| Message ID | 20240421175248.1692552-1-edumazet@google.com (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | 3584718cf2ec7e79b6814f2596dcf398c5fb2eca |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [net] net: fix sk_memory_allocated_{add|sub} vs softirqs | expand |
On Sun, Apr 21, 2024 at 1:52 PM Eric Dumazet <edumazet@google.com> wrote: > > Jonathan Heathcote reported a regression caused by blamed commit > on aarch64 architecture. > > x86 happens to have irq-safe __this_cpu_add_return() > and __this_cpu_sub(), but this is not generic. > > I think my confusion came from "struct sock" argument, > because these helpers are called with a locked socket. > But the memory accounting is per-proto (and per-cpu after > the blamed commit). We might cleanup these helpers later > to directly accept a "struct proto *proto" argument. > > Switch to this_cpu_add_return() and this_cpu_xchg() > operations, and get rid of preempt_disable()/preempt_enable() pairs. > > Fast path becomes a bit faster as a result :) > > Many thanks to Jonathan Heathcote for his awesome report and > investigations. > > Fixes: 3cd3399dd7a8 ("net: implement per-cpu reserves for memory_allocated") > Reported-by: Jonathan Heathcote <jonathan.heathcote@bbc.co.uk> > Closes: https://lore.kernel.org/netdev/VI1PR01MB42407D7947B2EA448F1E04EFD10D2@VI1PR01MB4240.eurprd01.prod.exchangelabs.com/ > Signed-off-by: Eric Dumazet <edumazet@google.com> > Cc: Soheil Hassas Yeganeh <soheil@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Very nice catch! and thank you for the fix! > --- > include/net/sock.h | 38 ++++++++++++++++++++------------------ > 1 file changed, 20 insertions(+), 18 deletions(-) > > diff --git a/include/net/sock.h b/include/net/sock.h > index f57bfd8a2ad2deaedf3f351325ab9336ae040504..b4b553df7870c0290ae632c51828ad7161ba332d 100644 > --- a/include/net/sock.h > +++ b/include/net/sock.h > @@ -1410,32 +1410,34 @@ sk_memory_allocated(const struct sock *sk) > #define SK_MEMORY_PCPU_RESERVE (1 << (20 - PAGE_SHIFT)) > extern int sysctl_mem_pcpu_rsv; > > +static inline void proto_memory_pcpu_drain(struct proto *proto) > +{ > + int val = this_cpu_xchg(*proto->per_cpu_fw_alloc, 0); > + > + if (val) > + atomic_long_add(val, proto->memory_allocated); > +} > + > static inline void > -sk_memory_allocated_add(struct sock *sk, int amt) > +sk_memory_allocated_add(const struct sock *sk, int val) > { > - int local_reserve; > + struct proto *proto = sk->sk_prot; > > - preempt_disable(); > - local_reserve = __this_cpu_add_return(*sk->sk_prot->per_cpu_fw_alloc, amt); > - if (local_reserve >= READ_ONCE(sysctl_mem_pcpu_rsv)) { > - __this_cpu_sub(*sk->sk_prot->per_cpu_fw_alloc, local_reserve); > - atomic_long_add(local_reserve, sk->sk_prot->memory_allocated); > - } > - preempt_enable(); > + val = this_cpu_add_return(*proto->per_cpu_fw_alloc, val); > + > + if (unlikely(val >= READ_ONCE(sysctl_mem_pcpu_rsv))) > + proto_memory_pcpu_drain(proto); > } > > static inline void > -sk_memory_allocated_sub(struct sock *sk, int amt) > +sk_memory_allocated_sub(const struct sock *sk, int val) > { > - int local_reserve; > + struct proto *proto = sk->sk_prot; > > - preempt_disable(); > - local_reserve = __this_cpu_sub_return(*sk->sk_prot->per_cpu_fw_alloc, amt); > - if (local_reserve <= -READ_ONCE(sysctl_mem_pcpu_rsv)) { > - __this_cpu_sub(*sk->sk_prot->per_cpu_fw_alloc, local_reserve); > - atomic_long_add(local_reserve, sk->sk_prot->memory_allocated); > - } > - preempt_enable(); > + val = this_cpu_sub_return(*proto->per_cpu_fw_alloc, val); > + > + if (unlikely(val <= -READ_ONCE(sysctl_mem_pcpu_rsv))) > + proto_memory_pcpu_drain(proto); > } > > #define SK_ALLOC_PERCPU_COUNTER_BATCH 16 > -- > 2.44.0.769.g3c40516874-goog >
On Sun, Apr 21, 2024 at 05:52:48PM +0000, Eric Dumazet wrote: > Jonathan Heathcote reported a regression caused by blamed commit > on aarch64 architecture. > > x86 happens to have irq-safe __this_cpu_add_return() > and __this_cpu_sub(), but this is not generic. > > I think my confusion came from "struct sock" argument, > because these helpers are called with a locked socket. > But the memory accounting is per-proto (and per-cpu after > the blamed commit). We might cleanup these helpers later > to directly accept a "struct proto *proto" argument. > > Switch to this_cpu_add_return() and this_cpu_xchg() > operations, and get rid of preempt_disable()/preempt_enable() pairs. > > Fast path becomes a bit faster as a result :) > > Many thanks to Jonathan Heathcote for his awesome report and > investigations. > > Fixes: 3cd3399dd7a8 ("net: implement per-cpu reserves for memory_allocated") > Reported-by: Jonathan Heathcote <jonathan.heathcote@bbc.co.uk> > Closes: https://lore.kernel.org/netdev/VI1PR01MB42407D7947B2EA448F1E04EFD10D2@VI1PR01MB4240.eurprd01.prod.exchangelabs.com/ > Signed-off-by: Eric Dumazet <edumazet@google.com> > Cc: Soheil Hassas Yeganeh <soheil@google.com> Reviewed-by: Shakeel Butt <shakeel.butt@linux.dev>
Hello: This patch was applied to netdev/net.git (main) by Jakub Kicinski <kuba@kernel.org>: On Sun, 21 Apr 2024 17:52:48 +0000 you wrote: > Jonathan Heathcote reported a regression caused by blamed commit > on aarch64 architecture. > > x86 happens to have irq-safe __this_cpu_add_return() > and __this_cpu_sub(), but this is not generic. > > I think my confusion came from "struct sock" argument, > because these helpers are called with a locked socket. > But the memory accounting is per-proto (and per-cpu after > the blamed commit). We might cleanup these helpers later > to directly accept a "struct proto *proto" argument. > > [...] Here is the summary with links: - [net] net: fix sk_memory_allocated_{add|sub} vs softirqs https://git.kernel.org/netdev/net/c/3584718cf2ec You are awesome, thank you!
