flush the mdlog before waiting on unsafe reqs
Subsystems (2)
v2 cover flush the mdlog before waiting on unsafe reqs
From: Xiubo Li <[email protected]>
For the client requests who will have unsafe and safe replies from MDS daemons, in the MDS side the MDS daemons won't flush the mdlog (journal log) immediatelly, because they think it's unnecessary. That's true for most cases but not all, likes the fsync request. The fsync will wait until all the unsafe replied requests to be safely replied.
Normally if there have multiple threads or clients are running, the whole mdlog in MDS daemons could be flushed in time if any request will trigger the mdlog submit thread. So usually we won't experience the normal operations will stuck for a long time. But in case there has only one client with only thread is running, the stuck phenomenon maybe obvious and the worst case it must wait at most 5 seconds to wait the mdlog to be flushed by the MDS's tick thread periodically.
This patch will trigger to flush the mdlog in all the relevant and auth MDSes manually just before waiting the unsafe requests to finish.
Changed in V2:
- send mdlog flush request to unsafe req relevant and auth MDSes only instead of all of them.
- rename the first two commits' subject.
- fix the log messages.
- remove the feature bits fixing patch.
- fix some comments.
- remove flush_mdlog() wrapper.
- update the ceph_session_op_name() for new _REQUEST_FLUSH_MDLOG.
Xiubo Li (4):
ceph: make ceph_create_session_msg a global symbol
ceph: make iterate_sessions a global symbol
ceph: flush mdlog before umounting
ceph: flush the mdlog before waiting on unsafe reqs
fs/ceph/caps.c | 104 ++++++++++++++++++++++++++--------- fs/ceph/mds_client.c | 90 ++++++++++++++++++++++-------- fs/ceph/mds_client.h | 5 ++ fs/ceph/strings.c | 1 + include/linux/ceph/ceph_fs.h | 1 + 5 files changed, 152 insertions(+), 49 deletions(-)
Signature · 1 line
2.27.0
PATCH v2 1/4 ceph: make ceph_create_session_msg a global symbol
From: Xiubo Li <[email protected]>
--- fs/ceph/mds_client.c | 16 +++++++++------- fs/ceph/mds_client.h | 1 + 2 files changed, 10 insertions(+), 7 deletions(-)
fs/ceph/mds_client.c +9 -7
@@ -1150,7 +1150,7 @@ static int __choose_mds(struct ceph_mds_client *mdsc,/* * session messages */ -static struct ceph_msg *create_session_msg (u32 op,u64 seq) +struct ceph_msg *ceph_create_session_msg (u32 op,u64 seq) {struct ceph_msg *msg;struct ceph_mds_session_head *h;
@@ -1158,7 +1158,8 @@ static struct ceph_msg *create_session_msg(u32 op, u64 seq) msg =ceph_msg_new (CEPH_MSG_CLIENT_SESSION ,sizeof (*h),GFP_NOFS , false);if (!msg) { -pr_err ("create_session_msg ENOMEM creating msg\n" ); +pr_err ("ENOMEM creating session %s msg\n" , +ceph_session_op_name (op));return NULL ; } h = msg->front.iov_base;
@@ -1289,7 +1290,7 @@ static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u6 msg =ceph_msg_new (CEPH_MSG_CLIENT_SESSION ,sizeof (*h) + extra_bytes,GFP_NOFS , false);if (!msg) { -pr_err ("create_session_msg ENOMEM creating msg\n" ); +pr_err ("ENOMEM creating session open msg\n" );return ERR_PTR (-ENOMEM ); } p = msg->front.iov_base;
@@ -1801,8 +1802,8 @@ static int send_renew_caps(struct ceph_mds_client *mdsc,dout ("send_renew_caps to mds%d (%s)\n" , session->s_mds,ceph_mds_state_name (state)); - msg =create_session_msg (CEPH_SESSION_REQUEST_RENEWCAPS , - ++session->s_renew_seq); + msg =ceph_create_session_msg (CEPH_SESSION_REQUEST_RENEWCAPS , + ++session->s_renew_seq);if (!msg)return -ENOMEM ;ceph_con_send (&session->s_con, msg);
@@ -1816,7 +1817,7 @@ static int send_flushmsg_ack(struct ceph_mds_client *mdsc,dout ("send_flushmsg_ack to mds%d (%s)s seq %lld\n" , session->s_mds,ceph_session_state_name (session->s_state), seq); - msg =create_session_msg (CEPH_SESSION_FLUSHMSG_ACK , seq); + msg =ceph_create_session_msg (CEPH_SESSION_FLUSHMSG_ACK , seq);if (!msg)return -ENOMEM ;ceph_con_send (&session->s_con, msg);
