nfsd: allocate/free session-based DRC slots on demand
Subsystems (2)
v2 cover nfsd: allocate/free session-based DRC slots on demand
Here is v2 of my series for on-demand allocation and freeing of session DRC slots.
- Now uses an xarray to store slots, and the limit is raised to 2048
- delays retiring a slot until the client has confirmed that it isn't using it as described in RFC:
The replier SHOULD retain the slots it wants to retire until the requester sends a request with a highest_slotid less than or equal to the replier's new enforced highest_slotid.
- When a retired slot is used, allow the seqid to be the next in sequence as required by the RFC:
Each time a slot is reused, the request MUST specify a sequence ID that is one greater than that of the previous request on the slot.
or "1" as (arguably) allowed by the RFC:
The first time a slot is used, the requester MUST specify a sequence ID of one
- current slot allocation is now reported in /proc/fs/nfsd/clients/*/info
This has been tested with highly aggressive shrinker settings:
nfsd_slot_shrinker->seeks = 0; nfsd_slot_shrinker->batch = 2;
and with periodic "echo 3 > drop_caches". The slot count drops as expected and then increases again.
NeilBrown
PATCH v2 1/6 nfsd: use an xarray to store v4.1 session slots
Using an xarray to store session slots will make it easier to change the number of active slots based on demand, and removes an unnecessary limit.
To achieve good throughput with a high-latency server it can be helpful to have hundreds of concurrent writes, which means hundreds of slots. So increase the limit to 2048 (twice what the Linux client will currently use).
--- fs/nfsd/nfs4state.c | 28 ++++++++++++++++++---------- fs/nfsd/state.h | 8 +++++--- 2 files changed, 23 insertions(+), 13 deletions(-)
fs/nfsd/nfs4state.c +18 -10
@@ -1915,8 +1915,11 @@ free_session_slots(struct nfsd4_session *ses)int i;for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { -free_svc_cred (&ses->se_slots[i]->sl_cred); -kfree (ses->se_slots[i]); +struct nfsd4_slot *slot =xa_load (&ses->se_slots, i); + +xa_erase (&ses->se_slots, i); +free_svc_cred (&slot->sl_cred); +kfree (slot); } }
@@ -1996,17 +1999,20 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs,struct nfsd4_session *new;int i; -BUILD_BUG_ON (struct_size (new, se_slots,NFSD_MAX_SLOTS_PER_SESSION ) - >PAGE_SIZE ); - - new =kzalloc (struct_size (new, se_slots, numslots),GFP_KERNEL ); + new =kzalloc (sizeof (*new),GFP_KERNEL );if (!new)return NULL ; +xa_init (&new->se_slots);/* allocate each struct nfsd4_slot and data cache in one piece */ for (i =0 ; i < numslots; i++) { - new->se_slots[i] =kzalloc (slotsize,GFP_KERNEL ); -if (!new->se_slots[i]) +struct nfsd4_slot *slot; + slot =kzalloc (slotsize,GFP_KERNEL ); +if (!slot) goto out_free; +if (xa_is_err (xa_store (&new->se_slots, i, slot,GFP_KERNEL ))) { +kfree (slot); + goto out_free; + } }memcpy (&new->se_fchannel, fattrs,sizeof (struct nfsd4_channel_attrs ));
@@ -2017,7 +2023,8 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs,return new; out_free:while (i--) -kfree (new->se_slots[i]); +kfree (xa_load (&new->se_slots, i)); +xa_destroy (&new->se_slots);kfree (new);return NULL ; }
@@ -2124,6 +2131,7 @@ static void nfsd4_del_conns(struct nfsd4_session *s)static void __free_session (struct nfsd4_session *ses) {free_session_slots (ses); +xa_destroy (&ses->se_slots);kfree (ses); }
@@ -4292,7 +4300,7 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,if (seq->slotid >= session->se_fchannel.maxreqs) goto out_put_session; - slot = session->se_slots[seq->slotid]; + slot =xa_load (&session->se_slots, seq->slotid);dprintk ("%s: slotid %d\n" , __func__, seq->slotid);/* We do not negotiate the number of slots yet, so set the
fs/nfsd/state.h +5 -3
@@ -245,8 +245,10 @@ static inline struct nfs4_delegation *delegstateid(struct nfs4_stid *s) return container_of(s, struct nfs4_delegation, dl_stid); } -/* Maximum number of slots per session. 160 is useful for long haul TCP */ -#define NFSD_MAX_SLOTS_PER_SESSION 160 +/* Maximum number of slots per session. A large number can be need to +* get good throughput on high-latency servers. +*/ +#define NFSD_MAX_SLOTS_PER_SESSION 2048/* Maximum session per slot cache size */ #define NFSD_SLOT_CACHE_SIZE 2048/* Maximum number of NFSD_SLOT_CACHE_SIZE slots per session */
@@ -345,7 +347,7 @@ struct nfsd4_session {struct nfsd4_cb_sec se_cb_sec;struct list_head se_conns;u32 se_cb_seq_nr[NFSD_BC_SLOT_TABLE_SIZE ]; -struct nfsd4_slot *se_slots[];/* forward channel slots */ +struct xarray se_slots;/* forward channel slots */ };/* formatted contents of nfs4_sessionid */
Signature · 1 line
2.47.0
PATCH v2 2/6 nfsd: remove artificial limits on the session-based DRC
In reply to NeilBrownRather than guessing how much space it might be safe to use for the DRC, simply try allocating slots and be prepared to accept failure.
The first slot for each session is allocated with GFP_KERNEL which is unlikely to fail. Subsequent slots are allocated with the addition of __GFP_NORETRY which is expected to fail if there isn't much free memory.
This is probably too aggressive but clears the way for adding a shrinker interface to free extra slots when memory is tight.
--- fs/nfsd/nfs4state.c | 91 +++++++-------------------------------------- fs/nfsd/nfsd.h | 3 -- fs/nfsd/nfssvc.c | 32 ---------------- 3 files changed, 14 insertions(+), 112 deletions(-)
fs/nfsd/nfs4state.c +14 -77
@@ -1938,65 +1938,13 @@ static inline u32 slot_bytes(struct nfsd4_channel_attrs *ca)return size +sizeof (struct nfsd4_slot ); } -/* -* XXX: If we run out of reserved DRC memory we could (up to a point) -* re-negotiate active sessions and reduce their slot usage to make -* room for new connections. For now we just fail the create session. -*/ -static u32 nfsd4_get_drc_mem (struct nfsd4_channel_attrs *ca,struct nfsd_net *nn) -{ -u32 slotsize =slot_bytes (ca); -u32 num = ca->maxreqs; -unsigned long avail, total_avail; -unsigned int scale_factor; - -spin_lock (&nfsd_drc_lock); -if (nfsd_drc_max_mem > nfsd_drc_mem_used) - total_avail = nfsd_drc_max_mem - nfsd_drc_mem_used; -else -/* We have handed out more space than we chose in -* set_max_drc() to allow. That isn't really a -* problem as long as that doesn't make us think we -* have lots more due to integer overflow. -*/ - total_avail =0 ; - avail =min ((unsigned long )NFSD_MAX_MEM_PER_SESSION , total_avail); -/* -* Never use more than a fraction of the remaining memory, -* unless it's the only way to give this client a slot. -* The chosen fraction is either 1/8 or 1/number of threads, -* whichever is smaller. This ensures there are adequate -* slots to support multiple clients per thread. -* Give the client one slot even if that would require -* over-allocation--it is better than failure. -*/ - scale_factor = max_t(unsigned int, 8, nn->nfsd_serv->sv_nrthreads); - - avail = clamp_t(unsigned long, avail, slotsize, - total_avail/scale_factor); - num =min_t (int, num, avail / slotsize); - num =max_t (int, num,1 ); - nfsd_drc_mem_used += num * slotsize; -spin_unlock (&nfsd_drc_lock); - -return num; -} - -static void nfsd4_put_drc_mem (struct nfsd4_channel_attrs *ca) -{ -int slotsize =slot_bytes (ca); - -spin_lock (&nfsd_drc_lock); - nfsd_drc_mem_used -= slotsize * ca->maxreqs; -spin_unlock (&nfsd_drc_lock); -} -static struct nfsd4_session *alloc_session (struct nfsd4_channel_attrs *fattrs,struct nfsd4_channel_attrs *battrs) {int numslots = fattrs->maxreqs;int slotsize =slot_bytes (fattrs);struct nfsd4_session *new; +struct nfsd4_slot *slot;int i; new =kzalloc (sizeof (*new),GFP_KERNEL );
@@ -2004,17 +1952,20 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs,return NULL ;xa_init (&new->se_slots);/* allocate each struct nfsd4_slot and data cache in one piece */ -for (i =0 ; i < numslots; i++) { -struct nfsd4_slot *slot; - slot =kzalloc (slotsize,GFP_KERNEL ); + slot =kzalloc (slotsize,GFP_KERNEL ); +if (!slot ||xa_is_err (xa_store (&new->se_slots,0 , slot,GFP_KERNEL ))) + goto out_free; + +for (i =1 ; i < numslots; i++) { + slot =kzalloc (slotsize,GFP_KERNEL | __GFP_NORETRY);if (!slot) - goto out_free; +break ;if (xa_is_err (xa_store (&new->se_slots, i, slot,GFP_KERNEL ))) {kfree (slot); - goto out_free; +break ; } } - + fattrs->maxreqs = i;memcpy (&new->se_fchannel, fattrs,sizeof (struct nfsd4_channel_attrs )); new->se_cb_slot_avail = ~0U ; new->se_cb_highest_slot = min(battrs->maxreqs - 1,
@@ -2022,8 +1973,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs,spin_lock_init (&new->se_lock);return new; out_free: -while (i--) -kfree (xa_load (&new->se_slots, i)); +kfree (slot);xa_destroy (&new->se_slots);kfree (new);return NULL ;
@@ -2138,7 +2088,6 @@ static void __free_session(struct nfsd4_session *ses)static void free_session (struct nfsd4_session *ses) {nfsd4_del_conns (ses); -nfsd4_put_drc_mem (&ses->se_fchannel);__free_session (ses); }
@@ -3800,17 +3749,6 @@ static __be32 check_forechannel_attrs(struct nfsd4_channel_attrs *ca, struct nfs ca->maxresp_cached =min_t (u32, ca->maxresp_cached,NFSD_SLOT_CACHE_SIZE +NFSD_MIN_HDR_SEQ_SZ ); ca->maxreqs =min_t (u32, ca->maxreqs,NFSD_MAX_SLOTS_PER_SESSION ); -/* -* Note decreasing slot size below client's request may make it -* difficult for client to function correctly, whereas -* decreasing the number of slots will (just?) affect -* performance. When short on memory we therefore prefer to -* decrease number of slots instead of their size. Clients that -* request larger slots than they need will get poor results: -* Note that we always allow at least one slot, because our -* accounting is soft and provides no guarantees either way. -*/ - ca->maxreqs =nfsd4_get_drc_mem (ca, nn);return nfs_ok; }
@@ -3888,11 +3826,11 @@ nfsd4_create_session(struct svc_rqst *rqstp,return status; status =check_backchannel_attrs (&cr_ses->back_channel);if (status) - goto out_release_drc_mem; + goto out_err; status = nfserr_jukebox; new =alloc_session (&cr_ses->fore_channel, &cr_ses->back_channel);if (!new) - goto out_release_drc_mem; + goto out_err; conn =alloc_conn_from_crses (rqstp, cr_ses);if (!conn) goto out_free_session;
@@ -4001,8 +3939,7 @@ nfsd4_create_session(struct svc_rqst *rqstp,free_conn (conn); out_free_session:__free_session (new); -out_release_drc_mem: -nfsd4_put_drc_mem (&cr_ses->fore_channel); +out_err:return status; }
fs/nfsd/nfsd.h +0 -3
@@ -88,9 +88,6 @@ struct nfsd_genl_rqstp {extern struct svc_program nfsd_programs[];extern const struct svc_version nfsd_version2, nfsd_version3, nfsd_version4;extern struct mutex nfsd_mutex; -extern spinlock_t nfsd_drc_lock; -extern unsigned long nfsd_drc_max_mem; -extern unsigned long nfsd_drc_mem_used;extern atomic_t nfsd_th_cnt;/* number of available threads */ extern const struct seq_operations nfs_exports_op;
fs/nfsd/nfssvc.c +0 -32
@@ -70,16 +70,6 @@ static __be32 nfsd_init_request(struct svc_rqst *,*/ DEFINE_MUTEX (nfsd_mutex); -/* -* nfsd_drc_lock protects nfsd_drc_max_pages and nfsd_drc_pages_used. -* nfsd_drc_max_pages limits the total amount of memory available for -* version 4.1 DRC caches. -* nfsd_drc_pages_used tracks the current version 4.1 DRC memory usage. -*/ -DEFINE_SPINLOCK (nfsd_drc_lock); -unsigned long nfsd_drc_max_mem; -unsigned long nfsd_drc_mem_used; -#if IS_ENABLED (CONFIG_NFS_LOCALIO )static const struct svc_version *localio_versions[] = { [1] = &localio_version1,
@@ -575,27 +565,6 @@ void nfsd_reset_versions(struct nfsd_net *nn) } } -/* -* Each session guarantees a negotiated per slot memory cache for replies -* which in turn consumes memory beyond the v2/v3/v4.0 server. A dedicated -* NFSv4.1 server might want to use more memory for a DRC than a machine -* with mutiple services. -* -* Impose a hard limit on the number of pages for the DRC which varies -* according to the machines free pages. This is of course only a default. -* -* For now this is a #defined shift which could be under admin control -* in the future. -*/ -static void set_max_drc (void ) -{ -#define NFSD_DRC_SIZE_SHIFT 7 - nfsd_drc_max_mem = (nr_free_buffer_pages () - >>NFSD_DRC_SIZE_SHIFT ) *PAGE_SIZE ; - nfsd_drc_mem_used =0 ; -dprintk ("%s nfsd_drc_max_mem %lu \n" , __func__, nfsd_drc_max_mem); -} -static int nfsd_get_default_max_blksize (void ) {struct sysinfo i;
@@ -678,7 +647,6 @@ int nfsd_create_serv(struct net *net) nn->nfsd_serv = serv;spin_unlock (&nfsd_notifier_lock); -set_max_drc ();/* check if the notifier is already set */ if (atomic_inc_return (&nfsd_notifier_refcount) ==1 ) {register_inetaddr_notifier (&nfsd_inetaddr_notifier);
Signature · 1 line
2.47.0
PATCH v2 3/6 nfsd: add session slot count to /proc/fs/nfsd/clients/*/info
In reply to NeilBrownEach client now reports the number of slots allocated in each session.
--- fs/nfsd/nfs4state.c | 8 ++++++++ 1 file changed, 8 insertions(+)
fs/nfsd/nfs4state.c +8 -0
@@ -2642,6 +2642,7 @@ static const char *cb_state2str(int state)static int client_info_show (struct seq_file *m,void *v) {struct inode *inode =file_inode (m->file); +struct nfsd4_session *ses;struct nfs4_client *clp;u64 clid;
@@ -2678,6 +2679,13 @@ static int client_info_show(struct seq_file *m, void *v)seq_printf (m,"callback address: \"%pISpc\"\n" , &clp->cl_cb_conn.cb_addr);seq_printf (m,"admin-revoked states: %d\n" ,atomic_read (&clp->cl_admin_revoked)); +seq_printf (m,"session slots:" ); +spin_lock (&clp->cl_lock); +list_for_each_entry (ses, &clp->cl_sessions, se_perclnt) +seq_printf (m," %u" , ses->se_fchannel.maxreqs); +spin_unlock (&clp->cl_lock); +seq_puts (m,"\n" ); +drop_client (clp);return 0 ;
Signature · 1 line
2.47.0
Re: [PATCH 3/6] nfsd: add session slot count to /proc/fs/nfsd/clients/*/info
Quoted 1 line · NeilBrown
Each client now reports the number of slots allocated in each session.
Can this file also report the target slot count? Ie, is the server matching the client's requested slot count, or is it over or under by some number?
Would it be useful for a server tester or administrator to poke a target slot count value into this file and watch the machinery adjust?
Quoted 34 lines
Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 8 ++++++++ 1 file changed, 8 insertions(+)fs/nfsd/nfs4state.c +8
-0@@ -2642,6 +2642,7 @@ static const char *cb_state2str(int state)static int client_info_show (struct seq_file *m,void *v) {struct inode *inode =file_inode (m->file); +struct nfsd4_session *ses;struct nfs4_client *clp;u64 clid;@@ -2678,6 +2679,13 @@ static int client_info_show(struct seq_file *m, void *v)seq_printf (m,"callback address: \"%pISpc\"\n" , &clp->cl_cb_conn.cb_addr);seq_printf (m,"admin-revoked states: %d\n" ,atomic_read (&clp->cl_admin_revoked)); +seq_printf (m,"session slots:" ); +spin_lock (&clp->cl_lock); +list_for_each_entry (ses, &clp->cl_sessions, se_perclnt) +seq_printf (m," %u" , ses->se_fchannel.maxreqs); +spin_unlock (&clp->cl_lock); +seq_puts (m,"\n" ); +drop_client (clp);return 0 ;Signature · 1 line
2.47.0
Signature · 1 line
Chuck Lever
Re: [PATCH 3/6] nfsd: add session slot count to /proc/fs/nfsd/clients/*/info
Quoted 6 lines · Chuck Lever
Quoted 1 line · NeilBrown
Each client now reports the number of slots allocated in each session.
