ceph: only send the metrices supported by the MDS for old cephs
cover ceph: only send the metrices supported by the MDS for old cephs
From: Xiubo Li <[email protected]>
This will fix the issue in [1], and only send metrics to MDS that supported, and for Quincy or higher ceph versions since it's safe will force to send all the metrics. This could make sure the early Quincy ceph versions still could get the metric which haven't backported [2], which will fill the metric bits supported by MDS when opening the sessions, no need to enable the force_ignore_metric_bits module parameter.
[1]: https://tracker.ceph.com/issues/54411
[2]: https://github.com/ceph/ceph/pull/45370
Xiubo Li (3):
ceph: add the Octopus,Pacific,Quency feature bits
ceph: only send the metrices supported by the MDS for old cephs
ceph: add force_ignore_metric_bits module parameter support
fs/ceph/mds_client.c | 19 +++- fs/ceph/mds_client.h | 10 ++- fs/ceph/metric.c | 207 ++++++++++++++++++++++++------------------- fs/ceph/metric.h | 1 + fs/ceph/super.c | 6 +- 5 files changed, 144 insertions(+), 99 deletions(-)
Signature · 1 line
2.27.0
PATCH 1/3 ceph: add the Octopus,Pacific,Quency feature bits
From: Xiubo Li <[email protected]>
URL: https://tracker.ceph.com/issues/54411
--- fs/ceph/mds_client.h | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)
fs/ceph/mds_client.h +6 -3
@@ -27,10 +27,13 @@ enum ceph_feature_type {CEPHFS_FEATURE_RECLAIM_CLIENT ,CEPHFS_FEATURE_LAZY_CAP_WANTED ,CEPHFS_FEATURE_MULTI_RECONNECT , -CEPHFS_FEATURE_DELEG_INO , -CEPHFS_FEATURE_METRIC_COLLECT , +CEPHFS_FEATURE_OCTOPUS , +CEPHFS_FEATURE_DELEG_INO =CEPHFS_FEATURE_OCTOPUS , +CEPHFS_FEATURE_PACIFIC , +CEPHFS_FEATURE_METRIC_COLLECT =CEPHFS_FEATURE_PACIFIC , +CEPHFS_FEATURE_QUINCY , - CEPHFS_FEATURE_MAX = CEPHFS_FEATURE_METRIC_COLLECT, + CEPHFS_FEATURE_MAX = CEPHFS_FEATURE_QUINCY, };/*
Signature · 1 line
2.27.0
Re: [PATCH 1/3] ceph: add the Octopus,Pacific,Quency feature bits
Quoted 19 lines · On Thu, 2022-03-31 at 14:52 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
URL: https://tracker.ceph.com/issues/54411
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/mds_client.h | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)fs/ceph/mds_client.h +1
-2@@ -27,10 +27,13 @@ enum ceph_feature_type {CEPHFS_FEATURE_RECLAIM_CLIENT ,CEPHFS_FEATURE_LAZY_CAP_WANTED ,CEPHFS_FEATURE_MULTI_RECONNECT , -CEPHFS_FEATURE_DELEG_INO , - CEPHFS_FEATURE_METRIC_COLLECT, + CEPHFS_FEATURE_OCTOPUS,
fs/ceph/mds_client.h enum ceph_feature_type { · line 39
Quoted 10 lines
fs/ceph/mds_client.h +5
-1+CEPHFS_FEATURE_DELEG_INO =CEPHFS_FEATURE_OCTOPUS , +CEPHFS_FEATURE_PACIFIC , +CEPHFS_FEATURE_METRIC_COLLECT =CEPHFS_FEATURE_PACIFIC , +CEPHFS_FEATURE_QUINCY , - CEPHFS_FEATURE_MAX = CEPHFS_FEATURE_METRIC_COLLECT, + CEPHFS_FEATURE_MAX = CEPHFS_FEATURE_QUINCY, };/*
(cc'ing Patrick)
I think we decided a while back to move away from "release" feature flags like this, because they're ambiguous. We do occasionally backport features to later stable versions and then the release flag becomes meaningless.
