Possible NFSD COPY clean-ups
Subsystems (3)
[email protected]
maintainercover Possible NFSD COPY clean-ups
From: Chuck Lever <[email protected]>
While working on OFFLOAD_STATUS and other potential improvements to Linux NFS's COPY offload implementation, I've come up with a few server-side observability enhancements and one or two possible bug fixes. These are candidates to merge for v6.12.
Comments welcome.
Chuck Lever (7):
NFSD: Async COPY result needs to return a write verifier
NFSD: Limit the number of concurrent async COPY operations
NFSD: Display copy stateids with conventional print formatting
NFSD: Record the callback stateid in copy tracepoints
NFSD: Clean up extra whitespace in trace_nfsd_copy_done
NFSD: Document callback stateid laundering
NFSD: Wrap async copy operations with trace points
fs/nfsd/netns.h | 1 + fs/nfsd/nfs4proc.c | 38 +++++++++--------- fs/nfsd/nfs4state.c | 37 ++++++++++++----- fs/nfsd/trace.h | 97 +++++++++++++++++++++++++++++++++++++++++---- fs/nfsd/xdr4.h | 1 + 5 files changed, 137 insertions(+), 37 deletions(-)
Signature · 1 line
2.46.0
[email protected]
maintainerRFC PATCH 1/7 NFSD: Async COPY result needs to return a write verifier
From: Chuck Lever <[email protected]>
Currently, when NFSD handles an asynchronous COPY, it returns a zero write verifier, relying on the subsequent CB_OFFLOAD callback to pass the write verifier and a stable_how4 value to the client.
However, if the CB_OFFLOAD never arrives at the client (for example, if a network partition occurs just as the server sends the CB_OFFLOAD operation), the client will never receive this verifier. Thus, if the client sends a follow-up COMMIT, there is no way for the client to assess the COMMIT result.
The usual recovery for a missing CB_OFFLOAD is for the client to send an OFFLOAD_STATUS operation, but that operation does not carry a write verifier in its result. Neither does it carry a stable_how4 value, so the client /must/ send a COMMIT in this case -- which will always fail because currently there's still no write verifier in the COPY result.
Thus the server needs to return a normal write verifier in its COPY result even if the COPY operation is to be performed asynchronously.
If the server recognizes the callback stateid in subsequent OFFLOAD_STATUS operations, then obviously it has not restarted, and the write verifier the client received in the COPY result is still valid and can be used to assess a COMMIT of the copied data, if one is needed.
--- fs/nfsd/nfs4proc.c | 25 +++++++++---------------- 1 file changed, 9 insertions(+), 16 deletions(-)
fs/nfsd/nfs4proc.c +9 -16
@@ -751,15 +751,6 @@ nfsd4_access(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, &access->ac_supported); } -static void gen_boot_verifier (nfs4_verifier *verifier,struct net *net) -{ -__be32 *verf = (__be32 *)verifier->data; - -BUILD_BUG_ON (2 *sizeof (*verf) !=sizeof (verifier->data)); - -nfsd_copy_write_verifier (verf,net_generic (net, nfsd_net_id)); -} -static __be32 nfsd4_commit (struct svc_rqst *rqstp,struct nfsd4_compound_state *cstate,union nfsd4_op_u *u)
@@ -1630,7 +1621,6 @@ static void nfsd4_init_copy_res(struct nfsd4_copy *copy, bool sync)test_bit (NFSD4_COPY_F_COMMITTED , ©->cp_flags) ?NFS_FILE_SYNC :NFS_UNSTABLE ;nfsd4_copy_set_sync (copy, sync); -gen_boot_verifier (©->cp_res.wr_verifier, copy->cp_clp->net); } static ssize_t _nfsd_copy_file_range(struct nfsd4_copy *copy,
@@ -1803,9 +1793,11 @@ static __be32 nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, union nfsd4_op_u *u) { - struct nfsd4_copy *copy = &u->copy; -__be32 status; + struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);struct nfsd4_copy *async_copy =NULL ; +struct nfsd4_copy *copy = &u->copy; +struct nfsd42_write_res *result; +__be32 status;/* * Currently, async COPY is not reliable. Force all COPY
@@ -1814,6 +1806,9 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,*/ nfsd4_copy_set_sync (copy, true); + result = ©->cp_res; +nfsd_copy_write_verifier ((__be32 *)&result->wr_verifier.data, nn); + copy->cp_clp = cstate->clp;if (nfsd4_ssc_is_inter (copy)) {trace_nfsd_copy_inter (copy);
@@ -1838,8 +1833,6 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,memcpy (©->fh, &cstate->current_fh.fh_handle,sizeof (struct knfsd_fh ));if (nfsd4_copy_is_async (copy)) { -struct nfsd_net *nn =net_generic (SVC_NET (rqstp), nfsd_net_id); - status =nfserrno (-ENOMEM ); async_copy =kzalloc (sizeof (struct nfsd4_copy ),GFP_KERNEL );if (!async_copy)
@@ -1851,8 +1844,8 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, goto out_err;if (!nfs4_init_copy_state (nn, copy)) goto out_err; -memcpy (©->cp_res.cb_stateid, ©->cp_stateid.cs_stid, -sizeof (copy->cp_res.cb_stateid)); +memcpy (&result->cb_stateid, ©->cp_stateid.cs_stid, +sizeof (result->cb_stateid));dup_copy_fields (copy, async_copy); async_copy->copy_task =kthread_create (nfsd4_do_async_copy, async_copy,"%s" ,"copy thread" );
Signature · 1 line
2.46.0
[email protected]
maintainerRFC PATCH 2/7 NFSD: Limit the number of concurrent async COPY operations
In reply to [email protected]From: Chuck Lever <[email protected]>
Nothing appears to limit the number of concurrent async COPY operations that clients can start. In addition, AFAICT each async COPY can copy an unlimited number of 4MB chunks, so can run for a long time. Thus IMO async COPY can become a DoS vector.
