This is merely a historical archive of years 2008-2021, before the migration to mailman3.
A maintained and still updated list archive can be found at https://lists.osmocom.org/hyperkitty/list/gerrit-log@lists.osmocom.org/.
laforge gerrit-no-reply at lists.osmocom.orglaforge has submitted this change. ( https://gerrit.osmocom.org/c/libosmo-netif/+/18628 )
Change subject: src/stream: Work around more Linux kernel ABI breakage
......................................................................
src/stream: Work around more Linux kernel ABI breakage
Back in Change-Id Ia95dd1f9ffed9f743c049e05797b1a6f1f9f8c69 we tried
to work-around kernel ABI breakage introduced in kernel >= v5.5, but it
seems that there have already been similar ABI breakages in v4.11 and v4.12.
See https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/include/uapi/linux/sctp.h?id=b6e6b5f1da7e8d092f86a4351802c27c0170c5a5
and https://marc.info/?l=linux-sctp&m=158729301516157&w=2 for the most
recent incarnation.
See https://osmocom.org/issues/4573#note-6 for all known cases
of SCTP_EVENTS ABI breakage.
Closes: OS#4573
Change-Id: Icc49f347cdc0bb77a5c0e230597d662ac35b4acc
---
M src/stream.c
1 file changed, 93 insertions(+), 20 deletions(-)
Approvals:
laforge: Looks good to me, approved
Jenkins Builder: Verified
diff --git a/src/stream.c b/src/stream.c
index 6e4c461..44ac1b4 100644
--- a/src/stream.c
+++ b/src/stream.c
@@ -73,6 +73,97 @@
#define MSG_NOSIGNAL 0
#endif
+/* is any of the bytes from offset .. u8_size in 'u8' non-zero? return offset or -1 if all zero */
+static int byte_nonzero(const uint8_t *u8, unsigned int offset, unsigned int u8_size)
+{
+ int j;
+
+ for (j = offset; j < u8_size; j++) {
+ if (u8[j] != 0)
+ return j;
+ }
+
+ return -1;
+}
+
+static int sctp_sockopt_event_subscribe_size = 0;
+
+static int determine_sctp_sockopt_event_subscribe_size(void)
+{
+ uint8_t buf[256];
+ socklen_t buf_len = sizeof(buf);
+ int sd, rc;
+
+ /* only do this once */
+ if (sctp_sockopt_event_subscribe_size != 0)
+ return 0;
+
+ sd = socket(AF_INET, SOCK_STREAM, IPPROTO_SCTP);
+ if (sd < 0)
+ return sd;
+
+ rc = getsockopt(sd, IPPROTO_SCTP, SCTP_EVENTS, buf, &buf_len);
+ if (rc < 0)
+ return rc;
+
+ sctp_sockopt_event_subscribe_size = buf_len;
+
+ LOGP(DLINP, LOGL_INFO, "sizes of 'struct sctp_event_subscribe': compile-time %zu, kernel: %u\n",
+ sizeof(struct sctp_event_subscribe), sctp_sockopt_event_subscribe_size);
+ return 0;
+}
+
+/* Attempt to work around Linux kernel ABI breakage
+ *
+ * The Linux kernel ABI for the SCTP_EVENTS socket option has been broken repeatedly.
+ * - until commit 35ea82d611da59f8bea44a37996b3b11bb1d3fd7 ( kernel < 4.11), the size is 10 bytes
+ * - in 4.11 it is 11 bytes
+ * - in 4.12 .. 5.4 it is 13 bytes
+ * - in kernels >= 5.5 it is 14 bytes
+ *
+ * This wouldn't be a problem if the kernel didn't have a "stupid" assumption that the structure
+ * size passed by userspace will match 1:1 the length of the structure at kernel compile time. In
+ * an ideal world, it would just use the known first bytes and assume the remainder is all zero.
