Merge tag 'blk-dim-v2' into rdma.git for-next

Generic DIM From: Tal Gilboa and Yamin Fridman Implement net DIM over a generic DIM library, add RDMA DIM dim.h lib exposes an implementation of the DIM algorithm for dynamically-tuned interrupt moderation for networking interfaces. We want a similar functionality for other protocols, which might need to optimize interrupts differently. Main motivation here is DIM for NVMf storage protocol. Current DIM implementation prioritizes reducing interrupt overhead over latency. Also, in order to reduce DIM's own overhead, the algorithm might take some time to identify it needs to change profiles. While this is acceptable for networking, it might not work well on other scenarios. Here we propose a new structure to DIM. The idea is to allow a slightly modified functionality without the risk of breaking Net DIM behavior for netdev. We verified there are no degradations in current DIM behavior with the modified solution. Suggested solution: - Common logic is implemented in lib/dim/dim.c - Net DIM (existing) logic is implemented in lib/dim/net_dim.c, which uses the common logic in dim.c - Any new DIM logic will be implemented in "lib/dim/new_dim.c". This new implementation will expose modified versions of profiles, dim_step() and dim_decision(). - DIM API is declared in include/linux/dim.h for all implementations. Pros for this solution are: - Zero impact on existing net_dim implementation and usage - Relatively more code reuse (compared to two separate solutions) - Increased extensibility Required for dependencies in the next series.

Jason Gunthorpe committed 7 years ago

Commit 2ef38e32 parents 89705e9 398c2b023 files · +728 −489Browse files

Unified · Split

23 files

MAINTAINERS +2 −1
drivers/net/ethernet/broadcom/Kconfig +1 −0
drivers/net/ethernet/broadcom/bcmsysport.c +10 −10
drivers/net/ethernet/broadcom/bcmsysport.h +2 −2
drivers/net/ethernet/broadcom/bnxt/bnxt.c +6 −6
drivers/net/ethernet/broadcom/bnxt/bnxt.h +2 −2
drivers/net/ethernet/broadcom/bnxt/bnxt_debugfs.c +3 −3
drivers/net/ethernet/broadcom/bnxt/bnxt_dim.c +4 −5
drivers/net/ethernet/broadcom/genet/bcmgenet.c +9 −9
drivers/net/ethernet/broadcom/genet/bcmgenet.h +2 −2
drivers/net/ethernet/mellanox/mlx5/core/Kconfig +1 −0
drivers/net/ethernet/mellanox/mlx5/core/en.h +5 −5
drivers/net/ethernet/mellanox/mlx5/core/en_dim.c +7 −7
drivers/net/ethernet/mellanox/mlx5/core/en_ethtool.c +2 −2
drivers/net/ethernet/mellanox/mlx5/core/en_main.c +11 −11
drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c +4 −6
include/linux/dim.h +366 −0
include/linux/net_dim.h +0 −418
lib/Kconfig +8 −0
lib/Makefile +1 −0
lib/dim/Makefile +9 −0
lib/dim/dim.c +83 −0
lib/dim/net_dim.c +190 −0

MAINTAINERS+2−1

@@ -5600,7 +5600,8 @@ F: include/linux/dynamic_debug.h · 5599 unmodified lines
5600	5600	DYNAMIC INTERRUPT MODERATION
5601	5601	M: Tal Gilboa <talgi@mellanox.com>
5602	5602	S: Maintained
5603		−F: include/linux/net_dim.h
	5603	+F: include/linux/dim.h
	5604	+F: lib/dim/
5604	5605
5605	5606	DZ DECSTATION DZ11 SERIAL DRIVER
5606	5607	M: "Maciej W. Rozycki" <macro@linux-mips.org>

drivers/net/ethernet/broadcom/Kconfig+1−0

@@ -8,6 +8,7 @@ config NET_VENDOR_BROADCOM · 7 unmodified lines
8	8	default y
9	9	depends on (SSB_POSSIBLE && HAS_DMA) \|\| PCI \|\| BCM63XX \|\| \
10	10	SIBYTE_SB1xxx_SOC
	11	+ select DIMLIB
11	12	---help---
12	13	If you have a network (Ethernet) chipset belonging to this class,
13	14	say Y.