@@ -1868,7 +1869,8 @@ static int request_close_session(struct ceph_mds_session *session)dout ("request_close_session mds%d state %s seq %lld\n" , session->s_mds,ceph_session_state_name (session->s_state), session->s_seq); - msg =create_session_msg (CEPH_SESSION_REQUEST_CLOSE , session->s_seq); + msg =ceph_create_session_msg (CEPH_SESSION_REQUEST_CLOSE , + session->s_seq);if (!msg)return -ENOMEM ;ceph_con_send (&session->s_con, msg);
fs/ceph/mds_client.h +1 -0
@@ -523,6 +523,7 @@ static inline void ceph_mdsc_put_request(struct ceph_mds_request *req)kref_put (&req->r_kref, ceph_mdsc_release_request); } +extern struct ceph_msg *ceph_create_session_msg (u32 op,u64 seq);extern void __ceph_queue_cap_release (struct ceph_mds_session *session,struct ceph_cap *cap); extern void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
Signature · 1 line
2.27.0
PATCH v2 2/4 ceph: make iterate_sessions a global symbol
In reply to [email protected]From: Xiubo Li <[email protected]>
--- fs/ceph/caps.c | 26 +---------------------- fs/ceph/mds_client.c | 49 +++++++++++++++++++++++++++++--------------- fs/ceph/mds_client.h | 3 +++ 3 files changed, 36 insertions(+), 42 deletions(-)
fs/ceph/caps.c +1 -25
@@ -4280,33 +4280,9 @@ static void flush_dirty_session_caps(struct ceph_mds_session *s)dout ("flush_dirty_caps done\n" ); } -static void iterate_sessions (struct ceph_mds_client *mdsc, -void (*cb)(struct ceph_mds_session *)) -{ -int mds; - -mutex_lock (&mdsc->mutex); -for (mds =0 ; mds < mdsc->max_sessions; ++mds) { -struct ceph_mds_session *s; - -if (!mdsc->sessions[mds]) -continue ; - - s =ceph_get_mds_session (mdsc->sessions[mds]); -if (!s) -continue ; - -mutex_unlock (&mdsc->mutex); -cb (s); -ceph_put_mds_session (s); -mutex_lock (&mdsc->mutex); - } -mutex_unlock (&mdsc->mutex); -} -void ceph_flush_dirty_caps (struct ceph_mds_client *mdsc) { -iterate_sessions (mdsc, flush_dirty_session_caps); +ceph_mdsc_iterate_sessions (mdsc, flush_dirty_session_caps, true); } void __ceph_touch_fmode(struct ceph_inode_info *ci,
fs/ceph/mds_client.c +32 -17
69 lines not on this page · raw message
fs/ceph/mds_client.h +3 -0
@@ -523,6 +523,9 @@ static inline void ceph_mdsc_put_request(struct ceph_mds_request *req)kref_put (&req->r_kref, ceph_mdsc_release_request); } +extern void ceph_mdsc_iterate_sessions (struct ceph_mds_client *mdsc, +void (*cb)(struct ceph_mds_session *), +bool check_state);extern struct ceph_msg *ceph_create_session_msg (u32 op,u64 seq);extern void __ceph_queue_cap_release (struct ceph_mds_session *session,struct ceph_cap *cap);
Signature · 1 line
2.27.0
PATCH v2 3/4 ceph: flush mdlog before umounting
In reply to [email protected]From: Xiubo Li <[email protected]>
--- fs/ceph/mds_client.c | 25 +++++++++++++++++++++++++ fs/ceph/mds_client.h | 1 + fs/ceph/strings.c | 1 + include/linux/ceph/ceph_fs.h | 1 + 4 files changed, 28 insertions(+)
fs/ceph/mds_client.c +25 -0
@@ -4690,6 +4690,30 @@ static void wait_requests(struct ceph_mds_client *mdsc)dout ("wait_requests done\n" ); } +void send_flush_mdlog (struct ceph_mds_session *s) +{ +struct ceph_msg *msg; + +/* +* Pre-luminous MDS crashes when it sees an unknown session request +*/ +if (!CEPH_HAVE_FEATURE (s->s_con.peer_features,SERVER_LUMINOUS )) +return ; + +mutex_lock (&s->s_mutex); +dout ("request mdlog flush to mds%d (%s)s seq %lld\n" , s->s_mds, +ceph_session_state_name (s->s_state), s->s_seq); + msg =ceph_create_session_msg (CEPH_SESSION_REQUEST_FLUSH_MDLOG , + s->s_seq); +if (!msg) { +pr_err ("failed to request mdlog flush to mds%d (%s) seq %lld\n" , + s->s_mds,ceph_session_state_name (s->s_state), s->s_seq); + }else { +ceph_con_send (&s->s_con, msg); + } +mutex_unlock (&s->s_mutex); +} +/* * called before mount is ro, and before dentries are torn down. * (hmm, does this still race with new lookups?)
