Re: [PATCH 2/3] ceph: only send the metrices supported by the MDS for old cephs

Xiubo Li <xiubli@xxxxxxxxxx> · Fri, 1 Apr 2022 22:11:17 +0800

On 4/1/22 5:41 PM, Jeff Layton wrote:
On Fri, 2022-04-01 at 09:18 +0800, Xiubo Li wrote:
On 3/31/22 8:11 PM, Jeff Layton wrote:
On Thu, 2022-03-31 at 14:52 +0800, xiubli@xxxxxxxxxx wrote:
From: Xiubo Li <xiubli@xxxxxxxxxx>

For some old ceph versions when receives unknown metrics it will
abort the MDS daemons. This will only send the metrics which are
supported by MDSes.

Defautly the MDS won't fill the s_metrics in the MClientSession
reply message, so with this patch will only force sending the
metrics to MDS since Quincy version, which is safe to receive
unknown metrics.

Next we will add one module option to force enable sending the
metrics if users think that is safe.

Is this really a problem we need to work around in the client?

This is an MDS bug and the patches to fix that abort are being
backported (or already have been). I think we shouldn't do this at all
and instead insist that this be fixed in the MDS.
Yes, though we have fixed that early in MDS there still have some use
cases which haven't backported yet.

Such as in the tracker#54411 we hit when upgrading from old ceph to new
ones when the new client will reconnect to the old ceph during the
upgrading, the mgr has two client instances which will reconnect to the
old ceph, and we cannot disable sending the metrics through the options,
because the mgr won't load them from ceph.conf for some reasons on purpose.

I am afraid there will have similar issue or strange use cases we
couldn't figure out like the tracker#54411. Such as when upgrading if
the kclients are upgraded first or a little early than ceph ?

Last week one guy pinged about the similar issue in Wechat in their own
upgrading test case they hit the metric crash.

As a workaround it's always better to crash the MDSes ?

We already have a kernel module option that disables sending metrics if
it's enabled. Why not just use that to disable metrics when dealing with
these problematic MDS versions that apparently can't be fixed?

Adding all of this complexity to avoid sending something to the MDS
because it might fall over really just seems like we're fixing this in
the wrong place.

If we were to take this, at what point can we drop this workaround? We
really don't want to carry this in perpetuity.

Okay, let's drop it for now.

-- Xiubo


URL: https://tracker.ceph.com/issues/54411
Signed-off-by: Xiubo Li <xiubli@xxxxxxxxxx>
---
   fs/ceph/mds_client.c |  19 +++-
   fs/ceph/mds_client.h |   1 +
   fs/ceph/metric.c     | 206 ++++++++++++++++++++++++-------------------
   3 files changed, 131 insertions(+), 95 deletions(-)

diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index f476c65fb985..65980ce97620 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -3422,7 +3422,7 @@ static void handle_session(struct ceph_mds_session *session,
   	void *end = p + msg->front.iov_len;
   	struct ceph_mds_session_head *h;
   	u32 op;
-	u64 seq, features = 0;
+	u64 seq, features = 0, metrics = 0;
   	int wake = 0;
   	bool blocklisted = false;
   
@@ -3452,11 +3452,21 @@ static void handle_session(struct ceph_mds_session *session,
   		}
   	}
   
+	/* version >= 4, metric bits */
+	if (msg_version >= 4) {
+		u32 len;
+		/* struct_v, struct_compat, and len */
+		ceph_decode_skip_n(&p, end, 2 + sizeof(u32), bad);
+		ceph_decode_32_safe(&p, end, len, bad);
+		if (len) {
+			ceph_decode_64_safe(&p, end, metrics, bad);
+			p += len - sizeof(metrics);
+		}
+	}
+
+	/* version >= 5, flags   */
   	if (msg_version >= 5) {
   		u32 flags;
-		/* version >= 4, struct_v, struct_cv, len, metric_spec */
-	        ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 2, bad);
-		/* version >= 5, flags   */
                   ceph_decode_32_safe(&p, end, flags, bad);
   		if (flags & CEPH_SESSION_BLOCKLISTED) {
   		        pr_warn("mds%d session blocklisted\n", session->s_mds);
@@ -3490,6 +3500,7 @@ static void handle_session(struct ceph_mds_session *session,
   			pr_info("mds%d reconnect success\n", session->s_mds);
   		session->s_state = CEPH_MDS_SESSION_OPEN;
   		session->s_features = features;
+		session->s_metrics = metrics;
   		renewed_caps(mdsc, session, 0);
   		if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
   			metric_schedule_delayed(&mdsc->metric);
diff --git a/fs/ceph/mds_client.h b/fs/ceph/mds_client.h
index 32107c26f50d..0f2061f5388d 100644
--- a/fs/ceph/mds_client.h
+++ b/fs/ceph/mds_client.h
@@ -188,6 +188,7 @@ struct ceph_mds_session {
   	int               s_state;
   	unsigned long     s_ttl;      /* time until mds kills us */
   	unsigned long	  s_features;
+	unsigned long	  s_metrics;
   	u64               s_seq;      /* incoming msg seq # */
   	struct mutex      s_mutex;    /* serialize session messages */
   