drivers/net/ethernet/broadcom/bcmsysport.c+10−10

@@ -609,7 +609,7 @@ static int bcm_sysport_set_coalesce(struct net_device *dev, · 608 unmodified lines
609	609	struct ethtool_coalesce *ec)
610	610	{
611	611	struct bcm_sysport_priv *priv = netdev_priv(dev);
612		− struct net_dim_cq_moder moder;
	612	+ struct dim_cq_moder moder;
613	613	u32 usecs, pkts;
614	614	unsigned int i;
615	615
@@ -992,7 +992,7 @@ static int bcm_sysport_poll(struct napi_struct *napi, int budget) · 376 unmodified lines
992	992	{
993	993	struct bcm_sysport_priv *priv =
994	994	container_of(napi, struct bcm_sysport_priv, napi);
995		− struct net_dim_sample dim_sample;
	995	+ struct dim_sample dim_sample;
996	996	unsigned int work_done = 0;
997	997
998	998	work_done = bcm_sysport_desc_rx(priv, budget);
@@ -1016,8 +1016,8 @@ static int bcm_sysport_poll(struct napi_struct *napi, int budget) · 17 unmodified lines
1016	1016	}
1017	1017
1018	1018	if (priv->dim.use_dim) {
1019		− net_dim_sample(priv->dim.event_ctr, priv->dim.packets,
1020		− priv->dim.bytes, &dim_sample);
	1019	+ dim_update_sample(priv->dim.event_ctr, priv->dim.packets,
	1020	+ priv->dim.bytes, &dim_sample);
1021	1021	net_dim(&priv->dim.dim, dim_sample);
1022	1022	}
1023	1023
@@ -1087,16 +1087,16 @@ static void bcm_sysport_resume_from_wol(struct bcm_sysport_priv *priv) · 63 unmodified lines
1087	1087
1088	1088	static void bcm_sysport_dim_work(struct work_struct *work)
1089	1089	{
1090		− struct net_dim *dim = container_of(work, struct net_dim, work);
	1090	+ struct dim *dim = container_of(work, struct dim, work);
1091	1091	struct bcm_sysport_net_dim *ndim =
1092	1092	container_of(dim, struct bcm_sysport_net_dim, dim);
1093	1093	struct bcm_sysport_priv *priv =
1094	1094	container_of(ndim, struct bcm_sysport_priv, dim);
1095		− struct net_dim_cq_moder cur_profile =
1096		− net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
	1095	+ struct dim_cq_moder cur_profile = net_dim_get_rx_moderation(dim->mode,
	1096	+ dim->profile_ix);
1097	1097
1098	1098	bcm_sysport_set_rx_coalesce(priv, cur_profile.usec, cur_profile.pkts);
1099		− dim->state = NET_DIM_START_MEASURE;
	1099	+ dim->state = DIM_START_MEASURE;
1100	1100	}
1101	1101
1102	1102	/* RX and misc interrupt routine */
@@ -1437,7 +1437,7 @@ static void bcm_sysport_init_dim(struct bcm_sysport_priv *priv, · 334 unmodified lines
1437	1437	struct bcm_sysport_net_dim *dim = &priv->dim;
1438	1438
1439	1439	INIT_WORK(&dim->dim.work, cb);
1440		− dim->dim.mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
	1440	+ dim->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
1441	1441	dim->event_ctr = 0;
1442	1442	dim->packets = 0;
1443	1443	dim->bytes = 0;
@@ -1446,7 +1446,7 @@ static void bcm_sysport_init_dim(struct bcm_sysport_priv *priv, · 2 unmodified lines
1446	1446	static void bcm_sysport_init_rx_coalesce(struct bcm_sysport_priv *priv)
1447	1447	{
1448	1448	struct bcm_sysport_net_dim *dim = &priv->dim;
1449		− struct net_dim_cq_moder moder;
	1449	+ struct dim_cq_moder moder;
1450	1450	u32 usecs, pkts;
1451	1451
1452	1452	usecs = priv->rx_coalesce_usecs;

drivers/net/ethernet/broadcom/bcmsysport.h+2−2

@@ -11,7 +11,7 @@ · 10 unmodified lines
11	11	#include <linux/bitmap.h>
12	12	#include <linux/ethtool.h>
13	13	#include <linux/if_vlan.h>
14		−#include <linux/net_dim.h>
	14	+#include <linux/dim.h>
15	15
16	16	/* Receive/transmit descriptor format */
17	17	#define DESC_ADDR_HI_STATUS_LEN 0x00
@@ -702,7 +702,7 @@ struct bcm_sysport_net_dim { · 684 unmodified lines
702	702	u16 event_ctr;
703	703	unsigned long packets;
704	704	unsigned long bytes;
705		− struct net_dim dim;
	705	+ struct dim dim;
706	706	};
707	707
708	708	/* Software view of the TX ring */

drivers/net/ethernet/broadcom/bnxt/bnxt.c+6−6

@@ -2130,12 +2130,12 @@ static int bnxt_poll(struct napi_struct *napi, int budget) · 2129 unmodified lines
2130	2130	}
2131	2131	}
2132	2132	if (bp->flags & BNXT_FLAG_DIM) {
2133		− struct net_dim_sample dim_sample;
	2133	+ struct dim_sample dim_sample;
2134	2134
2135		− net_dim_sample(cpr->event_ctr,
2136		− cpr->rx_packets,
2137		− cpr->rx_bytes,
2138		− &dim_sample);
	2135	+ dim_update_sample(cpr->event_ctr,
	2136	+ cpr->rx_packets,
	2137	+ cpr->rx_bytes,
	2138	+ &dim_sample);
2139	2139	net_dim(&cpr->dim, dim_sample);
2140	2140	}
2141	2141	return work_done;
@@ -7813,7 +7813,7 @@ static void bnxt_enable_napi(struct bnxt *bp) · 5671 unmodified lines
7813	7813
7814	7814	if (bp->bnapi[i]->rx_ring) {
7815	7815	INIT_WORK(&cpr->dim.work, bnxt_dim_work);
7816		− cpr->dim.mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
	7816	+ cpr->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
7817	7817	}
7818	7818	napi_enable(&bp->bnapi[i]->napi);
7819	7819	}

