Message ID | 172972079577.81717.6397186554656127040@noble.neil.brown.name (mailing list archive) |
---|---|
State | New |
Headers | show |
Series | [-,for,-next] nfsd: dynamically adjust per-client DRC slot limits. | expand |
On Thu, 24 Oct 2024 at 00:00, NeilBrown <neilb@suse.de> wrote: > > > Currently per-client DRC slot limits (for v4.1+) are calculated when the > client connects, and they are left unchanged. So earlier clients can > get a larger share when memory is tight. > > The heuristic for choosing a number includes the number of configured > server threads. This is problematic for 2 reasons. > 1/ sv_nrthreads is different in different net namespaces, but the > memory allocation is global across all namespaces. So different > namespaces get treated differently without good reason. > 2/ a future patch will auto-configure number of threads based on > load so that there will be no need to preconfigure a number. This will > make the current heuristic even more arbitrary. > > NFSv4.1 allows the number of slots to be varied dynamically - in the > reply to each SEQUENCE op. With this patch we provide a provisional > upper limit in the EXCHANGE_ID reply which might end up being too big, > and adjust it with each SEQUENCE reply. I have some concerns here: 1. How should this be tested? You are adding a random element to the test matrix, without an option to manually set a fixed value, or disable it. At least there should be a hook in /proc to set the number of DRC slots, so CI can simulate changes. 2. How do you provide backwards compatibility to clients which do not implement the number of slots adjustments per SEQUENCE op, or implement it wrong? Just claim "its the NFSv4.1 standard, and legacy clients should burn&die" is not an option, otherwise AWS&co will try a Jeanne d'Arc-style punishment on those responsible 3. How do you handle changes in memory, e.g. memory hotplug, or changes in mem assigned to the VM running the nfsd? Is swap space counted too? Finally: 4. Windows until Windows Server 2003 (leaked source code can be googled) had this kind of memory autoscaling for SMB. Each Windows release before Server 2003 had increasingly more complex code, with increasingly unpleasant curses in the source comments about unexpected and unwanted side effects. After Windows Server 2003 this whole machinery was yanked out, because it was much more harmful to customers than any savings in memory consumptions could ever bring. Why do you want to try to bring the same nightmarish experiment to the NFSv4 world, after Microsoft failed so horribly? Ced
> On Oct 28, 2024, at 3:45 AM, Cedric Blancher <cedric.blancher@gmail.com> wrote: > > 1. How should this be tested? You are adding a random element to the > test matrix, without an option to manually set a fixed value, or > disable it. At least there should be a hook in /proc to set the number > of DRC slots, so CI can simulate changes. Neil suggested exposing session parameters via /proc. I think I would prefer /sys instead. But perhaps we could make (some of) them writable as well as readable; Neil will have to let us know if that is feasible. pynfs already has some unit testing of SEQUENCE and CREATE_SESSION, but we don't (yet) have a lot of testing around server memory exhaustion. > 2. How do you provide backwards compatibility to clients which do not > implement the number of slots adjustments per SEQUENCE op, or > implement it wrong? Just claim "its the NFSv4.1 standard, and legacy > clients should burn&die" is not an option, otherwise AWS&co will try a > Jeanne d'Arc-style punishment on those responsible That's what a standard is for. Either a client implementation complies to the standard, and thus has interoperability guarantees; or it doesn't, and all bets are off. This isn't something unique to the NFS world, it's actually a very common approach to such issues. Do you know of any specific client implementation that would have difficulty with dynamic session resizing? Find an example and we can work through it. > 3. How do you handle changes in memory, e.g. memory hotplug, or > changes in mem assigned to the VM running the nfsd? It's certainly possible to add but has it been demonstrated that it is necessary on day zero? These are the corneriest of corner