Merge tag 'afs-fix-rotation-20240105' of git://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs

Pull afs updates from David Howells:
 "The majority of the patches are aimed at fixing and improving the AFS
  filesystem's rotation over server IP addresses, but there are also
  some fixes from Oleg Nesterov for the use of read_seqbegin_or_lock().

   - Fix fileserver probe handling so that the next round of probes
     doesn't break ongoing server/address rotation by clearing all the
     probe result tracking. This could occasionally cause the rotation
     algorithm to drop straight through, give a 'successful' result
     without actually emitting any RPC calls, leaving the reply buffer
     in an undefined state.

     Instead, detach the probe results into a separate struct and
     allocate a new one each time we start probing and update the
     pointer to it. Probes are also sent in order of address preference
     to try and improve the chance that the preferred one will complete
     first.

   - Fix server rotation so that it uses configurable address
     preferences across on the probes that have completed so far than
     ranking them by RTT as the latter doesn't necessarily give the best
     route. The preference list can be altered by writing into
     /proc/net/afs/addr_prefs.

   - Fix the handling of Read-Only (and Backup) volume callbacks as
     there is one per volume, not one per file, so if someone performs a
     command that, say, offlines the volume but doesn't change it, when
     it comes back online we don't spam the server with a status fetch
     for every vnode we're using. Instead, check the Creation timestamp
     in the VolSync record when prompted by a callback break.

   - Handle volume regression (ie. a RW volume being restored from a
     backup) by scrubbing all cache data for that volume. This is
     detected from the VolSync creation timestamp.

   - Adjust abort handling and abort -> error mapping to match better
     with what other AFS clients do.

   - Fix offline and busy volume state handling as they only apply to
     individual server instances and not entire volumes and the rotation
     algorithm should go and look at other servers if available. Also
     make it sleep briefly before each retry if all the volume instances
     are unavailable"

* tag 'afs-fix-rotation-20240105' of git://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs: (40 commits)
  afs: trace: Log afs_make_call(), including server address
  afs: Fix offline and busy message emission
  afs: Fix fileserver rotation
  afs: Overhaul invalidation handling to better support RO volumes
  afs: Parse the VolSync record in the reply of a number of RPC ops
  afs: Don't leave DONTUSE/NEWREPSITE servers out of server list
  afs: Fix comment in afs_do_lookup()
  afs: Apply server breaks to mmap'd files in the call processor
  afs: Move the vnode/volume validity checking code into its own file
  afs: Defer volume record destruction to a workqueue
  afs: Make it possible to find the volumes that are using a server
  afs: Combine the endpoint state bools into a bitmask
  afs: Keep a record of the current fileserver endpoint state
  afs: Dispatch vlserver probes in priority order
  afs: Dispatch fileserver probes in priority order
  afs: Mark address lists with configured priorities
  afs: Provide a way to configure address priorities
  afs: Remove the unimplemented afs_cmp_addr_list()
  afs: Add some more info to /proc/net/afs/servers
  rxrpc: Create a procfile to display outstanding client conn bundles
  ...

1 files changed, 39 insertions, 21 deletions

diff --git a/fs/afs/vl_probe.c b/fs/afs/vl_probe.c
index 58452b86e672..3d2e0c925460 100644
--- a/fs/afs/vl_probe.c
+++ b/fs/afs/vl_probe.c
@@ -46,11 +46,12 @@ static void afs_done_one_vl_probe(struct afs_vlserver *server, bool wake_up)
  */
 void afs_vlserver_probe_result(struct afs_call *call)
 {
-	struct afs_addr_list *alist = call->alist;
+	struct afs_addr_list *alist = call->vl_probe;
 	struct afs_vlserver *server = call->vlserver;
+	struct afs_address *addr = &alist->addrs[call->probe_index];
 	unsigned int server_index = call->server_index;
 	unsigned int rtt_us = 0;
-	unsigned int index = call->addr_ix;
+	unsigned int index = call->probe_index;
 	bool have_result = false;
 	int ret = call->error;
 
@@ -89,7 +90,7 @@ void afs_vlserver_probe_result(struct afs_call *call)
 	case -ETIME:
 	default:
 		clear_bit(index, &alist->responded);
-		set_bit(index, &alist->failed);
+		set_bit(index, &alist->probe_failed);
 		if (!(server->probe.flags & AFS_VLSERVER_PROBE_RESPONDED) &&
 		    (server->probe.error == 0 ||
 		     server->probe.error == -ETIMEDOUT ||
@@ -101,21 +102,21 @@ void afs_vlserver_probe_result(struct afs_call *call)
 
 responded:
 	set_bit(index, &alist->responded);
-	clear_bit(index, &alist->failed);
+	clear_bit(index, &alist->probe_failed);
 