Can this file also report the target slot count? Ie, is the server matching the client's requested slot count, or is it over or under by some number?
I could. Would you like to suggest a syntax?
Usually the numbers would be the same except for short transition periods, so I'm not convinced of the value.
Currently if the target is reduced while the client is idle there can be a longer delay before the slots are actually freed, but I think 2 lease-renewal SEQUENCE ops would do it. If/when we add use of the CB_RECALL_SLOT callback the delay should disappear.
Quoted 4 lines
Would it be useful for a server tester or administrator to poke a target slot count value into this file and watch the machinery adjust?
Maybe. By echo 3 > drop_caches does a pretty good job. I don't see that we need more.
Thanks,
NeilBrown
Re: [PATCH 3/6] nfsd: add session slot count to /proc/fs/nfsd/clients/*/info
Quoted 11 lines · NeilBrown
Quoted 6 lines · Chuck Lever
Quoted 1 line · NeilBrown
Each client now reports the number of slots allocated in each session.
Can this file also report the target slot count? Ie, is the server matching the client's requested slot count, or is it over or under by some number?
I could. Would you like to suggest a syntax?
Usually the numbers would be the same except for short transition periods, so I'm not convinced of the value.
That's precisely the kind of situation I would like to be able catch -- the two are unequal longer than expected.
Quoted 11 lines
Currently if the target is reduced while the client is idle there can be a longer delay before the slots are actually freed, but I think 2 lease-renewal SEQUENCE ops would do it. If/when we add use of the
CB_RECALL_SLOTcallback the delay should disappear.Quoted 3 lines
Would it be useful for a server tester or administrator to poke a target slot count value into this file and watch the machinery adjust?
Maybe. By echo 3 >
drop_cachesdoes a pretty good job. I don't see that we need more.
Fair enough.
Signature · 1 line
Chuck Lever
Re: [PATCH 3/6] nfsd: add session slot count to /proc/fs/nfsd/clients/*/info
In reply to NeilBrownQuoted 20 lines · NeilBrown
Each client now reports the number of slots allocated in each session.
Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 8 ++++++++ 1 file changed, 8 insertions(+)fs/nfsd/nfs4state.c +1
-0@@ -2642,6 +2642,7 @@ static const char *cb_state2str(int state)static int client_info_show (struct seq_file *m,void *v) {struct inode *inode =file_inode (m->file); +struct nfsd4_session *ses;struct nfs4_client *clp;u64 clid;
fs/nfsd/nfs4state.c static int client_info_show(struct seq_file *m, void *v) · line 2688
Quoted 10 lines
fs/nfsd/nfs4state.c +7
-0@@ -2678,6 +2679,13 @@ static int client_info_show(struct seq_file *m, void *v)seq_printf (m,"callback address: \"%pISpc\"\n" , &clp->cl_cb_conn.cb_addr);seq_printf (m,"admin-revoked states: %d\n" ,atomic_read (&clp->cl_admin_revoked)); +seq_printf (m,"session slots:" ); +spin_lock (&clp->cl_lock); +list_for_each_entry (ses, &clp->cl_sessions, se_perclnt) +seq_printf (m," %u" , ses->se_fchannel.maxreqs); +spin_unlock (&clp->cl_lock); +seq_puts (m,"\n" ); +
Also, I wonder if information about the backchannel session can be surfaced in this way?
Quoted 6 lines
fs/nfsd/nfs4state.c +0
-0drop_client (clp);return 0 ;Signature · 1 line
2.47.0
Signature · 1 line
Chuck Lever
Re: [PATCH 3/6] nfsd: add session slot count to /proc/fs/nfsd/clients/*/info
Quoted 20 lines · Chuck Lever
Quoted 19 lines · NeilBrown
Each client now reports the number of slots allocated in each session.
Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 8 ++++++++ 1 file changed, 8 insertions(+)fs/nfsd/nfs4state.c +1
-0@@ -2642,6 +2642,7 @@ static const char *cb_state2str(int state)static int client_info_show (struct seq_file *m,void *v) {struct inode *inode =file_inode (m->file); +struct nfsd4_session *ses;struct nfs4_client *clp;u64 clid;
Quoted 15 lines
fs/nfsd/nfs4state.c static int client_info_show(struct seq_file *m, void *v) · line 2688
Quoted 11 lines
fs/nfsd/nfs4state.c +7
-0@@ -2678,6 +2679,13 @@ static int client_info_show(struct seq_file *m, void *v)seq_printf (m,"callback address: \"%pISpc\"\n" , &clp->cl_cb_conn.cb_addr);seq_printf (m,"admin-revoked states: %d\n" ,atomic_read (&clp->cl_admin_revoked)); +seq_printf (m,"session slots:" ); +spin_lock (&clp->cl_lock); +list_for_each_entry (ses, &clp->cl_sessions, se_perclnt) +seq_printf (m," %u" , ses->se_fchannel.maxreqs); +spin_unlock (&clp->cl_lock); +seq_puts (m,"\n" ); +Also, I wonder if information about the backchannel session can be surfaced in this way?
Probably make sense. Maybe we should invent a syntax for reporting arbitrary info about each session.
session %d slots: %d
session %d cb-slots: %d
...
???
NeilBrown
Re: [PATCH 3/6] nfsd: add session slot count to /proc/fs/nfsd/clients/*/info
Quoted 37 lines · NeilBrown
Quoted 20 lines · Chuck Lever
Quoted 19 lines · NeilBrown
Each client now reports the number of slots allocated in each session.
Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 8 ++++++++ 1 file changed, 8 insertions(+)fs/nfsd/nfs4state.c +1
-0@@ -2642,6 +2642,7 @@ static const char *cb_state2str(int state)static int client_info_show (struct seq_file *m,void *v) {struct inode *inode =file_inode (m->file); +struct nfsd4_session *ses;struct nfs4_client *clp;u64 clid;Quoted 15 lines
fs/nfsd/nfs4state.c static int client_info_show(struct seq_file *m, void *v) · line 2688
Quoted 11 lines
fs/nfsd/nfs4state.c +7
-0@@ -2678,6 +2679,13 @@ static int client_info_show(struct seq_file *m, void *v)seq_printf (m,"callback address: \"%pISpc\"\n" , &clp->cl_cb_conn.cb_addr);seq_printf (m,"admin-revoked states: %d\n" ,atomic_read (&clp->cl_admin_revoked)); +seq_printf (m,"session slots:" ); +spin_lock (&clp->cl_lock); +list_for_each_entry (ses, &clp->cl_sessions, se_perclnt) +seq_printf (m," %u" , ses->se_fchannel.maxreqs); +spin_unlock (&clp->cl_lock); +seq_puts (m,"\n" ); +Also, I wonder if information about the backchannel session can be surfaced in this way?
Quoted 8 lines
Probably make sense. Maybe we should invent a syntax for reporting arbitrary info about each session.
session %d slots: %d
session %d cb-slots: %d
...???
If each client has a directory, then it should have a subdirectory called "sessions". Each subdirectory of "sessions" should be one session, named by its hex session ID (as it is presented by Wireshark). Each session directory could have a file for the forward channel, one for the backchannel, and maybe one for generic information like when the session was created and how many connections it has.
We don't need all of that in this patch set, but whatever is introduced here should be extensible to allow us to add more over time.
Signature · 1 line
Chuck Lever
Re: [PATCH 3/6] nfsd: add session slot count to /proc/fs/nfsd/clients/*/info
Quoted 55 lines · Chuck Lever
Quoted 37 lines · NeilBrown
Quoted 20 lines · Chuck Lever
Quoted 19 lines · NeilBrown
Each client now reports the number of slots allocated in each session.
Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 8 ++++++++ 1 file changed, 8 insertions(+)fs/nfsd/nfs4state.c +1
-0@@ -2642,6 +2642,7 @@ static const char *cb_state2str(int state)static int client_info_show (struct seq_file *m,void *v) {struct inode *inode =file_inode (m->file); +struct nfsd4_session *ses;struct nfs4_client *clp;u64 clid;Quoted 15 lines
fs/nfsd/nfs4state.c static int client_info_show(struct seq_file *m, void *v) · line 2688
Quoted 11 lines
fs/nfsd/nfs4state.c +7
-0@@ -2678,6 +2679,13 @@ static int client_info_show(struct seq_file *m, void *v)seq_printf (m,"callback address: \"%pISpc\"\n" , &clp->cl_cb_conn.cb_addr);seq_printf (m,"admin-revoked states: %d\n" ,atomic_read (&clp->cl_admin_revoked)); +seq_printf (m,"session slots:" ); +spin_lock (&clp->cl_lock); +list_for_each_entry (ses, &clp->cl_sessions, se_perclnt) +seq_printf (m," %u" , ses->se_fchannel.maxreqs); +spin_unlock (&clp->cl_lock); +seq_puts (m,"\n" ); +Also, I wonder if information about the backchannel session can be surfaced in this way?
Quoted 8 lines
Probably make sense. Maybe we should invent a syntax for reporting arbitrary info about each session.
session %d slots: %d
session %d cb-slots: %d
...???
If each client has a directory, then it should have a subdirectory called "sessions". Each subdirectory of "sessions" should be one session, named by its hex session ID (as it is presented by Wireshark). Each session directory could have a file for the forward channel, one for the backchannel, and maybe one for generic information like when the session was created and how many connections it has.
Quoted 3 lines
We don't need all of that in this patch set, but whatever is introduced here should be extensible to allow us to add more over time.
I cannot say I'm excited about the proliferation of tiny files. Your suggestion isn't quite as bad as sysfs which claims to want one file per value, but I think the sysfs approach provided more pain than gain and you seem to be heading that way. As evidence I present the rise of netlink. Netlink's main advantage is that it allows you to access a collection of data in a single syscall (or maybe pair of syscalls). If we had a standard format for doing that with open/read/close, the filesystem would be a much nicer interface. But the sysfs rules prevent that, so people who care avoid it.
We don't need to impose those same rules on nfsd-fs.
Having separate dirs for the clients makes some sense as the clients are quite independent. Sessions aren't - they are just part of the client. The *only* way session information is different from other client information is that there is more structure - an array of sessions each with detail. I don't think that justifies a new directory. I does justify a carefully designed (or chosen) format for representing structured data.
Thanks,
NeilBrown
Re: [PATCH 3/6] nfsd: add session slot count to /proc/fs/nfsd/clients/*/info
Quoted 64 lines · NeilBrown
Quoted 55 lines · Chuck Lever
Quoted 37 lines · NeilBrown
Quoted 20 lines · Chuck Lever
Quoted 19 lines · NeilBrown
Each client now reports the number of slots allocated in each session.
Signed-off-by NeilBrown <[email protected]>---
fs/nfsd/nfs4state.c| 8 ++++++++1 file changed, 8 insertions(+)fs/nfsd/nfs4state.c +1
-0@@ -2642,6 +2642,7 @@ static const char *cb_state2str(int state)static int client_info_show (struct seq_file *m,void *v) {struct inode *inode =file_inode (m->file); +struct nfsd4_session *ses;struct nfs4_client *clp;u64 clid;Quoted 15 lines
fs/nfsd/nfs4state.c static int client_info_show(struct seq_file *m, void *v) · line 2688
Quoted 11 lines
fs/nfsd/nfs4state.c +7
-0@@ -2678,6 +2679,13 @@ static int client_info_show(struct seq_file *m, void *v)seq_printf (m,"callback address: \"%pISpc\"\n" , &clp->cl_cb_conn.cb_addr);seq_printf (m,"admin-revoked states: %d\n" ,atomic_read (&clp->cl_admin_revoked)); +seq_printf (m,"session slots:" ); +spin_lock (&clp->cl_lock); +list_for_each_entry (ses, &clp->cl_sessions, se_perclnt) +seq_printf (m," %u" , ses->se_fchannel.maxreqs); +spin_unlock (&clp->cl_lock); +seq_puts (m,"\n" ); +Also, I wonder if information about the backchannel session can be surfaced in this way?
Quoted 8 lines
Probably make sense. Maybe we should invent a syntax for reporting arbitrary info about each session.
session %d slots: %d
session %d cb-slots: %d
...???
If each client has a directory, then it should have a subdirectory called "sessions". Each subdirectory of "sessions" should be one session, named by its hex session ID (as it is presented by Wireshark). Each session directory could have a file for the forward channel, one for the backchannel, and maybe one for generic information like when the session was created and how many connections it has.
Quoted 3 lines
We don't need all of that in this patch set, but whatever is introduced here should be extensible to allow us to add more over time.
Quoted 9 lines
I cannot say I'm excited about the proliferation of tiny files. Your suggestion isn't quite as bad as sysfs which claims to want one file per value, but I think the sysfs approach provided more pain than gain and you seem to be heading that way. As evidence I present the rise of netlink. Netlink's main advantage is that it allows you to access a collection of data in a single syscall (or maybe pair of syscalls). If we had a standard format for doing that with open/read/close, the filesystem would be a much nicer interface. But the sysfs rules prevent that, so people who care avoid it.
I don't see this set of information as being in a performance path. Needing multiple open/read/close iterations doesn't seem like an impediment to me.
The only possible issue is that user space might want a snapshot of certain related values, and having to get the values from multiple files means there's no guarantee that the values are consistent with each other.
Quoted 7 lines
We don't need to impose those same rules on nfsd-fs.
Having separate dirs for the clients makes some sense as the clients are quite independent. Sessions aren't - they are just part of the client. The *only* way session information is different from other client information is that there is more structure - an array of sessions each with detail. I don't think that justifies a new directory.
Hrm. IMHO a directory is exactly suited to this kind of information hierarchy. I can't say that I understand your view; perhaps you feel this way because the client implementations we are familiar with use only a single session. For that, of course, a directory is overkill.
Quoted 3 lines
It does
justify a carefully designed (or chosen) format for representing structured data.
That usually means JSON or XML, which also have their haters.
However, I don't feel strongly about this. You asked me for some thoughts, and here they are, at random.
My bottom line is reasonable extensibility -- the ability to provide more information in these files in the future without perturbing current consumers. IME that's been nearly impossible with designs that have one file full of fields that need to be parsed.
Should we expose session information via the new NFSD netlink protocol instead? Or a sessions/ directory with one formatted file per session? I'm open to discussion.
--
Chuck Lever
PATCH v2 4/6 nfsd: allocate new session-based DRC slots on demand.
In reply to NeilBrownIf a client ever uses the highest available slot for a given session, attempt to allocate another slot so there is room for the client to use more slots if wanted. GFP_NOWAIT is used so if there is not plenty of free memory, failure is expected - which is what we want. It also allows the allocation while holding a spinlock.
We would expect to stablise with one more slot available than the client actually uses.
Now that we grow the slot table on demand we can start with a smaller allocation. Define NFSD_MAX_INITIAL_SLOTS and allocate at most that many when session is created.
--- fs/nfsd/nfs4state.c | 32 ++++++++++++++++++++++++++------ fs/nfsd/state.h | 2 ++ 2 files changed, 28 insertions(+), 6 deletions(-)
fs/nfsd/nfs4state.c +26 -6
@@ -1956,7 +1956,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs,if (!slot ||xa_is_err (xa_store (&new->se_slots,0 , slot,GFP_KERNEL ))) goto out_free; -for (i =1 ; i < numslots; i++) { +for (i =1 ; i < numslots && i <NFSD_MAX_INITIAL_SLOTS ; i++) { slot =kzalloc (slotsize,GFP_KERNEL | __GFP_NORETRY);if (!slot)break ;
@@ -4248,11 +4248,6 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, slot =xa_load (&session->se_slots, seq->slotid);dprintk ("%s: slotid %d\n" , __func__, seq->slotid); -/* We do not negotiate the number of slots yet, so set the -* maxslots to the session maxreqs which is used to encode -* sr_highest_slotid and the sr_target_slot id to maxslots */ - seq->maxslots = session->se_fchannel.maxreqs; -trace_nfsd_slot_seqid_sequence (clp, seq, slot); status =check_slot_seqid (seq->seqid, slot->sl_seqid, slot->sl_flags &NFSD4_SLOT_INUSE );
@@ -4302,6 +4297,31 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, cstate->session = session; cstate->clp = clp; +/* +* If the client ever uses the highest available slot, +* gently try to allocate another one. +*/ +if (seq->slotid == session->se_fchannel.maxreqs -1 && + session->se_fchannel.maxreqs <NFSD_MAX_SLOTS_PER_SESSION ) { +int s = session->se_fchannel.maxreqs; + +/* +* GFP_NOWAIT is a low-priority non-blocking allocation +* which can be used under client_lock and only succeeds +* if there is plenty of memory. +* Use GFP_ATOMIC which is higher priority for xa_store() +* so we are less likely to waste the effort of the first +* allocation. +*/ + slot =kzalloc (slot_bytes (&session->se_fchannel),GFP_NOWAIT ); +if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot, +GFP_ATOMIC ))) + session->se_fchannel.maxreqs +=1 ; +else +kfree (slot); + } + seq->maxslots = session->se_fchannel.maxreqs; + out:switch (clp->cl_cb_state) {case NFSD4_CB_DOWN :
fs/nfsd/state.h +2 -0
@@ -249,6 +249,8 @@ static inline struct nfs4_delegation *delegstateid(struct nfs4_stid *s)* get good throughput on high-latency servers. */ #define NFSD_MAX_SLOTS_PER_SESSION 2048 +/* Maximum number of slots per session to allocate for CREATE_SESSION */ +#define NFSD_MAX_INITIAL_SLOTS 32/* Maximum session per slot cache size */ #define NFSD_SLOT_CACHE_SIZE 2048/* Maximum number of NFSD_SLOT_CACHE_SIZE slots per session */
Signature · 1 line
2.47.0
Re: [PATCH 4/6] nfsd: allocate new session-based DRC slots on demand.
