diff mbox series

[v2,9/9] tcp: Use per-vma locking for receive zerocopy

Message ID 20230711202047.3818697-10-willy@infradead.org (mailing list archive)
State New
Headers show
Series Avoid the mmap lock for fault-around | expand

Commit Message

Matthew Wilcox July 11, 2023, 8:20 p.m. UTC
From: Arjun Roy <arjunroy@google.com>

Per-VMA locking allows us to lock a struct vm_area_struct without
taking the process-wide mmap lock in read mode.

Consider a process workload where the mmap lock is taken constantly in
write mode. In this scenario, all zerocopy receives are periodically
blocked during that period of time - though in principle, the memory
ranges being used by TCP are not touched by the operations that need
the mmap write lock. This results in performance degradation.

Now consider another workload where the mmap lock is never taken in
write mode, but there are many TCP connections using receive zerocopy
that are concurrently receiving. These connections all take the mmap
lock in read mode, but this does induce a lot of contention and atomic
ops for this process-wide lock. This results in additional CPU
overhead caused by contending on the cache line for this lock.

However, with per-vma locking, both of these problems can be avoided.

As a test, I ran an RPC-style request/response workload with 4KB
payloads and receive zerocopy enabled, with 100 simultaneous TCP
connections. I measured perf cycles within the
find_tcp_vma/mmap_read_lock/mmap_read_unlock codepath, with and
without per-vma locking enabled.

When using process-wide mmap semaphore read locking, about 1% of
measured perf cycles were within this path. With per-VMA locking, this
value dropped to about 0.45%.

Signed-off-by: Arjun Roy <arjunroy@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
---
 net/ipv4/tcp.c | 39 ++++++++++++++++++++++++++++++++-------
 1 file changed, 32 insertions(+), 7 deletions(-)

Comments

Suren Baghdasaryan July 14, 2023, 3:40 a.m. UTC | #1
On Tue, Jul 11, 2023 at 1:21 PM Matthew Wilcox (Oracle)
<willy@infradead.org> wrote:
>
> From: Arjun Roy <arjunroy@google.com>
>
> Per-VMA locking allows us to lock a struct vm_area_struct without
> taking the process-wide mmap lock in read mode.
>
> Consider a process workload where the mmap lock is taken constantly in
> write mode. In this scenario, all zerocopy receives are periodically
> blocked during that period of time - though in principle, the memory
> ranges being used by TCP are not touched by the operations that need
> the mmap write lock. This results in performance degradation.
>
> Now consider another workload where the mmap lock is never taken in
> write mode, but there are many TCP connections using receive zerocopy
> that are concurrently receiving. These connections all take the mmap
> lock in read mode, but this does induce a lot of contention and atomic
> ops for this process-wide lock. This results in additional CPU
> overhead caused by contending on the cache line for this lock.
>
> However, with per-vma locking, both of these problems can be avoided.
>
> As a test, I ran an RPC-style request/response workload with 4KB
> payloads and receive zerocopy enabled, with 100 simultaneous TCP
> connections. I measured perf cycles within the
> find_tcp_vma/mmap_read_lock/mmap_read_unlock codepath, with and
> without per-vma locking enabled.
>
> When using process-wide mmap semaphore read locking, about 1% of
> measured perf cycles were within this path. With per-VMA locking, this
> value dropped to about 0.45%.
>
> Signed-off-by: Arjun Roy <arjunroy@google.com>
> Reviewed-by: Eric Dumazet <edumazet@google.com>
> Signed-off-by: David S. Miller <davem@davemloft.net>
> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>

Seems to match the original version with less churn.

Reviewed-by: Suren Baghdasaryan <surenb@google.com>

> ---
>  net/ipv4/tcp.c | 39 ++++++++++++++++++++++++++++++++-------
>  1 file changed, 32 insertions(+), 7 deletions(-)
>
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index 1542de3f66f7..7118ec6cf886 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -2038,6 +2038,30 @@ static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
>         }
>  }
>

nit: Maybe add a comment that mmap_locked value is
undefined/meaningless if the function returns NULL?