Add a restriction mechanism that bounds the number of concurrent background COPY operations. Start simple and try to be fair -- this patch implements a per-namespace limit.
An async COPY request that occurs while this limit is exceeded gets NFS4ERR_DELAY. The requesting client can choose to send the request again after a delay or fall back to a traditional read/write style copy.
If there is need to make the mechanism more sophisticated, we can visit that in future patches.
--- fs/nfsd/netns.h | 1 + fs/nfsd/nfs4proc.c | 12 ++++++++++-- fs/nfsd/nfs4state.c | 1 + fs/nfsd/xdr4.h | 1 + 4 files changed, 13 insertions(+), 2 deletions(-)
fs/nfsd/netns.h +1 -0
@@ -148,6 +148,7 @@ struct nfsd_net {u32 s2s_cp_cl_id;struct idr s2s_cp_stateids;spinlock_t s2s_cp_lock; +atomic_t pending_async_copies;/* * Version information
fs/nfsd/nfs4proc.c +10 -2
@@ -1279,6 +1279,7 @@ static void nfs4_put_copy(struct nfsd4_copy *copy) {if (!refcount_dec_and_test (©->refcount))return ; +atomic_dec (©->cp_nn->pending_async_copies);kfree (copy->cp_src);kfree (copy); }
@@ -1833,10 +1834,17 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,memcpy (©->fh, &cstate->current_fh.fh_handle,sizeof (struct knfsd_fh ));if (nfsd4_copy_is_async (copy)) { - status =nfserrno (-ENOMEM ); +/* Arbitrary cap on number of pending async copy operations */ +int nrthreads =atomic_read (&rqstp->rq_pool->sp_nrthreads); + async_copy =kzalloc (sizeof (struct nfsd4_copy ),GFP_KERNEL );if (!async_copy) goto out_err; + async_copy->cp_nn = nn; +if (atomic_inc_return (&nn->pending_async_copies) > nrthreads) { +atomic_dec (&nn->pending_async_copies); + goto out_err; + }INIT_LIST_HEAD (&async_copy->copies);refcount_set (&async_copy->refcount,1 ); async_copy->cp_src =kmalloc (sizeof (*async_copy->cp_src),GFP_KERNEL );
@@ -1876,7 +1884,7 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, }if (async_copy)cleanup_async_copy (async_copy); - status =nfserrno (-ENOMEM ); + status = nfserr_jukebox; goto out; }
fs/nfsd/nfs4state.c +1 -0
@@ -8554,6 +8554,7 @@ static int nfs4_state_create_net(struct net *net)spin_lock_init (&nn->client_lock);spin_lock_init (&nn->s2s_cp_lock);idr_init (&nn->s2s_cp_stateids); +atomic_set (&nn->pending_async_copies,0 );spin_lock_init (&nn->blocked_locks_lock);INIT_LIST_HEAD (&nn->blocked_locks_lru);
fs/nfsd/xdr4.h +1 -0
@@ -713,6 +713,7 @@ struct nfsd4_copy {struct nfsd4_ssc_umount_item *ss_nsui;struct nfs_fh c_fh;nfs4_stateid stateid; +struct nfsd_net *cp_nn; };static inline void nfsd4_copy_set_sync (struct nfsd4_copy *copy,bool sync)
Signature · 1 line
2.46.0
[email protected]
maintainerRFC PATCH 3/7 NFSD: Display copy stateids with conventional print formatting
In reply to [email protected]From: Chuck Lever <[email protected]>
Make it easier to grep for s2s COPY stateids in trace logs: Use the same display format in nfsd_copy_class as is used to display other stateids.