@@ -4699,6 +4723,7 @@ void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)dout ("pre_umount\n" ); mdsc->stopping =1 ; +ceph_mdsc_iterate_sessions (mdsc, send_flush_mdlog, true);ceph_mdsc_iterate_sessions (mdsc, lock_unlock_session, false);ceph_flush_dirty_caps (mdsc);wait_requests (mdsc);
fs/ceph/mds_client.h +1 -0
@@ -523,6 +523,7 @@ static inline void ceph_mdsc_put_request(struct ceph_mds_request *req)kref_put (&req->r_kref, ceph_mdsc_release_request); } +extern void send_flush_mdlog (struct ceph_mds_session *s);extern void ceph_mdsc_iterate_sessions (struct ceph_mds_client *mdsc,void (*cb)(struct ceph_mds_session *),bool check_state);
fs/ceph/strings.c +1 -0
@@ -46,6 +46,7 @@ const char *ceph_session_op_name(int op)case CEPH_SESSION_FLUSHMSG_ACK :return "flushmsg_ack" ;case CEPH_SESSION_FORCE_RO :return "force_ro" ;case CEPH_SESSION_REJECT :return "reject" ; +case CEPH_SESSION_REQUEST_FLUSH_MDLOG :return "flush_mdlog" ; }return "???" ; }
include/linux/ceph/ceph_fs.h +1 -0
@@ -300,6 +300,7 @@ enum {CEPH_SESSION_FLUSHMSG_ACK ,CEPH_SESSION_FORCE_RO ,CEPH_SESSION_REJECT , + CEPH_SESSION_REQUEST_FLUSH_MDLOG, };extern const char *ceph_session_op_name (int op);
Signature · 1 line
2.27.0
PATCH v2 4/4 ceph: flush the mdlog before waiting on unsafe reqs
In reply to [email protected]From: Xiubo Li <[email protected]>
For the client requests who will have unsafe and safe replies from MDS daemons, in the MDS side the MDS daemons won't flush the mdlog (journal log) immediatelly, because they think it's unnecessary. That's true for most cases but not all, likes the fsync request. The fsync will wait until all the unsafe replied requests to be safely replied.
Normally if there have multiple threads or clients are running, the whole mdlog in MDS daemons could be flushed in time if any request will trigger the mdlog submit thread. So usually we won't experience the normal operations will stuck for a long time. But in case there has only one client with only thread is running, the stuck phenomenon maybe obvious and the worst case it must wait at most 5 seconds to wait the mdlog to be flushed by the MDS's tick thread periodically.
This patch will trigger to flush the mdlog in the relevant and auth MDSes to which the in-flight requests are sent just before waiting the unsafe requests to finish.
--- fs/ceph/caps.c | 78 ++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 78 insertions(+)
fs/ceph/caps.c +78 -0
@@ -2286,6 +2286,7 @@ static int caps_are_flushed(struct inode *inode, u64 flush_tid)*/ static int unsafe_request_wait (struct inode *inode) { +struct ceph_mds_client *mdsc =ceph_sb_to_client (inode->i_sb)->mdsc;struct ceph_inode_info *ci =ceph_inode (inode);struct ceph_mds_request *req1 =NULL , *req2 =NULL ;int ret, err =0 ;
@@ -2305,6 +2306,82 @@ static int unsafe_request_wait(struct inode *inode) }spin_unlock (&ci->i_unsafe_lock); +/* +* Trigger to flush the journal logs in all the relevant MDSes +* manually, or in the worst case we must wait at most 5 seconds +* to wait the journal logs to be flushed by the MDSes periodically. +*/ +if (req1 || req2) { +struct ceph_mds_session **sessions =NULL ; +struct ceph_mds_session *s; +struct ceph_mds_request *req; +unsigned int max; +int i; + +/* +* The mdsc->max_sessions is unlikely to be changed +* mostly, here we will retry it by reallocating the +* sessions arrary memory to get rid of the mdsc->mutex +* lock. +*/ +retry: + max = mdsc->max_sessions; + sessions =krealloc (sessions, max *sizeof (s), __GFP_ZERO); +if (!sessions) { + err = -ENOMEM ; + goto out; + } +spin_lock (&ci->i_unsafe_lock); +if (req1) { +list_for_each_entry (req, &ci->i_unsafe_dirops, + r_unsafe_dir_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s; + } + } + } +if (req2) { +list_for_each_entry (req, &ci->i_unsafe_iops, + r_unsafe_target_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s; + } + } + } +spin_unlock (&ci->i_unsafe_lock); + +/* the auth MDS */ +spin_lock (&ci->i_ceph_lock); +if (ci->i_auth_cap) { + s = ci->i_auth_cap->session; +if (!sessions[s->s_mds]) + sessions[s->s_mds] =ceph_get_mds_session (s); + } +spin_unlock (&ci->i_ceph_lock); + +/* send flush mdlog request to MDSes */ +for (i =0 ; i < max; i++) { + s = sessions[i]; +if (s) { +send_flush_mdlog (s); +ceph_put_mds_session (s); + } + } +kfree (sessions); + } +dout ("unsafe_request_wait %p wait on tid %llu %llu\n" , inode, req1 ? req1->r_tid :0ULL , req2 ? req2->r_tid :0ULL );if (req1) {
@@ -2321,6 +2398,7 @@ static int unsafe_request_wait(struct inode *inode) err = -EIO ;ceph_mdsc_put_request (req2); } +out:return err; }
Signature · 1 line
2.27.0
Re: [PATCH v2 4/4] ceph: flush the mdlog before waiting on unsafe reqs
Quoted 53 lines · On Mon, 2021-07-05 at 09:22 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
For the client requests who will have unsafe and safe replies from MDS daemons, in the MDS side the MDS daemons won't flush the mdlog (journal log) immediatelly, because they think it's unnecessary. That's true for most cases but not all, likes the fsync request. The fsync will wait until all the unsafe replied requests to be safely replied.
Normally if there have multiple threads or clients are running, the whole mdlog in MDS daemons could be flushed in time if any request will trigger the mdlog submit thread. So usually we won't experience the normal operations will stuck for a long time. But in case there has only one client with only thread is running, the stuck phenomenon maybe obvious and the worst case it must wait at most 5 seconds to wait the mdlog to be flushed by the MDS's tick thread periodically.
This patch will trigger to flush the mdlog in the relevant and auth MDSes to which the in-flight requests are sent just before waiting the unsafe requests to finish.
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/caps.c | 78 ++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 78 insertions(+)fs/ceph/caps.c +12
-0@@ -2286,6 +2286,7 @@ static int caps_are_flushed(struct inode *inode, u64 flush_tid)*/ static int unsafe_request_wait (struct inode *inode) { +struct ceph_mds_client *mdsc =ceph_sb_to_client (inode->i_sb)->mdsc;struct ceph_inode_info *ci =ceph_inode (inode);struct ceph_mds_request *req1 =NULL , *req2 =NULL ;int ret, err =0 ;@@ -2305,6 +2306,82 @@ static int unsafe_request_wait(struct inode *inode) }spin_unlock (&ci->i_unsafe_lock); +/* +* Trigger to flush the journal logs in all the relevant MDSes +* manually, or in the worst case we must wait at most 5 seconds +* to wait the journal logs to be flushed by the MDSes periodically. +*/ +if (req1 || req2) { +struct ceph_mds_session **sessions =NULL ; +struct ceph_mds_session *s; +struct ceph_mds_request *req; +unsigned int max; +int i;
fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2329
Quoted 10 lines
fs/ceph/caps.c +10
-0+ +/* +* The mdsc->max_sessions is unlikely to be changed +* mostly, here we will retry it by reallocating the +* sessions arrary memory to get rid of the mdsc->mutex +* lock. +*/ +retry: + max = mdsc->max_sessions; + sessions =krealloc (sessions, max *sizeof (s), __GFP_ZERO);
The kerneldoc over krealloc() says:
* The contents of the object pointed to are preserved up to the
* lesser of the new and old sizes (__GFP_ZERO flag is effectively
ignored).
This code however relies on krealloc zeroing out the new part of the allocation. Do you know for certain that that works?