diff --git a/fs/ceph/metric.c b/fs/ceph/metric.c
index c47347d2e84e..f01c1f4e6b89 100644
--- a/fs/ceph/metric.c
+++ b/fs/ceph/metric.c
@@ -31,6 +31,7 @@ static bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc,
   	struct ceph_client_metric *m = &mdsc->metric;
   	u64 nr_caps = atomic64_read(&m->total_caps);
   	u32 header_len = sizeof(struct ceph_metric_header);
+	bool force = test_bit(CEPHFS_FEATURE_QUINCY, &s->s_features);
I don't necessarily have a problem with adding extra CEPHFS_FEATURE_*
enum values for different releases, as they're nice for documentation
purposes. In the actual client code however, we should ensure that we
only test for the _actual_ feature flag, and not the one corresponding
to a particular release.
Yeah, sounds good.

-- Xiubo

   	struct ceph_msg *msg;
   	s64 sum;
   	s32 items = 0;
@@ -51,117 +52,140 @@ static bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc,
   	head = msg->front.iov_base;
   
   	/* encode the cap metric */
-	cap = (struct ceph_metric_cap *)(head + 1);
-	cap->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_CAP_INFO);
-	cap->header.ver = 1;
-	cap->header.compat = 1;
-	cap->header.data_len = cpu_to_le32(sizeof(*cap) - header_len);
-	cap->hit = cpu_to_le64(percpu_counter_sum(&m->i_caps_hit));
-	cap->mis = cpu_to_le64(percpu_counter_sum(&m->i_caps_mis));
-	cap->total = cpu_to_le64(nr_caps);
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_CAP_INFO, &s->s_metrics)) {
+		cap = (struct ceph_metric_cap *)(head + 1);
+		cap->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_CAP_INFO);
+		cap->header.ver = 1;
+		cap->header.compat = 1;
+		cap->header.data_len = cpu_to_le32(sizeof(*cap) - header_len);
+		cap->hit = cpu_to_le64(percpu_counter_sum(&m->i_caps_hit));
+		cap->mis = cpu_to_le64(percpu_counter_sum(&m->i_caps_mis));
+		cap->total = cpu_to_le64(nr_caps);
+		items++;
+	}
   
   	/* encode the read latency metric */
-	read = (struct ceph_metric_read_latency *)(cap + 1);
-	read->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_READ_LATENCY);
-	read->header.ver = 2;
-	read->header.compat = 1;
-	read->header.data_len = cpu_to_le32(sizeof(*read) - header_len);
-	sum = m->metric[METRIC_READ].latency_sum;
-	ktime_to_ceph_timespec(&read->lat, sum);
-	ktime_to_ceph_timespec(&read->avg, m->metric[METRIC_READ].latency_avg);
-	read->sq_sum = cpu_to_le64(m->metric[METRIC_READ].latency_sq_sum);
-	read->count = cpu_to_le64(m->metric[METRIC_READ].total);
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_READ_LATENCY, &s->s_metrics)) {
+		read = (struct ceph_metric_read_latency *)(cap + 1);
+		read->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_READ_LATENCY);
+		read->header.ver = 2;
+		read->header.compat = 1;
+		read->header.data_len = cpu_to_le32(sizeof(*read) - header_len);
+		sum = m->metric[METRIC_READ].latency_sum;
+		ktime_to_ceph_timespec(&read->lat, sum);
+		ktime_to_ceph_timespec(&read->avg, m->metric[METRIC_READ].latency_avg);
+		read->sq_sum = cpu_to_le64(m->metric[METRIC_READ].latency_sq_sum);
+		read->count = cpu_to_le64(m->metric[METRIC_READ].total);
+		items++;
+	}
   