--- fs/nfsd/trace.h | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-)
fs/nfsd/trace.h +6 -6
@@ -2153,14 +2153,14 @@ DECLARE_EVENT_CLASS(nfsd_copy_class, sizeof(struct sockaddr_in6)); ),TP_printk ("client=%pISpc intra=%d async=%d " -"src_stateid[si_generation:0x%x cl_boot:0x%x cl_id:0x%x so_id:0x%x] " -"dst_stateid[si_generation:0x%x cl_boot:0x%x cl_id:0x%x so_id:0x%x] " +"src_client %08x:%08x src_stateid %08x:%08x " +"dst_client %08x:%08x dst_stateid %08x:%08x " "cp_src_pos=%llu cp_dst_pos=%llu cp_count=%llu" ,__get_sockaddr (addr), __entry->intra, __entry->async, - __entry->src_si_generation, __entry->src_cl_boot, - __entry->src_cl_id, __entry->src_so_id, - __entry->dst_si_generation, __entry->dst_cl_boot, - __entry->dst_cl_id, __entry->dst_so_id, + __entry->src_cl_boot, __entry->src_cl_id, + __entry->src_so_id, __entry->src_si_generation, + __entry->dst_cl_boot, __entry->dst_cl_id, + __entry->dst_so_id, __entry->dst_si_generation, __entry->src_cp_pos, __entry->dst_cp_pos, __entry->cp_count ) );
Signature · 1 line
2.46.0
[email protected]
maintainerRFC PATCH 4/7 NFSD: Record the callback stateid in copy tracepoints
In reply to [email protected]From: Chuck Lever <[email protected]>
Match COPY operations up with CB_OFFLOAD operations.
--- fs/nfsd/trace.h | 12 ++++++++++++ 1 file changed, 12 insertions(+)
fs/nfsd/trace.h +12 -0
@@ -2127,6 +2127,10 @@ DECLARE_EVENT_CLASS(nfsd_copy_class, __field(u32, dst_cl_id) __field(u32, dst_so_id) __field(u32, dst_si_generation) + __field(u32, cb_cl_boot) + __field(u32, cb_cl_id) + __field(u32, cb_so_id) + __field(u32, cb_si_generation) __field(u64, src_cp_pos) __field(u64, dst_cp_pos)__field (u64, cp_count)
@@ -2135,6 +2139,7 @@ DECLARE_EVENT_CLASS(nfsd_copy_class,TP_fast_assign ( const stateid_t *src_stp = ©->cp_src_stateid;const stateid_t *dst_stp = ©->cp_dst_stateid; +const stateid_t *cb_stp = ©->cp_res.cb_stateid; __entry->intra =test_bit (NFSD4_COPY_F_INTRA , ©->cp_flags); __entry->async = !test_bit (NFSD4_COPY_F_SYNCHRONOUS , ©->cp_flags);
@@ -2146,6 +2151,10 @@ DECLARE_EVENT_CLASS(nfsd_copy_class, __entry->dst_cl_id = dst_stp->si_opaque.so_clid.cl_id; __entry->dst_so_id = dst_stp->si_opaque.so_id; __entry->dst_si_generation = dst_stp->si_generation; + __entry->cb_cl_boot = cb_stp->si_opaque.so_clid.cl_boot; + __entry->cb_cl_id = cb_stp->si_opaque.so_clid.cl_id; + __entry->cb_so_id = cb_stp->si_opaque.so_id; + __entry->cb_si_generation = cb_stp->si_generation; __entry->src_cp_pos = copy->cp_src_pos; __entry->dst_cp_pos = copy->cp_dst_pos; __entry->cp_count = copy->cp_count;
@@ -2155,12 +2164,15 @@ DECLARE_EVENT_CLASS(nfsd_copy_class,TP_printk ("client=%pISpc intra=%d async=%d " "src_client %08x:%08x src_stateid %08x:%08x " "dst_client %08x:%08x dst_stateid %08x:%08x " +"cb_client %08x:%08x cb_stateid %08x:%08x " "cp_src_pos=%llu cp_dst_pos=%llu cp_count=%llu" ,__get_sockaddr (addr), __entry->intra, __entry->async, __entry->src_cl_boot, __entry->src_cl_id, __entry->src_so_id, __entry->src_si_generation, __entry->dst_cl_boot, __entry->dst_cl_id, __entry->dst_so_id, __entry->dst_si_generation, + __entry->cb_cl_boot, __entry->cb_cl_id, + __entry->cb_so_id, __entry->cb_si_generation, __entry->src_cp_pos, __entry->dst_cp_pos, __entry->cp_count ) );
Signature · 1 line
2.46.0
[email protected]
maintainerRFC PATCH 5/7 NFSD: Clean up extra whitespace in trace_nfsd_copy_done
In reply to [email protected]From: Chuck Lever <[email protected]>
--- fs/nfsd/trace.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
fs/nfsd/trace.h +1 -1
@@ -2205,7 +2205,7 @@ TRACE_EVENT(nfsd_copy_done,__assign_sockaddr (addr, ©->cp_clp->cl_addr,sizeof (struct sockaddr_in6 )); ), -TP_printk ("addr=%pISpc status=%d intra=%d async=%d " , +TP_printk ("addr=%pISpc status=%d intra=%d async=%d" ,__get_sockaddr (addr), __entry->status, __entry->intra, __entry->async ) );
Signature · 1 line
2.46.0
[email protected]
maintainerRFC PATCH 6/7 NFSD: Document callback stateid laundering
In reply to [email protected]From: Chuck Lever <[email protected]>
NFSD removes COPY callback stateids after once lease period. This practice keeps the list of callback stateids limited to prevent a DoS possibility, but doesn't comply with the spirit of RFC 7862 Section 4.8, which says:
Quoted 4 lines
A copy offload stateid will be valid until either (A) the client or server restarts or (B) the client returns the resource by issuing an
OFFLOAD_CANCELoperation or the client replies to aCB_OFFLOADoperation.