If the "feature" here is extended metrics, then this should be something like CEPHFS_FEATURE_METRIC_V2 or METRIC_EXTENDED or something. IOW, the flag name should describe the feature that we're advertising.
Signature · 1 line
Jeff Layton <[email protected]>
PATCH 2/3 ceph: only send the metrices supported by the MDS for old cephs
In reply to [email protected]From: Xiubo Li <[email protected]>
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.
URL: https://tracker.ceph.com/issues/54411
--- fs/ceph/mds_client.c | 19 +++- fs/ceph/mds_client.h | 1 + fs/ceph/metric.c | 206 ++++++++++++++++++++++++------------------- 3 files changed, 131 insertions(+), 95 deletions(-)
fs/ceph/mds_client.c +15 -4
@@ -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);
fs/ceph/mds_client.h +1 -0
@@ -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 */
fs/ceph/metric.c +115 -91
238 lines not on this page · raw message
Signature · 1 line
2.27.0
Quoted 6 lines · On Thu, 2022-03-31 at 14:52 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
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.
Quoted 8 lines
Defautly the MDS won't fill the
s_metricsin 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.
Quoted 71 lines
URL: https://tracker.ceph.com/issues/54411
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/mds_client.c | 19 +++- fs/ceph/mds_client.h | 1 + fs/ceph/metric.c | 206 ++++++++++++++++++++++++------------------- 3 files changed, 131 insertions(+), 95 deletions(-)fs/ceph/mds_client.c +15
-4@@ -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);fs/ceph/mds_client.h +1
-0@@ -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 */ fs/ceph/metric.c +0
-0
fs/ceph/metric.c static bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc, · line 34
Quoted 4 lines
fs/ceph/metric.c +1
-0@@ -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.
Quoted 235 lines
fs/ceph/metric.c +114
-91struct 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;
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 16 lines · Jeff Layton
Quoted 6 lines · On Thu, 2022-03-31 at 14:52 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
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.
Quoted 8 lines
Defautly the MDS won't fill the
s_metricsin 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.
Quoted 5 lines
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 ?
Quoted 66 lines
Quoted 66 lines
URL: https://tracker.ceph.com/issues/54411
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/mds_client.c | 19 +++- fs/ceph/mds_client.h | 1 + fs/ceph/metric.c | 206 ++++++++++++++++++++++++------------------- 3 files changed, 131 insertions(+), 95 deletions(-)fs/ceph/mds_client.c +15
-4@@ -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);fs/ceph/mds_client.h +1
-0@@ -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 */
Quoted 14 lines
fs/ceph/metric.c static bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc, · line 34
Quoted 9 lines
fs/ceph/metric.c +1
-0@@ -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
Quoted 236 lines
Quoted 236 lines
fs/ceph/metric.c +114
-91struct 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;
Quoted 32 lines · Xiubo Li
Quoted 16 lines · Jeff Layton
Quoted 6 lines · On Thu, 2022-03-31 at 14:52 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
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.
Quoted 8 lines
Defautly the MDS won't fill the
s_metricsin 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.
Quoted 5 lines
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.
Quoted 9 lines
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.
Quoted 322 lines
Quoted 66 lines
Quoted 66 lines
URL: https://tracker.ceph.com/issues/54411
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/mds_client.c | 19 +++- fs/ceph/mds_client.h | 1 + fs/ceph/metric.c | 206 ++++++++++++++++++++++++------------------- 3 files changed, 131 insertions(+), 95 deletions(-)fs/ceph/mds_client.c +15
-4@@ -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);fs/ceph/mds_client.h +1
-0@@ -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 */ Quoted 14 lines
fs/ceph/metric.c static bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc, · line 34
Quoted 9 lines
fs/ceph/metric.c +1
-0@@ -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
Quoted 236 lines
Quoted 236 lines
fs/ceph/metric.c +114
-91struct 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;
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 44 lines · Jeff Layton
Quoted 30 lines · Xiubo Li
Quoted 12 lines · Jeff Layton
Quoted 11 lines · On Thu, 2022-03-31 at 14:52 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
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_metricsin 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.Quoted 8 lines
Quoted 3 lines
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.