drivers/net/ethernet/broadcom/bnxt/bnxt.h+2−2

@@ -24,7 +24,7 @@ · 23 unmodified lines
24	24	#include <net/devlink.h>
25	25	#include <net/dst_metadata.h>
26	26	#include <net/xdp.h>
27		−#include <linux/net_dim.h>
	27	+#include <linux/dim.h>
28	28
29	29	struct tx_bd {
30	30	__le32 tx_bd_len_flags_type;
@@ -810,7 +810,7 @@ struct bnxt_cp_ring_info { · 779 unmodified lines
810	810	u64 rx_bytes;
811	811	u64 event_ctr;
812	812
813		− struct net_dim dim;
	813	+ struct dim dim;
814	814
815	815	union {
816	816	struct tx_cmp *cp_desc_ring[MAX_CP_PAGES];

drivers/net/ethernet/broadcom/bnxt/bnxt_debugfs.c+3−3

@@ -11,7 +11,7 @@ · 10 unmodified lines
11	11	#include <linux/module.h>
12	12	#include <linux/pci.h>
13	13	#include "bnxt_hsi.h"
14		−#include <linux/net_dim.h>
	14	+#include <linux/dim.h>
15	15	#include "bnxt.h"
16	16	#include "bnxt_debugfs.h"
17	17
@@ -21,7 +21,7 @@ static ssize_t debugfs_dim_read(struct file *filep, · 3 unmodified lines
21	21	char __user *buffer,
22	22	size_t count, loff_t *ppos)
23	23	{
24		− struct net_dim *dim = filep->private_data;
	24	+ struct dim *dim = filep->private_data;
25	25	int len;
26	26	char *buf;
27	27
@@ -61,7 +61,7 @@ static const struct file_operations debugfs_dim_fops = { · 33 unmodified lines
61	61	.read = debugfs_dim_read,
62	62	};
63	63
64		−static struct dentry debugfs_dim_ring_init(struct net_dim dim, int ring_idx,
	64	+static struct dentry debugfs_dim_ring_init(struct dim dim, int ring_idx,
65	65	struct dentry *dd)
66	66	{
67	67	static char qname[16];

drivers/net/ethernet/broadcom/bnxt/bnxt_dim.c+4−5

@@ -7,26 +7,25 @@ · 6 unmodified lines
7	7	* the Free Software Foundation.
8	8	*/
9	9
10		−#include <linux/net_dim.h>
	10	+#include <linux/dim.h>
11	11	#include "bnxt_hsi.h"
12	12	#include "bnxt.h"
13	13
14	14	void bnxt_dim_work(struct work_struct *work)
15	15	{
16		− struct net_dim *dim = container_of(work, struct net_dim,
17		− work);
	16	+ struct dim *dim = container_of(work, struct dim, work);
18	17	struct bnxt_cp_ring_info *cpr = container_of(dim,
19	18	struct bnxt_cp_ring_info,
20	19	dim);
21	20	struct bnxt_napi *bnapi = container_of(cpr,
22	21	struct bnxt_napi,
23	22	cp_ring);
24		− struct net_dim_cq_moder cur_moder =
	23	+ struct dim_cq_moder cur_moder =
25	24	net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
26	25
27	26	cpr->rx_ring_coal.coal_ticks = cur_moder.usec;
28	27	cpr->rx_ring_coal.coal_bufs = cur_moder.pkts;
29	28
30	29	bnxt_hwrm_set_ring_coal(bnapi->bp, bnapi);
31		− dim->state = NET_DIM_START_MEASURE;
	30	+ dim->state = DIM_START_MEASURE;
32	31	}