Note there are no BCP 14 compliance keywords in this text, so NFSD is free to ignore this stateid lifetime guideline without becoming non-compliant.
Nevertheless, this behavior variance should be explicitly documented in the code.
--- fs/nfsd/nfs4state.c | 36 +++++++++++++++++++++++++----------- 1 file changed, 25 insertions(+), 11 deletions(-)
fs/nfsd/nfs4state.c +25 -11
@@ -6324,6 +6324,29 @@ static void nfsd4_ssc_expire_umount(struct nfsd_net *nn) }#endif +/* +* RFC 7862 Section 4.8 says that, if the client hasn't replied to a +* CB_OFFLOAD, that COPY callback stateid will live until the client or +* server restarts. To prevent a DoS resulting from a pile of these +* stateids accruing over time, NFSD purges them after one lease period. +*/ +static void nfs4_launder_cpntf_statelist (struct nfsd_net *nn, +struct laundry_time *lt) +{ +struct nfs4_cpntf_state *cps; +copy_stateid_t *cps_t; +int i; + +spin_lock (&nn->s2s_cp_lock); +idr_for_each_entry (&nn->s2s_cp_stateids, cps_t, i) { + cps = container_of(cps_t, struct nfs4_cpntf_state, cp_stateid); +if (cps->cp_stateid.cs_type ==NFS4_COPYNOTIFY_STID && +state_expired (lt, cps->cpntf_time)) +_free_cpntf_state_locked (nn, cps); + } +spin_unlock (&nn->s2s_cp_lock); +} +/* Check if any lock belonging to this lockowner has any blockers */ static bool nfs4_lockowner_has_blockers (struct nfs4_lockowner *lo)
@@ -6495,9 +6518,6 @@ nfs4_laundromat(struct nfsd_net *nn) .cutoff = ktime_get_boottime_seconds() - nn->nfsd4_lease, .new_timeo = nn->nfsd4_lease }; -struct nfs4_cpntf_state *cps; -copy_stateid_t *cps_t; -int i;if (clients_still_reclaiming (nn)) { lt.new_timeo =0 ;
@@ -6505,14 +6525,8 @@ nfs4_laundromat(struct nfsd_net *nn) }nfsd4_end_grace (nn); -spin_lock (&nn->s2s_cp_lock); -idr_for_each_entry (&nn->s2s_cp_stateids, cps_t, i) { - cps = container_of(cps_t, struct nfs4_cpntf_state, cp_stateid); -if (cps->cp_stateid.cs_type ==NFS4_COPYNOTIFY_STID && -state_expired (<, cps->cpntf_time)) -_free_cpntf_state_locked (nn, cps); - } -spin_unlock (&nn->s2s_cp_lock); +nfs4_launder_cpntf_statelist (nn, <); +nfs4_get_client_reaplist (nn, &reaplist, <);nfs4_process_client_reaplist (&reaplist);
Signature · 1 line
2.46.0
[email protected]
maintainerRFC PATCH 7/7 NFSD: Wrap async copy operations with trace points
In reply to [email protected]From: Chuck Lever <[email protected]>
Add an nfsd_copy_async_done to record the timestamp, the final status code, and the callback stateid of an async copy.
Rename the nfsd_copy_do_async tracepoint to match that naming convention to make it easier to enable both of these with a single glob.
--- fs/nfsd/nfs4proc.c | 3 +- fs/nfsd/trace.h | 71 +++++++++++++++++++++++++++++++++++++++++++++- 2 files changed, 72 insertions(+), 2 deletions(-)
fs/nfsd/nfs4proc.c +2 -1
@@ -1758,7 +1758,7 @@ static int nfsd4_do_async_copy(void *data) {struct nfsd4_copy *copy = (struct nfsd4_copy *)data; -trace_nfsd_copy_do_async (copy); +trace_nfsd_copy_async (copy);if (nfsd4_ssc_is_inter (copy)) {struct file *filp;
@@ -1785,6 +1785,7 @@ static int nfsd4_do_async_copy(void *data) do_callback:set_bit (NFSD4_COPY_F_COMPLETED , ©->cp_flags); +trace_nfsd_copy_async_done (copy);nfsd4_send_cb_offload (copy);cleanup_async_copy (copy);return 0 ;
fs/nfsd/trace.h +70 -1
83 lines not on this page · raw message
Signature · 1 line
2.46.0
Re: [RFC PATCH 6/7] NFSD: Document callback stateid laundering
In reply to [email protected]Quoted 30 lines · On Wed, Aug 28, 2024 at 1:40 PM <[email protected]> wrote:
From: Chuck Lever <[email protected]>
NFSD removes COPY callback stateids after once lease period. This practice keeps the list of callback stateids limited to prevent a DoS possibility, but doesn't comply with the spirit of RFC 7862 Section 4.8, which says:
Quoted 4 lines
A copy offload stateid will be valid until either (A) the client or server restarts or (B) the client returns the resource by issuing an
OFFLOAD_CANCELoperation or the client replies to aCB_OFFLOADoperation.Note there are no BCP 14 compliance keywords in this text, so NFSD is free to ignore this stateid lifetime guideline without becoming non-compliant.