Quoted 8 lines · NeilBrown
If a client ever uses the highest available slot for a given session, attempt to allocate another slot so there is room for the client to use more slots if wanted.
GFP_NOWAITis used so if there is not plenty of free memory, failure is expected - which is what we want. It also allows the allocation while holding a spinlock.We would expect to stablise with one more slot available than the client actually uses.
Which begs the question "why have a 2048 slot maximum session slot table size?" 1025 might work too. But is there a need for any maximum at all, or is this just a sanity check?
Quoted 3 lines
Now that we grow the slot table on demand we can start with a smaller allocation. Define
NFSD_MAX_INITIAL_SLOTSand allocate at most that many when session is created.
Maybe NFSD_DEFAULT_INITIAL_SLOTS is more descriptive?
Quoted 40 lines
Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 32 ++++++++++++++++++++++++++------ fs/nfsd/state.h | 2 ++ 2 files changed, 28 insertions(+), 6 deletions(-)fs/nfsd/nfs4state.c +6
-6@@ -1956,7 +1956,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs,if (!slot ||xa_is_err (xa_store (&new->se_slots,0 , slot,GFP_KERNEL ))) goto out_free; -for (i =1 ; i < numslots; i++) { +for (i =1 ; i < numslots && i <NFSD_MAX_INITIAL_SLOTS ; i++) { slot =kzalloc (slotsize,GFP_KERNEL | __GFP_NORETRY);if (!slot)break ;@@ -4248,11 +4248,6 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, slot =xa_load (&session->se_slots, seq->slotid);dprintk ("%s: slotid %d\n" , __func__, seq->slotid); -/* We do not negotiate the number of slots yet, so set the -* maxslots to the session maxreqs which is used to encode -* sr_highest_slotid and the sr_target_slot id to maxslots */ - seq->maxslots = session->se_fchannel.maxreqs; -trace_nfsd_slot_seqid_sequence (clp, seq, slot); status =check_slot_seqid (seq->seqid, slot->sl_seqid, slot->sl_flags &NFSD4_SLOT_INUSE );@@ -4302,6 +4297,31 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, cstate->session = session; cstate->clp = clp; +/* +* If the client ever uses the highest available slot, +* gently try to allocate another one. +*/ + if (seq->slotid == session->se_fchannel.maxreqs - 1 &&
fs/nfsd/nfs4state.c nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, · line 4314
Quoted 10 lines
fs/nfsd/nfs4state.c +10
-0+ session->se_fchannel.maxreqs < NFSD_MAX_SLOTS_PER_SESSION) { +int s = session->se_fchannel.maxreqs; + +/* +* GFP_NOWAIT is a low-priority non-blocking allocation +* which can be used under client_lock and only succeeds +* if there is plenty of memory. +* Use GFP_ATOMIC which is higher priority for xa_store() +* so we are less likely to waste the effort of the first +* allocation.
IIUC, GFP_ATOMIC allocations come from a special pool. I don't think we want that here. I'd rather stick with NORETRY or KERNEL.
Quoted 18 lines
fs/nfsd/nfs4state.c +10
-0+*/ + slot =kzalloc (slot_bytes (&session->se_fchannel),GFP_NOWAIT ); +if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot, +GFP_ATOMIC ))) + session->se_fchannel.maxreqs +=1 ; +else +kfree (slot); + } + seq->maxslots = session->se_fchannel.maxreqs; + out:switch (clp->cl_cb_state) {case NFSD4_CB_DOWN :fs/nfsd/state.h +0
-0
fs/nfsd/state.h static inline struct nfs4_delegation *delegstateid(struct nfs4_stid *s) · line 253
Quoted 5 lines
fs/nfsd/state.h +2
-0@@ -249,6 +249,8 @@ static inline struct nfs4_delegation *delegstateid(struct nfs4_stid *s)* get good throughput on high-latency servers. */ #define NFSD_MAX_SLOTS_PER_SESSION 2048 +/* Maximum number of slots per session to allocate for CREATE_SESSION */ +#define NFSD_MAX_INITIAL_SLOTS 32
The first couple of patches did so nicely at ruthlessly discarding a lot of arbitrary logic. I'm not convinced by the patch description that the INITIAL_SLOTS complexity is needed...
Quoted 6 lines
fs/nfsd/state.h +0
-0/* Maximum session per slot cache size */ #define NFSD_SLOT_CACHE_SIZE 2048/* Maximum number of NFSD_SLOT_CACHE_SIZE slots per session */ Signature · 1 line
2.47.0
Signature · 1 line
Chuck Lever
Re: [PATCH 4/6] nfsd: allocate new session-based DRC slots on demand.
Quoted 7 lines · Chuck Lever
Quoted 6 lines · NeilBrown
If a client ever uses the highest available slot for a given session, attempt to allocate another slot so there is room for the client to use more slots if wanted.
GFP_NOWAITis used so if there is not plenty of free memory, failure is expected - which is what we want. It also allows the allocation while holding a spinlock.
Quoted 6 lines
Quoted 2 lines
We would expect to stablise with one more slot available than the client actually uses.
Which begs the question "why have a 2048 slot maximum session slot table size?" 1025 might work too. But is there a need for any maximum at all, or is this just a sanity check?
Linux NFS presumably isn't the only client, and it might change in the future. Maybe there is no need for a maximum. It was mostly as a sanity check.
It wouldn't take much to convince me to remove the limit.
Quoted 7 lines
Quoted 5 lines
Now that we grow the slot table on demand we can start with a smaller allocation. Define
NFSD_MAX_INITIAL_SLOTSand allocate at most that many when session is created.Maybe
NFSD_DEFAULT_INITIAL_SLOTSis more descriptive?
I don't think "DEFAULT" is the right word. The client requests a number of slots. That is the "Default". The server can impose a limit - a maximum.
Maybe we don't need a limit here either?
Thanks,
NeilBrown
Quoted 96 lines
Quoted 42 lines
Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 32 ++++++++++++++++++++++++++------ fs/nfsd/state.h | 2 ++ 2 files changed, 28 insertions(+), 6 deletions(-)fs/nfsd/nfs4state.c +6
-6@@ -1956,7 +1956,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs,if (!slot ||xa_is_err (xa_store (&new->se_slots,0 , slot,GFP_KERNEL ))) goto out_free; -for (i =1 ; i < numslots; i++) { +for (i =1 ; i < numslots && i <NFSD_MAX_INITIAL_SLOTS ; i++) { slot =kzalloc (slotsize,GFP_KERNEL | __GFP_NORETRY);if (!slot)break ;@@ -4248,11 +4248,6 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, slot =xa_load (&session->se_slots, seq->slotid);dprintk ("%s: slotid %d\n" , __func__, seq->slotid); -/* We do not negotiate the number of slots yet, so set the -* maxslots to the session maxreqs which is used to encode -* sr_highest_slotid and the sr_target_slot id to maxslots */ - seq->maxslots = session->se_fchannel.maxreqs; -trace_nfsd_slot_seqid_sequence (clp, seq, slot); status =check_slot_seqid (seq->seqid, slot->sl_seqid, slot->sl_flags &NFSD4_SLOT_INUSE );@@ -4302,6 +4297,31 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, cstate->session = session; cstate->clp = clp; +/* +* If the client ever uses the highest available slot, +* gently try to allocate another one. +*/ + if (seq->slotid == session->se_fchannel.maxreqs - 1 &&fs/nfsd/nfs4state.c nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, · line 4314
Quoted 10 lines
fs/nfsd/nfs4state.c +10
-0+ session->se_fchannel.maxreqs < NFSD_MAX_SLOTS_PER_SESSION) { +int s = session->se_fchannel.maxreqs; + +/* +* GFP_NOWAIT is a low-priority non-blocking allocation +* which can be used under client_lock and only succeeds +* if there is plenty of memory. +* Use GFP_ATOMIC which is higher priority for xa_store() +* so we are less likely to waste the effort of the first +* allocation. IIUC,
GFP_ATOMICallocations come from a special pool. I don't think we want that here. I'd rather stick with NORETRY or KERNEL.Quoted 18 lines
fs/nfsd/nfs4state.c +10
-0+*/ + slot =kzalloc (slot_bytes (&session->se_fchannel),GFP_NOWAIT ); +if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot, +GFP_ATOMIC ))) + session->se_fchannel.maxreqs +=1 ; +else +kfree (slot); + } + seq->maxslots = session->se_fchannel.maxreqs; + out:switch (clp->cl_cb_state) {case NFSD4_CB_DOWN :fs/nfsd/state.h +0
-0fs/nfsd/state.h static inline struct nfs4_delegation *delegstateid(struct nfs4_stid *s) · line 253
Quoted 5 lines
fs/nfsd/state.h +2
-0@@ -249,6 +249,8 @@ static inline struct nfs4_delegation *delegstateid(struct nfs4_stid *s)* get good throughput on high-latency servers. */ #define NFSD_MAX_SLOTS_PER_SESSION 2048 +/* Maximum number of slots per session to allocate for CREATE_SESSION */ +#define NFSD_MAX_INITIAL_SLOTS 32The first couple of patches did so nicely at ruthlessly discarding a lot of arbitrary logic. I'm not convinced by the patch description that the
INITIAL_SLOTScomplexity is needed...Quoted 6 lines
fs/nfsd/state.h +0
-0/* Maximum session per slot cache size */ #define NFSD_SLOT_CACHE_SIZE 2048/* Maximum number of NFSD_SLOT_CACHE_SIZE slots per session */ Signature · 1 line
2.47.0Signature · 1 line
Chuck Lever
Re: [PATCH 4/6] nfsd: allocate new session-based DRC slots on demand.
Quoted 15 lines · NeilBrown
Quoted 7 lines · Chuck Lever
Quoted 6 lines · NeilBrown
If a client ever uses the highest available slot for a given session, attempt to allocate another slot so there is room for the client to use more slots if wanted.
GFP_NOWAITis used so if there is not plenty of free memory, failure is expected - which is what we want. It also allows the allocation while holding a spinlock.Quoted 6 lines
Quoted 2 lines
We would expect to stablise with one more slot available than the client actually uses.
Which begs the question "why have a 2048 slot maximum session slot table size?" 1025 might work too. But is there a need for any maximum at all, or is this just a sanity check?
Quoted 5 lines
Linux NFS presumably isn't the only client, and it might change in the future. Maybe there is no need for a maximum. It was mostly as a sanity check.
It wouldn't take much to convince me to remove the limit.
What's the worse that might happen if there is no cap? Can this be used as a DoS vector?
If a maximum should be necessary, its value should be clearly labeled as "not an architectural limit -- for sanity checking only".
Quoted 10 lines
Quoted 5 lines
Quoted 3 lines
Now that we grow the slot table on demand we can start with a smaller allocation. Define
NFSD_MAX_INITIAL_SLOTSand allocate at most that many when session is created.Maybe
NFSD_DEFAULT_INITIAL_SLOTSis more descriptive?I don't think "DEFAULT" is the right word. The client requests a number of slots. That is the "Default". The server can impose a limit - a maximum.
Maybe we don't need a limit here either?
I see. Well I don't think there needs to be a "maximum" number of initial slots. NFSD can try to allocate the number the client requested as best it can, until it hits our sane maximum above.
I think sessions should have a minimum number of slots to guarantee forward progress (or IOW prevent a deadlock). I would say that number should be larger than 1 -- perhaps 2 or even 4.
The problem with a small initial slot count is that means the session has a slow start heuristic. That might or might not be desirable here.
Signature · 1 line
Chuck Lever
Re: [PATCH 4/6] nfsd: allocate new session-based DRC slots on demand.
Quoted 16 lines · Chuck Lever
Quoted 15 lines · NeilBrown
Quoted 13 lines · Chuck Lever
Quoted 8 lines · NeilBrown
If a client ever uses the highest available slot for a given session, attempt to allocate another slot so there is room for the client to use more slots if wanted.
GFP_NOWAITis used so if there is not plenty of free memory, failure is expected - which is what we want. It also allows the allocation while holding a spinlock.We would expect to stablise with one more slot available than the client actually uses.
Which begs the question "why have a 2048 slot maximum session slot table size?" 1025 might work too. But is there a need for any maximum at all, or is this just a sanity check?
Quoted 8 lines
Quoted 5 lines
Linux NFS presumably isn't the only client, and it might change in the future. Maybe there is no need for a maximum. It was mostly as a sanity check.
It wouldn't take much to convince me to remove the limit.
What's the worse that might happen if there is no cap? Can this be used as a DoS vector?
It depends on how much you trust the clients that you have decided to trust. Probably we want the option of a "public" NFS server (read only probably) so we cannot assume much trust in the implementation of the client.
Certainly a client could only ever use the highest slot number available
- though the RFC prefers lowest - and that could push allocating through the roof. We could defend against that in more subtle ways, but a hard upper limit is easy.
Quoted 3 lines
If a maximum should be necessary, its value should be clearly labeled as "not an architectural limit -- for sanity checking only".
That is certainly sensible.
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Quoted 5 lines
Now that we grow the slot table on demand we can start with a smaller allocation. Define
NFSD_MAX_INITIAL_SLOTSand allocate at most that many when session is created.Maybe
NFSD_DEFAULT_INITIAL_SLOTSis more descriptive?
Quoted 8 lines
Quoted 4 lines
I don't think "DEFAULT" is the right word. The client requests a number of slots. That is the "Default". The server can impose a limit - a maximum.
Maybe we don't need a limit here either?I see. Well I don't think there needs to be a "maximum" number of initial slots. NFSD can try to allocate the number the client requested as best it can, until it hits our sane maximum above.
Given that we have a shrinker to discard them if they ever become a problem, that makes sense.
Quoted 4 lines
I think sessions should have a minimum number of slots to guarantee forward progress (or IOW prevent a deadlock). I would say that number should be larger than 1 -- perhaps 2 or even 4.
I think one is enough to ensure forward progress. Otherwise the RFC would have something to say about this.
Quoted 4 lines
The problem with a small initial slot count is that means the session has a slow start heuristic. That might or might not be desirable here.
The question of how quickly to increase slot count can be relevant at any time, not just at session creation time. If there is a bust of activity after a quite time during which the shrinker discarded a lot of slots - how quickly should we rebuild?
My current approach is effectively one new slot per requests round-trip. So there might be 1 request in flight. Then 2. Then 3. etc.
We could aim for exponential rather than linear growth. Maybe when the highest slot is used, add 20% of the current number of slots - rounded up.
So 1,2,3,4,5,6,8,10,12,15,18,22,26,31,37,44,52,62,74,88,105,126,
??
Thanks,
NeilBrown
Jeff Layton
maintainerRe: [PATCH 4/6] nfsd: allocate new session-based DRC slots on demand.
In reply to NeilBrownQuoted 24 lines · NeilBrown
If a client ever uses the highest available slot for a given session, attempt to allocate another slot so there is room for the client to use more slots if wanted.
GFP_NOWAITis used so if there is not plenty of free memory, failure is expected - which is what we want. It also allows the allocation while holding a spinlock.We would expect to stablise with one more slot available than the client actually uses.
Now that we grow the slot table on demand we can start with a smaller allocation. Define
NFSD_MAX_INITIAL_SLOTSand allocate at most that many when session is created.Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 32 ++++++++++++++++++++++++++------ fs/nfsd/state.h | 2 ++ 2 files changed, 28 insertions(+), 6 deletions(-)fs/nfsd/nfs4state.c +0
-0
fs/nfsd/nfs4state.c static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, · line 1959
Quoted 5 lines
fs/nfsd/nfs4state.c +1
-1@@ -1956,7 +1956,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs,if (!slot ||xa_is_err (xa_store (&new->se_slots,0 , slot,GFP_KERNEL ))) goto out_free; -for (i =1 ; i < numslots; i++) { +for (i =1 ; i < numslots && i <NFSD_MAX_INITIAL_SLOTS ; i++) {
nit: maybe just clamp numslots at NFSD_MAX_INITIAL_SLOTS?