drivers/net/ethernet/broadcom/genet/bcmgenet.c+9−9

@@ -640,7 +640,7 @@ static void bcmgenet_set_rx_coalesce(struct bcmgenet_rx_ring *ring, · 639 unmodified lines
640	640	static void bcmgenet_set_ring_rx_coalesce(struct bcmgenet_rx_ring *ring,
641	641	struct ethtool_coalesce *ec)
642	642	{
643		− struct net_dim_cq_moder moder;
	643	+ struct dim_cq_moder moder;
644	644	u32 usecs, pkts;
645	645
646	646	ring->rx_coalesce_usecs = ec->rx_coalesce_usecs;
@@ -1895,7 +1895,7 @@ static int bcmgenet_rx_poll(struct napi_struct *napi, int budget) · 1248 unmodified lines
1895	1895	{
1896	1896	struct bcmgenet_rx_ring *ring = container_of(napi,
1897	1897	struct bcmgenet_rx_ring, napi);
1898		− struct net_dim_sample dim_sample;
	1898	+ struct dim_sample dim_sample;
1899	1899	unsigned int work_done;
1900	1900
1901	1901	work_done = bcmgenet_desc_rx(ring, budget);
@@ -1906,8 +1906,8 @@ static int bcmgenet_rx_poll(struct napi_struct *napi, int budget) · 4 unmodified lines
1906	1906	}
1907	1907
1908	1908	if (ring->dim.use_dim) {
1909		− net_dim_sample(ring->dim.event_ctr, ring->dim.packets,
1910		− ring->dim.bytes, &dim_sample);
	1909	+ dim_update_sample(ring->dim.event_ctr, ring->dim.packets,
	1910	+ ring->dim.bytes, &dim_sample);
1911	1911	net_dim(&ring->dim.dim, dim_sample);
1912	1912	}
1913	1913
@@ -1916,16 +1916,16 @@ static int bcmgenet_rx_poll(struct napi_struct *napi, int budget) · 2 unmodified lines
1916	1916
1917	1917	static void bcmgenet_dim_work(struct work_struct *work)
1918	1918	{
1919		− struct net_dim *dim = container_of(work, struct net_dim, work);
	1919	+ struct dim *dim = container_of(work, struct dim, work);
1920	1920	struct bcmgenet_net_dim *ndim =
1921	1921	container_of(dim, struct bcmgenet_net_dim, dim);
1922	1922	struct bcmgenet_rx_ring *ring =
1923	1923	container_of(ndim, struct bcmgenet_rx_ring, dim);
1924		− struct net_dim_cq_moder cur_profile =
	1924	+ struct dim_cq_moder cur_profile =
1925	1925	net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
1926	1926
1927	1927	bcmgenet_set_rx_coalesce(ring, cur_profile.usec, cur_profile.pkts);
1928		− dim->state = NET_DIM_START_MEASURE;
	1928	+ dim->state = DIM_START_MEASURE;
1929	1929	}
1930	1930
1931	1931	/* Assign skb to RX DMA descriptor. */
@@ -2082,7 +2082,7 @@ static void bcmgenet_init_dim(struct bcmgenet_rx_ring *ring, · 150 unmodified lines
2082	2082	struct bcmgenet_net_dim *dim = &ring->dim;
2083	2083
2084	2084	INIT_WORK(&dim->dim.work, cb);
2085		− dim->dim.mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
	2085	+ dim->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
2086	2086	dim->event_ctr = 0;
2087	2087	dim->packets = 0;
2088	2088	dim->bytes = 0;
@@ -2091,7 +2091,7 @@ static void bcmgenet_init_dim(struct bcmgenet_rx_ring *ring, · 2 unmodified lines
2091	2091	static void bcmgenet_init_rx_coalesce(struct bcmgenet_rx_ring *ring)
2092	2092	{
2093	2093	struct bcmgenet_net_dim *dim = &ring->dim;
2094		− struct net_dim_cq_moder moder;
	2094	+ struct dim_cq_moder moder;
2095	2095	u32 usecs, pkts;
2096	2096
2097	2097	usecs = ring->rx_coalesce_usecs;

drivers/net/ethernet/broadcom/genet/bcmgenet.h+2−2

@@ -13,7 +13,7 @@ · 12 unmodified lines
13	13	#include <linux/mii.h>
14	14	#include <linux/if_vlan.h>
15	15	#include <linux/phy.h>
16		−#include <linux/net_dim.h>
	16	+#include <linux/dim.h>
17	17
18	18	/* total number of Buffer Descriptors, same for Rx/Tx */
19	19	#define TOTAL_DESC 256
@@ -578,7 +578,7 @@ struct bcmgenet_net_dim { · 558 unmodified lines
578	578	u16 event_ctr;
579	579	unsigned long packets;
580	580	unsigned long bytes;
581		− struct net_dim dim;
	581	+ struct dim dim;
582	582	};
583	583
584	584	struct bcmgenet_rx_ring {

drivers/net/ethernet/mellanox/mlx5/core/Kconfig+1−0

@@ -34,6 +34,7 @@ config MLX5_CORE_EN · 33 unmodified lines
34	34	depends on NETDEVICES && ETHERNET && INET && PCI && MLX5_CORE
35	35	depends on IPV6=y \|\| IPV6=n \|\| MLX5_CORE=m
36	36	select PAGE_POOL
	37	+ select DIMLIB
37	38	default n
38	39	---help---
39	40	Ethernet support in Mellanox Technologies ConnectX-4 NIC.

drivers/net/ethernet/mellanox/mlx5/core/en.h+5−5

@@ -48,7 +48,7 @@ · 47 unmodified lines
48	48	#include <linux/rhashtable.h>
49	49	#include <net/switchdev.h>
50	50	#include <net/xdp.h>
51		−#include <linux/net_dim.h>
	51	+#include <linux/dim.h>
52	52	#include <linux/bits.h>
53	53	#include "wq.h"
54	54	#include "mlx5_core.h"
@@ -238,9 +238,9 @@ struct mlx5e_params { · 183 unmodified lines
238	238	u16 num_channels;
239	239	u8 num_tc;
240	240	bool rx_cqe_compress_def;
241		− struct net_dim_cq_moder rx_cq_moderation;
242		− struct net_dim_cq_moder tx_cq_moderation;
243	241	bool tunneled_offload_en;
	242	+ struct dim_cq_moder rx_cq_moderation;
	243	+ struct dim_cq_moder tx_cq_moderation;
244	244	bool lro_en;
245	245	u8 tx_min_inline_mode;
246	246	bool vlan_strip_disable;
@@ -356,7 +356,7 @@ struct mlx5e_txqsq { · 109 unmodified lines
356	356	/* dirtied @completion */
357	357	u16 cc;
358	358	u32 dma_fifo_cc;
359		− struct net_dim dim; /* Adaptive Moderation */
	359	+ struct dim dim; /* Adaptive Moderation */
360	360
361	361	/* dirtied @xmit */
362	362	u16 pc ____cacheline_aligned_in_smp;
@@ -596,7 +596,7 @@ struct mlx5e_rq { · 233 unmodified lines
596	596	int ix;
597	597	unsigned int hw_mtu;
598	598
599		− struct net_dim dim; /* Dynamic Interrupt Moderation */
	599	+ struct dim dim; /* Dynamic Interrupt Moderation */
600	600
601	601	/* XDP */
602	602	struct bpf_prog *xdp_prog;

drivers/net/ethernet/mellanox/mlx5/core/en_dim.c+7−7

@@ -30,22 +30,22 @@ · 29 unmodified lines
30	30	* SOFTWARE.
31	31	*/
32	32
33		−#include <linux/net_dim.h>
	33	+#include <linux/dim.h>
34	34	#include "en.h"
35	35
36	36	static void
37		−mlx5e_complete_dim_work(struct net_dim *dim, struct net_dim_cq_moder moder,
	37	+mlx5e_complete_dim_work(struct dim *dim, struct dim_cq_moder moder,
38	38	struct mlx5_core_dev mdev, struct mlx5_core_cq mcq)
39	39	{
40	40	mlx5_core_modify_cq_moderation(mdev, mcq, moder.usec, moder.pkts);
41		− dim->state = NET_DIM_START_MEASURE;
	41	+ dim->state = DIM_START_MEASURE;
42	42	}
43	43
44	44	void mlx5e_rx_dim_work(struct work_struct *work)
45	45	{
46		− struct net_dim *dim = container_of(work, struct net_dim, work);
	46	+ struct dim *dim = container_of(work, struct dim, work);
47	47	struct mlx5e_rq *rq = container_of(dim, struct mlx5e_rq, dim);
48		− struct net_dim_cq_moder cur_moder =
	48	+ struct dim_cq_moder cur_moder =
49	49	net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
50	50
51	51	mlx5e_complete_dim_work(dim, cur_moder, rq->mdev, &rq->cq.mcq);
@@ -53,9 +53,9 @@ void mlx5e_rx_dim_work(struct work_struct *work) · 1 unmodified line
53	53
54	54	void mlx5e_tx_dim_work(struct work_struct *work)
55	55	{
56		− struct net_dim *dim = container_of(work, struct net_dim, work);
	56	+ struct dim *dim = container_of(work, struct dim, work);
57	57	struct mlx5e_txqsq *sq = container_of(dim, struct mlx5e_txqsq, dim);
58		− struct net_dim_cq_moder cur_moder =
	58	+ struct dim_cq_moder cur_moder =
59	59	net_dim_get_tx_moderation(dim->mode, dim->profile_ix);
60	60
61	61	mlx5e_complete_dim_work(dim, cur_moder, sq->cq.mdev, &sq->cq.mcq);

drivers/net/ethernet/mellanox/mlx5/core/en_ethtool.c+2−2

@@ -466,7 +466,7 @@ static int mlx5e_set_channels(struct net_device *dev, · 465 unmodified lines
466	466	int mlx5e_ethtool_get_coalesce(struct mlx5e_priv *priv,
467	467	struct ethtool_coalesce *coal)
468	468	{
469		− struct net_dim_cq_moder rx_moder, tx_moder;
	469	+ struct dim_cq_moder rx_moder, tx_moder;
470	470
471	471	if (!MLX5_CAP_GEN(priv->mdev, cq_moderation))
472	472	return -EOPNOTSUPP;
@@ -521,7 +521,7 @@ mlx5e_set_priv_channels_coalesce(struct mlx5e_priv *priv, struct ethtool_coalesc · 48 unmodified lines
521	521	int mlx5e_ethtool_set_coalesce(struct mlx5e_priv *priv,
522	522	struct ethtool_coalesce *coal)
523	523	{
524		− struct net_dim_cq_moder rx_moder, tx_moder;
	524	+ struct dim_cq_moder rx_moder, tx_moder;
525	525	struct mlx5_core_dev *mdev = priv->mdev;
526	526	struct mlx5e_channels new_channels = {};
527	527	int err = 0;

drivers/net/ethernet/mellanox/mlx5/core/en_main.c+11−11

@@ -584,11 +584,11 @@ static int mlx5e_alloc_rq(struct mlx5e_channel *c, · 583 unmodified lines
584	584
585	585	switch (params->rx_cq_moderation.cq_period_mode) {
586	586	case MLX5_CQ_PERIOD_MODE_START_FROM_CQE:
587		− rq->dim.mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE;
	587	+ rq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_CQE;
588	588	break;
589	589	case MLX5_CQ_PERIOD_MODE_START_FROM_EQE:
590	590	default:
591		− rq->dim.mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
	591	+ rq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
592	592	}
593	593
594	594	rq->page_cache.head = 0;
@@ -1571,7 +1571,7 @@ static void mlx5e_destroy_cq(struct mlx5e_cq *cq) · 976 unmodified lines
1571	1571	}
1572	1572
1573	1573	static int mlx5e_open_cq(struct mlx5e_channel *c,
1574		− struct net_dim_cq_moder moder,
	1574	+ struct dim_cq_moder moder,
1575	1575	struct mlx5e_cq_param *param,
1576	1576	struct mlx5e_cq *cq)
1577	1577	{
@@ -1776,7 +1776,7 @@ static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix, · 198 unmodified lines
1776	1776	struct mlx5e_channel **cp)
1777	1777	{
1778	1778	int cpu = cpumask_first(mlx5_comp_irq_get_affinity_mask(priv->mdev, ix));
1779		− struct net_dim_cq_moder icocq_moder = {0, 0};
	1779	+ struct dim_cq_moder icocq_moder = {0, 0};
1780	1780	struct net_device *netdev = priv->netdev;
1781	1781	struct mlx5e_channel *c;
1782	1782	unsigned int irq;
@@ -2153,7 +2153,7 @@ static void mlx5e_build_ico_cq_param(struct mlx5e_priv *priv, · 370 unmodified lines
2153	2153
2154	2154	mlx5e_build_common_cq_param(priv, param);
2155	2155
2156		− param->cq_period_mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
	2156	+ param->cq_period_mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
2157	2157	}
2158	2158
2159	2159	static void mlx5e_build_icosq_param(struct mlx5e_priv *priv,
@@ -4421,9 +4421,9 @@ static bool slow_pci_heuristic(struct mlx5_core_dev *mdev) · 2261 unmodified lines
4421	4421	link_speed > MLX5E_SLOW_PCI_RATIO * pci_bw;
4422	4422	}
4423	4423
4424		−static struct net_dim_cq_moder mlx5e_get_def_tx_moderation(u8 cq_period_mode)
	4424	+static struct dim_cq_moder mlx5e_get_def_tx_moderation(u8 cq_period_mode)
4425	4425	{
4426		− struct net_dim_cq_moder moder;
	4426	+ struct dim_cq_moder moder;
4427	4427
4428	4428	moder.cq_period_mode = cq_period_mode;
4429	4429	moder.pkts = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
@@ -4434,9 +4434,9 @@ static struct net_dim_cq_moder mlx5e_get_def_tx_moderation(u8 cq_period_mode) · 4 unmodified lines
4434	4434	return moder;
4435	4435	}
4436	4436
4437		−static struct net_dim_cq_moder mlx5e_get_def_rx_moderation(u8 cq_period_mode)
	4437	+static struct dim_cq_moder mlx5e_get_def_rx_moderation(u8 cq_period_mode)
4438	4438	{
4439		− struct net_dim_cq_moder moder;
	4439	+ struct dim_cq_moder moder;
4440	4440
4441	4441	moder.cq_period_mode = cq_period_mode;
4442	4442	moder.pkts = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
@@ -4450,8 +4450,8 @@ static struct net_dim_cq_moder mlx5e_get_def_rx_moderation(u8 cq_period_mode) · 7 unmodified lines
4450	4450	static u8 mlx5_to_net_dim_cq_period_mode(u8 cq_period_mode)
4451	4451	{
4452	4452	return cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE ?
4453		− NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE :
4454		− NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
	4453	+ DIM_CQ_PERIOD_MODE_START_FROM_CQE :
	4454	+ DIM_CQ_PERIOD_MODE_START_FROM_EQE;
4455	4455	}
4456	4456
4457	4457	void mlx5e_set_tx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)

drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c+4−6

@@ -48,26 +48,24 @@ static inline bool mlx5e_channel_no_affinity_change(struct mlx5e_channel *c) · 47 unmodified lines
48	48	static void mlx5e_handle_tx_dim(struct mlx5e_txqsq *sq)
49	49	{
50	50	struct mlx5e_sq_stats *stats = sq->stats;
51		− struct net_dim_sample dim_sample;
	51	+ struct dim_sample dim_sample;
52	52
53	53	if (unlikely(!test_bit(MLX5E_SQ_STATE_AM, &sq->state)))
54	54	return;
55	55
56		− net_dim_sample(sq->cq.event_ctr, stats->packets, stats->bytes,
57		− &dim_sample);
	56	+ dim_update_sample(sq->cq.event_ctr, stats->packets, stats->bytes, &dim_sample);
58	57	net_dim(&sq->dim, dim_sample);
59	58	}
60	59
61	60	static void mlx5e_handle_rx_dim(struct mlx5e_rq *rq)
62	61	{
63	62	struct mlx5e_rq_stats *stats = rq->stats;
64		− struct net_dim_sample dim_sample;
	63	+ struct dim_sample dim_sample;
65	64
66	65	if (unlikely(!test_bit(MLX5E_RQ_STATE_AM, &rq->state)))
67	66	return;
68	67
69		− net_dim_sample(rq->cq.event_ctr, stats->packets, stats->bytes,
70		− &dim_sample);
	68	+ dim_update_sample(rq->cq.event_ctr, stats->packets, stats->bytes, &dim_sample);
71	69	net_dim(&rq->dim, dim_sample);
72	70	}
73	71

include/linux/dim.h+366−0

@@ -0,0 +1,366 @@
	1	+/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
	2	+/* Copyright (c) 2019 Mellanox Technologies. */
	3	+
	4	+#ifndef DIM_H
	5	+#define DIM_H
	6	+
	7	+#include <linux/module.h>
	8	+
	9	+/**
	10	+ * Number of events between DIM iterations.
	11	+ * Causes a moderation of the algorithm run.
	12	+ */
	13	+#define DIM_NEVENTS 64
	14	+
	15	+/**
	16	+ * Is a difference between values justifies taking an action.
	17	+ * We consider 10% difference as significant.
	18	+ */
	19	+#define IS_SIGNIFICANT_DIFF(val, ref) \
	20	+ (((100UL * abs((val) - (ref))) / (ref)) > 10)
	21	+
	22	+/**
	23	+ * Calculate the gap between two values.
	24	+ * Take wrap-around and variable size into consideration.
	25	+ */
	26	+#define BIT_GAP(bits, end, start) ((((end) - (start)) + BIT_ULL(bits)) \
	27	+ & (BIT_ULL(bits) - 1))
	28	+
	29	+/**
	30	+ * Structure for CQ moderation values.
	31	+ * Used for communications between DIM and its consumer.
	32	+ *
	33	+ * @usec: CQ timer suggestion (by DIM)
	34	+ * @pkts: CQ packet counter suggestion (by DIM)
	35	+ * @cq_period_mode: CQ priod count mode (from CQE/EQE)
	36	+ */
	37	+struct dim_cq_moder {
	38	+ u16 usec;
	39	+ u16 pkts;
	40	+ u16 comps;
	41	+ u8 cq_period_mode;
	42	+};
	43	+
	44	+/**
	45	+ * Structure for DIM sample data.
	46	+ * Used for communications between DIM and its consumer.
	47	+ *
	48	+ * @time: Sample timestamp
	49	+ * @pkt_ctr: Number of packets
	50	+ * @byte_ctr: Number of bytes
	51	+ * @event_ctr: Number of events
	52	+ */
	53	+struct dim_sample {
	54	+ ktime_t time;
	55	+ u32 pkt_ctr;
	56	+ u32 byte_ctr;
	57	+ u16 event_ctr;
	58	+ u32 comp_ctr;
	59	+};
	60	+
	61	+/**
	62	+ * Structure for DIM stats.
	63	+ * Used for holding current measured rates.
	64	+ *
	65	+ * @ppms: Packets per msec
	66	+ * @bpms: Bytes per msec
	67	+ * @epms: Events per msec
	68	+ */
	69	+struct dim_stats {
	70	+ int ppms; /* packets per msec */
	71	+ int bpms; /* bytes per msec */
	72	+ int epms; /* events per msec */
	73	+ int cpms; /* completions per msec */
	74	+ int cpe_ratio; /* ratio of completions to events */
	75	+};
	76	+
	77	+/**
	78	+ * Main structure for dynamic interrupt moderation (DIM).
	79	+ * Used for holding all information about a specific DIM instance.
	80	+ *
	81	+ * @state: Algorithm state (see below)
	82	+ * @prev_stats: Measured rates from previous iteration (for comparison)
	83	+ * @start_sample: Sampled data at start of current iteration
	84	+ * @work: Work to perform on action required
	85	+ * @profile_ix: Current moderation profile
	86	+ * @mode: CQ period count mode
	87	+ * @tune_state: Algorithm tuning state (see below)
	88	+ * @steps_right: Number of steps taken towards higher moderation
	89	+ * @steps_left: Number of steps taken towards lower moderation
	90	+ * @tired: Parking depth counter
	91	+ */
	92	+struct dim {
	93	+ u8 state;
	94	+ struct dim_stats prev_stats;
	95	+ struct dim_sample start_sample;
	96	+ struct dim_sample measuring_sample;
	97	+ struct work_struct work;
	98	+ u8 profile_ix;
	99	+ u8 mode;
	100	+ u8 tune_state;
	101	+ u8 steps_right;
	102	+ u8 steps_left;
	103	+ u8 tired;
	104	+};
	105	+
	106	+/**
	107	+ * enum dim_cq_period_mode
	108	+ *
	109	+ * These are the modes for CQ period count.
	110	+ *
	111	+ * @DIM_CQ_PERIOD_MODE_START_FROM_EQE: Start counting from EQE
	112	+ * @DIM_CQ_PERIOD_MODE_START_FROM_CQE: Start counting from CQE (implies timer reset)
	113	+ * @DIM_CQ_PERIOD_NUM_MODES: Number of modes
	114	+ */
	115	+enum {
	116	+ DIM_CQ_PERIOD_MODE_START_FROM_EQE = 0x0,
	117	+ DIM_CQ_PERIOD_MODE_START_FROM_CQE = 0x1,
	118	+ DIM_CQ_PERIOD_NUM_MODES
	119	+};
	120	+
	121	+/**
	122	+ * enum dim_state
	123	+ *
	124	+ * These are the DIM algorithm states.
	125	+ * These will determine if the algorithm is in a valid state to start an iteration.
	126	+ *
	127	+ * @DIM_START_MEASURE: This is the first iteration (also after applying a new profile)
	128	+ * @DIM_MEASURE_IN_PROGRESS: Algorithm is already in progress - check if
	129	+ * need to perform an action
	130	+ * @DIM_APPLY_NEW_PROFILE: DIM consumer is currently applying a profile - no need to measure
	131	+ */
	132	+enum {
	133	+ DIM_START_MEASURE,
	134	+ DIM_MEASURE_IN_PROGRESS,
	135	+ DIM_APPLY_NEW_PROFILE,
	136	+};
	137	+
	138	+/**
	139	+ * enum dim_tune_state
	140	+ *
	141	+ * These are the DIM algorithm tune states.
	142	+ * These will determine which action the algorithm should perform.
	143	+ *