Quoted 6 lines
fs/ceph/caps.c +6
-0+if (!sessions) { + err = -ENOMEM ; + goto out; + } +spin_lock (&ci->i_unsafe_lock); +if (req1) {
fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2345
Quoted 10 lines
fs/ceph/caps.c +10
-0+list_for_each_entry (req, &ci->i_unsafe_dirops, + r_unsafe_dir_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s;
nit: maybe just do:
sessions[s->s_mds] = ceph_get_mds_session(s);
Quoted 43 lines
fs/ceph/caps.c +39
-0+ } + } + } +if (req2) { +list_for_each_entry (req, &ci->i_unsafe_iops, + r_unsafe_target_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s; + } + } + } +spin_unlock (&ci->i_unsafe_lock); + +/* the auth MDS */ +spin_lock (&ci->i_ceph_lock); +if (ci->i_auth_cap) { + s = ci->i_auth_cap->session; +if (!sessions[s->s_mds]) + sessions[s->s_mds] =ceph_get_mds_session (s); + } +spin_unlock (&ci->i_ceph_lock); + +/* send flush mdlog request to MDSes */ +for (i =0 ; i < max; i++) { + s = sessions[i]; +if (s) { +send_flush_mdlog (s); +ceph_put_mds_session (s); + } + } +kfree (sessions); + } +dout ("unsafe_request_wait %p wait on tid %llu %llu\n" , inode, req1 ? req1->r_tid :0ULL , req2 ? req2->r_tid :0ULL );if (req1) {
fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2404
Quoted 7 lines
fs/ceph/caps.c +1
-0@@ -2321,6 +2398,7 @@ static int unsafe_request_wait(struct inode *inode) err = -EIO ;ceph_mdsc_put_request (req2); } +out:return err; }
Otherwise the whole set looks pretty reasonable.
Thanks,
Signature · 1 line
Jeff Layton <[email protected]>
Re: [PATCH v2 4/4] ceph: flush the mdlog before waiting on unsafe reqs
Quoted 66 lines · Jeff Layton
Quoted 53 lines · On Mon, 2021-07-05 at 09:22 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
For the client requests who will have unsafe and safe replies from MDS daemons, in the MDS side the MDS daemons won't flush the mdlog (journal log) immediatelly, because they think it's unnecessary. That's true for most cases but not all, likes the fsync request. The fsync will wait until all the unsafe replied requests to be safely replied.
Normally if there have multiple threads or clients are running, the whole mdlog in MDS daemons could be flushed in time if any request will trigger the mdlog submit thread. So usually we won't experience the normal operations will stuck for a long time. But in case there has only one client with only thread is running, the stuck phenomenon maybe obvious and the worst case it must wait at most 5 seconds to wait the mdlog to be flushed by the MDS's tick thread periodically.
This patch will trigger to flush the mdlog in the relevant and auth MDSes to which the in-flight requests are sent just before waiting the unsafe requests to finish.
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/caps.c | 78 ++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 78 insertions(+)fs/ceph/caps.c +12
-0@@ -2286,6 +2286,7 @@ static int caps_are_flushed(struct inode *inode, u64 flush_tid)*/ static int unsafe_request_wait (struct inode *inode) { +struct ceph_mds_client *mdsc =ceph_sb_to_client (inode->i_sb)->mdsc;struct ceph_inode_info *ci =ceph_inode (inode);struct ceph_mds_request *req1 =NULL , *req2 =NULL ;int ret, err =0 ;@@ -2305,6 +2306,82 @@ static int unsafe_request_wait(struct inode *inode) }spin_unlock (&ci->i_unsafe_lock); +/* +* Trigger to flush the journal logs in all the relevant MDSes +* manually, or in the worst case we must wait at most 5 seconds +* to wait the journal logs to be flushed by the MDSes periodically. +*/ +if (req1 || req2) { +struct ceph_mds_session **sessions =NULL ; +struct ceph_mds_session *s; +struct ceph_mds_request *req; +unsigned int max; +int i;fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2329
Quoted 10 lines
fs/ceph/caps.c +10
-0+ +/* +* The mdsc->max_sessions is unlikely to be changed +* mostly, here we will retry it by reallocating the +* sessions arrary memory to get rid of the mdsc->mutex +* lock. +*/ +retry: + max = mdsc->max_sessions; + sessions =krealloc (sessions, max *sizeof (s), __GFP_ZERO);The kerneldoc over
krealloc()says:
Quoted 6 lines
* The contents of the object pointed to are preserved up to the
* lesser of the new and old sizes (__GFP_ZEROflag is effectively
ignored).This code however relies on krealloc zeroing out the new part of the allocation. Do you know for certain that that works?
I readed the krealloc() code, the "__GFP_ZERO flag will be ignored" only for the preserved contents. If the slab really needs to allocate a new object, the slab will help zero it first and then copy the old contents to it, the new part will keep zeroed.