+ * But as it doesn't do that, let's try to work around this */
+static int sctp_setsockopt_events_linux_workaround(int fd, const struct sctp_event_subscribe *event)
+{
+
+ const unsigned int compiletime_size = sizeof(*event);
+ int rc;
+
+ if (determine_sctp_sockopt_event_subscribe_size() < 0) {
+ LOGP(DLINP, LOGL_ERROR, "Cannot determine SCTP_EVENTS socket option size\n");
+ return -1;
+ }
+
+ if (compiletime_size == sctp_sockopt_event_subscribe_size) {
+ /* no kernel workaround needed */
+ return setsockopt(fd, IPPROTO_SCTP, SCTP_EVENTS, event, compiletime_size);
+ } else if (compiletime_size < sctp_sockopt_event_subscribe_size) {
+ /* we are using an older userspace with a more modern kernel and hence need
+ * to pad the data */
+ uint8_t buf[sctp_sockopt_event_subscribe_size];
+
+ memcpy(buf, event, compiletime_size);
+ memset(buf + sizeof(*event), 0, sctp_sockopt_event_subscribe_size - compiletime_size);
+ return setsockopt(fd, IPPROTO_SCTP, SCTP_EVENTS, buf, sctp_sockopt_event_subscribe_size);
+ } else /* if (compiletime_size > sctp_sockopt_event_subscribe_size) */ {
+ /* we are using a newer userspace with an older kernel and hence need to truncate
+ * the data - but only if the caller didn't try to enable any of the events of the
+ * truncated portion */
+ rc = byte_nonzero((const uint8_t *)event, sctp_sockopt_event_subscribe_size,
+ compiletime_size);
+ if (rc >= 0) {
+ LOGP(DLINP, LOGL_ERROR, "Kernel only supports sctp_event_subscribe of %u bytes, "
+ "but caller tried to enable more modern event at offset %u\n",
+ sctp_sockopt_event_subscribe_size, rc);
+ return -1;
+ }
+
+ return setsockopt(fd, IPPROTO_SCTP, SCTP_EVENTS, event, sctp_sockopt_event_subscribe_size);
+ }
+}
+
static int sctp_sock_activate_events(int fd)
{
#ifdef HAVE_LIBSCTP
@@ -89,28 +180,10 @@
event.sctp_shutdown_event = 1;
/* IMPORTANT: Do NOT enable sender_dry_event here, see
* https://bugzilla.redhat.com/show_bug.cgi?id=1442784 */
- rc = setsockopt(fd, IPPROTO_SCTP, SCTP_EVENTS,
- &event, sizeof(event));
- /*
- * Attempt to work around kernel ABI breakage
- *
- * In kernel commit b6e6b5f1da7e8d092f86a4351802c27c0170c5a5, the
- * struct sctp_event_subscribe had a u8 field added to it at the end, thus
- * breaking ABI.
- * See https://marc.info/?l=linux-sctp&m=158729301516157&w=2 for discussion.
- *
- * We attempt to work around the issue where the kernel header are new
- * and running kernel is old, by forcing the size of the struct to 13 which
- * is the "old" size
- */
- if ((rc < 0) && (sizeof(event) != 13))
- rc = setsockopt(fd, IPPROTO_SCTP, SCTP_EVENTS,
- &event, 13);
-
+ rc = sctp_setsockopt_events_linux_workaround(fd, &event);
if (rc < 0)
- LOGP(DLINP, LOGL_ERROR, "couldn't activate SCTP events "
- "on FD %u\n", fd);
+ LOGP(DLINP, LOGL_ERROR, "couldn't activate SCTP events on FD %u\n", fd);
return rc;
#else
return -1;
--
To view, visit https://gerrit.osmocom.org/c/libosmo-netif/+/18628
To unsubscribe, or for help writing mail filters, visit https://gerrit.osmocom.org/settings
Gerrit-Project: libosmo-netif
Gerrit-Branch: master
Gerrit-Change-Id: Icc49f347cdc0bb77a5c0e230597d662ac35b4acc
Gerrit-Change-Number: 18628
Gerrit-PatchSet: 3
Gerrit-Owner: laforge <laforge at osmocom.org>
Gerrit-Reviewer: Jenkins Builder
Gerrit-Reviewer: laforge <laforge at osmocom.org>
Gerrit-CC: pespin <pespin at sysmocom.de>
Gerrit-MessageType: merged
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osmocom.org/pipermail/gerrit-log/attachments/20200602/22c40935/attachment.htm>