Quoted 9 lines
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?
Quoted 6 lines
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
Quoted 319 lines
Quoted 319 lines
Quoted 66 lines
Quoted 66 lines
URL: https://tracker.ceph.com/issues/54411
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/mds_client.c | 19 +++- fs/ceph/mds_client.h | 1 + fs/ceph/metric.c | 206 ++++++++++++++++++++++++------------------- 3 files changed, 131 insertions(+), 95 deletions(-)fs/ceph/mds_client.c +15
-4@@ -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);fs/ceph/mds_client.h +1
-0@@ -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 */ Quoted 14 lines
fs/ceph/metric.c static bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc, · line 34
Quoted 9 lines
fs/ceph/metric.c +1
-0@@ -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
Quoted 235 lines
Quoted 235 lines
fs/ceph/metric.c +114
-91struct 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;
Jeff,
I think this still makes sense, more detail please see [1], which is as I worried for the older ceph and when users want to upgrade the cluster.
Though this is a MDS side bug and have already been fixed in the MDS, but if there are existing clusters running before that fix, so when upgrading the kclient only or if the upgrade the kclient first, they will complain the same issue in [1].
[1] https://tracker.ceph.com/issues/56529
-- Xiubo
Quoted 339 lines · Jeff Layton
Quoted 6 lines · On Thu, 2022-03-31 at 14:52 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
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.
Quoted 8 lines
Defautly the MDS won't fill the
s_metricsin 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.
Quoted 71 lines
URL: https://tracker.ceph.com/issues/54411
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/mds_client.c | 19 +++- fs/ceph/mds_client.h | 1 + fs/ceph/metric.c | 206 ++++++++++++++++++++++++------------------- 3 files changed, 131 insertions(+), 95 deletions(-)fs/ceph/mds_client.c +15
-4@@ -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);fs/ceph/mds_client.h +1
-0@@ -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 */ fs/ceph/metric.c +0
-0fs/ceph/metric.c static bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc, · line 34
Quoted 4 lines
fs/ceph/metric.c +1
-0@@ -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.
Quoted 235 lines
fs/ceph/metric.c +114
-91struct 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;
More info.
Check the upstream ceph code:
The commit c669de12e076bcedf3fe6cfe52377713f06d7ceb introduced the ceph_abort() code when received an unknown metric from clients.
$ git tag --contain c669de12e076bcedf3fe6cfe52377713f06d7ceb v16.1.0 v16.2.0 v16.2.1 v16.2.2 v16.2.3 v16.2.4 v16.2.5 v16.2.6 v16.2.7 v16.2.8 v16.2.9 v17.0.0 v17.1.0 v17.2.0 v17.2.1
And a following commit b1b44d775df3160d937c068d5e1079e24199ed6b has fixed it:
src/include/cephfs/metrics/Types.h +0 -0
@@ -53,7 +53,7 @@ inline std::ostream &operator<<(std::ostream &os, const ClientMetricType &type)
os << "OPENED_INODES";
break;
default:
- ceph_abort();
+ os << "Unknown metric type: " << type;
}
return os;
$ git tag --contain b1b44d775df3160d937c068d5e1079e24199ed6b v17.1.0 v17.2.0 v17.2.1
For the backported commit for the above patch in Pacific:
$ git tag --contain 48396a2a6effc09768e83e7b2709a0dc42d08199 v16.2.5 v16.2.6 v16.2.7 v16.2.8 v16.2.9
That means for all the Pacific version of:
v16.1.0
v16.2.0
v16.2.1
v16.2.2
v16.2.3
v16.2.4
We must disable the client metrics in kclient if users are using these version with newer kernels, or if they are using this ceph versions the couldn't see anything about the client metrics even for those they support.
Checked the downstream ceph versions, there are 159 tags gap between the bug and fix.
-- Xiubo
Quoted 397 lines · Xiubo Li
Jeff,
I think this still makes sense, more detail please see [1], which is as I worried for the older ceph and when users want to upgrade the cluster.