Quoted 18 lines
Quoted 6 lines
fs/ceph/caps.c +6
-0+if (!sessions) { + err = -ENOMEM ; + goto out; + } +spin_lock (&ci->i_unsafe_lock); +if (req1) {fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2345
Quoted 10 lines
fs/ceph/caps.c +10
-0+list_for_each_entry (req, &ci->i_unsafe_dirops, + r_unsafe_dir_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s;nit: maybe just do:
Quoted 1 line
sessions[s->s_mds] = ceph_get_mds_session(s);
Then it will exceed 80 chars for this line. Should we ignore it here ?
Thanks.
Quoted 55 lines
Quoted 45 lines
fs/ceph/caps.c +39
-0+ } + } + } +if (req2) { +list_for_each_entry (req, &ci->i_unsafe_iops, + r_unsafe_target_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s; + } + } + } +spin_unlock (&ci->i_unsafe_lock); + +/* the auth MDS */ +spin_lock (&ci->i_ceph_lock); +if (ci->i_auth_cap) { + s = ci->i_auth_cap->session; +if (!sessions[s->s_mds]) + sessions[s->s_mds] =ceph_get_mds_session (s); + } +spin_unlock (&ci->i_ceph_lock); + +/* send flush mdlog request to MDSes */ +for (i =0 ; i < max; i++) { + s = sessions[i]; +if (s) { +send_flush_mdlog (s); +ceph_put_mds_session (s); + } + } +kfree (sessions); + } +dout ("unsafe_request_wait %p wait on tid %llu %llu\n" , inode, req1 ? req1->r_tid :0ULL , req2 ? req2->r_tid :0ULL );if (req1) {fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2404
Quoted 7 lines
fs/ceph/caps.c +1
-0@@ -2321,6 +2398,7 @@ static int unsafe_request_wait(struct inode *inode) err = -EIO ;ceph_mdsc_put_request (req2); } +out:return err; }Otherwise the whole set looks pretty reasonable.
Thanks,
Re: [PATCH v2 4/4] ceph: flush the mdlog before waiting on unsafe reqs
Quoted 71 lines · Xiubo Li
Quoted 70 lines · Jeff Layton
Quoted 53 lines · On Mon, 2021-07-05 at 09:22 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
For the client requests who will have unsafe and safe replies from MDS daemons, in the MDS side the MDS daemons won't flush the mdlog (journal log) immediatelly, because they think it's unnecessary. That's true for most cases but not all, likes the fsync request. The fsync will wait until all the unsafe replied requests to be safely replied.
Normally if there have multiple threads or clients are running, the whole mdlog in MDS daemons could be flushed in time if any request will trigger the mdlog submit thread. So usually we won't experience the normal operations will stuck for a long time. But in case there has only one client with only thread is running, the stuck phenomenon maybe obvious and the worst case it must wait at most 5 seconds to wait the mdlog to be flushed by the MDS's tick thread periodically.
This patch will trigger to flush the mdlog in the relevant and auth MDSes to which the in-flight requests are sent just before waiting the unsafe requests to finish.
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/caps.c | 78 ++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 78 insertions(+)fs/ceph/caps.c +12
-0@@ -2286,6 +2286,7 @@ static int caps_are_flushed(struct inode *inode, u64 flush_tid)*/ static int unsafe_request_wait (struct inode *inode) { +struct ceph_mds_client *mdsc =ceph_sb_to_client (inode->i_sb)->mdsc;struct ceph_inode_info *ci =ceph_inode (inode);struct ceph_mds_request *req1 =NULL , *req2 =NULL ;int ret, err =0 ;@@ -2305,6 +2306,82 @@ static int unsafe_request_wait(struct inode *inode) }spin_unlock (&ci->i_unsafe_lock); +/* +* Trigger to flush the journal logs in all the relevant MDSes +* manually, or in the worst case we must wait at most 5 seconds +* to wait the journal logs to be flushed by the MDSes periodically. +*/ +if (req1 || req2) { +struct ceph_mds_session **sessions =NULL ; +struct ceph_mds_session *s; +struct ceph_mds_request *req; +unsigned int max; +int i;fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2329
Quoted 10 lines
fs/ceph/caps.c +10
-0+ +/* +* The mdsc->max_sessions is unlikely to be changed +* mostly, here we will retry it by reallocating the +* sessions arrary memory to get rid of the mdsc->mutex +* lock. +*/ +retry: + max = mdsc->max_sessions; + sessions =krealloc (sessions, max *sizeof (s), __GFP_ZERO);The kerneldoc over
krealloc()says:* The contents of the object pointed to are preserved up to the
* lesser of the new and old sizes (__GFP_ZEROflag is effectively
ignored).
