On Sun, 2024-11-10 at 21:19 -0500, Olga Kornievskaia wrote: > On Sat, Nov 9, 2024 at 2:26 PM Jeff Layton <jlayton@xxxxxxxxxx> wrote: > > > > On Sat, 2024-11-09 at 13:50 -0500, Olga Kornievskaia wrote: > > > On Tue, Nov 5, 2024 at 7:31 PM Jeff Layton <jlayton@xxxxxxxxxx> wrote: > > > > > > > > nfsd currently only uses a single slot in the callback channel, which is > > > > proving to be a bottleneck in some cases. Widen the callback channel to > > > > a max of 32 slots (subject to the client's target_maxreqs value). > > > > > > > > Change the cb_holds_slot boolean to an integer that tracks the current > > > > slot number (with -1 meaning "unassigned"). Move the callback slot > > > > tracking info into the session. Add a new u32 that acts as a bitmap to > > > > track which slots are in use, and a u32 to track the latest callback > > > > target_slotid that the client reports. To protect the new fields, add > > > > a new per-session spinlock (the se_lock). Fix nfsd41_cb_get_slot to always > > > > search for the lowest slotid (using ffs()). > > > > > > > > Finally, convert the session->se_cb_seq_nr field into an array of > > > > counters and add the necessary handling to ensure that the seqids get > > > > reset at the appropriate times. > > > > > > > > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx> > > > > Signed-off-by: Chuck Lever <chuck.lever@xxxxxxxxxx> > > > > --- > > > > v3 has a bug that Olga hit in testing. This version should fix the wait > > > > when the slot table is full. Olga, if you're able to test this one, it > > > > would be much appreciated. > > > > --- > > > > Changes in v4: > > > > - Fix the wait for a slot in nfsd41_cb_get_slot() > > > > - Link to v3: https://lore.kernel.org/r/20241030-bcwide-v3-0-c2df49a26c45@xxxxxxxxxx > > > > > > > > Changes in v3: > > > > - add patch to convert se_flags to single se_dead bool > > > > - fix off-by-one bug in handling of NFSD_BC_SLOT_TABLE_MAX > > > > - don't reject target highest slot value of 0 > > > > - Link to v2: https://lore.kernel.org/r/20241029-bcwide-v2-1-e9010b6ef55d@xxxxxxxxxx > > > > > > > > Changes in v2: > > > > - take cl_lock when fetching fields from session to be encoded > > > > - use fls() instead of bespoke highest_unset_index() > > > > - rename variables in several functions with more descriptive names > > > > - clamp limit of for loop in update_cb_slot_table() > > > > - re-add missing rpc_wake_up_queued_task() call > > > > - fix slotid check in decode_cb_sequence4resok() > > > > - add new per-session spinlock > > > > --- > > > > fs/nfsd/nfs4callback.c | 113 ++++++++++++++++++++++++++++++++++++------------- > > > > fs/nfsd/nfs4state.c | 11 +++-- > > > > fs/nfsd/state.h | 15 ++++--- > > > > fs/nfsd/trace.h | 2 +- > > > > 4 files changed, 101 insertions(+), 40 deletions(-) > > > > > > > > diff --git a/fs/nfsd/nfs4callback.c b/fs/nfsd/nfs4callback.c > > > > index e38fa834b3d91333acf1425eb14c644e5d5f2601..47a678333907eaa92db305dada503704c34c15b2 100644 > > > > --- a/fs/nfsd/nfs4callback.c > > > > +++ b/fs/nfsd/nfs4callback.c > > > > @@ -406,6 +406,19 @@ encode_cb_getattr4args(struct xdr_stream *xdr, struct nfs4_cb_compound_hdr *hdr, > > > > hdr->nops++; > > > > } > > > > > > > > +static u32 highest_slotid(struct nfsd4_session *ses) > > > > +{ > > > > + u32 idx; > > > > + > > > > + spin_lock(&ses->se_lock); > > > > + idx = fls(~ses->se_cb_slot_avail); > > > > + if (idx > 0) > > > > + --idx; > > > > + idx = max(idx, ses->se_cb_highest_slot); > > > > + spin_unlock(&ses->se_lock); > > > > + return idx; > > > > +} > > > > + > > > > /* > > > > * CB_SEQUENCE4args > > > > * > > > > @@ -432,15 +445,35 @@ static void encode_cb_sequence4args(struct xdr_stream *xdr, > > > > encode_sessionid4(xdr, session); > > > > > > > > p = xdr_reserve_space(xdr, 4 + 4 + 4 + 4 + 4); > > > > - *p++ = cpu_to_be32(session->se_cb_seq_nr); /* csa_sequenceid */ > > > > - *p++ = xdr_zero; /* csa_slotid */ > > > > - *p++ = xdr_zero; /* csa_highest_slotid */ > > > > + *p++ = cpu_to_be32(session->se_cb_seq_nr[cb->cb_held_slot]); /* csa_sequenceid */ > > > > + *p++ = cpu_to_be32(cb->cb_held_slot); /* csa_slotid */ > > > > + *p++ = cpu_to_be32(highest_slotid(session)); /* csa_highest_slotid */ > > > > *p++ = xdr_zero; /* csa_cachethis */ > > > > xdr_encode_empty_array(p); /* csa_referring_call_lists */ > > > > > > > > hdr->nops++; > > > > } > > > > > > > > +static void update_cb_slot_table(struct nfsd4_session *ses, u32 target) > > > > +{ > > > > + /* No need to do anything if nothing changed */ > > > > + if (likely(target == READ_ONCE(ses->se_cb_highest_slot))) > > > > + return; > > > > + > > > > + spin_lock(&ses->se_lock); > > > > + if (target > ses->se_cb_highest_slot) { > > > > + int i; > > > > + > > > > + target = min(target, NFSD_BC_SLOT_TABLE_MAX); > > > > + > > > > + /* Growing the slot table. Reset any new sequences to 1 */ > > > > + for (i = ses->se_cb_highest_slot + 1; i <= target; ++i) > > > > + ses->se_cb_seq_nr[i] = 1; > > > > + } > > > > + ses->se_cb_highest_slot = target; > > > > + spin_unlock(&ses->se_lock); > > > > +} > > > > + > > > > /* > > > > * CB_SEQUENCE4resok > > > > * > > > > @@ -468,7 +501,7 @@ static int decode_cb_sequence4resok(struct xdr_stream *xdr, > > > > struct nfsd4_session *session = cb->cb_clp->cl_cb_session; > > > > int status = -ESERVERFAULT; > > > > __be32 *p; > > > > - u32 dummy; > > > > + u32 seqid, slotid, target; > > > > > > > > /* > > > > * If the server returns different values for sessionID, slotID or > > > > @@ -484,21 +517,22 @@ static int decode_cb_sequence4resok(struct xdr_stream *xdr, > > > > } > > > > p += XDR_QUADLEN(NFS4_MAX_SESSIONID_LEN); > > > > > > > > - dummy = be32_to_cpup(p++); > > > > - if (dummy != session->se_cb_seq_nr) { > > > > + seqid = be32_to_cpup(p++); > > > > + if (seqid != session->se_cb_seq_nr[cb->cb_held_slot]) { > > > > dprintk("NFS: %s Invalid sequence number\n", __func__); > > > > goto out; > > > > } > > > > > > > > - dummy = be32_to_cpup(p++); > > > > - if (dummy != 0) { > > > > + slotid = be32_to_cpup(p++); > > > > + if (slotid != cb->cb_held_slot) { > > > > dprintk("NFS: %s Invalid slotid\n", __func__); > > > > goto out; > > > > } > > > > > > > > - /* > > > > - * FIXME: process highest slotid and target highest slotid > > > > - */ > > > > + p++; // ignore current highest slot