[PATCH] dma-debug: add a per-cpu cache to avoid lock contention

[PATCH] dma-debug: add a per-cpu cache to avoid lock contention

[Eric Dumazet]

Networking drivers very often have to replace one page with
another for their RX ring buffers.

A multi-queue NIC will severly hit a contention point
in dma-debug while grabbing free_entries_lock spinlock.

Adding a one entry per-cpu cache removes the need
to grab this spinlock twice per page replacement.

Tested on a 40Gbit mlx4 NIC, with 16 RX queues and about
1,000,000 replacements per second.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Christoph Hellwig <hch@lst.de>
---
 kernel/dma/debug.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index a310dbb1515e92c081f8f3f9a7290dd5e53fc889..b7221426ef49cf640db5bcb261b0817d714a3033 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -97,6 +97,8 @@ static LIST_HEAD(free_entries);
 /* Lock for the list above */
 static DEFINE_SPINLOCK(free_entries_lock);
 
+static DEFINE_PER_CPU(struct dma_debug_entry *, dma_debug_entry_cache);
+
 /* Global disable flag - will be set in case of an error */
 static bool global_disable __read_mostly;
 
@@ -676,6 +678,10 @@ static struct dma_debug_entry *dma_entry_alloc(void)
 	struct dma_debug_entry *entry;
 	unsigned long flags;
 
+	entry = this_cpu_xchg(dma_debug_entry_cache, NULL);
+	if (entry)
+		goto end;
+
 	spin_lock_irqsave(&free_entries_lock, flags);
 	if (num_free_entries == 0) {
 		if (dma_debug_create_entries(GFP_ATOMIC)) {
@@ -690,7 +696,7 @@ static struct dma_debug_entry *dma_entry_alloc(void)
 	entry = __dma_entry_alloc();
 
 	spin_unlock_irqrestore(&free_entries_lock, flags);
-
+end:
 #ifdef CONFIG_STACKTRACE
 	entry->stack_len = stack_trace_save(entry->stack_entries,
 					    ARRAY_SIZE(entry->stack_entries),
@@ -705,6 +711,9 @@ static void dma_entry_free(struct dma_debug_entry *entry)
 
 	active_cacheline_remove(entry);
 
+	if (!this_cpu_cmpxchg(dma_debug_entry_cache, NULL, entry))
+		return;
+
 	/*
 	 * add to beginning of the list - this way the entries are
 	 * more likely cache hot when they are reallocated.
-- 
2.25.0.341.g760bfbb309-goog

Re: [PATCH] dma-debug: add a per-cpu cache to avoid lock contention

[Robin Murphy]

On 2020-01-30 8:14 pm, Eric Dumazet via iommu wrote:
> Networking drivers very often have to replace one page with
> another for their RX ring buffers.
> 
> A multi-queue NIC will severly hit a contention point
> in dma-debug while grabbing free_entries_lock spinlock.
> 
> Adding a one entry per-cpu cache removes the need
> to grab this spinlock twice per page replacement.
> 
> Tested on a 40Gbit mlx4 NIC, with 16 RX queues and about
> 1,000,000 replacements per second.

Looks like a decent benefit for minimal collateral impact - from what 
I've seen, networking appears to be a particularly easy place to run 
into awkward races, so mitigating even a little of the debug-induced 
serialisation for the common case can only be good.

Reviewed-by: Robin Murphy <robin.murphy@arm.com>

Thanks,
Robin.

> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Christoph Hellwig <hch@lst.de>
> ---
>   kernel/dma/debug.c | 11 ++++++++++-
>   1 file changed, 10 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
> index a310dbb1515e92c081f8f3f9a7290dd5e53fc889..b7221426ef49cf640db5bcb261b0817d714a3033 100644
> --- a/kernel/dma/debug.c
> +++ b/kernel/dma/debug.c
> @@ -97,6 +97,8 @@ static LIST_HEAD(free_entries);
>   /* Lock for the list above */
>   static DEFINE_SPINLOCK(free_entries_lock);
>   
> +static DEFINE_PER_CPU(struct dma_debug_entry *, dma_debug_entry_cache);
> +
>   /* Global disable flag - will be set in case of an error */
>   static bool global_disable __read_mostly;
>   
> @@ -676,6 +678,10 @@ static struct dma_debug_entry *dma_entry_alloc(void)
>   	struct dma_debug_entry *entry;
>   	unsigned long flags;
>   
> +	entry = this_cpu_xchg(dma_debug_entry_cache, NULL);
> +	if (entry)
> +		goto end;
> +
>   	spin_lock_irqsave(&free_entries_lock, flags);
>   	if (num_free_entries == 0) {
>   		if (dma_debug_create_entries(GFP_ATOMIC)) {
> @@ -690,7 +696,7 @@ static struct dma_debug_entry *dma_entry_alloc(void)
>   	entry = __dma_entry_alloc();
>   
>   	spin_unlock_irqrestore(&free_entries_lock, flags);
> -
> +end:
>   #ifdef CONFIG_STACKTRACE
>   	entry->stack_len = stack_trace_save(entry->stack_entries,
>   					    ARRAY_SIZE(entry->stack_entries),
> @@ -705,6 +711,9 @@ static void dma_entry_free(struct dma_debug_entry *entry)
>   
>   	active_cacheline_remove(entry);
>   
> +	if (!this_cpu_cmpxchg(dma_debug_entry_cache, NULL, entry))
> +		return;
> +
>   	/*
>   	 * add to beginning of the list - this way the entries are
>   	 * more likely cache hot when they are reallocated.
>