Quoted 9 lines
Quoted 2 lines
This code however relies on krealloc zeroing out the new part of the allocation. Do you know for certain that that works?
I readed the
krealloc()code, the "__GFP_ZEROflag will be ignored" only for the preserved contents. If the slab really needs to allocate a new object, the slab will help zero it first and then copy the old contents to it, the new part will keep zeroed.
Ok, and in the case where it's an initial kmalloc, that will be done with __GFP_ZERO so any remaining space in the allocation will already be zeroed. That works.
Quoted 18 lines
Quoted 18 lines
Quoted 6 lines
fs/ceph/caps.c +6
-0+if (!sessions) { + err = -ENOMEM ; + goto out; + } +spin_lock (&ci->i_unsafe_lock); +if (req1) {fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2345
Quoted 10 lines
fs/ceph/caps.c +10
-0+list_for_each_entry (req, &ci->i_unsafe_dirops, + r_unsafe_dir_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s;nit: maybe just do:
Quoted 4 lines
Quoted 1 line
sessions[s->s_mds] = ceph_get_mds_session(s);Then it will exceed 80 chars for this line. Should we ignore it here ?
I probably would have but it's not worth respinning over all by itself.
It might also be possible to do all of this without taking the i_unsafe_lock twice, but that too probably won't make much difference.
I'll give these a closer look and probably merge into testing branch later today unless I see a problem.
Thanks!
Jeff
Quoted 55 lines
Quoted 54 lines
Quoted 44 lines
fs/ceph/caps.c +39
-0+ } + } + } +if (req2) { +list_for_each_entry (req, &ci->i_unsafe_iops, + r_unsafe_target_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s; + } + } + } +spin_unlock (&ci->i_unsafe_lock); + +/* the auth MDS */ +spin_lock (&ci->i_ceph_lock); +if (ci->i_auth_cap) { + s = ci->i_auth_cap->session; +if (!sessions[s->s_mds]) + sessions[s->s_mds] =ceph_get_mds_session (s); + } +spin_unlock (&ci->i_ceph_lock); + +/* send flush mdlog request to MDSes */ +for (i =0 ; i < max; i++) { + s = sessions[i]; +if (s) { +send_flush_mdlog (s); +ceph_put_mds_session (s); + } + } +kfree (sessions); + } +dout ("unsafe_request_wait %p wait on tid %llu %llu\n" , inode, req1 ? req1->r_tid :0ULL , req2 ? req2->r_tid :0ULL );if (req1) {fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2404
Quoted 7 lines
fs/ceph/caps.c +1
-0@@ -2321,6 +2398,7 @@ static int unsafe_request_wait(struct inode *inode) err = -EIO ;ceph_mdsc_put_request (req2); } +out:return err; }Otherwise the whole set looks pretty reasonable.
Thanks,
Signature · 1 line
Jeff Layton <[email protected]>
Re: [PATCH v2 4/4] ceph: flush the mdlog before waiting on unsafe reqs
Quoted 80 lines · Jeff Layton
Quoted 71 lines · Xiubo Li
Quoted 70 lines · Jeff Layton
Quoted 53 lines · On Mon, 2021-07-05 at 09:22 +0800, [email protected] wrote:
From: Xiubo Li <[email protected]>
For the client requests who will have unsafe and safe replies from MDS daemons, in the MDS side the MDS daemons won't flush the mdlog (journal log) immediatelly, because they think it's unnecessary. That's true for most cases but not all, likes the fsync request. The fsync will wait until all the unsafe replied requests to be safely replied.
Normally if there have multiple threads or clients are running, the whole mdlog in MDS daemons could be flushed in time if any request will trigger the mdlog submit thread. So usually we won't experience the normal operations will stuck for a long time. But in case there has only one client with only thread is running, the stuck phenomenon maybe obvious and the worst case it must wait at most 5 seconds to wait the mdlog to be flushed by the MDS's tick thread periodically.
This patch will trigger to flush the mdlog in the relevant and auth MDSes to which the in-flight requests are sent just before waiting the unsafe requests to finish.