Quoted 60 lines
fs/nfsd/nfs4state.c +25
-5slot =kzalloc (slotsize,GFP_KERNEL | __GFP_NORETRY);if (!slot)break ;@@ -4248,11 +4248,6 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, slot =xa_load (&session->se_slots, seq->slotid);dprintk ("%s: slotid %d\n" , __func__, seq->slotid); -/* We do not negotiate the number of slots yet, so set the -* maxslots to the session maxreqs which is used to encode -* sr_highest_slotid and the sr_target_slot id to maxslots */ - seq->maxslots = session->se_fchannel.maxreqs; -trace_nfsd_slot_seqid_sequence (clp, seq, slot); status =check_slot_seqid (seq->seqid, slot->sl_seqid, slot->sl_flags &NFSD4_SLOT_INUSE );@@ -4302,6 +4297,31 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, cstate->session = session; cstate->clp = clp; +/* +* If the client ever uses the highest available slot, +* gently try to allocate another one. +*/ +if (seq->slotid == session->se_fchannel.maxreqs -1 && + session->se_fchannel.maxreqs <NFSD_MAX_SLOTS_PER_SESSION ) { +int s = session->se_fchannel.maxreqs; + +/* +* GFP_NOWAIT is a low-priority non-blocking allocation +* which can be used under client_lock and only succeeds +* if there is plenty of memory. +* Use GFP_ATOMIC which is higher priority for xa_store() +* so we are less likely to waste the effort of the first +* allocation. +*/ + slot =kzalloc (slot_bytes (&session->se_fchannel),GFP_NOWAIT ); +if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot, +GFP_ATOMIC ))) + session->se_fchannel.maxreqs +=1 ; +else +kfree (slot); + } + seq->maxslots = session->se_fchannel.maxreqs; + out:switch (clp->cl_cb_state) {case NFSD4_CB_DOWN :fs/nfsd/state.h +2
-0@@ -249,6 +249,8 @@ static inline struct nfs4_delegation *delegstateid(struct nfs4_stid *s)* get good throughput on high-latency servers. */ #define NFSD_MAX_SLOTS_PER_SESSION 2048 +/* Maximum number of slots per session to allocate for CREATE_SESSION */ +#define NFSD_MAX_INITIAL_SLOTS 32/* Maximum session per slot cache size */ #define NFSD_SLOT_CACHE_SIZE 2048/* Maximum number of NFSD_SLOT_CACHE_SIZE slots per session */
Signature · 1 line
Jeff Layton <[email protected]>
PATCH v2 5/6 nfsd: add support for freeing unused session-DRC slots
In reply to NeilBrownReducing the number of slots in the session slot table requires confirmation from the client. This patch adds reduce_session_slots() which starts the process of getting confirmation, but never calls it. That will come in a later patch.
Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower cr_maxrequests in a SEQUENCE reply and then seeing a ca_maxrequests on the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reduced cr_maxrequests.
To achieve this we introduce a concept of request "generations". Each time we decide to reduce cr_maxrequests we increment the generation number, and record this when we return the lower cr_maxrequests to the client. When a slot with the current generation reports a low ca_maxrequests, we commit to that level and free extra slots.
We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
When decoding sa_highest_slotid into maxslots we need to add 1 - this matches how it is encoded for the reply.
--- fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)
fs/nfsd/nfs4state.c +69 -12
@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void -free_session_slots (struct nfsd4_session *ses) +free_session_slots (struct nfsd4_session *ses,int from) {int i; -for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { +if (from >= ses->se_fchannel.maxreqs) +return ; + +for (i = from; i < ses->se_fchannel.maxreqs; i++) {struct nfsd4_slot *slot =xa_load (&ses->se_slots, i); -xa_erase (&ses->se_slots, i); +/* +* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );free_svc_cred (&slot->sl_cred);kfree (slot); } + ses->se_fchannel.maxreqs = from; +if (ses->se_target_maxslots > from) + ses->se_target_maxslots = from; +} + +static int __maybe_unused +reduce_session_slots (struct nfsd4_session *ses,int dec) +{ +struct nfsd_net *nn =net_generic (ses->se_client->net, + nfsd_net_id); +int ret =0 ; + +if (ses->se_target_maxslots <=1 ) +return ret; +if (!spin_trylock (&nn->client_lock)) +return ret; + ret =min (dec, ses->se_target_maxslots-1 ); + ses->se_target_maxslots -= ret; + ses->se_slot_gen +=1 ; +if (ses->se_slot_gen ==0 ) { +int i; + ses->se_slot_gen =1 ; +for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { +struct nfsd4_slot *slot =xa_load (&ses->se_slots, i); + slot->sl_generation =0 ; + } + } +spin_unlock (&nn->client_lock); +return ret; }/*
@@ -1967,6 +2005,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, } fattrs->maxreqs = i;memcpy (&new->se_fchannel, fattrs,sizeof (struct nfsd4_channel_attrs )); + new->se_target_maxslots = i; new->se_cb_slot_avail = ~0U ; new->se_cb_highest_slot =min (battrs->maxreqs -1 ,NFSD_BC_SLOT_TABLE_SIZE -1 );
@@ -2080,7 +2119,7 @@ static void nfsd4_del_conns(struct nfsd4_session *s)static void __free_session (struct nfsd4_session *ses) { -free_session_slots (ses); +free_session_slots (ses,0 );xa_destroy (&ses->se_slots);kfree (ses); }
@@ -3687,10 +3726,10 @@ nfsd4_exchange_id_release(union nfsd4_op_u *u)kfree (exid->server_impl_name); } -static __be32 check_slot_seqid (u32 seqid,u32 slot_seqid,bool slot_inuse) +static __be32 check_slot_seqid (u32 seqid,u32 slot_seqid,u8 flags) {/* The slot is in use, and no response has been sent. */ -if (slot_inuse) { +if (flags &NFSD4_SLOT_INUSE ) {if (seqid == slot_seqid)return nfserr_jukebox;else
@@ -3699,6 +3738,8 @@ static __be32 check_slot_seqid(u32 seqid, u32 slot_seqid, bool slot_inuse)/* Note unsigned 32-bit arithmetic handles wraparound: */ if (likely (seqid == slot_seqid +1 ))return nfs_ok; +if ((flags &NFSD4_SLOT_REUSED ) && seqid ==1 ) +return nfs_ok;if (seqid == slot_seqid)return nfserr_replay_cache;return nfserr_seq_misordered;
@@ -4249,8 +4290,7 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,dprintk ("%s: slotid %d\n" , __func__, seq->slotid);trace_nfsd_slot_seqid_sequence (clp, seq, slot); - status =check_slot_seqid (seq->seqid, slot->sl_seqid, - slot->sl_flags &NFSD4_SLOT_INUSE ); + status =check_slot_seqid (seq->seqid, slot->sl_seqid, slot->sl_flags);if (status == nfserr_replay_cache) { status = nfserr_seq_misordered;if (!(slot->sl_flags &NFSD4_SLOT_INITIALIZED ))
@@ -4275,6 +4315,12 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,if (status) goto out_put_session; +if (session->se_target_maxslots < session->se_fchannel.maxreqs && + slot->sl_generation == session->se_slot_gen && + seq->maxslots <= session->se_target_maxslots) +/* Client acknowledged our reduce maxreqs */ +free_session_slots (session, session->se_target_maxslots); + buflen = (seq->cachethis) ? session->se_fchannel.maxresp_cached : session->se_fchannel.maxresp_sz;
@@ -4285,8 +4331,9 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,svc_reserve (rqstp, buflen); status = nfs_ok; -/* Success! bump slot seqid */ +/* Success! accept new slot seqid */ slot->sl_seqid = seq->seqid; + slot->sl_flags &= ~NFSD4_SLOT_REUSED ; slot->sl_flags |=NFSD4_SLOT_INUSE ;if (seq->cachethis) slot->sl_flags |=NFSD4_SLOT_CACHETHIS ;
@@ -4302,8 +4349,10 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,* gently try to allocate another one. */ if (seq->slotid == session->se_fchannel.maxreqs -1 && + session->se_target_maxslots >= session->se_fchannel.maxreqs && session->se_fchannel.maxreqs <NFSD_MAX_SLOTS_PER_SESSION ) {int s = session->se_fchannel.maxreqs; +void *prev_slot;/* * GFP_NOWAIT is a low-priority non-blocking allocation
@@ -4314,13 +4363,21 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,* allocation. */ slot =kzalloc (slot_bytes (&session->se_fchannel),GFP_NOWAIT ); + prev_slot =xa_load (&session->se_slots, s); +if (xa_is_value (prev_slot) && slot) { + slot->sl_seqid =xa_to_value (prev_slot); + slot->sl_flags |=NFSD4_SLOT_REUSED ; + }if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot, -GFP_ATOMIC ))) +GFP_ATOMIC ))) { session->se_fchannel.maxreqs +=1 ; -else + session->se_target_maxslots = session->se_fchannel.maxreqs; + }else {kfree (slot); + } } - seq->maxslots = session->se_fchannel.maxreqs; + seq->maxslots =max (session->se_target_maxslots, seq->maxslots); + seq->target_maxslots = session->se_target_maxslots; out:switch (clp->cl_cb_state) {
fs/nfsd/nfs4xdr.c +3 -2
@@ -1905,7 +1905,8 @@ nfsd4_decode_sequence(struct nfsd4_compoundargs *argp,return nfserr_bad_xdr; seq->seqid =be32_to_cpup (p++); seq->slotid =be32_to_cpup (p++); - seq->maxslots =be32_to_cpup (p++); +/* sa_highest_slotid counts from 0 but maxslots counts from 1 ... */ + seq->maxslots =be32_to_cpup (p++) +1 ; seq->cachethis =be32_to_cpup (p); seq->status_flags =0 ;
@@ -5054,7 +5055,7 @@ nfsd4_encode_sequence(struct nfsd4_compoundres *resp, __be32 nfserr,if (nfserr != nfs_ok)return nfserr;/* sr_target_highest_slotid */ - nfserr =nfsd4_encode_slotid4 (xdr, seq->maxslots -1 ); + nfserr =nfsd4_encode_slotid4 (xdr, seq->target_maxslots -1 );if (nfserr != nfs_ok)return nfserr;/* sr_status_flags */
fs/nfsd/state.h +4 -0
@@ -268,7 +268,9 @@ struct nfsd4_slot {#define NFSD4_SLOT_CACHETHIS (1 << 1)#define NFSD4_SLOT_INITIALIZED (1 << 2)#define NFSD4_SLOT_CACHED (1 << 3) +#define NFSD4_SLOT_REUSED (1 << 4)u8 sl_flags; +u8 sl_generation;char sl_data[]; };
@@ -350,6 +352,8 @@ struct nfsd4_session {struct list_head se_conns;u32 se_cb_seq_nr[NFSD_BC_SLOT_TABLE_SIZE ];struct xarray se_slots;/* forward channel slots */ +u8 se_slot_gen; +u32 se_target_maxslots; };/* formatted contents of nfs4_sessionid */
fs/nfsd/xdr4.h +0 -2
@@ -576,9 +576,7 @@ struct nfsd4_sequence {u32 slotid;/* request/response */ u32 maxslots;/* request/response */ u32 cachethis;/* request */ -#if 0 u32 target_maxslots;/* response */ -#endif /* not yet */ u32 status_flags;/* response */ };
Signature · 1 line
2.47.0
Re: [PATCH 5/6] nfsd: add support for freeing unused session-DRC slots
Quoted 56 lines · NeilBrown
Reducing the number of slots in the session slot table requires confirmation from the client. This patch adds
reduce_session_slots()which starts the process of getting confirmation, but never calls it. That will come in a later patch.Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower
cr_maxrequestsin a SEQUENCE reply and then seeing aca_maxrequestson the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reducedcr_maxrequests.To achieve this we introduce a concept of request "generations". Each time we decide to reduce
cr_maxrequestswe increment the generation number, and record this when we return the lowercr_maxrequeststo the client. When a slot with the current generation reports a lowca_maxrequests, we commit to that level and free extra slots.We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
When decoding
sa_highest_slotidinto maxslots we need to add 1 - this matches how it is encoded for the reply.Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)fs/nfsd/nfs4state.c +5
-2@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void -free_session_slots (struct nfsd4_session *ses) +free_session_slots (struct nfsd4_session *ses,int from) {int i; -for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { +if (from >= ses->se_fchannel.maxreqs) +return ; + +for (i = from; i < ses->se_fchannel.maxreqs; i++) {
fs/nfsd/nfs4state.c gen_sessionid(struct nfsd4_session *ses) · line 1929
Quoted 10 lines
fs/nfsd/nfs4state.c +7
-1struct nfsd4_slot *slot =xa_load (&ses->se_slots, i); -xa_erase (&ses->se_slots, i); +/* +* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );
Again... ATOMIC is probably not what we want here, even if it is only documentary.
And, I thought we determined that an unretired slot had a sequence number that is reset. Why save the slot's seqid? If I'm missing something, the comment here should be bolstered to explain it.
Quoted 207 lines
fs/nfsd/nfs4state.c +57
-9free_svc_cred (&slot->sl_cred);kfree (slot); } + ses->se_fchannel.maxreqs = from; +if (ses->se_target_maxslots > from) + ses->se_target_maxslots = from; +} + +static int __maybe_unused +reduce_session_slots (struct nfsd4_session *ses,int dec) +{ +struct nfsd_net *nn =net_generic (ses->se_client->net, + nfsd_net_id); +int ret =0 ; + +if (ses->se_target_maxslots <=1 ) +return ret; +if (!spin_trylock (&nn->client_lock)) +return ret; + ret =min (dec, ses->se_target_maxslots-1 ); + ses->se_target_maxslots -= ret; + ses->se_slot_gen +=1 ; +if (ses->se_slot_gen ==0 ) { +int i; + ses->se_slot_gen =1 ; +for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { +struct nfsd4_slot *slot =xa_load (&ses->se_slots, i); + slot->sl_generation =0 ; + } + } +spin_unlock (&nn->client_lock); +return ret; }/* @@ -1967,6 +2005,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, } fattrs->maxreqs = i;memcpy (&new->se_fchannel, fattrs,sizeof (struct nfsd4_channel_attrs )); + new->se_target_maxslots = i; new->se_cb_slot_avail = ~0U ; new->se_cb_highest_slot =min (battrs->maxreqs -1 ,NFSD_BC_SLOT_TABLE_SIZE -1 );@@ -2080,7 +2119,7 @@ static void nfsd4_del_conns(struct nfsd4_session *s)static void __free_session (struct nfsd4_session *ses) { -free_session_slots (ses); +free_session_slots (ses,0 );xa_destroy (&ses->se_slots);kfree (ses); }@@ -3687,10 +3726,10 @@ nfsd4_exchange_id_release(union nfsd4_op_u *u)kfree (exid->server_impl_name); } -static __be32 check_slot_seqid (u32 seqid,u32 slot_seqid,bool slot_inuse) +static __be32 check_slot_seqid (u32 seqid,u32 slot_seqid,u8 flags) {/* The slot is in use, and no response has been sent. */ -if (slot_inuse) { +if (flags &NFSD4_SLOT_INUSE ) {if (seqid == slot_seqid)return nfserr_jukebox;else @@ -3699,6 +3738,8 @@ static __be32 check_slot_seqid(u32 seqid, u32 slot_seqid, bool slot_inuse)/* Note unsigned 32-bit arithmetic handles wraparound: */ if (likely (seqid == slot_seqid +1 ))return nfs_ok; +if ((flags &NFSD4_SLOT_REUSED ) && seqid ==1 ) +return nfs_ok;if (seqid == slot_seqid)return nfserr_replay_cache;return nfserr_seq_misordered;@@ -4249,8 +4290,7 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,dprintk ("%s: slotid %d\n" , __func__, seq->slotid);trace_nfsd_slot_seqid_sequence (clp, seq, slot); - status =check_slot_seqid (seq->seqid, slot->sl_seqid, - slot->sl_flags &NFSD4_SLOT_INUSE ); + status =check_slot_seqid (seq->seqid, slot->sl_seqid, slot->sl_flags);if (status == nfserr_replay_cache) { status = nfserr_seq_misordered;if (!(slot->sl_flags &NFSD4_SLOT_INITIALIZED ))@@ -4275,6 +4315,12 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,if (status) goto out_put_session; +if (session->se_target_maxslots < session->se_fchannel.maxreqs && + slot->sl_generation == session->se_slot_gen && + seq->maxslots <= session->se_target_maxslots) +/* Client acknowledged our reduce maxreqs */ +free_session_slots (session, session->se_target_maxslots); + buflen = (seq->cachethis) ? session->se_fchannel.maxresp_cached : session->se_fchannel.maxresp_sz;@@ -4285,8 +4331,9 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,svc_reserve (rqstp, buflen); status = nfs_ok; -/* Success! bump slot seqid */ +/* Success! accept new slot seqid */ slot->sl_seqid = seq->seqid; + slot->sl_flags &= ~NFSD4_SLOT_REUSED ; slot->sl_flags |=NFSD4_SLOT_INUSE ;if (seq->cachethis) slot->sl_flags |=NFSD4_SLOT_CACHETHIS ;@@ -4302,8 +4349,10 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,* gently try to allocate another one. */ if (seq->slotid == session->se_fchannel.maxreqs -1 && + session->se_target_maxslots >= session->se_fchannel.maxreqs && session->se_fchannel.maxreqs <NFSD_MAX_SLOTS_PER_SESSION ) {int s = session->se_fchannel.maxreqs; +void *prev_slot;/* * GFP_NOWAIT is a low-priority non-blocking allocation @@ -4314,13 +4363,21 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,* allocation. */ slot =kzalloc (slot_bytes (&session->se_fchannel),GFP_NOWAIT ); + prev_slot =xa_load (&session->se_slots, s); +if (xa_is_value (prev_slot) && slot) { + slot->sl_seqid =xa_to_value (prev_slot); + slot->sl_flags |=NFSD4_SLOT_REUSED ; + }if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot, -GFP_ATOMIC ))) +GFP_ATOMIC ))) { session->se_fchannel.maxreqs +=1 ; -else + session->se_target_maxslots = session->se_fchannel.maxreqs; + }else {kfree (slot); + } } - seq->maxslots = session->se_fchannel.maxreqs; + seq->maxslots =max (session->se_target_maxslots, seq->maxslots); + seq->target_maxslots = session->se_target_maxslots; out:switch (clp->cl_cb_state) {fs/nfsd/nfs4xdr.c +3
-2@@ -1905,7 +1905,8 @@ nfsd4_decode_sequence(struct nfsd4_compoundargs *argp,return nfserr_bad_xdr; seq->seqid =be32_to_cpup (p++); seq->slotid =be32_to_cpup (p++); - seq->maxslots =be32_to_cpup (p++); +/* sa_highest_slotid counts from 0 but maxslots counts from 1 ... */ + seq->maxslots =be32_to_cpup (p++) +1 ; seq->cachethis =be32_to_cpup (p); seq->status_flags =0 ;@@ -5054,7 +5055,7 @@ nfsd4_encode_sequence(struct nfsd4_compoundres *resp, __be32 nfserr,if (nfserr != nfs_ok)return nfserr;/* sr_target_highest_slotid */ - nfserr =nfsd4_encode_slotid4 (xdr, seq->maxslots -1 ); + nfserr =nfsd4_encode_slotid4 (xdr, seq->target_maxslots -1 );if (nfserr != nfs_ok)return nfserr;/* sr_status_flags */ fs/nfsd/state.h +4
-0@@ -268,7 +268,9 @@ struct nfsd4_slot {#define NFSD4_SLOT_CACHETHIS (1 << 1)#define NFSD4_SLOT_INITIALIZED (1 << 2)#define NFSD4_SLOT_CACHED (1 << 3) +#define NFSD4_SLOT_REUSED (1 << 4)u8 sl_flags; +u8 sl_generation;char sl_data[]; };@@ -350,6 +352,8 @@ struct nfsd4_session {struct list_head se_conns;u32 se_cb_seq_nr[NFSD_BC_SLOT_TABLE_SIZE ];struct xarray se_slots;/* forward channel slots */ +u8 se_slot_gen; +u32 se_target_maxslots; };/* formatted contents of nfs4_sessionid */ fs/nfsd/xdr4.h +0
-2@@ -576,9 +576,7 @@ struct nfsd4_sequence {u32 slotid;/* request/response */ u32 maxslots;/* request/response */ u32 cachethis;/* request */ -#if 0 u32 target_maxslots;/* response */ -#endif /* not yet */ u32 status_flags;/* response */ };Signature · 1 line
2.47.0
Signature · 1 line
Chuck Lever
Re: [PATCH 5/6] nfsd: add support for freeing unused session-DRC slots
Quoted 59 lines · Chuck Lever
Quoted 58 lines · NeilBrown
Reducing the number of slots in the session slot table requires confirmation from the client. This patch adds
reduce_session_slots()which starts the process of getting confirmation, but never calls it. That will come in a later patch.Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower
cr_maxrequestsin a SEQUENCE reply and then seeing aca_maxrequestson the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reducedcr_maxrequests.To achieve this we introduce a concept of request "generations". Each time we decide to reduce
cr_maxrequestswe increment the generation number, and record this when we return the lowercr_maxrequeststo the client. When a slot with the current generation reports a lowca_maxrequests, we commit to that level and free extra slots.We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
When decoding
sa_highest_slotidinto maxslots we need to add 1 - this matches how it is encoded for the reply.Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)fs/nfsd/nfs4state.c +5
-2@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void -free_session_slots (struct nfsd4_session *ses) +free_session_slots (struct nfsd4_session *ses,int from) {int i; -for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { +if (from >= ses->se_fchannel.maxreqs) +return ; + +for (i = from; i < ses->se_fchannel.maxreqs; i++) {struct nfsd4_slot *slot =xa_load (&ses->se_slots, i);
Quoted 11 lines
fs/nfsd/nfs4state.c gen_sessionid(struct nfsd4_session *ses) · line 1929
Quoted 8 lines
fs/nfsd/nfs4state.c +7
-1-xa_erase (&ses->se_slots, i); +/* +* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );Again... ATOMIC is probably not what we want here, even if it is only documentary.