   	/* encode the write latency metric */
-	write = (struct ceph_metric_write_latency *)(read + 1);
-	write->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_WRITE_LATENCY);
-	write->header.ver = 2;
-	write->header.compat = 1;
-	write->header.data_len = cpu_to_le32(sizeof(*write) - header_len);
-	sum = m->metric[METRIC_WRITE].latency_sum;
-	ktime_to_ceph_timespec(&write->lat, sum);
-	ktime_to_ceph_timespec(&write->avg, m->metric[METRIC_WRITE].latency_avg);
-	write->sq_sum = cpu_to_le64(m->metric[METRIC_WRITE].latency_sq_sum);
-	write->count = cpu_to_le64(m->metric[METRIC_WRITE].total);
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_WRITE_LATENCY, &s->s_metrics)) {
+		write = (struct ceph_metric_write_latency *)(read + 1);
+		write->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_WRITE_LATENCY);
+		write->header.ver = 2;
+		write->header.compat = 1;
+		write->header.data_len = cpu_to_le32(sizeof(*write) - header_len);
+		sum = m->metric[METRIC_WRITE].latency_sum;
+		ktime_to_ceph_timespec(&write->lat, sum);
+		ktime_to_ceph_timespec(&write->avg, m->metric[METRIC_WRITE].latency_avg);
+		write->sq_sum = cpu_to_le64(m->metric[METRIC_WRITE].latency_sq_sum);
+		write->count = cpu_to_le64(m->metric[METRIC_WRITE].total);
+		items++;
+	}
   
   	/* encode the metadata latency metric */
-	meta = (struct ceph_metric_metadata_latency *)(write + 1);
-	meta->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_METADATA_LATENCY);
-	meta->header.ver = 2;
-	meta->header.compat = 1;
-	meta->header.data_len = cpu_to_le32(sizeof(*meta) - header_len);
-	sum = m->metric[METRIC_METADATA].latency_sum;
-	ktime_to_ceph_timespec(&meta->lat, sum);
-	ktime_to_ceph_timespec(&meta->avg, m->metric[METRIC_METADATA].latency_avg);
-	meta->sq_sum = cpu_to_le64(m->metric[METRIC_METADATA].latency_sq_sum);
-	meta->count = cpu_to_le64(m->metric[METRIC_METADATA].total);
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_METADATA_LATENCY, &s->s_metrics)) {
+		meta = (struct ceph_metric_metadata_latency *)(write + 1);
+		meta->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_METADATA_LATENCY);
+		meta->header.ver = 2;
+		meta->header.compat = 1;
+		meta->header.data_len = cpu_to_le32(sizeof(*meta) - header_len);
+		sum = m->metric[METRIC_METADATA].latency_sum;
+		ktime_to_ceph_timespec(&meta->lat, sum);
+		ktime_to_ceph_timespec(&meta->avg, m->metric[METRIC_METADATA].latency_avg);
+		meta->sq_sum = cpu_to_le64(m->metric[METRIC_METADATA].latency_sq_sum);
+		meta->count = cpu_to_le64(m->metric[METRIC_METADATA].total);
+		items++;
+	}
   
   	/* encode the dentry lease metric */
-	dlease = (struct ceph_metric_dlease *)(meta + 1);
-	dlease->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_DENTRY_LEASE);
-	dlease->header.ver = 1;
-	dlease->header.compat = 1;
-	dlease->header.data_len = cpu_to_le32(sizeof(*dlease) - header_len);
-	dlease->hit = cpu_to_le64(percpu_counter_sum(&m->d_lease_hit));
-	dlease->mis = cpu_to_le64(percpu_counter_sum(&m->d_lease_mis));
-	dlease->total = cpu_to_le64(atomic64_read(&m->total_dentries));
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_DENTRY_LEASE, &s->s_metrics)) {
+		dlease = (struct ceph_metric_dlease *)(meta + 1);
+		dlease->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_DENTRY_LEASE);
+		dlease->header.ver = 1;
+		dlease->header.compat = 1;
+		dlease->header.data_len = cpu_to_le32(sizeof(*dlease) - header_len);
+		dlease->hit = cpu_to_le64(percpu_counter_sum(&m->d_lease_hit));
+		dlease->mis = cpu_to_le64(percpu_counter_sum(&m->d_lease_mis));
+		dlease->total = cpu_to_le64(atomic64_read(&m->total_dentries));
+		items++;
+	}
   
   	sum = percpu_counter_sum(&m->total_inodes);
   
   	/* encode the opened files metric */
-	files = (struct ceph_opened_files *)(dlease + 1);
-	files->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_OPENED_FILES);
-	files->header.ver = 1;
-	files->header.compat = 1;
-	files->header.data_len = cpu_to_le32(sizeof(*files) - header_len);
-	files->opened_files = cpu_to_le64(atomic64_read(&m->opened_files));
-	files->total = cpu_to_le64(sum);
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_OPENED_FILES, &s->s_metrics)) {
+		files = (struct ceph_opened_files *)(dlease + 1);
+		files->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_OPENED_FILES);
+		files->header.ver = 1;
+		files->header.compat = 1;
+		files->header.data_len = cpu_to_le32(sizeof(*files) - header_len);
+		files->opened_files = cpu_to_le64(atomic64_read(&m->opened_files));
+		files->total = cpu_to_le64(sum);
+		items++;
+	}
   