> +static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
> +               unsigned long address, bool *mmap_locked)
> +{
> +       struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
> +
> +       if (vma) {
> +               if (vma->vm_ops != &tcp_vm_ops) {
> +                       vma_end_read(vma);
> +                       return NULL;
> +               }
> +               *mmap_locked = false;
> +               return vma;
> +       }
> +
> +       mmap_read_lock(mm);
> +       vma = vma_lookup(mm, address);
> +       if (!vma || vma->vm_ops != &tcp_vm_ops) {
> +               mmap_read_unlock(mm);
> +               return NULL;
> +       }
> +       *mmap_locked = true;
> +       return vma;
> +}
> +
>  #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
>  static int tcp_zerocopy_receive(struct sock *sk,
>                                 struct tcp_zerocopy_receive *zc,
> @@ -2055,6 +2079,7 @@ static int tcp_zerocopy_receive(struct sock *sk,
>         u32 seq = tp->copied_seq;
>         u32 total_bytes_to_map;
>         int inq = tcp_inq(sk);
> +       bool mmap_locked;
>         int ret;
>
>         zc->copybuf_len = 0;
> @@ -2079,13 +2104,10 @@ static int tcp_zerocopy_receive(struct sock *sk,
>                 return 0;
>         }
>
> -       mmap_read_lock(current->mm);
> -
> -       vma = vma_lookup(current->mm, address);
> -       if (!vma || vma->vm_ops != &tcp_vm_ops) {
> -               mmap_read_unlock(current->mm);
> +       vma = find_tcp_vma(current->mm, address, &mmap_locked);
> +       if (!vma)
>                 return -EINVAL;
> -       }
> +
>         vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
>         avail_len = min_t(u32, vma_len, inq);
>         total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
> @@ -2159,7 +2181,10 @@ static int tcp_zerocopy_receive(struct sock *sk,
>                                                    zc, total_bytes_to_map);
>         }
>  out:
> -       mmap_read_unlock(current->mm);
> +       if (mmap_locked)
> +               mmap_read_unlock(current->mm);
> +       else
> +               vma_end_read(vma);
>         /* Try to copy straggler data. */
>         if (!ret)
>                 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
> --
> 2.39.2
>
Matthew Wilcox July 21, 2023, 6:48 p.m. UTC | #2
Eric?  Arjun?  Any comments?

On Tue, Jul 11, 2023 at 09:20:47PM +0100, Matthew Wilcox (Oracle) wrote:
> From: Arjun Roy <arjunroy@google.com>
> 
> Per-VMA locking allows us to lock a struct vm_area_struct without
> taking the process-wide mmap lock in read mode.
> 
> Consider a process workload where the mmap lock is taken constantly in
> write mode. In this scenario, all zerocopy receives are periodically
> blocked during that period of time - though in principle, the memory
> ranges being used by TCP are not touched by the operations that need
> the mmap write lock. This results in performance degradation.
> 
> Now consider another workload where the mmap lock is never taken in
> write mode, but there are many TCP connections using receive zerocopy
> that are concurrently receiving. These connections all take the mmap
> lock in read mode, but this does induce a lot of contention and atomic
> ops for this process-wide lock. This results in additional CPU
> overhead caused by contending on the cache line for this lock.
> 
> However, with per-vma locking, both of these problems can be avoided.
> 
> As a test, I ran an RPC-style request/response workload with 4KB
> payloads and receive zerocopy enabled, with 100 simultaneous TCP
> connections. I measured perf cycles within the
> find_tcp_vma/mmap_read_lock/mmap_read_unlock codepath, with and
> without per-vma locking enabled.
> 
> When using process-wide mmap semaphore read locking, about 1% of
> measured perf cycles were within this path. With per-VMA locking, this
> value dropped to about 0.45%.
> 
> Signed-off-by: Arjun Roy <arjunroy@google.com>
> Reviewed-by: Eric Dumazet <edumazet@google.com>
> Signed-off-by: David S. Miller <davem@davemloft.net>
> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
> ---
>  net/ipv4/tcp.c | 39 ++++++++++++++++++++++++++++++++-------
>  1 file changed, 32 insertions(+), 7 deletions(-)
> 
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index 1542de3f66f7..7118ec6cf886 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -2038,6 +2038,30 @@ static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
>  	}
>  }
>  
> +static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
> +		unsigned long address, bool *mmap_locked)
> +{
> +	struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
> +
> +	if (vma) {
> +		if (vma->vm_ops != &tcp_vm_ops) {
> +			vma_end_read(vma);
> +			return NULL;
> +		}
> +		*mmap_locked = false;
> +		return vma;
> +	}
> +
> +	mmap_read_lock(mm);
> +	vma = vma_lookup(mm, address);
> +	if (!vma || vma->vm_ops != &tcp_vm_ops) {
> +		mmap_read_unlock(mm);
> +		return NULL;
> +	}
> +	*mmap_locked = true;
> +	return vma;
> +}
> +
>  #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
>  static int tcp_zerocopy_receive(struct sock *sk,
>  				struct tcp_zerocopy_receive *zc,
> @@ -2055,6 +2079,7 @@ static int tcp_zerocopy_receive(struct sock *sk,
>  	u32 seq = tp->copied_seq;
>  	u32 total_bytes_to_map;
>  	int inq = tcp_inq(sk);
> +	bool mmap_locked;
>  	int ret;
>  
>  	zc->copybuf_len = 0;
> @@ -2079,13 +2104,10 @@ static int tcp_zerocopy_receive(struct sock *sk,
>  		return 0;
>  	}
>  
> -	mmap_read_lock(current->mm);
> -
> -	vma = vma_lookup(current->mm, address);
> -	if (!vma || vma->vm_ops != &tcp_vm_ops) {
> -		mmap_read_unlock(current->mm);
> +	vma = find_tcp_vma(current->mm, address, &mmap_locked);
> +	if (!vma)
>  		return -EINVAL;
> -	}
> +
>  	vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
>  	avail_len = min_t(u32, vma_len, inq);
>  	total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
> @@ -2159,7 +2181,10 @@ static int tcp_zerocopy_receive(struct sock *sk,
>  						   zc, total_bytes_to_map);
>  	}
>  out:
> -	mmap_read_unlock(current->mm);
> +	if (mmap_locked)
> +		mmap_read_unlock(current->mm);
> +	else
> +		vma_end_read(vma);
>  	/* Try to copy straggler data. */
>  	if (!ret)
>  		copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
> -- 
> 2.39.2
>
diff mbox series