Nevertheless, this behavior variance should be explicitly documented in the code.
Signed-off-by Chuck Lever <[email protected]>--- fs/nfsd/nfs4state.c | 36 +++++++++++++++++++++++++----------- 1 file changed, 25 insertions(+), 11 deletions(-)fs/nfsd/nfs4state.c +0
-0
fs/nfsd/nfs4state.c static void nfsd4_ssc_expire_umount(struct nfsd_net *nn) · line 6332
Quoted 9 lines
fs/nfsd/nfs4state.c +6
-0@@ -6324,6 +6324,29 @@ static void nfsd4_ssc_expire_umount(struct nfsd_net *nn) }#endif +/* +* RFC 7862 Section 4.8 says that, if the client hasn't replied to a +* CB_OFFLOAD, that COPY callback stateid will live until the client or +* server restarts. To prevent a DoS resulting from a pile of these +* stateids accruing over time, NFSD purges them after one lease period. +*/
I don't believe this is correct documentation for this piece of code. There are two kinds of stateids in the copy offload world: one is issued by the source server "cnr_stateid" in the response of the COPY_NOTIFY and given to be client (to be given to the destination server to use for the read) and those are the ones kept in the knfsd's s2s_cp_stateids list and then cleaned up by the laundry thread when copy isn't cleaned up on the source server. The text from the RFC quoted here is for copy's callback stateids. In the current implementation, we don't keep callback stateids around past when the async copy is done. I agree that needs to be changed for OFFLOAD_STATUS op and then we can add the wording of how long we are keeping those.
Quoted 51 lines
fs/nfsd/nfs4state.c +19
-12+static void nfs4_launder_cpntf_statelist (struct nfsd_net *nn, +struct laundry_time *lt) +{ +struct nfs4_cpntf_state *cps; +copy_stateid_t *cps_t; +int i; + +spin_lock (&nn->s2s_cp_lock); +idr_for_each_entry (&nn->s2s_cp_stateids, cps_t, i) { + cps = container_of(cps_t, struct nfs4_cpntf_state, cp_stateid); +if (cps->cp_stateid.cs_type ==NFS4_COPYNOTIFY_STID && +state_expired (lt, cps->cpntf_time)) +_free_cpntf_state_locked (nn, cps); + } +spin_unlock (&nn->s2s_cp_lock); +} +/* Check if any lock belonging to this lockowner has any blockers */ static bool nfs4_lockowner_has_blockers (struct nfs4_lockowner *lo)@@ -6495,9 +6518,6 @@ nfs4_laundromat(struct nfsd_net *nn) .cutoff = ktime_get_boottime_seconds() - nn->nfsd4_lease, .new_timeo = nn->nfsd4_lease }; -struct nfs4_cpntf_state *cps; -copy_stateid_t *cps_t; -int i;if (clients_still_reclaiming (nn)) { lt.new_timeo =0 ;@@ -6505,14 +6525,8 @@ nfs4_laundromat(struct nfsd_net *nn) }nfsd4_end_grace (nn); -spin_lock (&nn->s2s_cp_lock); -idr_for_each_entry (&nn->s2s_cp_stateids, cps_t, i) { - cps = container_of(cps_t, struct nfs4_cpntf_state, cp_stateid); -if (cps->cp_stateid.cs_type ==NFS4_COPYNOTIFY_STID && -state_expired (<, cps->cpntf_time)) -_free_cpntf_state_locked (nn, cps); - } -spin_unlock (&nn->s2s_cp_lock); +nfs4_launder_cpntf_statelist (nn, <); +nfs4_get_client_reaplist (nn, &reaplist, <);nfs4_process_client_reaplist (&reaplist); --2.46.0
Re: [RFC PATCH 1/7] NFSD: Async COPY result needs to return a write verifier
In reply to [email protected]Quoted 95 lines · On Wed, 2024-08-28 at 13:40 -0400, [email protected] wrote:
From: Chuck Lever <[email protected]>
Currently, when NFSD handles an asynchronous COPY, it returns a zero write verifier, relying on the subsequent
CB_OFFLOADcallback to pass the write verifier and astable_how4value to the client.However, if the
CB_OFFLOADnever arrives at the client (for example, if a network partition occurs just as the server sends theCB_OFFLOADoperation), the client will never receive this verifier. Thus, if the client sends a follow-up COMMIT, there is no way for the client to assess the COMMIT result.The usual recovery for a missing
CB_OFFLOADis for the client to send anOFFLOAD_STATUSoperation, but that operation does not carry a write verifier in its result. Neither does it carry astable_how4value, so the client /must/ send a COMMIT in this case -- which will always fail because currently there's still no write verifier in the COPY result.Thus the server needs to return a normal write verifier in its COPY result even if the COPY operation is to be performed asynchronously.