Though this is a MDS side bug and have already been fixed in the MDS, but if there are existing clusters running before that fix, so when upgrading the kclient only or if the upgrade the kclient first, they will complain the same issue in [1].
[1] https://tracker.ceph.com/issues/56529
-- Xiubo
Quoted 385 lines · Jeff Layton
Quoted 6 lines · On Thu, 2022-03-31 at 14:52 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
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.
Quoted 8 lines
Defautly the MDS won't fill the
s_metricsin 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.
Quoted 71 lines
URL: https://tracker.ceph.com/issues/54411
Signed-off-by Xiubo Li <[email protected]>---
fs/ceph/mds_client.c| 19 +++-
fs/ceph/mds_client.h| 1 +
fs/ceph/metric.c | 206
++++++++++++++++++++++++-------------------3 files changed, 131 insertions(+), 95 deletions(-)fs/ceph/mds_client.c +1
-1@@ -3422,7 +3422,7 @@ static void handle_session(structceph_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);fs/ceph/mds_client.c +1
-0@@ -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);fs/ceph/mds_client.h +0
-0@@ -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 longs_metrics;
u64 s_seq; /* incoming msg seq # */ struct mutex s_mutex; /* serialize session messages */ diff --gita/fs/ceph/metric.cb/fs/ceph/metric.c
indexc47347d2e84e..f01c1f4e6b89100644fs/ceph/metric.c static bool ceph_mdsc_send_metrics(struct · line 35
Quoted 5 lines
fs/ceph/metric.c +1
-0@@ -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.
Quoted 280 lines
fs/ceph/metric.c +0
-0struct ceph_msg *msg; s64 sum;s32 items = 0;
fs/ceph/metric.c +81
-69@@ -51,117 +52,140 @@ static bool ceph_mdsc_send_metrics(structceph_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 = (structceph_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 = (structceph_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 = (structceph_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 = (structceph_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 = (structceph_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;
PATCH 3/3 ceph: add force_ignore_metric_bits module parameter support
In reply to [email protected]From: Xiubo Li <[email protected]>
This parameter will force ignoring the metric bits from the MDS, and will force sending all the metrics kernel supports. It's dangerous for some old ceph clusters which will crash the MDSes when receive unknown metrics.
URL: https://tracker.ceph.com/issues/54411
--- fs/ceph/metric.c | 1 + fs/ceph/metric.h | 1 + fs/ceph/super.c | 6 +++++- 3 files changed, 7 insertions(+), 1 deletion(-)
fs/ceph/metric.c +1 -0
@@ -51,6 +51,7 @@ static bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc, head = msg->front.iov_base; + force = force || force_ignore_metric_bits;/* encode the cap metric */ if (force ||test_bit (CLIENT_METRIC_TYPE_CAP_INFO , &s->s_metrics)) { cap = (struct ceph_metric_cap *)(head +1 );
fs/ceph/metric.h +1 -0
@@ -7,6 +7,7 @@#include <linux/ktime.h> extern bool disable_send_metrics; +extern bool force_ignore_metric_bits;enum ceph_metric_type {CLIENT_METRIC_TYPE_CAP_INFO ,
fs/ceph/super.c +5 -1
@@ -1485,7 +1485,11 @@ static const struct kernel_param_ops param_ops_metrics = {bool disable_send_metrics = false;module_param_cb (disable_send_metrics, ¶m_ops_metrics, &disable_send_metrics,0644 ); -MODULE_PARM_DESC (disable_send_metrics,"Enable sending perf metrics to ceph cluster (default: on)" ); +MODULE_PARM_DESC (disable_send_metrics,"Disable sending perf metrics to ceph cluster (default: off)" ); + +bool force_ignore_metric_bits = false; +module_param_cb (force_ignore_metric_bits, ¶m_ops_bool, &force_ignore_metric_bits,0644 ); +MODULE_PARM_DESC (disable_send_metrics,"Force ignoring session's metric bits from MDS (default: off)" );/* for both v1 and v2 syntax */ static bool mount_support = true;
Signature · 1 line
2.27.0