diff --git a/include/net/sock.h b/include/net/sock.h index f57bfd8a2ad2deaedf3f351325ab9336ae040504..b4b553df7870c0290ae632c51828ad7161ba332d 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -1410,32 +1410,34 @@ sk_memory_allocated(const struct sock *sk) #define SK_MEMORY_PCPU_RESERVE (1 << (20 - PAGE_SHIFT)) extern int sysctl_mem_pcpu_rsv; +static inline void proto_memory_pcpu_drain(struct proto *proto) +{ + int val = this_cpu_xchg(*proto->per_cpu_fw_alloc, 0); + + if (val) + atomic_long_add(val, proto->memory_allocated); +} + static inline void -sk_memory_allocated_add(struct sock *sk, int amt) +sk_memory_allocated_add(const struct sock *sk, int val) { - int local_reserve; + struct proto *proto = sk->sk_prot; - preempt_disable(); - local_reserve = __this_cpu_add_return(*sk->sk_prot->per_cpu_fw_alloc, amt); - if (local_reserve >= READ_ONCE(sysctl_mem_pcpu_rsv)) { - __this_cpu_sub(*sk->sk_prot->per_cpu_fw_alloc, local_reserve); - atomic_long_add(local_reserve, sk->sk_prot->memory_allocated); - } - preempt_enable(); + val = this_cpu_add_return(*proto->per_cpu_fw_alloc, val); + + if (unlikely(val >= READ_ONCE(sysctl_mem_pcpu_rsv))) + proto_memory_pcpu_drain(proto); } static inline void -sk_memory_allocated_sub(struct sock *sk, int amt) +sk_memory_allocated_sub(const struct sock *sk, int val) { - int local_reserve; + struct proto *proto = sk->sk_prot; - preempt_disable(); - local_reserve = __this_cpu_sub_return(*sk->sk_prot->per_cpu_fw_alloc, amt); - if (local_reserve <= -READ_ONCE(sysctl_mem_pcpu_rsv)) { - __this_cpu_sub(*sk->sk_prot->per_cpu_fw_alloc, local_reserve); - atomic_long_add(local_reserve, sk->sk_prot->memory_allocated); - } - preempt_enable(); + val = this_cpu_sub_return(*proto->per_cpu_fw_alloc, val); + + if (unlikely(val <= -READ_ONCE(sysctl_mem_pcpu_rsv))) + proto_memory_pcpu_drain(proto); } #define SK_ALLOC_PERCPU_COUNTER_BATCH 16
Jonathan Heathcote reported a regression caused by blamed commit on aarch64 architecture. x86 happens to have irq-safe __this_cpu_add_return() and __this_cpu_sub(), but this is not generic. I think my confusion came from "struct sock" argument, because these helpers are called with a locked socket. But the memory accounting is per-proto (and per-cpu after the blamed commit). We might cleanup these helpers later to directly accept a "struct proto *proto" argument. Switch to this_cpu_add_return() and this_cpu_xchg() operations, and get rid of preempt_disable()/preempt_enable() pairs. Fast path becomes a bit faster as a result :) Many thanks to Jonathan Heathcote for his awesome report and investigations. Fixes: 3cd3399dd7a8 ("net: implement per-cpu reserves for memory_allocated") Reported-by: Jonathan Heathcote <jonathan.heathcote@bbc.co.uk> Closes: https://lore.kernel.org/netdev/VI1PR01MB42407D7947B2EA448F1E04EFD10D2@VI1PR01MB4240.eurprd01.prod.exchangelabs.com/ Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> --- include/net/sock.h | 38 ++++++++++++++++++++------------------ 1 file changed, 20 insertions(+), 18 deletions(-)