File 3028-snapshot-upgrade-fix.patch of Package ceph-ceph-20.2.0+20260319.5bb32787
diff --git a/src/mds/MDSMap.cc b/src/mds/MDSMap.cc
index e12a97486..af6798401 100644
--- a/src/mds/MDSMap.cc
+++ b/src/mds/MDSMap.cc
@@ -984,11 +984,123 @@ void MDSMap::decode(bufferlist::const_iterator& p)
}
if (ev >= 17) {
- decode(bal_rank_mask, p);
- }
+ /*
+ * This is a workaround for https://bugs.launchpad.net/ubuntu/+source/ceph/+bug/2089565.
+ *
+ * The issue is that bal_rank_mask is stored before max_xattr_size in squid
+ * and in reverse order in the squid from Ubuntu 24.04 release (it has ceph
+ * squid from a git snapshot which was taken before the squid release).
+ *
+ * bal_rank_mask is stored as a text string with a decimal representation
+ * of a number inside. The string is stored as length of the string (4
+ * bytes, little endian) and then the string itself (without trailing 0,
+ * just the string itself).
+ *
+ * max_xattr_size is stored as uint64_t little endian integer.
+ *
+ * The following patch was used to get the hexdumps below:
+ *
+ * diff --git a/src/mon/MDSMonitor.cc b/src/mon/MDSMonitor.cc
+ * index 76a57ac443..d36bed2257 100644
+ * --- a/src/mon/MDSMonitor.cc
+ * +++ b/src/mon/MDSMonitor.cc
+ * @@ -143,6 +143,7 @@ void MDSMonitor::update_from_paxos(bool *need_bootstrap)
+ *
+ * ceph_assert(fsmap_bl.length() > 0);
+ * dout(10) << __func__ << " got " << version << dendl;
+ * + fsmap_bl.hexdump(std::cout);
+ * try {
+ * PaxosFSMap::decode(fsmap_bl);
+ * } catch (const ceph::buffer::malformed_input& e) {
+ *
+ *
+ * This is how the relevant part of the bufferlist looks like for the
+ * squid from the 24.04 release 19.2.0~git20240301.4c76c50-0ubuntu6:
+ *
+ * ceph-mon[...]: 00000620 00 00 00 00 00 00 00 00 00 00 00 00 01 07 00 00 |................|
+ * ceph-mon[...]: 00000630 00 63 65 70 68 2d 66 73 00 00 00 00 00 00 00 00 |.ceph-fs........|
+ * ceph-mon[...]: 00000640 ff ff ff ff 00 00 00 00 00 00 00 00 00 02 00 00 |................|
+ * ******** <<< bal_rank_mask
+ * string length
+ * the top byte of the (4 bytes, the value is 2)
+ * string length >>>>>>>>> ** ##### <<< bal_rank_mask itself (it's string "-1" here)
+ * ceph-mon[...]: 00000650 00 2d 31 00 00 01 00 00 00 00 00 00 00 00 00 00 |.-1.............|
+ * ^^^^^^^^^^^^^^^^^^^^^^^^ max_xattr_size
+ * (the default value 65536)
+ * ceph-mon[...]: 00000660 00 00 00 00 00 00 00 01 01 05 00 00 00 00 00 00 |................|
+ * ceph-mon[...]: 00000670 00 00 01 00 00 00 b1 17 00 00 00 00 00 00 01 00 |................|
+ * ceph-mon[...]: 00000680 00 00 00 00 00 00 00 00 00 00 01 a1 6e ad 67 53 |............n.gS|
+ *
+ *
+ * And this is how the relevant part of the bufferlist looks like for
+ * the squid release 19.2.0-0ubuntu0.24.04.1:
+ *
+ * ceph-mon[...]: 000003a0 00 00 00 00 00 00 00 00 01 07 00 00 00 63 65 70 |.............cep|
+ * ceph-mon[...]: 000003b0 68 2d 66 73 00 00 00 00 00 00 00 00 ff ff ff ff |h-fs............|
+ * ceph-mon[...]: 000003c0 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00 00 |................|
+ * ^^^^^^^^^^^^^^^^^^^^
+ * vv <<<<<<<<<<<<<<<<<<<<<<<< max_xattr_size (8 bytes, the value is 65536)
+ * ceph-mon[...]: 000003d0 00 02 00 00 00 2d 31 01 01 05 00 00 00 00 00 00 |.....-1.........|
+ * ^^^^^^^^^^^ ##### <<< bal_rank_mask (string, the value is "-1")
+ * bal_rank_mask string length (2)
+ *
+ * The fix for the bug looks at the byte 4 bytes ahead (if the current
+ * position is 0x3C9, then the code would look at the byte at 0x3CD). In
+ * the squid release the byte most likely would be 0 (it could be non-zero
+ * for 4GiB+ extended attributes, which is highly unlikely). In the
+ * squid git snapshot from 24.04 release it would be the first char of a
+ * decimal representation of a number, which is either "-" (0x2D) or a
+ * number (a value from 0x30 to 0x39). This patch assumes that
+ * max_xattr_size is less than 64GiB and checks the byte against 0x10, and
+ * then uses the correct decoding order for bal_rank_mask and
+ * max_xattr_size for each situation.
+ *
+ * Reef 18.2.4 has bal_rank_mask and then max_xattr_size, the same order as
+ * in the squid release.
+ *
+ * Quincy 17.2.7 has neither of the 2 fields.
+ */
+ dout(20) << "ev=" << ev << dendl;
+
+ unsigned off = p.get_off();
+ dout(20) << "off=" << off << dendl;
+
+ /*
+ * The smallest size is a single decimal digit bal_rank_mask,
+ * which takes 5 bytes.
+ */
+ if (p.get_remaining() < 5) {
+ throw ceph::buffer::malformed_input("the buffer is too small");
+ }
+
+ p.seek(off + 4);
+ uint8_t byte;
+ decode(byte, p);
+ p.seek(off);
+
+ dout(20) << "byte=" << static_cast<int>(byte) << dendl;
+
+ if (byte < 0x10) {
+ dout(20) << "decoding in the normal squid order" << dendl;
+
+ if (ev >= 17) {
+ decode(max_xattr_size, p);
+ }
- if (ev >= 18) {
- decode(max_xattr_size, p);
+ if (ev >= 18) {
+ decode(bal_rank_mask, p);
+ }
+ } else {
+ dout(20) << "decoding in the squid git snapshot order" << dendl;
+
+ if (ev >= 17) {
+ decode(bal_rank_mask, p);
+ }
+
+ if (ev >= 18) {
+ decode(max_xattr_size, p);
+ }
+ }
}
if (ev >= 19) {