Signed-off-by Xiubo Li <[email protected]>--- fs/ceph/caps.c | 78 ++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 78 insertions(+)fs/ceph/caps.c +12
-0@@ -2286,6 +2286,7 @@ static int caps_are_flushed(struct inode *inode, u64 flush_tid)*/ static int unsafe_request_wait (struct inode *inode) { +struct ceph_mds_client *mdsc =ceph_sb_to_client (inode->i_sb)->mdsc;struct ceph_inode_info *ci =ceph_inode (inode);struct ceph_mds_request *req1 =NULL , *req2 =NULL ;int ret, err =0 ;@@ -2305,6 +2306,82 @@ static int unsafe_request_wait(struct inode *inode) }spin_unlock (&ci->i_unsafe_lock); +/* +* Trigger to flush the journal logs in all the relevant MDSes +* manually, or in the worst case we must wait at most 5 seconds +* to wait the journal logs to be flushed by the MDSes periodically. +*/ +if (req1 || req2) { +struct ceph_mds_session **sessions =NULL ; +struct ceph_mds_session *s; +struct ceph_mds_request *req; +unsigned int max; +int i;fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2329
Quoted 10 lines
fs/ceph/caps.c +10
-0+ +/* +* The mdsc->max_sessions is unlikely to be changed +* mostly, here we will retry it by reallocating the +* sessions arrary memory to get rid of the mdsc->mutex +* lock. +*/ +retry: + max = mdsc->max_sessions; + sessions =krealloc (sessions, max *sizeof (s), __GFP_ZERO);The kerneldoc over
krealloc()says:* The contents of the object pointed to are preserved up to the
* lesser of the new and old sizes (__GFP_ZEROflag is effectively
ignored).Quoted 8 lines
Quoted 2 lines
This code however relies on krealloc zeroing out the new part of the allocation. Do you know for certain that that works?
I readed the
krealloc()code, the "__GFP_ZEROflag will be ignored" only for the preserved contents. If the slab really needs to allocate a new object, the slab will help zero it first and then copy the old contents to it, the new part will keep zeroed.
Quoted 3 lines
Ok, and in the case where it's an initial kmalloc, that will be done with
__GFP_ZEROso any remaining space in the allocation will already be zeroed. That works.
Yeah, it is.
Quoted 22 lines
Quoted 19 lines
Quoted 19 lines
Quoted 7 lines
fs/ceph/caps.c +6
-0+if (!sessions) { + err = -ENOMEM ; + goto out; + } +spin_lock (&ci->i_unsafe_lock); +if (req1) {fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2345
Quoted 10 lines
fs/ceph/caps.c +10
-0+list_for_each_entry (req, &ci->i_unsafe_dirops, + r_unsafe_dir_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s;nit: maybe just do:
Quoted 3 lines
Quoted 1 line
sessions[s->s_mds] = ceph_get_mds_session(s);Then it will exceed 80 chars for this line. Should we ignore it here ?
Quoted 7 lines
I probably would have but it's not worth respinning over all by itself.
It might also be possible to do all of this without taking the
i_unsafe_locktwice, but that too probably won't make much difference.I'll give these a closer look and probably merge into testing branch later today unless I see a problem.
Sure, thanks Jeff.
Quoted 57 lines
Thanks!
JeffQuoted 53 lines
Quoted 53 lines
Quoted 43 lines
fs/ceph/caps.c +39
-0+ } + } + } +if (req2) { +list_for_each_entry (req, &ci->i_unsafe_iops, + r_unsafe_target_item) { + s = req->r_session; +if (unlikely (s->s_mds > max)) { +spin_unlock (&ci->i_unsafe_lock); + goto retry; + } +if (!sessions[s->s_mds]) { + s =ceph_get_mds_session (s); + sessions[s->s_mds] = s; + } + } + } +spin_unlock (&ci->i_unsafe_lock); + +/* the auth MDS */ +spin_lock (&ci->i_ceph_lock); +if (ci->i_auth_cap) { + s = ci->i_auth_cap->session; +if (!sessions[s->s_mds]) + sessions[s->s_mds] =ceph_get_mds_session (s); + } +spin_unlock (&ci->i_ceph_lock); + +/* send flush mdlog request to MDSes */ +for (i =0 ; i < max; i++) { + s = sessions[i]; +if (s) { +send_flush_mdlog (s); +ceph_put_mds_session (s); + } + } +kfree (sessions); + } +dout ("unsafe_request_wait %p wait on tid %llu %llu\n" , inode, req1 ? req1->r_tid :0ULL , req2 ? req2->r_tid :0ULL );if (req1) {fs/ceph/caps.c static int unsafe_request_wait(struct inode *inode) · line 2404
Quoted 7 lines
fs/ceph/caps.c +1
-0@@ -2321,6 +2398,7 @@ static int unsafe_request_wait(struct inode *inode) err = -EIO ;ceph_mdsc_put_request (req2); } +out:return err; }Otherwise the whole set looks pretty reasonable.
Thanks,