Why not? It might be called under a spinlock so GFP_KERNEL might trigger a warning.
Quoted 4 lines
And, I thought we determined that an unretired slot had a sequence number that is reset. Why save the slot's seqid? If I'm missing something, the comment here should be bolstered to explain it.
It isn't clear to me that we determined that - only the some people asserted it. Until the spec is clarified I think it is safest to be cautious.
Thanks,
NeilBrown
Re: [PATCH 5/6] nfsd: add support for freeing unused session-DRC slots
Quoted 72 lines · NeilBrown
Quoted 59 lines · Chuck Lever
Quoted 58 lines · NeilBrown
Reducing the number of slots in the session slot table requires confirmation from the client. This patch adds
reduce_session_slots()which starts the process of getting confirmation, but never calls it. That will come in a later patch.Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower
cr_maxrequestsin a SEQUENCE reply and then seeing aca_maxrequestson the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reducedcr_maxrequests.To achieve this we introduce a concept of request "generations". Each time we decide to reduce
cr_maxrequestswe increment the generation number, and record this when we return the lowercr_maxrequeststo the client. When a slot with the current generation reports a lowca_maxrequests, we commit to that level and free extra slots.We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
When decoding
sa_highest_slotidinto maxslots we need to add 1 - this matches how it is encoded for the reply.Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)fs/nfsd/nfs4state.c +5
-2@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void -free_session_slots (struct nfsd4_session *ses) +free_session_slots (struct nfsd4_session *ses,int from) {int i; -for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { +if (from >= ses->se_fchannel.maxreqs) +return ; + +for (i = from; i < ses->se_fchannel.maxreqs; i++) {struct nfsd4_slot *slot =xa_load (&ses->se_slots, i);Quoted 11 lines
fs/nfsd/nfs4state.c gen_sessionid(struct nfsd4_session *ses) · line 1929
Quoted 8 lines
fs/nfsd/nfs4state.c +7
-1-xa_erase (&ses->se_slots, i); +/* +* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );Again... ATOMIC is probably not what we want here, even if it is only documentary.
Quoted 2 lines
Why not? It might be called under a spinlock so
GFP_KERNELmight trigger a warning.
I find using GFP_ATOMIC here to be confusing -- it requests allocation from special memory reserves and is to be used in situations where allocation might result in system failure. That is clearly not the case here, and the resulting memory allocation might be long-lived.
I see the comment that says memory won't actually be allocated. I'm not sure that's the way xa_store() works, however.
I don't immediately see another good choice, however. I can reach out to Matthew and Liam and see if they have a better idea.
Quoted 6 lines
Quoted 3 lines
And, I thought we determined that an unretired slot had a sequence number that is reset. Why save the slot's seqid? If I'm missing something, the comment here should be bolstered to explain it.
It isn't clear to me that we determined that - only the some people asserted it.
From what I've read, everyone else who responded has said "use one". And they have provided enough spec quotations that 1 seems like the right initial slot sequence number value, always.
You should trust Tom Talpey's opinion on this. He was directly involved 25 years ago when sessions were invented in DAFS and then transferred into the NFSv4.1 protocol.
Quoted 1 line
Until the spec is clarified I think it is safest to be cautious.
The usual line we draw for adding code/features/complexity is the proposer must demonstrate a use case for it. So far I have not seen a client implementation that needs a server to remember the sequence number in a slot that has been shrunken and then re-activated.
Will this dead slot be subject to being freed by the session shrinker?
But the proposed implementation accepts 1 in this case, and it doesn't seem tremendously difficult to remove the "remember the seqid" mechanism once it has been codified to everyone's satisfaction. So I won't belabor the point.
Signature · 1 line
Chuck Lever
Re: [PATCH 5/6] nfsd: add support for freeing unused session-DRC slots
Quoted 73 lines · Chuck Lever
Quoted 72 lines · NeilBrown
Quoted 70 lines · Chuck Lever
Quoted 56 lines · NeilBrown
Reducing the number of slots in the session slot table requires confirmation from the client. This patch adds
reduce_session_slots()which starts the process of getting confirmation, but never calls it. That will come in a later patch.Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower
cr_maxrequestsin a SEQUENCE reply and then seeing aca_maxrequestson the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reducedcr_maxrequests.To achieve this we introduce a concept of request "generations". Each time we decide to reduce
cr_maxrequestswe increment the generation number, and record this when we return the lowercr_maxrequeststo the client. When a slot with the current generation reports a lowca_maxrequests, we commit to that level and free extra slots.We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
When decoding
sa_highest_slotidinto maxslots we need to add 1 - this matches how it is encoded for the reply.Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)fs/nfsd/nfs4state.c +5
-2@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void -free_session_slots (struct nfsd4_session *ses) +free_session_slots (struct nfsd4_session *ses,int from) {int i; -for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { +if (from >= ses->se_fchannel.maxreqs) +return ; + +for (i = from; i < ses->se_fchannel.maxreqs; i++) {fs/nfsd/nfs4state.c gen_sessionid(struct nfsd4_session *ses) · line 1929
Quoted 10 lines
fs/nfsd/nfs4state.c +7
-1struct nfsd4_slot *slot =xa_load (&ses->se_slots, i); -xa_erase (&ses->se_slots, i); +/* +* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );Again... ATOMIC is probably not what we want here, even if it is only documentary.
Quoted 8 lines
Quoted 2 lines
Why not? It might be called under a spinlock so
GFP_KERNELmight trigger a warning.I find using
GFP_ATOMIChere to be confusing -- it requests allocation from special memory reserves and is to be used in situations where allocation might result in system failure. That is clearly not the case here, and the resulting memory allocation might be long-lived.
Would you be comfortable with GFP_NOWAIT which leaves out __GFP_HIGH ??
My understanding of how GFP_ATOMIC is used is it is what people choose when they have to allocate in a no-sleep context. It can fail and there must always be a fall-back option. In many cases GFP_NOWAIT could possibly be used when it isn't a high priority, but there are 430 uses for GFP_NOWAIT compared with over 5000 of GFP_ATOMIC.
Quoted 3 lines
I see the comment that says memory won't actually be allocated. I'm not sure that's the way
xa_store()works, however.
xarray.rst says:
The xa_store(), xa_cmpxchg(), xa_alloc(),
xa_reserve() and xa_insert() functions take a gfp_t
parameter in case the XArray needs to allocate memory to store this entry. If the entry is being deleted, no memory allocation needs to be performed, and the GFP flags specified will be ignored.`
The particular context is that a normal pointer is currently stored a the given index, and we are replacing that with a number. The above doesn't explicitly say that won't require a memory allocation, but I think it is reasonable to say it won't need "to allocate memory to store this entry" - as an entry is already stored - so allocation should not be needed.
Quoted 12 lines
I don't immediately see another good choice, however. I can reach out to Matthew and Liam and see if they have a better idea.
Quoted 6 lines
Quoted 3 lines
And, I thought we determined that an unretired slot had a sequence number that is reset. Why save the slot's seqid? If I'm missing something, the comment here should be bolstered to explain it.
It isn't clear to me that we determined that - only the some people asserted it.
Quoted 7 lines
From what I've read, everyone else who responded has said "use one". And they have provided enough spec quotations that 1 seems like the right initial slot sequence number value, always.
You should trust Tom Talpey's opinion on this. He was directly involved 25 years ago when sessions were invented in DAFS and then transferred into the NFSv4.1 protocol.
Dave Noveck (also deeply involved) say:
It does. The problem is that it undercuts the core goal of the slot-based approach
In that it makes it possible to have multiple requests with the same session id/ slot ID / sequence triple.
i.e. resetting to 1 undercuts the core goal.... That is not a resounding endorsement.
While I respect the people, I prefer to trust reasoned arguments.
What exactly is the sequence number protecting against? It must be protecting against a request being sent, not reply received, connection closed, request sent on another connection, reply received. Original request getting to the server before the "close" message. Server must be sure not to handle this request. sequence number provides that.
But what if the above all happens for request with seqno of 1, then the server and client negotiate a "retiring" of slot 1 and then it's reuse before the original request arrives. How does the server know to ignore it?
And Tom said that the handling of maxreq etc is "optional" for both server and client. So how can the server know if the client has retired the slot when doing so is optional???
I really don't think we have clarity on this at all.
Quoted 8 lines
Quoted 3 lines
Until the spec is clarified I think it is safest to be cautious.
The usual line we draw for adding code/features/complexity is the proposer must demonstrate a use case for it. So far I have not seen a client implementation that needs a server to remember the sequence number in a slot that has been shrunken and then re-activated.
And I cannot in good faith submit code that I think violates the spec.
Quoted 3 lines
Will this dead slot be subject to being freed by the session shrinker?
No, but it uses much much less space than a slot.
Quoted 5 lines
But the proposed implementation accepts 1 in this case, and it doesn't seem tremendously difficult to remove the "remember the seqid" mechanism once it has been codified to everyone's satisfaction. So I won't belabor the point.
Thank you!
NeilBrown
Re: [PATCH 5/6] nfsd: add support for freeing unused session-DRC slots
Quoted 81 lines · NeilBrown
Quoted 76 lines · Chuck Lever
Quoted 74 lines · NeilBrown
Quoted 59 lines · Chuck Lever
Quoted 58 lines · NeilBrown
Reducing the number of slots in the session slot table requires confirmation from the client. This patch adds
reduce_session_slots()which starts the process of getting confirmation, but never calls it. That will come in a later patch.Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower
cr_maxrequestsin a SEQUENCE reply and then seeing aca_maxrequestson the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reducedcr_maxrequests.To achieve this we introduce a concept of request "generations". Each time we decide to reduce
cr_maxrequestswe increment the generation number, and record this when we return the lowercr_maxrequeststo the client. When a slot with the current generation reports a lowca_maxrequests, we commit to that level and free extra slots.We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
When decoding
sa_highest_slotidinto maxslots we need to add 1 - this matches how it is encoded for the reply.Signed-off-by NeilBrown <[email protected]>---
fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)fs/nfsd/nfs4state.c +0
-0@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void
fs/nfsd/nfs4state.c +1
-1-free_session_slots(struct nfsd4_session *ses) +free_session_slots(struct nfsd4_session *ses, int from){
int i;- for (i = 0; i <
ses->se_fchannel.maxreqs; i++) {
+ if (from >=ses->se_fchannel.maxreqs)
+ return;
+
+ for (i = from; i <ses->se_fchannel.maxreqs; i++) { structnfsd4_slot*slot =xa_load(&ses->se_slots, i);Quoted 11 lines
fs/nfsd/nfs4state.c gen_sessionid(struct nfsd4_session *ses) · line 1929
Quoted 8 lines
-
xa_erase(&ses->se_slots, i);
+ /*fs/nfsd/nfs4state.c +6
-0+* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );Again... ATOMIC is probably not what we want here, even if it is only documentary.
Why not? It might be called under a spinlock so
GFP_KERNELmight trigger a warning.
Quoted 7 lines
Quoted 5 lines
I find using
GFP_ATOMIChere to be confusing -- it requests allocation from special memory reserves and is to be used in situations where allocation might result in system failure. That is clearly not the case here, and the resulting memory allocation might be long-lived.Would you be comfortable with
GFP_NOWAITwhich leaves out__GFP_HIGH??
I will be comfortable when I hear back from Matthew and Liam.
:-)
Quoted 11 lines
Quoted 2 lines
I see the comment that says memory won't actually be allocated. I'm not sure that's the way
xa_store()works, however.
xarray.rstsays:The
xa_store(),xa_cmpxchg(),xa_alloc(),
xa_reserve()andxa_insert()functions take agfp_t
parameter in case the XArray needs to allocate memory to store this entry. If the entry is being deleted, no memory allocation needs to be performed, and the GFP flags specified will be ignored.`
Quoted 6 lines
The particular context is that a normal pointer is currently stored a the given index, and we are replacing that with a number. The above doesn't explicitly say that won't require a memory allocation, but I think it is reasonable to say it won't need "to allocate memory to store this entry" - as an entry is already stored - so allocation should not be needed.
xa_mk_value() converts a number to a pointer, and xa_store stores that pointer.
I suspect that xa_store() is allowed to rebalance the xarray's internal data structures, and that could result in memory release or allocation. That's why a GFP flag is one of the arguments.
