that.
+ */
+ if (ctf_type_encoding(ctfp, ctm.ctm_type, &cte) != CTF_ERR)
+ offset += cte.cte_offset;
if (sizep || ldopp) {
uint_t ldop;
@@ -1982,7 +1995,7 @@ dt_cg_ctf_offsetof(const char *structname, const char *membername,
*ldopp = ldop;
}
- return (ctm.ctm_offset / NBBY);
+ return (offset / NBBY);
}
static void
dt_cg_act_breakpoint(dt_pcb_t *pcb, dt_node_t *dnp, dtrace_actkind_t kind)
--
2.43.5
^ permalink raw reply related [flat|nested] 2+ messages in thread
* Re: [PATCH v2] cg: fix offset for > 8 bit bitfields in dt_cg_ctf_offsetof()
2025-09-04 9:08 [PATCH v2] cg: fix offset for > 8 bit bitfields in dt_cg_ctf_offsetof() Alan Maguire
@ 2025-09-05 1:39 ` Eugene Loh
0 siblings, 0 replies; 2+ messages in thread
From: Eugene Loh @ 2025-09-05 1:39 UTC (permalink / raw)
To: Alan Maguire, dtrace; +Cc: dtrace-devel
Reviewed-by: Eugene Loh <eugene.loh@oracle.com>
On 9/4/25 05:08, Alan Maguire wrote:
> The tcp provider uses dt_cg_tramp_get_member() to retrieve the
> offset of the sk_protocol field in struct sock. However it
> returns the wrong value on UEK6 since it is an 8-bit bitfield.
> From pahole we see:
>
> unsigned int __sk_flags_offset[0]; /* 560 0 */
> unsigned int sk_padding:1; /* 560: 0 4 */
> unsigned int sk_kern_sock:1; /* 560: 1 4 */
> unsigned int sk_no_check_tx:1; /* 560: 2 4 */
> unsigned int sk_no_check_rx:1; /* 560: 3 4 */
> unsigned int sk_userlocks:4; /* 560: 4 4 */
> unsigned int sk_protocol:8; /* 560: 8 4 */
>
> In other words it is really at offset 561 but because we just
> lookup the member offset and not the member type offset we get the
> wrong value for the sk_protoocol.
>
> This in turn causes tcp state-change probes (and in-progress UDP
> probes) to not fire since we verify that sk_protocol == IPPROTO_TCP.
>
> The fix is to look up the member _type_ offset and add it to the
> bit offset we get for the member itself. With this in place the
> state-change probes fire, but the local tcp tests still fail due
> to separate issues with the tcp:::accept-established probe.
>
> This issue is not seen on more recent kernels because sk_protocol
> becomes a __u16 as the number of protocols exceeds 256.
>
> Signed-off-by: Alan Maguire <alan.maguire@oracle.com>
>
> Changes since v1:
>
> - added comment reflecting need to handle more complex bitfield values
> in future (Nick)
> ---
> libdtrace/dt_cg.c | 15 ++++++++++++++-
> 1 file changed, 14 insertions(+), 1 deletion(-)
>
> diff --git a/libdtrace/dt_cg.c b/libdtrace/dt_cg.c
> index cd9e7f4e..e378cf96 100644
> --- a/libdtrace/dt_cg.c
> +++ b/libdtrace/dt_cg.c
> @@ -1959,6 +1959,8 @@ dt_cg_ctf_offsetof(const char *structname, const char *membername,
> dtrace_typeinfo_t sym;
> ctf_file_t *ctfp;
> ctf_membinfo_t ctm;
> + ctf_encoding_t cte;
> + int offset;
>
> if (dtrace_lookup_by_type(yypcb->pcb_hdl, DTRACE_OBJ_EVERY, structname,
> &sym))
> @@ -1973,6 +1975,17 @@ dt_cg_ctf_offsetof(const char *structname, const char *membername,
>
> longjmp(yypcb->pcb_jmpbuf, EDT_NOCTF);
> }
> + offset = ctm.ctm_offset;
> +
> + /* a bitfield may have an additional bit offset which means we need
> + * to adjust the reported byte offset. Currently there is no need
> + * to bitshift values as they always both fall on 8-bit boundaries
> + * and are sizes which are multiples of 8 bits, but in future if
> + * more complex bitfield retrieval is required, changes such as those
> + * in dt_cg_field_get() will be needed to support that.
> + */
> + if (ctf_type_encoding(ctfp, ctm.ctm_type, &cte) != CTF_ERR)
> + offset += cte.cte_offset;
>
> if (sizep || ldopp) {
> uint_t ldop;
> @@ -1982,7 +1995,7 @@ dt_cg_ctf_offsetof(const char *structname, const char *membername,
> *ldopp = ldop;
> }
>
> - return (ctm.ctm_offset / NBBY);
> + return (offset / NBBY);
> }
> static void
> dt_cg_act_breakpoint(dt_pcb_t *pcb, dt_node_t *dnp, dtrace_actkind_t kind)
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2025-09-05 1:40 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2025-09-04 9:08 [PATCH v2] cg: fix offset for > 8 bit bitfields in dt_cg_ctf_offsetof() Alan Maguire
2025-09-05 1:39 ` Eugene Loh
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox