Received-SPF: pass (google.com: domain of linux-kernel+bounces-53278-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
From: Loic Poulain <loic.poulain@linaro.org>
To: rafael@kernel.org,
	len.brown@intel.com,
	pavel@ucw.cz
Cc: linux-pm@vger.kernel.org,
	linux-kernel@vger.kernel.org,
	Loic Poulain <loic.poulain@linaro.org>
Subject: [RESEND PATCH] PM: hibernate: Request 0 CPU latency during image decompression
Date: Mon,  5 Feb 2024 20:01:57 +0100
Message-Id: <20240205190157.50278-1-loic.poulain@linaro.org>
Precedence: bulk
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit

The (SMP) load_image_lzo procedure relies on multiple threads and
their synchronization for decompressing the hibernate image. for
performance reasons, it is important to keep the CPUs responsive
so that the synchronization overhead stays minimal. This overhead
roughly corresponds to thread wakeup latency, which is linked to
the CPU idle exit time.

By requesting 0 CPU latency, we prevent CPUs to enter their deepest
idle states, ensuring that they will be as responsive as possible
during the whole decompression procedure, and by extension improve
the hibernate resume time (CPU dependent).

On iMX8M mini SoC, that gives a ~40% boost for the decompression time:
PM: hibernation: Read 365640 kbytes in 1.88 seconds (194.48 MB/s)
vs
PM: hibernation: Read 363476 kbytes in 1.26 seconds (288.47 MB/s)

Note: 40% diff sounds huge, and may point to iMX specific cpuidle
issue. Anyhow, we know that during this bottleneck procedure, the CPUs
will be fully dedicated to the decompress task, and must be kept ready
for it.

Signed-off-by: Loic Poulain <loic.poulain@linaro.org>
---
 kernel/power/swap.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/kernel/power/swap.c b/kernel/power/swap.c
index f6ebcd00c410..d2cd53497c4e 100644
--- a/kernel/power/swap.c
+++ b/kernel/power/swap.c
@@ -22,6 +22,7 @@
 #include <linux/swap.h>
 #include <linux/swapops.h>
 #include <linux/pm.h>
+#include <linux/pm_qos.h>
 #include <linux/slab.h>
 #include <linux/lzo.h>
 #include <linux/vmalloc.h>
@@ -1180,6 +1181,7 @@ static int load_image_lzo(struct swap_map_handle *handle,
 	unsigned char **page = NULL;
 	struct dec_data *data = NULL;
 	struct crc_data *crc = NULL;
+	struct pm_qos_request qos;
 
 	hib_init_batch(&hb);
 
@@ -1190,6 +1192,8 @@ static int load_image_lzo(struct swap_map_handle *handle,
 	nr_threads = num_online_cpus() - 1;
 	nr_threads = clamp_val(nr_threads, 1, LZO_THREADS);
 
+	cpu_latency_qos_add_request(&qos, 0);
+
 	page = vmalloc(array_size(LZO_MAX_RD_PAGES, sizeof(*page)));
 	if (!page) {
 		pr_err("Failed to allocate LZO page\n");
@@ -1470,6 +1474,8 @@ static int load_image_lzo(struct swap_map_handle *handle,
 	}
 	vfree(page);
 
+	cpu_latency_qos_remove_request(&qos);
+
 	return ret;
 }
 
-- 
2.34.1