value > > > > + > > > > + target = be32_to_cpup(p++); > > > > + update_cb_slot_table(session, target); > > > > status = 0; > > > > out: > > > > cb->cb_seq_status = status; > > > > @@ -1203,6 +1237,22 @@ void nfsd4_change_callback(struct nfs4_client *clp, struct nfs4_cb_conn *conn) > > > > spin_unlock(&clp->cl_lock); > > > > } > > > > > > > > +static int grab_slot(struct nfsd4_session *ses) > > > > +{ > > > > + int idx; > > > > + > > > > + spin_lock(&ses->se_lock); > > > > + idx = ffs(ses->se_cb_slot_avail) - 1; > > > > + if (idx < 0 || idx > ses->se_cb_highest_slot) { > > > > + spin_unlock(&ses->se_lock); > > > > + return -1; > > > > + } > > > > + /* clear the bit for the slot */ > > > > + ses->se_cb_slot_avail &= ~BIT(idx); > > > > + spin_unlock(&ses->se_lock); > > > > + return idx; > > > > +} > > > > + > > > > /* > > > > * There's currently a single callback channel slot. > > > > * If the slot is available, then mark it busy. Otherwise, set the > > > > @@ -1211,28 +1261,32 @@ void nfsd4_change_callback(struct nfs4_client *clp, struct nfs4_cb_conn *conn) > > > > static bool nfsd41_cb_get_slot(struct nfsd4_callback *cb, struct rpc_task *task) > > > > { > > > > struct nfs4_client *clp = cb->cb_clp; > > > > + struct nfsd4_session *ses = clp->cl_cb_session; > > > > > > > > - if (!cb->cb_holds_slot && > > > > - test_and_set_bit(0, &clp->cl_cb_slot_busy) != 0) { > > > > + if (cb->cb_held_slot >= 0) > > > > + return true; > > > > + cb->cb_held_slot = grab_slot(ses); > > > > + if (cb->cb_held_slot < 0) { > > > > rpc_sleep_on(&clp->cl_cb_waitq, task, NULL); > > > > /* Race breaker */ > > > > - if (test_and_set_bit(0, &clp->cl_cb_slot_busy) != 0) { > > > > - dprintk("%s slot is busy\n", __func__); > > > > + cb->cb_held_slot = grab_slot(ses); > > > > + if (cb->cb_held_slot < 0) > > > > return false; > > > > - } > > > > rpc_wake_up_queued_task(&clp->cl_cb_waitq, task); > > > > } > > > > - cb->cb_holds_slot = true; > > > > return true; > > > > } > > > > > > > > static void nfsd41_cb_release_slot(struct nfsd4_callback *cb) > > > > { > > > > struct nfs4_client *clp = cb->cb_clp; > > > > + struct nfsd4_session *ses = clp->cl_cb_session; > > > > > > > > - if (cb->cb_holds_slot) { > > > > - cb->cb_holds_slot = false; > > > > - clear_bit(0, &clp->cl_cb_slot_busy); > > > > + if (cb->cb_held_slot >= 0) { > > > > + spin_lock(&ses->se_lock); > > > > + ses->se_cb_slot_avail |= BIT(cb->cb_held_slot); > > > > + spin_unlock(&ses->se_lock); > > > > + cb->cb_held_slot = -1; > > > > rpc_wake_up_next(&clp->cl_cb_waitq); > > > > } > > > > } > > > > @@ -1249,8 +1303,8 @@ static void nfsd41_destroy_cb(struct nfsd4_callback *cb) > > > > } > > > > > > > > /* > > > > - * TODO: cb_sequence should support referring call lists, cachethis, multiple > > > > - * slots, and mark callback channel down on communication errors. > > > > + * TODO: cb_sequence should support referring call lists, cachethis, > > > > + * and mark callback channel down on communication errors. > > > > */ > > > > static void nfsd4_cb_prepare(struct rpc_task *task, void *calldata) > > > > { > > > > @@ -1292,7 +1346,7 @@ static bool nfsd4_cb_sequence_done(struct rpc_task *task, struct nfsd4_callback > > > > return true; > > > > } > > > > > > > > - if (!cb->cb_holds_slot) > > > > + if (cb->cb_held_slot < 0) > > > > goto need_restart; > > > > > > > > /* This is the operation status code for CB_SEQUENCE */ > > > > @@ -1306,10 +1360,10 @@ static bool nfsd4_cb_sequence_done(struct rpc_task *task, struct nfsd4_callback > > > > * If CB_SEQUENCE returns an error, then the state of the slot > > > > * (sequence ID, cached reply) MUST NOT change. > > > > */ > > > > - ++session->se_cb_seq_nr; > > > > + ++session->se_cb_seq_nr[cb->cb_held_slot]; > > > > break; > > > > case -ESERVERFAULT: > > > > - ++session->se_cb_seq_nr; > > > > + ++session->se_cb_seq_nr[cb->cb_held_slot]; > > > > nfsd4_mark_cb_fault(cb->cb_clp); > > > > ret = false; > > > > break; > > > > @@ -1335,17 +1389,16 @@ static bool nfsd4_cb_sequence_done(struct rpc_task *task, struct nfsd4_callback > > > > case -NFS4ERR_BADSLOT: > > > > goto retry_nowait; > > > > case -NFS4ERR_SEQ_MISORDERED: > > > > - if (session->se_cb_seq_nr != 1) { > > > > - session->se_cb_seq_nr = 1; > > > > + if (session->se_cb_seq_nr[cb->cb_held_slot] != 1) { > > > > + session->se_cb_seq_nr[cb->cb_held_slot] = 1; > > > > goto retry_nowait; > > > > } > > > > break; > > > > default: > > > > nfsd4_mark_cb_fault(cb->cb_clp); > > > > } > > > > - nfsd41_cb_release_slot(cb); > > > > - > > > > trace_nfsd_cb_free_slot(task, cb); > > > > + nfsd41_cb_release_slot(cb); > > > > > > > > if (RPC_SIGNALLED(task)) > > > > goto need_restart; > > > > @@ -1565,7 +1618,7 @@ void nfsd4_init_cb(struct nfsd4_callback *cb, struct nfs4_client *clp, > > > > INIT_WORK(&cb->cb_work, nfsd4_run_cb_work); > > > > cb->cb_status = 0; > > > > cb->cb_need_restart = false; > > > > - cb->cb_holds_slot = false; > > > > + cb->cb_held_slot = -1; > > > > } > > > > > > > > /** > > > > diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c > > > > index baf7994131fe1b0a4715174ba943fd2a9882aa12..75557e7cc9265517f51952563beaa4cfe8adcc3f 100644 > > > > --- a/fs/nfsd/nfs4state.c > > > > +++ b/fs/nfsd/nfs4state.c > > > > @@ -2002,6 +2002,9 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, > > > > } > > > > > > > > memcpy(&new->se_fchannel, fattrs, sizeof(struct nfsd4_channel_attrs)); > > > > + new->se_cb_slot_avail = ~0U; > > > > + new->se_cb_highest_slot = battrs->maxreqs - 1; > > > > + spin_lock_init(&new->se_lock); > > > > return new; > > > > out_free: > > > > while (i--) > > > > @@ -2132,11 +2135,14 @@ static void init_session(struct svc_rqst *rqstp, struct nfsd4_session *new, stru > > > > > > > > INIT_LIST_HEAD(&new->se_conns); > > > > > > > > - new->se_cb_seq_nr = 1; > > > > + atomic_set(&new->se_ref, 0); > > > > new->se_dead = false; > > > > new->se_cb_prog = cses->callback_prog; > > > > new->se_cb_sec = cses->cb_sec; > > > > - atomic_set(&new->se_ref, 0); > > > > + > > > > + for (idx = 0; idx < NFSD_BC_SLOT_TABLE_MAX; ++idx) > > > > + new->se_cb_seq_nr[idx] = 1; > > > > + > > > > idx = hash_sessionid(&new->se_sessionid); > > > > list_add(&new->se_hash, &nn->sessionid_hashtbl[idx]); > > > > spin_lock(&clp->cl_lock); > > > > @@ -3159,7 +3165,6 @@ static struct nfs4_client *create_client(struct xdr_netobj name, > > > > kref_init(&clp->cl_nfsdfs.cl_ref); > > > > nfsd4_init_cb(&clp->cl_cb_null, clp, NULL, NFSPROC4_CLNT_CB_NULL); > > > > clp->cl_time = ktime_get_boottime_seconds(); > > > > - clear_bit(0, &clp->cl_cb_slot_busy); > > > > copy_verf(clp, verf); > > > > memcpy(&clp->cl_addr, sa, sizeof(struct sockaddr_storage)); > > > > clp->cl_cb_session = NULL; > > > > diff --git a/fs/nfsd/state.h b/fs/nfsd/state.h > > > > index d22e4f2c9039324a0953a9e15a3c255fb8ee1a44..848d023cb308f0b69916c4ee34b09075708f0de3 100644 > > > > --- a/fs/nfsd/state.h > > > > +++ b/fs/nfsd/state.h > > > > @@ -71,8 +71,8 @@ struct nfsd4_callback { > > > > struct work_struct cb_work; > > > > int cb_seq_status; > > > > int cb_status; > > > > + int cb_held_slot; > > > > bool cb_need_restart; > > > > - bool cb_holds_slot; > > > > }; > > > > > > > > struct nfsd4_callback_ops { > > > > @@ -307,6 +307,9 @@ struct nfsd4_conn { > > > > unsigned char cn_flags; > > > > }; > > > > > > > > +/* Highest slot index that nfsd implements in NFSv4.1+ backchannel */ > > > > +#define NFSD_BC_SLOT_TABLE_MAX (sizeof(u32) * 8 - 1) > > > > > > Are there some values that are known not to work? I was experimenting > > > with values and set it to 2 and 4 and the kernel oopsed. I understand > > > it's not a configurable value but it would still be good to know the > > > expectations... > > > > > > [ 198.625021] Unable to handle kernel paging request at virtual > > > address dfff800020000000 > > > [ 198.625870] KASAN: probably user-memory-access in range > > > [0x0000000100000000-0x0000000100000007] > > > [ 198.626444] Mem abort info: > > > [ 198.626630] ESR = 0x0000000096000005 > > > [ 198.626882] EC = 0x25: DABT (current EL), IL = 32 bits > > > [ 198.627234] SET = 0, FnV = 0 > > > [ 198.627441] EA = 0, S1PTW = 0 > > > [ 198.627627] FSC = 0x05: level 1 translation fault > > > [ 198.627859] Data abort info: > > > [ 198.628000] ISV = 0, ISS = 0x00000005, ISS2 = 0x00000000 > > > [ 198.628272] CM = 0, WnR = 0, TnD = 0, TagAccess = 0 > > > [ 198.628619] GCS = 0, Overlay = 0, DirtyBit = 0, Xs = 0 > > > [ 198.628967] [dfff800020000000] address between user and kernel address ranges > > > [ 198.629438] Internal error: Oops: 0000000096000005 [#1] SMP > > > [ 198.629806] Modules linked in: rpcsec_gss_krb5 nfsv4 dns_resolver > > > nfs netfs nfnetlink_queue nfnetlink_log nfnetlink bluetooth cfg80211 > > > rpcrdma rdma_cm iw_cm ib_cm ib_core nfsd auth_rpcgss nfs_acl lockd > > > grace isofs uinput snd_seq_dummy snd_hrtimer vsock_loopback > > > vmw_vsock_virtio_transport_common qrtr rfkill vmw_vsock_vmci_transport > > > vsock sunrpc vfat fat snd_hda_codec_generic snd_hda_intel > > > snd_intel_dspcfg snd_hda_codec snd_hda_core snd_hwdep snd_seq uvcvideo > > > videobuf2_vmalloc snd_seq_device videobuf2_memops uvc videobuf2_v4l2 > > > videodev snd_pcm videobuf2_common mc snd_timer snd vmw_vmci soundcore > > > xfs libcrc32c vmwgfx drm_ttm_helper ttm nvme drm_kms_helper > > > crct10dif_ce nvme_core ghash_ce sha2_ce sha256_arm64 sha1_ce drm > > > nvme_auth sr_mod cdrom e1000e sg fuse > > > [ 198.633799] CPU: 5 UID: 0 PID: 6081 Comm: nfsd Kdump: loaded