--
Chuck Lever
Re: [PATCH 5/6] nfsd: add support for freeing unused session-DRC slots
Quoted 116 lines · Chuck Lever III
Quoted 82 lines · NeilBrown
Quoted 76 lines · Chuck Lever
Quoted 74 lines · NeilBrown
Quoted 59 lines · Chuck Lever
Quoted 58 lines · NeilBrown
Reducing the number of slots in the session slot table requires confirmation from the client. This patch adds
reduce_session_slots()which starts the process of getting confirmation, but never calls it. That will come in a later patch.Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower
cr_maxrequestsin a SEQUENCE reply and then seeing aca_maxrequestson the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reducedcr_maxrequests.To achieve this we introduce a concept of request "generations". Each time we decide to reduce
cr_maxrequestswe increment the generation number, and record this when we return the lowercr_maxrequeststo the client. When a slot with the current generation reports a lowca_maxrequests, we commit to that level and free extra slots.We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
When decoding
sa_highest_slotidinto maxslots we need to add 1 - this matches how it is encoded for the reply.Signed-off-by NeilBrown <[email protected]>---
fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)fs/nfsd/nfs4state.c +0
-0@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void
fs/nfsd/nfs4state.c +1
-1-free_session_slots(struct nfsd4_session *ses) +free_session_slots(struct nfsd4_session *ses, int from){
int i;- for (i = 0; i <
ses->se_fchannel.maxreqs; i++) {
+ if (from >=ses->se_fchannel.maxreqs)
+ return;
+
+ for (i = from; i <ses->se_fchannel.maxreqs; i++) { structnfsd4_slot*slot =xa_load(&ses->se_slots, i);Quoted 11 lines
fs/nfsd/nfs4state.c gen_sessionid(struct nfsd4_session *ses) · line 1929
Quoted 8 lines
-
xa_erase(&ses->se_slots, i);
+ /*fs/nfsd/nfs4state.c +6
-0+* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );Again... ATOMIC is probably not what we want here, even if it is only documentary.
Why not? It might be called under a spinlock so
GFP_KERNELmight trigger a warning.Quoted 7 lines
Quoted 5 lines
I find using
GFP_ATOMIChere to be confusing -- it requests allocation from special memory reserves and is to be used in situations where allocation might result in system failure. That is clearly not the case here, and the resulting memory allocation might be long-lived.Would you be comfortable with
GFP_NOWAITwhich leaves out__GFP_HIGH??I will be comfortable when I hear back from Matthew and Liam.
:-)
Quoted 11 lines
Quoted 2 lines
I see the comment that says memory won't actually be allocated. I'm not sure that's the way
xa_store()works, however.
xarray.rstsays:The
xa_store(),xa_cmpxchg(),xa_alloc(),xa_reserve()andxa_insert()functions take agfp_t
parameter in case the XArray needs to allocate memory to store this entry. If the entry is being deleted, no memory allocation needs to be performed, and the GFP flags specified will be ignored.`Quoted 6 lines
The particular context is that a normal pointer is currently stored a the given index, and we are replacing that with a number. The above doesn't explicitly say that won't require a memory allocation, but I think it is reasonable to say it won't need "to allocate memory to store this entry" - as an entry is already stored - so allocation should not be needed.
Quoted 7 lines
xa_mk_value()converts a number to a pointer, andxa_storestores that pointer.I suspect that
xa_store()is allowed to rebalance the xarray's internal data structures, and that could result in memory release or allocation. That's why a GFP flag is one of the arguments.
Matthew says the xa_store() is guaranteed not to do a memory allocation in this case. However, they prefer an annotation of the call site with a "0" GFP argument to show that the allocation flags are not relevant.
Does this:
xa_store(&ses->se_slots, i, xa_mk_value(slot->sl_seqid), 0);
work for you?
--
Chuck Lever
Re: [PATCH 5/6] nfsd: add support for freeing unused session-DRC slots
Quoted 129 lines · Chuck Lever III
Quoted 116 lines · Chuck Lever III
Quoted 82 lines · NeilBrown
Quoted 76 lines · Chuck Lever
Quoted 74 lines · NeilBrown
Quoted 59 lines · Chuck Lever
Quoted 58 lines · NeilBrown
Reducing the number of slots in the session slot table requires confirmation from the client. This patch adds
reduce_session_slots()which starts the process of getting confirmation, but never calls it. That will come in a later patch.Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower
cr_maxrequestsin a SEQUENCE reply and then seeing aca_maxrequestson the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reducedcr_maxrequests.To achieve this we introduce a concept of request "generations". Each time we decide to reduce
cr_maxrequestswe increment the generation number, and record this when we return the lowercr_maxrequeststo the client. When a slot with the current generation reports a lowca_maxrequests, we commit to that level and free extra slots.We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
When decoding
sa_highest_slotidinto maxslots we need to add 1 - this matches how it is encoded for the reply.Signed-off-by NeilBrown <[email protected]>---
fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)fs/nfsd/nfs4state.c +0
-0@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void
fs/nfsd/nfs4state.c +1
-1-free_session_slots(struct nfsd4_session *ses) +free_session_slots(struct nfsd4_session *ses, int from){
int i;- for (i = 0; i <
ses->se_fchannel.maxreqs; i++) {
+ if (from >=ses->se_fchannel.maxreqs)
+ return;
+
+ for (i = from; i <ses->se_fchannel.maxreqs; i++) { structnfsd4_slot*slot =xa_load(&ses->se_slots, i);Quoted 11 lines
fs/nfsd/nfs4state.c
Quoted 8 lines
-
xa_erase(&ses->se_slots, i);
+ /*fs/nfsd/nfs4state.c +6
-0+* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );Again... ATOMIC is probably not what we want here, even if it is only documentary.
Why not? It might be called under a spinlock so
GFP_KERNELmight trigger a warning.Quoted 7 lines
Quoted 5 lines
I find using
GFP_ATOMIChere to be confusing -- it requests allocation from special memory reserves and is to be used in situations where allocation might result in system failure. That is clearly not the case here, and the resulting memory allocation might be long-lived.Would you be comfortable with
GFP_NOWAITwhich leaves out__GFP_HIGH??I will be comfortable when I hear back from Matthew and Liam.
:-)
Quoted 11 lines
Quoted 2 lines
I see the comment that says memory won't actually be allocated. I'm not sure that's the way
xa_store()works, however.
xarray.rstsays:The
xa_store(),xa_cmpxchg(),xa_alloc(),xa_reserve()andxa_insert()functions take agfp_t
parameter in case the XArray needs to allocate memory to store this entry. If the entry is being deleted, no memory allocation needs to be performed, and the GFP flags specified will be ignored.`Quoted 6 lines
The particular context is that a normal pointer is currently stored a the given index, and we are replacing that with a number. The above doesn't explicitly say that won't require a memory allocation, but I think it is reasonable to say it won't need "to allocate memory to store this entry" - as an entry is already stored - so allocation should not be needed.
Quoted 7 lines
xa_mk_value()converts a number to a pointer, andxa_storestores that pointer.I suspect that
xa_store()is allowed to rebalance the xarray's internal data structures, and that could result in memory release or allocation. That's why a GFP flag is one of the arguments.Matthew says the
xa_store()is guaranteed not to do a memory allocation in this case. However, they prefer an annotation of the call site with a "0" GFP argument to show that the allocation flags are not relevant.
Quoted 5 lines
Does this:
xa_store(&ses->se_slots, i, xa_mk_value(slot->sl_seqid), 0);work for you?
Sure.
And it looks like sparse will be happy even though "0" isn't explicitly "gfp_t" because 0 is "special".
I might prefer GFP_NULL or similar, but 0 certainly works for me. I'll include that when I resend.
Thanks
NeilBrown
Re: [PATCH 5/6] nfsd: add support for freeing unused session-DRC slots
Quoted 138 lines · NeilBrown
Quoted 133 lines · Chuck Lever III
Quoted 116 lines · Chuck Lever III
Quoted 82 lines · NeilBrown
Quoted 76 lines · Chuck Lever
Quoted 74 lines · NeilBrown
Quoted 59 lines · Chuck Lever
Quoted 58 lines · NeilBrown
Reducing the number of slots in the session slot table requires confirmation from the client. This patch adds
reduce_session_slots()which starts the process of getting confirmation, but never calls it. That will come in a later patch.Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower
cr_maxrequestsin a SEQUENCE reply and then seeing aca_maxrequestson the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reducedcr_maxrequests.To achieve this we introduce a concept of request "generations". Each time we decide to reduce
cr_maxrequestswe increment the generation number, and record this when we return the lowercr_maxrequeststo the client. When a slot with the current generation reports a lowca_maxrequests, we commit to that level and free extra slots.We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
When decoding
sa_highest_slotidinto maxslots we need to add 1 - this matches how it is encoded for the reply.Signed-off-by NeilBrown <[email protected]>---
fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)fs/nfsd/nfs4state.c +0
-0@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void
fs/nfsd/nfs4state.c +1
-1-free_session_slots(struct nfsd4_session *ses) +free_session_slots(struct nfsd4_session *ses, int from){
int i;- for (i = 0; i <
ses->se_fchannel.maxreqs; i++) {
+ if (from >=ses->se_fchannel.maxreqs)
+ return;
+
+ for (i = from; i <ses->se_fchannel.maxreqs; i++) { structnfsd4_slot*slot =xa_load(&ses->se_slots, i);Quoted 11 lines
fs/nfsd/nfs4state.c
Quoted 8 lines
-
xa_erase(&ses->se_slots, i);
+ /*fs/nfsd/nfs4state.c +6
-0+* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );Again... ATOMIC is probably not what we want here, even if it is only documentary.
Why not? It might be called under a spinlock so
GFP_KERNELmight trigger a warning.Quoted 7 lines
Quoted 5 lines
I find using
GFP_ATOMIChere to be confusing -- it requests allocation from special memory reserves and is to be used in situations where allocation might result in system failure. That is clearly not the case here, and the resulting memory allocation might be long-lived.Would you be comfortable with
GFP_NOWAITwhich leaves out__GFP_HIGH??I will be comfortable when I hear back from Matthew and Liam.
:-)
Quoted 11 lines
Quoted 2 lines
I see the comment that says memory won't actually be allocated. I'm not sure that's the way
xa_store()works, however.
xarray.rstsays:The
xa_store(),xa_cmpxchg(),xa_alloc(),xa_reserve()andxa_insert()functions take agfp_t
parameter in case the XArray needs to allocate memory to store this entry. If the entry is being deleted, no memory allocation needs to be performed, and the GFP flags specified will be ignored.`Quoted 6 lines
The particular context is that a normal pointer is currently stored a the given index, and we are replacing that with a number. The above doesn't explicitly say that won't require a memory allocation, but I think it is reasonable to say it won't need "to allocate memory to store this entry" - as an entry is already stored - so allocation should not be needed.
Quoted 7 lines
xa_mk_value()converts a number to a pointer, andxa_storestores that pointer.I suspect that
xa_store()is allowed to rebalance the xarray's internal data structures, and that could result in memory release or allocation. That's why a GFP flag is one of the arguments.Matthew says the
xa_store()is guaranteed not to do a memory allocation in this case. However, they prefer an annotation of the call site with a "0" GFP argument to show that the allocation flags are not relevant.Does this:
xa_store(&ses->se_slots, i, xa_mk_value(slot->sl_seqid), 0);
Quoted 8 lines
Quoted 1 line
work for you?
Sure.
And it looks like sparse will be happy even though "0" isn't explicitly "gfp_t" because 0 is "special".I might prefer
GFP_NULLor similar, but 0 certainly works for me. I'll include that when I resend.
Matthew suggested GFP_NOALLOC. But neither of these symbolic flags exist yet. I'd rather not hold up this series behind the bikeshedding of the flag name ;-)
--
Chuck Lever
Jeff Layton
maintainerRe: [PATCH 5/6] nfsd: add support for freeing unused session-DRC slots
In reply to NeilBrownQuoted 20 lines · NeilBrown
Reducing the number of slots in the session slot table requires confirmation from the client. This patch adds
reduce_session_slots()which starts the process of getting confirmation, but never calls it. That will come in a later patch.Before we can free a slot we need to confirm that the client won't try to use it again. This involves returning a lower
cr_maxrequestsin a SEQUENCE reply and then seeing aca_maxrequestson the same slot which is not larger than we limit we are trying to impose. So for each slot we need to remember that we have sent a reducedcr_maxrequests.To achieve this we introduce a concept of request "generations". Each time we decide to reduce
cr_maxrequestswe increment the generation number, and record this when we return the lowercr_maxrequeststo the client. When a slot with the current generation reports a lowca_maxrequests, we commit to that level and free extra slots.We use an 8 bit generation number (64 seems wasteful) and if it cycles we iterate all slots and reset the generation number to avoid false matches.
Quoted 7 lines
When we free a slot we store the seqid in the slot pointer so that it can be restored when we reactivate the slot. The RFC can be read as suggesting that the slot number could restart from one after a slot is retired and reactivated, but also suggests that retiring slots is not required. So when we reactive a slot we accept with the next seqid in sequence, or 1.
Personally, I think that resetting to 1 is the only sane choice. After shrinking the slot table, either side is free to forget the slot information. When the slot is resurrected, we need to treat it as a new slot. Expecting the server to remember all seqids, and their cached replies for all slots ever used on a session seems like an open-ended mandate.
That said, I'm ok with the server being accepting here, in case there are client implementations that have done it the other way. Some clear guidance from the RFCs would sure be nice though.
Quoted 243 lines
When decoding
sa_highest_slotidinto maxslots we need to add 1 - this matches how it is encoded for the reply.Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++++++++++++++++------- fs/nfsd/nfs4xdr.c | 5 +-- fs/nfsd/state.h | 4 +++ fs/nfsd/xdr4.h | 2 -- 4 files changed, 76 insertions(+), 16 deletions(-)fs/nfsd/nfs4state.c +69
-12@@ -1910,17 +1910,55 @@ gen_sessionid(struct nfsd4_session *ses)#define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44)static void -free_session_slots (struct nfsd4_session *ses) +free_session_slots (struct nfsd4_session *ses,int from) {int i; -for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { +if (from >= ses->se_fchannel.maxreqs) +return ; + +for (i = from; i < ses->se_fchannel.maxreqs; i++) {struct nfsd4_slot *slot =xa_load (&ses->se_slots, i); -xa_erase (&ses->se_slots, i); +/* +* Save the seqid in case we reactivate this slot. +* This will never require a memory allocation so GFP +* flag is irrelevant +*/ +xa_store (&ses->se_slots, i,xa_mk_value (slot->sl_seqid), +GFP_ATOMIC );free_svc_cred (&slot->sl_cred);kfree (slot); } + ses->se_fchannel.maxreqs = from; +if (ses->se_target_maxslots > from) + ses->se_target_maxslots = from; +} + +static int __maybe_unused +reduce_session_slots (struct nfsd4_session *ses,int dec) +{ +struct nfsd_net *nn =net_generic (ses->se_client->net, + nfsd_net_id); +int ret =0 ; + +if (ses->se_target_maxslots <=1 ) +return ret; +if (!spin_trylock (&nn->client_lock)) +return ret; + ret =min (dec, ses->se_target_maxslots-1 ); + ses->se_target_maxslots -= ret; + ses->se_slot_gen +=1 ; +if (ses->se_slot_gen ==0 ) { +int i; + ses->se_slot_gen =1 ; +for (i =0 ; i < ses->se_fchannel.maxreqs; i++) { +struct nfsd4_slot *slot =xa_load (&ses->se_slots, i); + slot->sl_generation =0 ; + } + } +spin_unlock (&nn->client_lock); +return ret; }/* @@ -1967,6 +2005,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, } fattrs->maxreqs = i;memcpy (&new->se_fchannel, fattrs,sizeof (struct nfsd4_channel_attrs )); + new->se_target_maxslots = i; new->se_cb_slot_avail = ~0U ; new->se_cb_highest_slot =min (battrs->maxreqs -1 ,NFSD_BC_SLOT_TABLE_SIZE -1 );@@ -2080,7 +2119,7 @@ static void nfsd4_del_conns(struct nfsd4_session *s)static void __free_session (struct nfsd4_session *ses) { -free_session_slots (ses); +free_session_slots (ses,0 );xa_destroy (&ses->se_slots);kfree (ses); }@@ -3687,10 +3726,10 @@ nfsd4_exchange_id_release(union nfsd4_op_u *u)kfree (exid->server_impl_name); } -static __be32 check_slot_seqid (u32 seqid,u32 slot_seqid,bool slot_inuse) +static __be32 check_slot_seqid (u32 seqid,u32 slot_seqid,u8 flags) {/* The slot is in use, and no response has been sent. */ -if (slot_inuse) { +if (flags &NFSD4_SLOT_INUSE ) {if (seqid == slot_seqid)return nfserr_jukebox;else @@ -3699,6 +3738,8 @@ static __be32 check_slot_seqid(u32 seqid, u32 slot_seqid, bool slot_inuse)/* Note unsigned 32-bit arithmetic handles wraparound: */ if (likely (seqid == slot_seqid +1 ))return nfs_ok; +if ((flags &NFSD4_SLOT_REUSED ) && seqid ==1 ) +return nfs_ok;if (seqid == slot_seqid)return nfserr_replay_cache;return nfserr_seq_misordered;@@ -4249,8 +4290,7 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,dprintk ("%s: slotid %d\n" , __func__, seq->slotid);trace_nfsd_slot_seqid_sequence (clp, seq, slot); - status =check_slot_seqid (seq->seqid, slot->sl_seqid, - slot->sl_flags &NFSD4_SLOT_INUSE ); + status =check_slot_seqid (seq->seqid, slot->sl_seqid, slot->sl_flags);if (status == nfserr_replay_cache) { status = nfserr_seq_misordered;if (!(slot->sl_flags &NFSD4_SLOT_INITIALIZED ))@@ -4275,6 +4315,12 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,if (status) goto out_put_session; +if (session->se_target_maxslots < session->se_fchannel.maxreqs && + slot->sl_generation == session->se_slot_gen && + seq->maxslots <= session->se_target_maxslots) +/* Client acknowledged our reduce maxreqs */ +free_session_slots (session, session->se_target_maxslots); + buflen = (seq->cachethis) ? session->se_fchannel.maxresp_cached : session->se_fchannel.maxresp_sz;@@ -4285,8 +4331,9 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,svc_reserve (rqstp, buflen); status = nfs_ok; -/* Success! bump slot seqid */ +/* Success! accept new slot seqid */ slot->sl_seqid = seq->seqid; + slot->sl_flags &= ~NFSD4_SLOT_REUSED ; slot->sl_flags |=NFSD4_SLOT_INUSE ;if (seq->cachethis) slot->sl_flags |=NFSD4_SLOT_CACHETHIS ;@@ -4302,8 +4349,10 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,* gently try to allocate another one. */ if (seq->slotid == session->se_fchannel.maxreqs -1 && + session->se_target_maxslots >= session->se_fchannel.maxreqs && session->se_fchannel.maxreqs <NFSD_MAX_SLOTS_PER_SESSION ) {int s = session->se_fchannel.maxreqs; +void *prev_slot;/* * GFP_NOWAIT is a low-priority non-blocking allocation @@ -4314,13 +4363,21 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,* allocation. */ slot =kzalloc (slot_bytes (&session->se_fchannel),GFP_NOWAIT ); + prev_slot =xa_load (&session->se_slots, s); +if (xa_is_value (prev_slot) && slot) { + slot->sl_seqid =xa_to_value (prev_slot); + slot->sl_flags |=NFSD4_SLOT_REUSED ; + }if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot, -GFP_ATOMIC ))) +GFP_ATOMIC ))) { session->se_fchannel.maxreqs +=1 ; -else + session->se_target_maxslots = session->se_fchannel.maxreqs; + }else {kfree (slot); + } } - seq->maxslots = session->se_fchannel.maxreqs; + seq->maxslots =max (session->se_target_maxslots, seq->maxslots); + seq->target_maxslots = session->se_target_maxslots; out:switch (clp->cl_cb_state) {fs/nfsd/nfs4xdr.c +3
-2@@ -1905,7 +1905,8 @@ nfsd4_decode_sequence(struct nfsd4_compoundargs *argp,return nfserr_bad_xdr; seq->seqid =be32_to_cpup (p++); seq->slotid =be32_to_cpup (p++); - seq->maxslots =be32_to_cpup (p++); +/* sa_highest_slotid counts from 0 but maxslots counts from 1 ... */ + seq->maxslots =be32_to_cpup (p++) +1 ; seq->cachethis =be32_to_cpup (p); seq->status_flags =0 ;@@ -5054,7 +5055,7 @@ nfsd4_encode_sequence(struct nfsd4_compoundres *resp, __be32 nfserr,if (nfserr != nfs_ok)return nfserr;/* sr_target_highest_slotid */ - nfserr =nfsd4_encode_slotid4 (xdr, seq->maxslots -1 ); + nfserr =nfsd4_encode_slotid4 (xdr, seq->target_maxslots -1 );if (nfserr != nfs_ok)return nfserr;/* sr_status_flags */ fs/nfsd/state.h +4
-0@@ -268,7 +268,9 @@ struct nfsd4_slot {#define NFSD4_SLOT_CACHETHIS (1 << 1)#define NFSD4_SLOT_INITIALIZED (1 << 2)#define NFSD4_SLOT_CACHED (1 << 3) +#define NFSD4_SLOT_REUSED (1 << 4)u8 sl_flags; +u8 sl_generation;char sl_data[]; };@@ -350,6 +352,8 @@ struct nfsd4_session {struct list_head se_conns;u32 se_cb_seq_nr[NFSD_BC_SLOT_TABLE_SIZE ];struct xarray se_slots;/* forward channel slots */ +u8 se_slot_gen; +u32 se_target_maxslots; };/* formatted contents of nfs4_sessionid */ fs/nfsd/xdr4.h +0
-2@@ -576,9 +576,7 @@ struct nfsd4_sequence {u32 slotid;/* request/response */ u32 maxslots;/* request/response */ u32 cachethis;/* request */ -#if 0 u32 target_maxslots;/* response */ -#endif /* not yet */ u32 status_flags;/* response */ };
Signature · 1 line
Jeff Layton <[email protected]>
PATCH v2 6/6 nfsd: add shrinker to reduce number of slots allocated per session
In reply to NeilBrownAdd a shrinker which frees unused slots and may ask the clients to use fewer slots on each session.