Patch

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 1542de3f66f7..7118ec6cf886 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -2038,6 +2038,30 @@  static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
 	}
 }
 
+static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
+		unsigned long address, bool *mmap_locked)
+{
+	struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
+
+	if (vma) {
+		if (vma->vm_ops != &tcp_vm_ops) {
+			vma_end_read(vma);
+			return NULL;
+		}
+		*mmap_locked = false;
+		return vma;
+	}
+
+	mmap_read_lock(mm);
+	vma = vma_lookup(mm, address);
+	if (!vma || vma->vm_ops != &tcp_vm_ops) {
+		mmap_read_unlock(mm);
+		return NULL;
+	}
+	*mmap_locked = true;
+	return vma;
+}
+
 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
 static int tcp_zerocopy_receive(struct sock *sk,
 				struct tcp_zerocopy_receive *zc,
@@ -2055,6 +2079,7 @@  static int tcp_zerocopy_receive(struct sock *sk,
 	u32 seq = tp->copied_seq;
 	u32 total_bytes_to_map;
 	int inq = tcp_inq(sk);
+	bool mmap_locked;
 	int ret;
 
 	zc->copybuf_len = 0;
@@ -2079,13 +2104,10 @@  static int tcp_zerocopy_receive(struct sock *sk,
 		return 0;
 	}
 
-	mmap_read_lock(current->mm);
-
-	vma = vma_lookup(current->mm, address);
-	if (!vma || vma->vm_ops != &tcp_vm_ops) {
-		mmap_read_unlock(current->mm);
+	vma = find_tcp_vma(current->mm, address, &mmap_locked);
+	if (!vma)
 		return -EINVAL;
-	}
+
 	vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
 	avail_len = min_t(u32, vma_len, inq);
 	total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
@@ -2159,7 +2181,10 @@  static int tcp_zerocopy_receive(struct sock *sk,
 						   zc, total_bytes_to_map);
 	}
 out:
-	mmap_read_unlock(current->mm);
+	if (mmap_locked)
+		mmap_read_unlock(current->mm);
+	else
+		vma_end_read(vma);
 	/* Try to copy straggler data. */
 	if (!ret)
 		copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);