If the server recognizes the callback stateid in subsequent
OFFLOAD_STATUSoperations, then obviously it has not restarted, and the write verifier the client received in the COPY result is still valid and can be used to assess a COMMIT of the copied data, if one is needed.Signed-off-by Chuck Lever <[email protected]>--- fs/nfsd/nfs4proc.c | 25 +++++++++---------------- 1 file changed, 9 insertions(+), 16 deletions(-)fs/nfsd/nfs4proc.c +7
-14@@ -751,15 +751,6 @@ nfsd4_access(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, &access->ac_supported); } -static void gen_boot_verifier (nfs4_verifier *verifier,struct net *net) -{ -__be32 *verf = (__be32 *)verifier->data; - -BUILD_BUG_ON (2 *sizeof (*verf) !=sizeof (verifier->data)); - -nfsd_copy_write_verifier (verf,net_generic (net, nfsd_net_id)); -} -static __be32 nfsd4_commit (struct svc_rqst *rqstp,struct nfsd4_compound_state *cstate,union nfsd4_op_u *u)@@ -1630,7 +1621,6 @@ static void nfsd4_init_copy_res(struct nfsd4_copy *copy, bool sync)test_bit (NFSD4_COPY_F_COMMITTED , ©->cp_flags) ?NFS_FILE_SYNC :NFS_UNSTABLE ;nfsd4_copy_set_sync (copy, sync); -gen_boot_verifier (©->cp_res.wr_verifier, copy->cp_clp->net); } static ssize_t _nfsd_copy_file_range(struct nfsd4_copy *copy,@@ -1803,9 +1793,11 @@ static __be32 nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, union nfsd4_op_u *u) { - struct nfsd4_copy *copy = &u->copy; -__be32 status; + struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);struct nfsd4_copy *async_copy =NULL ; +struct nfsd4_copy *copy = &u->copy; +struct nfsd42_write_res *result; +__be32 status;/* * Currently, async COPY is not reliable. Force all COPY @@ -1814,6 +1806,9 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,*/ nfsd4_copy_set_sync (copy, true); + result = ©->cp_res; +nfsd_copy_write_verifier ((__be32 *)&result->wr_verifier.data, nn); + copy->cp_clp = cstate->clp;if (nfsd4_ssc_is_inter (copy)) {trace_nfsd_copy_inter (copy);@@ -1838,8 +1833,6 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,memcpy (©->fh, &cstate->current_fh.fh_handle,sizeof (struct knfsd_fh ));if (nfsd4_copy_is_async (copy)) { -struct nfsd_net *nn =net_generic (SVC_NET (rqstp), nfsd_net_id); - status =nfserrno (-ENOMEM ); async_copy =kzalloc (sizeof (struct nfsd4_copy ),GFP_KERNEL );if (!async_copy)
fs/nfsd/nfs4proc.c nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, · line 1851
Quoted 10 lines
fs/nfsd/nfs4proc.c +2
-2@@ -1851,8 +1844,8 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, goto out_err;if (!nfs4_init_copy_state (nn, copy)) goto out_err; -memcpy (©->cp_res.cb_stateid, ©->cp_stateid.cs_stid, -sizeof (copy->cp_res.cb_stateid)); +memcpy (&result->cb_stateid, ©->cp_stateid.cs_stid, +sizeof (result->cb_stateid));dup_copy_fields (copy, async_copy); async_copy->copy_task =kthread_create (nfsd4_do_async_copy, async_copy,"%s" ,"copy thread" );
Re: [RFC PATCH 2/7] NFSD: Limit the number of concurrent async COPY operations
In reply to [email protected]Quoted 72 lines · On Wed, 2024-08-28 at 13:40 -0400, [email protected] wrote:
From: Chuck Lever <[email protected]>
Nothing appears to limit the number of concurrent async COPY operations that clients can start. In addition, AFAICT each async COPY can copy an unlimited number of 4MB chunks, so can run for a long time. Thus IMO async COPY can become a DoS vector.
Add a restriction mechanism that bounds the number of concurrent background COPY operations. Start simple and try to be fair -- this patch implements a per-namespace limit.
An async COPY request that occurs while this limit is exceeded gets
NFS4ERR_DELAY. The requesting client can choose to send the request again after a delay or fall back to a traditional read/write style copy.If there is need to make the mechanism more sophisticated, we can visit that in future patches.