Each session now tracks se_client_maxreqs which is the most recent max-requests-in-use reported by the client, and se_target_maxreqs which is a target number of requests which is reduced by the shrinker.
The shrinker iterates over all sessions on all client in all net-namespaces and reduces the target by 1 for each. The shrinker may get called multiple times to reduce by more than 1 each.
If se_target_maxreqs is above se_client_maxreqs, those slots can be freed immediately. If not the client will be ask to reduce its usage and as the usage goes down slots will be freed.
Once the usage has dropped to match the target, the target can be increased if the client uses all available slots and if a GFP_NOWAIT allocation succeeds.
--- fs/nfsd/nfs4state.c | 72 ++++++++++++++++++++++++++++++++++++++++++--- fs/nfsd/state.h | 1 + 2 files changed, 69 insertions(+), 4 deletions(-)
fs/nfsd/nfs4state.c +68 -4
@@ -1909,6 +1909,16 @@ gen_sessionid(struct nfsd4_session *ses)*/ #define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44) +static struct shrinker *nfsd_slot_shrinker; +static DEFINE_SPINLOCK (nfsd_session_list_lock); +static LIST_HEAD (nfsd_session_list); +/* The sum of "target_slots-1" on every session. The shrinker can push this +* down, though it can take a little while for the memory to actually +* be freed. The "-1" is because we can never free slot 0 while the +* session is active. +*/ +static atomic_t nfsd_total_target_slots =ATOMIC_INIT (0 ); +static void free_session_slots (struct nfsd4_session *ses,int from) {
@@ -1931,11 +1941,14 @@ free_session_slots(struct nfsd4_session *ses, int from)kfree (slot); } ses->se_fchannel.maxreqs = from; -if (ses->se_target_maxslots > from) - ses->se_target_maxslots = from; +if (ses->se_target_maxslots > from) { +int new_target = from ?:1 ; +atomic_sub (ses->se_target_maxslots - new_target, &nfsd_total_target_slots); + ses->se_target_maxslots = new_target; + } } -static int __maybe_unused +static int reduce_session_slots (struct nfsd4_session *ses,int dec) { struct nfsd_net *nn = net_generic(ses->se_client->net,
@@ -1948,6 +1961,7 @@ reduce_session_slots(struct nfsd4_session *ses, int dec)return ret; ret =min (dec, ses->se_target_maxslots-1 ); ses->se_target_maxslots -= ret; +atomic_sub (ret, &nfsd_total_target_slots); ses->se_slot_gen +=1 ;if (ses->se_slot_gen ==0 ) {int i;
@@ -2006,6 +2020,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, fattrs->maxreqs = i;memcpy (&new->se_fchannel, fattrs,sizeof (struct nfsd4_channel_attrs )); new->se_target_maxslots = i; +atomic_add (i -1 , &nfsd_total_target_slots); new->se_cb_slot_avail = ~0U ; new->se_cb_highest_slot =min (battrs->maxreqs -1 ,NFSD_BC_SLOT_TABLE_SIZE -1 );
@@ -2130,6 +2145,36 @@ static void free_session(struct nfsd4_session *ses)__free_session (ses); } +static unsigned long +nfsd_slot_count (struct shrinker *s,struct shrink_control *sc) +{ +unsigned long cnt =atomic_read (&nfsd_total_target_slots); + +return cnt ? cnt :SHRINK_EMPTY ; +} + +static unsigned long +nfsd_slot_scan (struct shrinker *s,struct shrink_control *sc) +{ +struct nfsd4_session *ses; +unsigned long scanned =0 ; +unsigned long freed =0 ; + +spin_lock (&nfsd_session_list_lock); +list_for_each_entry (ses, &nfsd_session_list, se_all_sessions) { + freed +=reduce_session_slots (ses,1 ); + scanned +=1 ; +if (scanned >= sc->nr_to_scan) { +/* Move starting point for next scan */ +list_move (&nfsd_session_list, &ses->se_all_sessions); +break ; + } + } +spin_unlock (&nfsd_session_list_lock); + sc->nr_scanned = scanned; +return freed; +} +static void init_session (struct svc_rqst *rqstp,struct nfsd4_session *new,struct nfs4_client *clp,struct nfsd4_create_session *cses) {int idx;
@@ -2154,6 +2199,10 @@ static void init_session(struct svc_rqst *rqstp, struct nfsd4_session *new, strulist_add (&new->se_perclnt, &clp->cl_sessions);spin_unlock (&clp->cl_lock); +spin_lock (&nfsd_session_list_lock); +list_add_tail (&new->se_all_sessions, &nfsd_session_list); +spin_unlock (&nfsd_session_list_lock); + {struct sockaddr *sa =svc_addr (rqstp);/*
@@ -2223,6 +2272,9 @@ unhash_session(struct nfsd4_session *ses)spin_lock (&ses->se_client->cl_lock);list_del (&ses->se_perclnt);spin_unlock (&ses->se_client->cl_lock); +spin_lock (&nfsd_session_list_lock); +list_del (&ses->se_all_sessions); +spin_unlock (&nfsd_session_list_lock); }/* SETCLIENTID and SETCLIENTID_CONFIRM Helper functions */
@@ -4335,6 +4387,7 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, slot->sl_seqid = seq->seqid; slot->sl_flags &= ~NFSD4_SLOT_REUSED ; slot->sl_flags |=NFSD4_SLOT_INUSE ; + slot->sl_generation = session->se_slot_gen;if (seq->cachethis) slot->sl_flags |=NFSD4_SLOT_CACHETHIS ;else
@@ -4371,6 +4424,8 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot,GFP_ATOMIC ))) { session->se_fchannel.maxreqs +=1 ; +atomic_add (session->se_fchannel.maxreqs - session->se_target_maxslots, + &nfsd_total_target_slots); session->se_target_maxslots = session->se_fchannel.maxreqs; }else {kfree (slot);
@@ -8779,7 +8834,6 @@ nfs4_state_start_net(struct net *net) }/* initialization to perform when the nfsd service is started: */ -int nfs4_state_start (void ) {
@@ -8789,6 +8843,15 @@ nfs4_state_start(void)if (ret)return ret; + nfsd_slot_shrinker =shrinker_alloc (0 ,"nfsd-DRC-slot" ); +if (!nfsd_slot_shrinker) { +rhltable_destroy (&nfs4_file_rhltable); +return -ENOMEM ; + } + nfsd_slot_shrinker->count_objects = nfsd_slot_count; + nfsd_slot_shrinker->scan_objects = nfsd_slot_scan; +shrinker_register (nfsd_slot_shrinker); +set_max_delegations ();return 0 ; }
@@ -8830,6 +8893,7 @@ voidnfs4_state_shutdown (void) {rhltable_destroy (&nfs4_file_rhltable); +shrinker_free (nfsd_slot_shrinker); } static void
fs/nfsd/state.h +1 -0
@@ -345,6 +345,7 @@ struct nfsd4_session {bool se_dead;struct list_head se_hash;/* hash by sessionid */ struct list_head se_perclnt; +struct list_head se_all_sessions;/* global list of sessions */ struct nfs4_client *se_client;struct nfs4_sessionid se_sessionid;struct nfsd4_channel_attrs se_fchannel;
Signature · 1 line
2.47.0
Re: [PATCH 6/6] nfsd: add shrinker to reduce number of slots allocated per session
Quoted 192 lines · NeilBrown
Add a shrinker which frees unused slots and may ask the clients to use fewer slots on each session.
Each session now tracks
se_client_maxreqswhich is the most recent max-requests-in-use reported by the client, andse_target_maxreqswhich is a target number of requests which is reduced by the shrinker.The shrinker iterates over all sessions on all client in all net-namespaces and reduces the target by 1 for each. The shrinker may get called multiple times to reduce by more than 1 each.
If
se_target_maxreqsis abovese_client_maxreqs, those slots can be freed immediately. If not the client will be ask to reduce its usage and as the usage goes down slots will be freed.Once the usage has dropped to match the target, the target can be increased if the client uses all available slots and if a
GFP_NOWAITallocation succeeds.Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 72 ++++++++++++++++++++++++++++++++++++++++++--- fs/nfsd/state.h | 1 + 2 files changed, 69 insertions(+), 4 deletions(-)fs/nfsd/nfs4state.c +68
-4@@ -1909,6 +1909,16 @@ gen_sessionid(struct nfsd4_session *ses)*/ #define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44) +static struct shrinker *nfsd_slot_shrinker; +static DEFINE_SPINLOCK (nfsd_session_list_lock); +static LIST_HEAD (nfsd_session_list); +/* The sum of "target_slots-1" on every session. The shrinker can push this +* down, though it can take a little while for the memory to actually +* be freed. The "-1" is because we can never free slot 0 while the +* session is active. +*/ +static atomic_t nfsd_total_target_slots =ATOMIC_INIT (0 ); +static void free_session_slots (struct nfsd4_session *ses,int from) {@@ -1931,11 +1941,14 @@ free_session_slots(struct nfsd4_session *ses, int from)kfree (slot); } ses->se_fchannel.maxreqs = from; -if (ses->se_target_maxslots > from) - ses->se_target_maxslots = from; +if (ses->se_target_maxslots > from) { +int new_target = from ?:1 ; +atomic_sub (ses->se_target_maxslots - new_target, &nfsd_total_target_slots); + ses->se_target_maxslots = new_target; + } } -static int __maybe_unused +static int reduce_session_slots (struct nfsd4_session *ses,int dec) { struct nfsd_net *nn = net_generic(ses->se_client->net,@@ -1948,6 +1961,7 @@ reduce_session_slots(struct nfsd4_session *ses, int dec)return ret; ret =min (dec, ses->se_target_maxslots-1 ); ses->se_target_maxslots -= ret; +atomic_sub (ret, &nfsd_total_target_slots); ses->se_slot_gen +=1 ;if (ses->se_slot_gen ==0 ) {int i;@@ -2006,6 +2020,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, fattrs->maxreqs = i;memcpy (&new->se_fchannel, fattrs,sizeof (struct nfsd4_channel_attrs )); new->se_target_maxslots = i; +atomic_add (i -1 , &nfsd_total_target_slots); new->se_cb_slot_avail = ~0U ; new->se_cb_highest_slot =min (battrs->maxreqs -1 ,NFSD_BC_SLOT_TABLE_SIZE -1 );@@ -2130,6 +2145,36 @@ static void free_session(struct nfsd4_session *ses)__free_session (ses); } +static unsigned long +nfsd_slot_count (struct shrinker *s,struct shrink_control *sc) +{ +unsigned long cnt =atomic_read (&nfsd_total_target_slots); + +return cnt ? cnt :SHRINK_EMPTY ; +} + +static unsigned long +nfsd_slot_scan (struct shrinker *s,struct shrink_control *sc) +{ +struct nfsd4_session *ses; +unsigned long scanned =0 ; +unsigned long freed =0 ; + +spin_lock (&nfsd_session_list_lock); +list_for_each_entry (ses, &nfsd_session_list, se_all_sessions) { + freed +=reduce_session_slots (ses,1 ); + scanned +=1 ; +if (scanned >= sc->nr_to_scan) { +/* Move starting point for next scan */ +list_move (&nfsd_session_list, &ses->se_all_sessions); +break ; + } + } +spin_unlock (&nfsd_session_list_lock); + sc->nr_scanned = scanned; +return freed; +} +static void init_session (struct svc_rqst *rqstp,struct nfsd4_session *new,struct nfs4_client *clp,struct nfsd4_create_session *cses) {int idx;@@ -2154,6 +2199,10 @@ static void init_session(struct svc_rqst *rqstp, struct nfsd4_session *new, strulist_add (&new->se_perclnt, &clp->cl_sessions);spin_unlock (&clp->cl_lock); +spin_lock (&nfsd_session_list_lock); +list_add_tail (&new->se_all_sessions, &nfsd_session_list); +spin_unlock (&nfsd_session_list_lock); + {struct sockaddr *sa =svc_addr (rqstp);/* @@ -2223,6 +2272,9 @@ unhash_session(struct nfsd4_session *ses)spin_lock (&ses->se_client->cl_lock);list_del (&ses->se_perclnt);spin_unlock (&ses->se_client->cl_lock); +spin_lock (&nfsd_session_list_lock); +list_del (&ses->se_all_sessions); +spin_unlock (&nfsd_session_list_lock); }/* SETCLIENTID and SETCLIENTID_CONFIRM Helper functions */ @@ -4335,6 +4387,7 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, slot->sl_seqid = seq->seqid; slot->sl_flags &= ~NFSD4_SLOT_REUSED ; slot->sl_flags |=NFSD4_SLOT_INUSE ; + slot->sl_generation = session->se_slot_gen;if (seq->cachethis) slot->sl_flags |=NFSD4_SLOT_CACHETHIS ;else @@ -4371,6 +4424,8 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot,GFP_ATOMIC ))) { session->se_fchannel.maxreqs +=1 ; +atomic_add (session->se_fchannel.maxreqs - session->se_target_maxslots, + &nfsd_total_target_slots); session->se_target_maxslots = session->se_fchannel.maxreqs; }else {kfree (slot);@@ -8779,7 +8834,6 @@ nfs4_state_start_net(struct net *net) }/* initialization to perform when the nfsd service is started: */ -int nfs4_state_start (void ) {@@ -8789,6 +8843,15 @@ nfs4_state_start(void)if (ret)return ret; + nfsd_slot_shrinker =shrinker_alloc (0 ,"nfsd-DRC-slot" ); +if (!nfsd_slot_shrinker) { +rhltable_destroy (&nfs4_file_rhltable); +return -ENOMEM ; + } + nfsd_slot_shrinker->count_objects = nfsd_slot_count; + nfsd_slot_shrinker->scan_objects = nfsd_slot_scan; +shrinker_register (nfsd_slot_shrinker); +set_max_delegations ();return 0 ; }@@ -8830,6 +8893,7 @@ voidnfs4_state_shutdown (void) {rhltable_destroy (&nfs4_file_rhltable); +shrinker_free (nfsd_slot_shrinker); } static voidfs/nfsd/state.h +0
-0
fs/nfsd/state.h struct nfsd4_session { · line 348
Quoted 4 lines
fs/nfsd/state.h +1
-0@@ -345,6 +345,7 @@ struct nfsd4_session {bool se_dead;struct list_head se_hash;/* hash by sessionid */ struct list_head se_perclnt; +struct list_head se_all_sessions;/* global list of sessions */
I think my only minor issue here is whether we truly want an "all_sessions" list. Since we don't expect the shrinker to run very often, isn't there another mechanism that can already iterate all clients and their sessions?