Not > > > tainted 6.12.0-rc6+ #47 > > > [ 198.634345] Hardware name: VMware, Inc. VMware20,1/VBSA, BIOS > > > VMW201.00V.21805430.BA64.2305221830 05/22/2023 > > > [ 198.635014] pstate: 11400005 (nzcV daif +PAN -UAO -TCO +DIT -SSBS BTYPE=--) > > > [ 198.635492] pc : nfsd4_sequence+0x5a0/0x1f60 [nfsd] > > > [ 198.635798] lr : nfsd4_sequence+0x340/0x1f60 [nfsd] > > > [ 198.636065] sp : ffff8000884977e0 > > > [ 198.636234] x29: ffff800088497910 x28: ffff0000b1b39280 x27: ffff0000ab508128 > > > [ 198.636624] x26: ffff0000b1b39298 x25: ffff0000b1b39290 x24: ffff0000a65e1c64 > > > [ 198.637049] x23: 1fffe000212e6804 x22: ffff000109734024 x21: 1ffff00011092f16 > > > [ 198.637472] x20: ffff00010aed8000 x19: ffff000109734000 x18: 1fffe0002de20c8b > > > [ 198.637883] x17: 0100000000000000 x16: 1ffff0000fcef234 x15: 1fffe000212e600f > > > [ 198.638286] x14: ffff80007e779000 x13: ffff80007e7791a0 x12: 0000000000000000 > > > [ 198.638697] x11: ffff0000a65e1c38 x10: ffff00010aedaca0 x9 : 1fffe000215db594 > > > [ 198.639110] x8 : 1fffe00014cbc387 x7 : ffff0000a65e1c03 x6 : ffff0000a65e1c00 > > > [ 198.639541] x5 : ffff0000a65e1c00 x4 : 0000000020000000 x3 : 0000000100000001 > > > [ 198.639962] x2 : ffff000109730060 x1 : 0000000000000003 x0 : dfff800000000000 > > > [ 198.640332] Call trace: > > > [ 198.640460] nfsd4_sequence+0x5a0/0x1f60 [nfsd] > > > [ 198.640715] nfsd4_proc_compound+0xb94/0x23b0 [nfsd] > > > [ 198.640997] nfsd_dispatch+0x22c/0x718 [nfsd] > > > [ 198.641260] svc_process_common+0x8e8/0x1968 [sunrpc] > > > [ 198.641566] svc_process+0x3d4/0x7e0 [sunrpc] > > > [ 198.641827] svc_handle_xprt+0x828/0xe10 [sunrpc] > > > [ 198.642108] svc_recv+0x2cc/0x6a8 [sunrpc] > > > [ 198.642346] nfsd+0x270/0x400 [nfsd] > > > [ 198.642562] kthread+0x288/0x310 > > > [ 198.642745] ret_from_fork+0x10/0x20 > > > [ 198.642937] Code: f2fbffe0 f9003be4 f94007e2 52800061 (38e06880) > > > [ 198.643267] SMP: stopping secondary CPUs > > > > > > > > > > > > > > > Good catch. I think the problem here is that we don't currently cap the > > initial value of se_cb_highest_slot at NFSD_BC_SLOT_TABLE_MAX. Does > > this patch prevent the panic? > > > > diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c > > index 3afe56ab9e0a..839be4ba765a 100644 > > --- a/fs/nfsd/nfs4state.c > > +++ b/fs/nfsd/nfs4state.c > > @@ -2011,7 +2011,7 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, > > > > memcpy(&new->se_fchannel, fattrs, sizeof(struct nfsd4_channel_attrs)); > > new->se_cb_slot_avail = ~0U; > > - new->se_cb_highest_slot = battrs->maxreqs - 1; > > + new->se_cb_highest_slot = min(battrs->maxreqs - 1, NFSD_BC_SLOT_TABLE_MAX); > > spin_lock_init(&new->se_lock); > > return new; > > out_free: > > It does help. I thought that the CREATE_SESSION reply for the > backchannel would be guided by the NFSD_BC_SLOT_TABLE_MAX value but > instead it seems like it's not. But yes I can see that the highest > slot used by the server is capped by the NFSD_BC_SLOT_TABLE_MAX value. Thanks for testing it, Olga. Chuck, would you be OK with folding the above delta into 9ab4c4077de9, or would you rather I resend the patch? -- Jeff Layton <jlayton@xxxxxxxxxx>