Signed-off-by Chuck Lever <[email protected]>--- fs/nfsd/netns.h | 1 + fs/nfsd/nfs4proc.c | 12 ++++++++++-- fs/nfsd/nfs4state.c | 1 + fs/nfsd/xdr4.h | 1 + 4 files changed, 13 insertions(+), 2 deletions(-)fs/nfsd/netns.h +1
-0@@ -148,6 +148,7 @@ struct nfsd_net {u32 s2s_cp_cl_id;struct idr s2s_cp_stateids;spinlock_t s2s_cp_lock; +atomic_t pending_async_copies;/* * Version information fs/nfsd/nfs4proc.c +9
-1@@ -1279,6 +1279,7 @@ static void nfs4_put_copy(struct nfsd4_copy *copy) {if (!refcount_dec_and_test (©->refcount))return ; +atomic_dec (©->cp_nn->pending_async_copies);kfree (copy->cp_src);kfree (copy); }@@ -1833,10 +1834,17 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,memcpy (©->fh, &cstate->current_fh.fh_handle,sizeof (struct knfsd_fh ));if (nfsd4_copy_is_async (copy)) { - status =nfserrno (-ENOMEM ); +/* Arbitrary cap on number of pending async copy operations */ +int nrthreads =atomic_read (&rqstp->rq_pool->sp_nrthreads); + async_copy =kzalloc (sizeof (struct nfsd4_copy ),GFP_KERNEL );if (!async_copy) goto out_err; + async_copy->cp_nn = nn; +if (atomic_inc_return (&nn->pending_async_copies) > nrthreads) { +atomic_dec (&nn->pending_async_copies); + goto out_err; + }INIT_LIST_HEAD (&async_copy->copies);refcount_set (&async_copy->refcount,1 ); async_copy->cp_src =kmalloc (sizeof (*async_copy->cp_src),GFP_KERNEL );
fs/nfsd/nfs4proc.c nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, · line 1887
Quoted 5 lines
fs/nfsd/nfs4proc.c +1
-1@@ -1876,7 +1884,7 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, }if (async_copy)cleanup_async_copy (async_copy); - status =nfserrno (-ENOMEM ); + status = nfserr_jukebox;
ENOMEM gets translated to nfserr_jukebox anyway, so this doesn't change anything (good).
Quoted 20 lines
fs/nfsd/nfs4proc.c +0
-0goto out; }fs/nfsd/nfs4state.c +1
-0@@ -8554,6 +8554,7 @@ static int nfs4_state_create_net(struct net *net)spin_lock_init (&nn->client_lock);spin_lock_init (&nn->s2s_cp_lock);idr_init (&nn->s2s_cp_stateids); +atomic_set (&nn->pending_async_copies,0 );spin_lock_init (&nn->blocked_locks_lock);INIT_LIST_HEAD (&nn->blocked_locks_lru);fs/nfsd/xdr4.h +0
-0
fs/nfsd/xdr4.h struct nfsd4_copy { · line 719
Quoted 7 lines
fs/nfsd/xdr4.h +1
-0@@ -713,6 +713,7 @@ struct nfsd4_copy {struct nfsd4_ssc_umount_item *ss_nsui;struct nfs_fh c_fh;nfs4_stateid stateid; +struct nfsd_net *cp_nn; };static inline void nfsd4_copy_set_sync (struct nfsd4_copy *copy,bool sync)
A per-ns limit sounds like a reasonable place to start.
Quoted 18 lines · On Wed, 2024-08-28 at 13:40 -0400, [email protected] wrote:
From: Chuck Lever <[email protected]>
While working on
OFFLOAD_STATUSand other potential improvements to Linux NFS's COPY offload implementation, I've come up with a few server-side observability enhancements and one or two possible bug fixes. These are candidates to merge for v6.12.Comments welcome.
Chuck Lever (7):
NFSD: Async COPY result needs to return a write verifier
NFSD: Limit the number of concurrent async COPY operations
NFSD: Display copy stateids with conventional print formatting
NFSD: Record the callback stateid in copy tracepoints
NFSD: Clean up extra whitespace intrace_nfsd_copy_done
NFSD: Document callback stateid laundering
NFSD: Wrap async copy operations with trace points
Quoted 7 lines
fs/nfsd/netns.h | 1 + fs/nfsd/nfs4proc.c | 38 +++++++++--------- fs/nfsd/nfs4state.c | 37 ++++++++++++----- fs/nfsd/trace.h | 97 +++++++++++++++++++++++++++++++++++++++++---- fs/nfsd/xdr4.h | 1 + 5 files changed, 137 insertions(+), 37 deletions(-)
I sent some R-b's already but the rest of the series looks fine to me.
Re: [RFC PATCH 6/7] NFSD: Document callback stateid laundering
In reply to Olga KornievskaiaQuoted 38 lines · Olga Kornievskaia
Quoted 33 lines · On Wed, Aug 28, 2024 at 1:40 PM <[email protected]> wrote:
From: Chuck Lever <[email protected]>
NFSD removes COPY callback stateids after once lease period. This practice keeps the list of callback stateids limited to prevent a DoS possibility, but doesn't comply with the spirit of RFC 7862 Section 4.8, which says:
Quoted 4 lines
A copy offload stateid will be valid until either (A) the client or server restarts or (B) the client returns the resource by issuing an
OFFLOAD_CANCELoperation or the client replies to aCB_OFFLOADoperation.Note there are no BCP 14 compliance keywords in this text, so NFSD is free to ignore this stateid lifetime guideline without becoming non-compliant.
Nevertheless, this behavior variance should be explicitly documented in the code.
