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

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <87bde968-0103-c90d-380d-e0793cd39aa5@arm.com>
Date:   Fri, 31 Jan 2020 13:14:36 +0000
From:   Robin Murphy <robin.murphy@....com>
To:     Eric Dumazet <edumazet@...gle.com>, Christoph Hellwig <hch@....de>,
        Joerg Roedel <jroedel@...e.de>
Cc:     iommu@...ts.linux-foundation.org,
        Eric Dumazet <eric.dumazet@...il.com>,
        linux-kernel <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] dma-debug: add a per-cpu cache to avoid lock contention

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@....com>

Thanks,
Robin.

> Signed-off-by: Eric Dumazet <edumazet@...gle.com>
> Cc: Christoph Hellwig <hch@....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.
>