cases; I would prefer to avoid heroic interventions in this code until it just can't be avoided. I also note that a shrinker could signal NFSD when system memory conditions have changed. A shrinker callback does not have to release any memory, and can be used to opportunistically reduce future memory needs. > Is swap space counted too? It isn't currently. I can't think of a reason NFSD would need to do that. What were you thinking of? > 4. Windows until Windows Server 2003 (leaked source code can be > googled) had this kind of memory autoscaling for SMB. Each Windows > release before Server 2003 had increasingly more complex code, with > increasingly unpleasant curses in the source comments about unexpected > and unwanted side effects. After Windows Server 2003 this whole > machinery was yanked out, because it was much more harmful to > customers than any savings in memory consumptions could ever bring. > > Why do you want to try to bring the same nightmarish experiment to the > NFSv4 world, after Microsoft failed so horribly? You might be comparing apples and oranges, but I can't tell because there is so very little detail here that would enable us to say whether we are implementing anything remotely similar to what Microsoft might have done 20 years ago, nor do we know any specifics about the implementation issues they faced. We do have one or two former SMB folks on list who could provide some facts. But, suffice to say, other NFSD server implementations already implement dynamic session slot resizing, and there have been only minor issues with those. I do expect that ours will have to be tuned and tweaked over time. -- Chuck Lever
diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c index 56b261608af4..c92fc2029685 100644 --- a/fs/nfsd/nfs4state.c +++ b/fs/nfsd/nfs4state.c @@ -1909,44 +1909,26 @@ static inline u32 slot_bytes(struct nfsd4_channel_attrs *ca) } /* - * 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. + * When a client connects it gets a max_requests number that would allow + * it to use 1/8 of the memory we think can reasonably be used for the DRC. + * Later in response to SEQUENCE operations we further limit that when there + * are more than 8 concurrent clients. */ -static u32 nfsd4_get_drc_mem(struct nfsd4_channel_attrs *ca, struct nfsd_net *nn) +static u32 nfsd4_get_drc_mem(struct nfsd4_channel_attrs *ca) { u32 slotsize = slot_bytes(ca); u32 num = ca->maxreqs; - unsigned long avail, total_avail; - unsigned int scale_factor; + unsigned long avail; 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; + avail = min(NFSD_MAX_MEM_PER_SESSION, + nfsd_drc_max_mem / 8); + + num = clamp_t(int, num, 1, avail / slotsize); + + nfsd_drc_slotsize_sum += slotsize; + spin_unlock(&nfsd_drc_lock); return num; @@ -1957,10 +1939,33 @@ 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; + nfsd_drc_slotsize_sum -= slotsize; spin_unlock(&nfsd_drc_lock); } +/* + * Report the number of slots that we would like the client to limit + * itself to. When the number of clients is large, we start sharing + * memory so all clients get the same number of slots. + */ +static unsigned int nfsd4_get_drc_slots(struct nfsd4_session *session) +{ + u32 slotsize = slot_bytes(&session->se_fchannel); + unsigned long avail; + unsigned long slotsize_sum = READ_ONCE(nfsd_drc_slotsize_sum); + + if (slotsize_sum < slotsize) + slotsize_sum = slotsize; + + /* Find our share of avail mem across all active clients, + * then limit to 1/8 of total, and configured max + */ + avail = min3(nfsd_drc_max_mem * slotsize / slotsize_sum, + nfsd_drc_max_mem / 8, + NFSD_MAX_MEM_PER_SESSION); + return max3(1UL, avail / slotsize, session->se_fchannel.maxreqs); +} + static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs, struct nfsd4_channel_attrs *battrs) { @@ -3731,7 +3736,7 @@ static __be32 check_forechannel_attrs(struct nfsd4_channel_attrs *ca, struct nfs * 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); + ca->maxreqs = nfsd4_get_drc_mem(ca); return nfs_ok; } @@ -4225,10 +4230,12 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate, slot = 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; + /* Negotiate number of slots: set the target, and use the + * same for max as long as it doesn't decrease the max + * (that isn't allowed). + */ + seq->target_maxslots = nfsd4_get_drc_slots(session); + seq->maxslots = max(seq->maxslots, seq->target_maxslots); trace_nfsd_slot_seqid_sequence(clp, seq, slot); status = check_slot_seqid(seq->seqid, slot->sl_seqid, diff --git a/fs/nfsd/nfs4xdr.c b/fs/nfsd/nfs4xdr.c index f118921250c3..e4e706872e54 100644 --- a/fs/nfsd/nfs4xdr.c +++ b/fs/nfsd/nfs4xdr.c @@ -4955,7 +4955,7 @@ 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 */ diff --git a/fs/nfsd/nfsd.h b/fs/nfsd/nfsd.h index 4b56ba1e8e48..33c9db3ee8b6 100644 --- a/fs/nfsd/nfsd.h +++ b/fs/nfsd/nfsd.h @@ -90,7 +90,11 @@ 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; +/* Storing the sum of the slot sizes for all active clients (across + * all net-namespaces) allows a share of total available memory to + * be allocaed to each client based on its slot size. + */ +extern unsigned long nfsd_drc_slotsize_sum; extern atomic_t nfsd_th_cnt; /* number of available threads */ extern const struct seq_operations nfs_exports_op; diff --git a/fs/nfsd/nfssvc.c b/fs/nfsd/nfssvc.c index 49e2f32102ab..e596eb5d10db 100644 --- a/fs/nfsd/nfssvc.c +++ b/fs/nfsd/nfssvc.c @@ -78,7 +78,7 @@ DEFINE_MUTEX(nfsd_mutex); */ DEFINE_SPINLOCK(nfsd_drc_lock); unsigned long nfsd_drc_max_mem; -unsigned long nfsd_drc_mem_used; +unsigned long nfsd_drc_slotsize_sum; #if IS_ENABLED(CONFIG_NFS_LOCALIO) static const struct svc_version *localio_versions[] = { @@ -589,10 +589,13 @@ void nfsd_reset_versions(struct nfsd_net *nn) */ static void set_max_drc(void) { + if (nfsd_drc_max_mem) + return; + #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; + nfsd_drc_slotsize_sum = 0; dprintk("%s nfsd_drc_max_mem %lu \n", __func__, nfsd_drc_max_mem); } diff --git a/fs/nfsd/state.h b/fs/nfsd/state.h index 79c743c01a47..fe71ae3c577b 100644 --- a/fs/nfsd/state.h +++ b/fs/nfsd/state.h @@ -214,7 +214,7 @@ static inline struct nfs4_delegation *delegstateid(struct nfs4_stid *s) /* Maximum number of slots per session. 160 is useful for long haul TCP */ #define NFSD_MAX_SLOTS_PER_SESSION 160 /* Maximum session per slot cache size */ -#define NFSD_SLOT_CACHE_SIZE 2048 +#define NFSD_SLOT_CACHE_SIZE 2048UL /* Maximum number of NFSD_SLOT_CACHE_SIZE slots per session */ #define NFSD_CACHE_SIZE_SLOTS_PER_SESSION 32 #define NFSD_MAX_MEM_PER_SESSION \ diff --git a/fs/nfsd/xdr4.h b/fs/nfsd/xdr4.h index 2a21a7662e03..71b87190a4a6 100644 --- a/fs/nfsd/xdr4.h +++ b/fs/nfsd/xdr4.h @@ -575,9 +575,7 @@ 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 */ };
Currently per-client DRC slot limits (for v4.1+) are calculated when the client connects, and they are left unchanged. So earlier clients can get a larger share when memory is tight. The heuristic for choosing a number includes the number of configured server threads. This is problematic for 2 reasons. 1/ sv_nrthreads is different in different net namespaces, but the memory allocation is global across all namespaces. So different namespaces get treated differently without good reason. 2/ a future patch will auto-configure number of threads based on load so that there will be no need to preconfigure a number. This will make the current heuristic even more arbitrary. NFSv4.1 allows the number of slots to be varied dynamically - in the reply to each SEQUENCE op. With this patch we provide a provisional upper limit in the EXCHANGE_ID reply which might end up being too big, and adjust it with each SEQUENCE reply. The goal for when memory is tight is to allow each client to have a similar number of slots. So clients that ask for larger slots get more memory. This may not be ideal. It could be changed later. So we track the sum of the slot sizes of all active clients, and share memory out based on the ratio of the slot size for a given client with the sum of slot sizes. We never allow more in a SEQUENCE reply than we did in the EXCHANGE_ID reply. Signed-off-by: NeilBrown <neilb@suse.de> --- As requested, here is that dynamic-DRC patch against nfsd-next, with a small comment change suggested by Tom. fs/nfsd/nfs4state.c | 81 ++++++++++++++++++++++++--------------------- fs/nfsd/nfs4xdr.c | 2 +- fs/nfsd/nfsd.h | 6 +++- fs/nfsd/nfssvc.c | 7 ++-- fs/nfsd/state.h | 2 +- fs/nfsd/xdr4.h | 2 -- 6 files changed, 56 insertions(+), 44 deletions(-)