Signed-off-by Chuck Lever <[email protected]>---
fs/nfsd/nfs4state.c| 36 +++++++++++++++++++++++++-----------1 file changed, 25 insertions(+), 11 deletions(-)fs/nfsd/nfs4state.c +0
-0@@ -6324,6 +6324,29 @@ static void nfsd4_ssc_expire_umount(struct nfsd_net *nn) }#endif
Quoted 14 lines
fs/nfsd/nfs4state.c static void nfsd4_ssc_expire_umount(struct nfsd_net *nn) · line 6332
Quoted 6 lines
fs/nfsd/nfs4state.c +6
-0+/* +* RFC 7862 Section 4.8 says that, if the client hasn't replied to a +* CB_OFFLOAD, that COPY callback stateid will live until the client or +* server restarts. To prevent a DoS resulting from a pile of these +* stateids accruing over time, NFSD purges them after one lease period. +*/ I don't believe this is correct documentation for this piece of code. There are two kinds of stateids in the copy offload world: one is issued by the source server "
cnr_stateid" in the response of theCOPY_NOTIFYand given to be client (to be given to the destination server to use for the read) and those are the ones kept in the knfsd'ss2s_cp_stateidslist and then cleaned up by the laundry thread when copy isn't cleaned up on the source server.
Got it.
That detail was not included in what you mentioned to me before ;-) Either I can drop this patch, or please suggest a corrected text and I will replace the comment.
Quoted 2 lines
The text from the RFC
quoted here is for copy's callback stateids.
I'll look for something more relevant, if only for my own edification.
Quoted 5 lines
In the current
implementation, we don't keep callback stateids around past when the async copy is done. I agree that needs to be changed forOFFLOAD_STATUSop and then we can add the wording of how long we are keeping those.
Yep, as we discussed yesterday.
Quoted 47 lines
Quoted 47 lines
fs/nfsd/nfs4state.c +17
-0+static void nfs4_launder_cpntf_statelist (struct nfsd_net *nn, +struct laundry_time *lt) +{ +struct nfs4_cpntf_state *cps; +copy_stateid_t *cps_t; +int i; + +spin_lock (&nn->s2s_cp_lock); +idr_for_each_entry (&nn->s2s_cp_stateids, cps_t, i) { + cps = container_of(cps_t, struct nfs4_cpntf_state, cp_stateid); +if (cps->cp_stateid.cs_type ==NFS4_COPYNOTIFY_STID && +state_expired (lt, cps->cpntf_time)) +_free_cpntf_state_locked (nn, cps); + } +spin_unlock (&nn->s2s_cp_lock); +} +/* Check if any lock belonging to this lockowner has any blockers */ static bool
nfs4_lockowner_has_blockers(struct nfs4_lockowner *lo)fs/nfsd/nfs4state.c +2
-11@@ -6495,9 +6518,6 @@ nfs4_laundromat(struct nfsd_net *nn) .cutoff = ktime_get_boottime_seconds() - nn->nfsd4_lease, .new_timeo = nn->nfsd4_lease }; -struct nfs4_cpntf_state *cps; -copy_stateid_t *cps_t; -int i;if (clients_still_reclaiming (nn)) { lt.new_timeo =0 ;@@ -6505,14 +6525,8 @@ nfs4_laundromat(struct nfsd_net *nn) }nfsd4_end_grace (nn); -spin_lock (&nn->s2s_cp_lock); -idr_for_each_entry (&nn->s2s_cp_stateids, cps_t, i) { - cps = container_of(cps_t, struct nfs4_cpntf_state, cp_stateid); -if (cps->cp_stateid.cs_type ==NFS4_COPYNOTIFY_STID && -state_expired (<, cps->cpntf_time)) -_free_cpntf_state_locked (nn, cps); - } -spin_unlock (&nn->s2s_cp_lock); +nfs4_launder_cpntf_statelist (nn, <); +nfs4_get_client_reaplist (nn, &reaplist, <);nfs4_process_client_reaplist (&reaplist);
Quoted 2 lines
Quoted 2 lines
--
2.46.0
--
Chuck Lever
Quoted 22 lines · On Aug 28, 2024, at 1:40 PM, [email protected] wrote:
From: Chuck Lever <[email protected]>
While working on
OFFLOAD_STATUSand other potential improvements to Linux NFS's COPY offload implementation, I've come up with a few server-side observability enhancements and one or two possible bug fixes. These are candidates to merge for v6.12.Comments welcome.
Chuck Lever (7):
NFSD: Async COPY result needs to return a write verifier
NFSD: Limit the number of concurrent async COPY operations
NFSD: Display copy stateids with conventional print formatting
NFSD: Record the callback stateid in copy tracepoints
NFSD: Clean up extra whitespace intrace_nfsd_copy_done
NFSD: Document callback stateid laundering
NFSD: Wrap async copy operations with trace points
Quoted 9 lines
fs/nfsd/netns.h | 1 + fs/nfsd/nfs4proc.c | 38 +++++++++--------- fs/nfsd/nfs4state.c | 37 ++++++++++++----- fs/nfsd/trace.h | 97 +++++++++++++++++++++++++++++++++++++++++---- fs/nfsd/xdr4.h | 1 + 5 files changed, 137 insertions(+), 37 deletions(-)Signature · 1 line
2.46.0
I've applied all but "NFSD: Document callback stateid laundering" to nfsd-next.
--
Chuck Lever