RE: [PATCH] svcrdma: Avoid direct reclaim when allocating Read sink buffers

> -----Original Message-----
> From: Chuck Lever <[email protected]>
> Sent: Friday, June 5, 2026 4:31 PM
> To: Mike Snitzer <[email protected]>
> Cc: [email protected]; [email protected]; Chuck Lever
> <[email protected]>; Jonathan Flynn
> <jonathan.flynn-F/[email protected]>
> Subject: [PATCH] svcrdma: Avoid direct reclaim when allocating Read sink
> buffers
>
> From: Chuck Lever <[email protected]>
>
> svc_rdma_alloc_read_pages() passes __GFP_NORETRY, which limits the
> allocator to a single round of direct reclaim and asynchronous
compaction per
> attempt. Under memory pressure or fragmentation that round can take a
long
> time, and the fallback loop repeats it at each order, multiplying the
stall while
> the RPC waits for its Read sink buffer.
>
> The contiguous allocation is opportunistic: when it fails, Read sink
buffers
> come from the pages already in rq_pages[]. Direct reclaim effort buys
little
> here. Allocate with GFP_NOWAIT instead, which omits
> __GFP_DIRECT_RECLAIM so the allocator takes pages only from the free
lists
> and returns NULL immediately when none are available. GFP_NOWAIT retains
> __GFP_KSWAPD_RECLAIM, so a failed attempt still wakes kswapd to
replenish
> higher-order pages in the background, and it already includes
> __GFP_NOWARN. __GFP_NORETRY has no effect once direct reclaim is off.
> skb_page_frag_refill() takes the same approach for its opportunistic
high-
> order allocation.
>
> Reported-by: Jonathan Flynn <jonathan.flynn-F/[email protected]>
> Fixes: 18755b8c2f24 ("svcrdma: Use contiguous pages for RDMA Read sink
> buffers")
> Signed-off-by: Chuck Lever <[email protected]>
> ---
>  net/sunrpc/xprtrdma/svc_rdma_rw.c | 18 +++++++++---------
>  1 file changed, 9 insertions(+), 9 deletions(-)
>
>
> Given the perf symbol resolution inaccuracies I can't swear this will
fix the
> issue, but here's a stab at it.
>
>
> diff --git a/net/sunrpc/xprtrdma/svc_rdma_rw.c
> b/net/sunrpc/xprtrdma/svc_rdma_rw.c
> index 587e4cd29303..efde26cac961 100644
> --- a/net/sunrpc/xprtrdma/svc_rdma_rw.c
> +++ b/net/sunrpc/xprtrdma/svc_rdma_rw.c
> @@ -746,10 +746,9 @@ int svc_rdma_prepare_reply_chunk(struct
> svcxprt_rdma *rdma,  }
>
>  /*
> - * Cap contiguous RDMA Read sink allocations at order-4.
> - * Higher orders risk allocation failure under
> - * __GFP_NORETRY, which would negate the benefit of the
> - * contiguous fast path.
> + * Cap contiguous RDMA Read sink allocations at order-4. Higher orders
> + risk
> + * allocation failure under GFP_NOWAIT, which would negate the benefit
> + of
> + * the contiguous fast path.
>   */
>  #define SVC_RDMA_CONTIG_MAX_ORDER	4
>
> @@ -758,9 +757,11 @@ int svc_rdma_prepare_reply_chunk(struct
> svcxprt_rdma *rdma,
>   * @nr_pages: number of pages needed
>   * @order: on success, set to the allocation order
>   *
> - * Attempts a higher-order allocation, falling back to smaller orders.
> - * The returned pages are split immediately so each sub-page has its
> - * own refcount and can be freed independently.
> + * Attempts a higher-order allocation, falling back to smaller orders.
> + The
> + * allocation is opportunistic: it takes pages only from the free
> + lists,
> + * without direct reclaim, so it fails fast under memory pressure. The
> + * returned pages are split immediately so each sub-page has its own
> + * refcount and can be freed independently.
>   *
>   * Returns a pointer to the first page on success, or NULL if even
>   * order-1 allocation fails.
> @@ -775,8 +776,7 @@ svc_rdma_alloc_read_pages(unsigned int nr_pages,
> unsigned int *order)
>  		SVC_RDMA_CONTIG_MAX_ORDER);
>
>  	while (o >= 1) {
> -		page = alloc_pages(GFP_KERNEL | __GFP_NORETRY |
> __GFP_NOWARN,
> -				   o);
> +		page = alloc_pages(GFP_NOWAIT, o);
>  		if (page) {
>  			split_page(page, o);
>  			*order = o;
> --
> 2.54.0
Unfortunately, the GFP_NOWAIT change did not materially affect either
throughput or the perf profile. The allocator-heavy stack rooted at
svc_rdma_build_read_segment_contig() remains dominant, with
alloc_pages_noprof() and rmqueue_buddy() continuing to account for a
significant portion of the samples, similar to the original regressed
build.

I have added a gfp-nowait directory to the OneDrive link referenced in my
previous email. It contains the fio results, perf reports, and a
flamegraph for the GFP_NOWAIT test.

I have also added a flamegraph to:

rpcrdma-regression/regressed/phase2/server

for the original regressed configuration.

-Jon