Quoted 6 lines
fs/nfsd/state.h +0
-0struct nfs4_client *se_client;struct nfs4_sessionid se_sessionid;struct nfsd4_channel_attrs se_fchannel;Signature · 1 line
2.47.0
Signature · 1 line
Chuck Lever
Re: [PATCH 6/6] nfsd: add shrinker to reduce number of slots allocated per session
Quoted 15 lines · Chuck Lever
fs/nfsd/state.h struct nfsd4_session { · line 348
Quoted 9 lines · NeilBrown
fs/nfsd/state.h +1
-0@@ -345,6 +345,7 @@ struct nfsd4_session {bool se_dead;struct list_head se_hash;/* hash by sessionid */ struct list_head se_perclnt; +struct list_head se_all_sessions;/* global list of sessions */ I think my only minor issue here is whether we truly want an "
all_sessions" list. Since we don't expect the shrinker to run very often, isn't there another mechanism that can already iterate all clients and their sessions?
"all_sessions" certainly isn't my favourite part of the set. But I do think we need it.
We can iterate all sessions by iterating all net-namespaces, then all clients, then all sessions. But that isn't what we need.
The shrinker mechanism seems to assume an LRU. It makes "scan" requests one "batch" at a time, and may request several batches in sequence without telling you in advance how many batches to expect. So you need some concept of the "next" thing to free. Often this is the end of the LRU.
But we don't have an LRU because the slots aren't a cache.
An important detail is that when nfsd_slot_scan() has scanned all that it was asked, it moves the head to the current point in the list. So the next time it is called it will start with the correct next session.
This will only become important where there are more than 64 (default batch size) sessions.
NeilBrown
Quoted 12 lines
Quoted 8 lines
fs/nfsd/state.h +0
-0struct nfs4_client *se_client;struct nfs4_sessionid se_sessionid;struct nfsd4_channel_attrs se_fchannel;Signature · 1 line
2.47.0Signature · 1 line
Chuck Lever
Jeff Layton
maintainerRe: [PATCH 6/6] nfsd: add shrinker to reduce number of slots allocated per session
In reply to NeilBrownQuoted 47 lines · NeilBrown
Add a shrinker which frees unused slots and may ask the clients to use fewer slots on each session.
Each session now tracks
se_client_maxreqswhich is the most recent max-requests-in-use reported by the client, andse_target_maxreqswhich is a target number of requests which is reduced by the shrinker.The shrinker iterates over all sessions on all client in all net-namespaces and reduces the target by 1 for each. The shrinker may get called multiple times to reduce by more than 1 each.
If
se_target_maxreqsis abovese_client_maxreqs, those slots can be freed immediately. If not the client will be ask to reduce its usage and as the usage goes down slots will be freed.Once the usage has dropped to match the target, the target can be increased if the client uses all available slots and if a
GFP_NOWAITallocation succeeds.Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 72 ++++++++++++++++++++++++++++++++++++++++++--- fs/nfsd/state.h | 1 + 2 files changed, 69 insertions(+), 4 deletions(-)fs/nfsd/nfs4state.c +10
-0@@ -1909,6 +1909,16 @@ gen_sessionid(struct nfsd4_session *ses)*/ #define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44) +static struct shrinker *nfsd_slot_shrinker; +static DEFINE_SPINLOCK (nfsd_session_list_lock); +static LIST_HEAD (nfsd_session_list); +/* The sum of "target_slots-1" on every session. The shrinker can push this +* down, though it can take a little while for the memory to actually +* be freed. The "-1" is because we can never free slot 0 while the +* session is active. +*/ +static atomic_t nfsd_total_target_slots =ATOMIC_INIT (0 ); +static void free_session_slots (struct nfsd4_session *ses,int from) {
fs/nfsd/nfs4state.c free_session_slots(struct nfsd4_session *ses, int from) · line 1945
Quoted 7 lines
fs/nfsd/nfs4state.c +2
-2@@ -1931,11 +1941,14 @@ free_session_slots(struct nfsd4_session *ses, int from)kfree (slot); } ses->se_fchannel.maxreqs = from; -if (ses->se_target_maxslots > from) - ses->se_target_maxslots = from; +if (ses->se_target_maxslots > from) { +int new_target = from ?:1 ;
Let's make that "from ? from : 1". The above is a non-standard gcc-ism (AIUI).
Quoted 145 lines
fs/nfsd/nfs4state.c +56
-2+atomic_sub (ses->se_target_maxslots - new_target, &nfsd_total_target_slots); + ses->se_target_maxslots = new_target; + } } -static int __maybe_unused +static int reduce_session_slots (struct nfsd4_session *ses,int dec) { struct nfsd_net *nn = net_generic(ses->se_client->net,@@ -1948,6 +1961,7 @@ reduce_session_slots(struct nfsd4_session *ses, int dec)return ret; ret =min (dec, ses->se_target_maxslots-1 ); ses->se_target_maxslots -= ret; +atomic_sub (ret, &nfsd_total_target_slots); ses->se_slot_gen +=1 ;if (ses->se_slot_gen ==0 ) {int i;@@ -2006,6 +2020,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, fattrs->maxreqs = i;memcpy (&new->se_fchannel, fattrs,sizeof (struct nfsd4_channel_attrs )); new->se_target_maxslots = i; +atomic_add (i -1 , &nfsd_total_target_slots); new->se_cb_slot_avail = ~0U ; new->se_cb_highest_slot =min (battrs->maxreqs -1 ,NFSD_BC_SLOT_TABLE_SIZE -1 );@@ -2130,6 +2145,36 @@ static void free_session(struct nfsd4_session *ses)__free_session (ses); } +static unsigned long +nfsd_slot_count (struct shrinker *s,struct shrink_control *sc) +{ +unsigned long cnt =atomic_read (&nfsd_total_target_slots); + +return cnt ? cnt :SHRINK_EMPTY ; +} + +static unsigned long +nfsd_slot_scan (struct shrinker *s,struct shrink_control *sc) +{ +struct nfsd4_session *ses; +unsigned long scanned =0 ; +unsigned long freed =0 ; + +spin_lock (&nfsd_session_list_lock); +list_for_each_entry (ses, &nfsd_session_list, se_all_sessions) { + freed +=reduce_session_slots (ses,1 ); + scanned +=1 ; +if (scanned >= sc->nr_to_scan) { +/* Move starting point for next scan */ +list_move (&nfsd_session_list, &ses->se_all_sessions); +break ; + } + } +spin_unlock (&nfsd_session_list_lock); + sc->nr_scanned = scanned; +return freed; +} +static void init_session (struct svc_rqst *rqstp,struct nfsd4_session *new,struct nfs4_client *clp,struct nfsd4_create_session *cses) {int idx;@@ -2154,6 +2199,10 @@ static void init_session(struct svc_rqst *rqstp, struct nfsd4_session *new, strulist_add (&new->se_perclnt, &clp->cl_sessions);spin_unlock (&clp->cl_lock); +spin_lock (&nfsd_session_list_lock); +list_add_tail (&new->se_all_sessions, &nfsd_session_list); +spin_unlock (&nfsd_session_list_lock); + {struct sockaddr *sa =svc_addr (rqstp);/* @@ -2223,6 +2272,9 @@ unhash_session(struct nfsd4_session *ses)spin_lock (&ses->se_client->cl_lock);list_del (&ses->se_perclnt);spin_unlock (&ses->se_client->cl_lock); +spin_lock (&nfsd_session_list_lock); +list_del (&ses->se_all_sessions); +spin_unlock (&nfsd_session_list_lock); }/* SETCLIENTID and SETCLIENTID_CONFIRM Helper functions */ @@ -4335,6 +4387,7 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, slot->sl_seqid = seq->seqid; slot->sl_flags &= ~NFSD4_SLOT_REUSED ; slot->sl_flags |=NFSD4_SLOT_INUSE ; + slot->sl_generation = session->se_slot_gen;if (seq->cachethis) slot->sl_flags |=NFSD4_SLOT_CACHETHIS ;else @@ -4371,6 +4424,8 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,if (slot && !xa_is_err (xa_store (&session->se_slots, s, slot,GFP_ATOMIC ))) { session->se_fchannel.maxreqs +=1 ; +atomic_add (session->se_fchannel.maxreqs - session->se_target_maxslots, + &nfsd_total_target_slots); session->se_target_maxslots = session->se_fchannel.maxreqs; }else {kfree (slot);@@ -8779,7 +8834,6 @@ nfs4_state_start_net(struct net *net) }/* initialization to perform when the nfsd service is started: */ -int nfs4_state_start (void ) {@@ -8789,6 +8843,15 @@ nfs4_state_start(void)if (ret)return ret; + nfsd_slot_shrinker =shrinker_alloc (0 ,"nfsd-DRC-slot" ); +if (!nfsd_slot_shrinker) { +rhltable_destroy (&nfs4_file_rhltable); +return -ENOMEM ; + } + nfsd_slot_shrinker->count_objects = nfsd_slot_count; + nfsd_slot_shrinker->scan_objects = nfsd_slot_scan; +shrinker_register (nfsd_slot_shrinker); +set_max_delegations ();return 0 ; }@@ -8830,6 +8893,7 @@ voidnfs4_state_shutdown (void) {rhltable_destroy (&nfs4_file_rhltable); +shrinker_free (nfsd_slot_shrinker); } static voidfs/nfsd/state.h +1
-0@@ -345,6 +345,7 @@ struct nfsd4_session {bool se_dead;struct list_head se_hash;/* hash by sessionid */ struct list_head se_perclnt; +struct list_head se_all_sessions;/* global list of sessions */ struct nfs4_client *se_client;struct nfs4_sessionid se_sessionid;struct nfsd4_channel_attrs se_fchannel;
Signature · 1 line
Jeff Layton <[email protected]>
Re: [PATCH 6/6] nfsd: add shrinker to reduce number of slots allocated per session
Quoted 50 lines · Jeff Layton
Quoted 49 lines · NeilBrown
Add a shrinker which frees unused slots and may ask the clients to use fewer slots on each session.
Each session now tracks
se_client_maxreqswhich is the most recent max-requests-in-use reported by the client, andse_target_maxreqswhich is a target number of requests which is reduced by the shrinker.The shrinker iterates over all sessions on all client in all net-namespaces and reduces the target by 1 for each. The shrinker may get called multiple times to reduce by more than 1 each.
If
se_target_maxreqsis abovese_client_maxreqs, those slots can be freed immediately. If not the client will be ask to reduce its usage and as the usage goes down slots will be freed.Once the usage has dropped to match the target, the target can be increased if the client uses all available slots and if a
GFP_NOWAITallocation succeeds.Signed-off-by NeilBrown <[email protected]>--- fs/nfsd/nfs4state.c | 72 ++++++++++++++++++++++++++++++++++++++++++--- fs/nfsd/state.h | 1 + 2 files changed, 69 insertions(+), 4 deletions(-)fs/nfsd/nfs4state.c +10
-0@@ -1909,6 +1909,16 @@ gen_sessionid(struct nfsd4_session *ses)*/ #define NFSD_MIN_HDR_SEQ_SZ (24 + 12 + 44) +static struct shrinker *nfsd_slot_shrinker; +static DEFINE_SPINLOCK (nfsd_session_list_lock); +static LIST_HEAD (nfsd_session_list); +/* The sum of "target_slots-1" on every session. The shrinker can push this +* down, though it can take a little while for the memory to actually +* be freed. The "-1" is because we can never free slot 0 while the +* session is active. +*/ +static atomic_t nfsd_total_target_slots =ATOMIC_INIT (0 ); +static void free_session_slots (struct nfsd4_session *ses,int from) {@@ -1931,11 +1941,14 @@ free_session_slots(struct nfsd4_session *ses, int from)kfree (slot); }
Quoted 8 lines
fs/nfsd/nfs4state.c free_session_slots(struct nfsd4_session *ses, int from) · line 1945
Quoted 5 lines
fs/nfsd/nfs4state.c +2
-2ses->se_fchannel.maxreqs = from; -if (ses->se_target_maxslots > from) - ses->se_target_maxslots = from; +if (ses->se_target_maxslots > from) { +int new_target = from ?:1 ;Let's make that "from ? from : 1". The above is a non-standard gcc-ism (AIUI).
Let's not. There are currently 1926 lines in .c and .h files in the Linux kernel which contain "?:" and another 848 which contain "? :". I think it is an established part of the kernel style.
This is admittedly dominated by bcachefs, but there is a long tail with tools, net, crypto, drivers all contributing. Outside of bcachefs, fs/ contributes only 102 in 29 different filesystems.
Thanks,
NeilBrown
Jeff Layton
maintainerQuoted 13 lines · NeilBrown
Here is v2 of my series for on-demand allocation and freeing of session DRC slots.
- Now uses an xarray to store slots, and the limit is raised to 2048
- delays retiring a slot until the client has confirmed that it isn't using it as described in RFC:The replier SHOULD retain the slots it wants to retire until the requester sends a request with a
highest_slotidless than or equal to the replier's new enforcedhighest_slotid.- When a retired slot is used, allow the seqid to be the next in sequence as required by the RFC:
Quoted 9 lines
Each time a slot is reused, the request MUST specify a sequence ID that is one greater than that of the previous request on the slot.
or "1" as (arguably) allowed by the RFC:
The first time a slot is used, the requester MUST specify a sequence ID of one
I thought that the conclusion of the IETF discussion was that we should reset this to 1. It'd be ideal to just do that, as then we wouldn't need NFSD4_SLOT_REUSED.
Are there any clients that expect to reuse the old seqid in this situation? I know the Linux client doesn't. Do Solaris or FreeBSD?
Quoted 9 lines
- current slot allocation is now reported in
/proc/fs/nfsd/clients/*/infoThis has been tested with highly aggressive shrinker settings:
nfsd_slot_shrinker->seeks = 0; nfsd_slot_shrinker->batch = 2;and with periodic "echo 3 >
drop_caches". The slot count drops as expected and then increases again.
This is really great work, Neil!
--
Jeff Layton <[email protected]>
Quoted 20 lines · Jeff Layton
Quoted 19 lines · NeilBrown
Here is v2 of my series for on-demand allocation and freeing of session DRC slots.
- Now uses an xarray to store slots, and the limit is raised to 2048
- delays retiring a slot until the client has confirmed that it isn't using it as described in RFC:The replier SHOULD retain the slots it wants to retire until the requester sends a request with a
highest_slotidless than or equal to the replier's new enforcedhighest_slotid.- When a retired slot is used, allow the seqid to be the next in sequence as required by the RFC:
Each time a slot is reused, the request MUST specify a sequence ID that is one greater than that of the previous request on the slot.
or "1" as (arguably) allowed by the RFC:
Quoted 7 lines
Quoted 3 lines
The first time a slot is used, the requester MUST specify a sequence ID of one
I thought that the conclusion of the IETF discussion was that we should reset this to 1. It'd be ideal to just do that, as then we wouldn't need
NFSD4_SLOT_REUSED.
I thought the conclusion was:
I'm convinced. The next draft of rfc5661bis will address this issue.
Until the issue is addressed I don't think it would be wise to preempt the result.
Quoted 3 lines
Are there any clients that expect to reuse the old seqid in this situation? I know the Linux client doesn't. Do Solaris or FreeBSD?
I don't know. But I tend to code the the spec, not to other clients. I still think the specs says
Each time a slot is reused, the request MUST specify a sequence ID that is one greater than that of the previous request on the slot.
and I don't see any good reason not to treat what we are doing here as "reuse".
So I'm making a concession to the linux client and to you by allowing '1'. I'm still not convinced that it is a good idea. At best I'm convinced that the spec can be read to suggest that might be an option.
Thanks,
NeilBrown