   	/* encode the pinned icaps metric */
-	icaps = (struct ceph_pinned_icaps *)(files + 1);
-	icaps->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_PINNED_ICAPS);
-	icaps->header.ver = 1;
-	icaps->header.compat = 1;
-	icaps->header.data_len = cpu_to_le32(sizeof(*icaps) - header_len);
-	icaps->pinned_icaps = cpu_to_le64(nr_caps);
-	icaps->total = cpu_to_le64(sum);
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_PINNED_ICAPS, &s->s_metrics)) {
+		icaps = (struct ceph_pinned_icaps *)(files + 1);
+		icaps->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_PINNED_ICAPS);
+		icaps->header.ver = 1;
+		icaps->header.compat = 1;
+		icaps->header.data_len = cpu_to_le32(sizeof(*icaps) - header_len);
+		icaps->pinned_icaps = cpu_to_le64(nr_caps);
+		icaps->total = cpu_to_le64(sum);
+		items++;
+	}
   
   	/* encode the opened inodes metric */
-	inodes = (struct ceph_opened_inodes *)(icaps + 1);
-	inodes->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_OPENED_INODES);
-	inodes->header.ver = 1;
-	inodes->header.compat = 1;
-	inodes->header.data_len = cpu_to_le32(sizeof(*inodes) - header_len);
-	inodes->opened_inodes = cpu_to_le64(percpu_counter_sum(&m->opened_inodes));
-	inodes->total = cpu_to_le64(sum);
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_OPENED_INODES, &s->s_metrics)) {
+		inodes = (struct ceph_opened_inodes *)(icaps + 1);
+		inodes->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_OPENED_INODES);
+		inodes->header.ver = 1;
+		inodes->header.compat = 1;
+		inodes->header.data_len = cpu_to_le32(sizeof(*inodes) - header_len);
+		inodes->opened_inodes = cpu_to_le64(percpu_counter_sum(&m->opened_inodes));
+		inodes->total = cpu_to_le64(sum);
+		items++;
+	}
   
   	/* encode the read io size metric */
-	rsize = (struct ceph_read_io_size *)(inodes + 1);
-	rsize->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_READ_IO_SIZES);
-	rsize->header.ver = 1;
-	rsize->header.compat = 1;
-	rsize->header.data_len = cpu_to_le32(sizeof(*rsize) - header_len);
-	rsize->total_ops = cpu_to_le64(m->metric[METRIC_READ].total);
-	rsize->total_size = cpu_to_le64(m->metric[METRIC_READ].size_sum);
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_READ_IO_SIZES, &s->s_metrics)) {
+		rsize = (struct ceph_read_io_size *)(inodes + 1);
+		rsize->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_READ_IO_SIZES);
+		rsize->header.ver = 1;
+		rsize->header.compat = 1;
+		rsize->header.data_len = cpu_to_le32(sizeof(*rsize) - header_len);
+		rsize->total_ops = cpu_to_le64(m->metric[METRIC_READ].total);
+		rsize->total_size = cpu_to_le64(m->metric[METRIC_READ].size_sum);
+		items++;
+	}
   
   	/* encode the write io size metric */
-	wsize = (struct ceph_write_io_size *)(rsize + 1);
-	wsize->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_WRITE_IO_SIZES);
-	wsize->header.ver = 1;
-	wsize->header.compat = 1;
-	wsize->header.data_len = cpu_to_le32(sizeof(*wsize) - header_len);
-	wsize->total_ops = cpu_to_le64(m->metric[METRIC_WRITE].total);
-	wsize->total_size = cpu_to_le64(m->metric[METRIC_WRITE].size_sum);
-	items++;
+	if (force || test_bit(CLIENT_METRIC_TYPE_WRITE_IO_SIZES, &s->s_metrics)) {
+		wsize = (struct ceph_write_io_size *)(rsize + 1);
+		wsize->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_WRITE_IO_SIZES);
+		wsize->header.ver = 1;
+		wsize->header.compat = 1;
+		wsize->header.data_len = cpu_to_le32(sizeof(*wsize) - header_len);
+		wsize->total_ops = cpu_to_le64(m->metric[METRIC_WRITE].total);
+		wsize->total_size = cpu_to_le64(m->metric[METRIC_WRITE].size_sum);
+		items++;
+	}
+
+	if (!items)
+		return true;
   
   	put_unaligned_le32(items, &head->num);
   	msg->front.iov_len = len;