 	if (call->service_id == YFS_VL_SERVICE) {
 		server->probe.flags |= AFS_VLSERVER_PROBE_IS_YFS;
 		set_bit(AFS_VLSERVER_FL_IS_YFS, &server->flags);
-		alist->addrs[index].srx_service = call->service_id;
+		server->service_id = call->service_id;
 	} else {
 		server->probe.flags |= AFS_VLSERVER_PROBE_NOT_YFS;
 		if (!(server->probe.flags & AFS_VLSERVER_PROBE_IS_YFS)) {
 			clear_bit(AFS_VLSERVER_FL_IS_YFS, &server->flags);
-			alist->addrs[index].srx_service = call->service_id;
+			server->service_id = call->service_id;
 		}
 	}
 
-	rxrpc_kernel_get_srtt(call->net->socket, call->rxcall, &rtt_us);
+	rtt_us = rxrpc_kernel_get_srtt(addr->peer);
 	if (rtt_us < server->probe.rtt) {
 		server->probe.rtt = rtt_us;
 		server->rtt = rtt_us;
@@ -130,8 +131,10 @@ responded:
 out:
 	spin_unlock(&server->probe_lock);
 
-	_debug("probe [%u][%u] %pISpc rtt=%u ret=%d",
-	       server_index, index, &alist->addrs[index].transport, rtt_us, ret);
+	trace_afs_vl_probe(server, false, alist, index, call->error, call->abort_code, rtt_us);
+	_debug("probe [%u][%u] %pISpc rtt=%d ret=%d",
+	       server_index, index, rxrpc_kernel_remote_addr(addr->peer),
+	       rtt_us, ret);
 
 	afs_done_one_vl_probe(server, have_result);
 }
@@ -146,35 +149,52 @@ static bool afs_do_probe_vlserver(struct afs_net *net,
 				  unsigned int server_index,
 				  struct afs_error *_e)
 {
-	struct afs_addr_cursor ac = {
-		.index = 0,
-	};
+	struct afs_addr_list *alist;
 	struct afs_call *call;
+	unsigned long unprobed;
+	unsigned int index, i;
 	bool in_progress = false;
+	int best_prio;
 
 	_enter("%s", server->name);
 
 	read_lock(&server->lock);
-	ac.alist = rcu_dereference_protected(server->addresses,
-					     lockdep_is_held(&server->lock));
+	alist = rcu_dereference_protected(server->addresses,
+					  lockdep_is_held(&server->lock));
+	afs_get_addrlist(alist, afs_alist_trace_get_vlprobe);
 	read_unlock(&server->lock);
 
-	atomic_set(&server->probe_outstanding, ac.alist->nr_addrs);
+	atomic_set(&server->probe_outstanding, alist->nr_addrs);
 	memset(&server->probe, 0, sizeof(server->probe));
 	server->probe.rtt = UINT_MAX;
 
-	for (ac.index = 0; ac.index < ac.alist->nr_addrs; ac.index++) {
-		call = afs_vl_get_capabilities(net, &ac, key, server,
+	unprobed = (1UL << alist->nr_addrs) - 1;
+	while (unprobed) {
+		best_prio = -1;
+		index = 0;
+		for (i = 0; i < alist->nr_addrs; i++) {
+			if (test_bit(i, &unprobed) &&
+			    alist->addrs[i].prio > best_prio) {
+				index = i;
+				best_prio = alist->addrs[i].prio;
+			}
+		}
+		__clear_bit(index, &unprobed);
+
+		trace_afs_vl_probe(server, true, alist, index, 0, 0, 0);
+		call = afs_vl_get_capabilities(net, alist, index, key, server,
 					       server_index);
 		if (!IS_ERR(call)) {
+			afs_prioritise_error(_e, call->error, call->abort_code);
 			afs_put_call(call);
 			in_progress = true;
 		} else {
-			afs_prioritise_error(_e, PTR_ERR(call), ac.abort_code);
+			afs_prioritise_error(_e, PTR_ERR(call), 0);
 			afs_done_one_vl_probe(server, false);
 		}
 	}
 
+	afs_put_addrlist(alist, afs_alist_trace_put_vlprobe);
 	return in_progress;
 }
 
@@ -185,12 +205,10 @@ int afs_send_vl_probes(struct afs_net *net, struct key *key,
 		       struct afs_vlserver_list *vllist)
 {
 	struct afs_vlserver *server;
-	struct afs_error e;
+	struct afs_error e = {};
 	bool in_progress = false;
 	int i;
 
-	e.error = 0;
-	e.responded = false;
 	for (i = 0; i < vllist->nr_servers; i++) {
 		server = vllist->servers[i].server;
 		if (test_bit(AFS_VLSERVER_FL_PROBED, &server->flags))