	144	+ * @DIM_PARKING_ON_TOP: Algorithm found a local top point - exit on significant difference
	145	+ * @DIM_PARKING_TIRED: Algorithm found a deep top point - don't exit if tired > 0
	146	+ * @DIM_GOING_RIGHT: Algorithm is currently trying higher moderation levels
	147	+ * @DIM_GOING_LEFT: Algorithm is currently trying lower moderation levels
	148	+ */
	149	+enum {
	150	+ DIM_PARKING_ON_TOP,
	151	+ DIM_PARKING_TIRED,
	152	+ DIM_GOING_RIGHT,
	153	+ DIM_GOING_LEFT,
	154	+};
	155	+
	156	+/**
	157	+ * enum dim_stats_state
	158	+ *
	159	+ * These are the DIM algorithm statistics states.
	160	+ * These will determine the verdict of current iteration.
	161	+ *
	162	+ * @DIM_STATS_WORSE: Current iteration shows worse performance than before
	163	+ * @DIM_STATS_WORSE: Current iteration shows same performance than before
	164	+ * @DIM_STATS_WORSE: Current iteration shows better performance than before
	165	+ */
	166	+enum {
	167	+ DIM_STATS_WORSE,
	168	+ DIM_STATS_SAME,
	169	+ DIM_STATS_BETTER,
	170	+};
	171	+
	172	+/**
	173	+ * enum dim_step_result
	174	+ *
	175	+ * These are the DIM algorithm step results.
	176	+ * These describe the result of a step.
	177	+ *
	178	+ * @DIM_STEPPED: Performed a regular step
	179	+ * @DIM_TOO_TIRED: Same kind of step was done multiple times - should go to
	180	+ * tired parking
	181	+ * @DIM_ON_EDGE: Stepped to the most left/right profile
	182	+ */
	183	+enum {
	184	+ DIM_STEPPED,
	185	+ DIM_TOO_TIRED,
	186	+ DIM_ON_EDGE,
	187	+};
	188	+
	189	+/**
	190	+ * dim_on_top - check if current state is a good place to stop (top location)
	191	+ * @dim: DIM context
	192	+ *
	193	+ * Check if current profile is a good place to park at.
	194	+ * This will result in reducing the DIM checks frequency as we assume we
	195	+ * shouldn't probably change profiles, unless traffic pattern wasn't changed.
	196	+ */
	197	+bool dim_on_top(struct dim *dim);
	198	+
	199	+/**
	200	+ * dim_turn - change profile alterning direction
	201	+ * @dim: DIM context
	202	+ *
	203	+ * Go left if we were going right and vice-versa.
	204	+ * Do nothing if currently parking.
	205	+ */
	206	+void dim_turn(struct dim *dim);
	207	+
	208	+/**
	209	+ * dim_park_on_top - enter a parking state on a top location
	210	+ * @dim: DIM context
	211	+ *
	212	+ * Enter parking state.
	213	+ * Clear all movement history.
	214	+ */
	215	+void dim_park_on_top(struct dim *dim);
	216	+
	217	+/**
	218	+ * dim_park_tired - enter a tired parking state
	219	+ * @dim: DIM context
	220	+ *
	221	+ * Enter parking state.
	222	+ * Clear all movement history and cause DIM checks frequency to reduce.
	223	+ */
	224	+void dim_park_tired(struct dim *dim);
	225	+
	226	+/**
	227	+ * dim_calc_stats - calculate the difference between two samples
	228	+ * @start: start sample
	229	+ * @end: end sample
	230	+ * @curr_stats: delta between samples
	231	+ *
	232	+ * Calculate the delta between two samples (in data rates).
	233	+ * Takes into consideration counter wrap-around.
	234	+ */
	235	+void dim_calc_stats(struct dim_sample start, struct dim_sample end,
	236	+ struct dim_stats *curr_stats);
	237	+
	238	+/**
	239	+ * dim_update_sample - set a sample's fields with give values
	240	+ * @event_ctr: number of events to set
	241	+ * @packets: number of packets to set
	242	+ * @bytes: number of bytes to set
	243	+ * @s: DIM sample
	244	+ */
	245	+static inline void
	246	+dim_update_sample(u16 event_ctr, u64 packets, u64 bytes, struct dim_sample *s)
	247	+{
	248	+ s->time = ktime_get();
	249	+ s->pkt_ctr = packets;
	250	+ s->byte_ctr = bytes;
	251	+ s->event_ctr = event_ctr;
	252	+}
	253	+
	254	+/**
	255	+ * dim_update_sample_with_comps - set a sample's fields with given
	256	+ * values including the completion parameter
	257	+ * @event_ctr: number of events to set
	258	+ * @packets: number of packets to set
	259	+ * @bytes: number of bytes to set
	260	+ * @comps: number of completions to set
	261	+ * @s: DIM sample
	262	+ */
	263	+static inline void
	264	+dim_update_sample_with_comps(u16 event_ctr, u64 packets, u64 bytes, u64 comps,
	265	+ struct dim_sample *s)
	266	+{
	267	+ dim_update_sample(event_ctr, packets, bytes, s);
	268	+ s->comp_ctr = comps;
	269	+}
	270	+
	271	+/* Net DIM */
	272	+
	273	+/*
	274	+ * Net DIM profiles:
	275	+ * There are different set of profiles for each CQ period mode.
	276	+ * There are different set of profiles for RX/TX CQs.
	277	+ * Each profile size must be of NET_DIM_PARAMS_NUM_PROFILES
	278	+ */
	279	+#define NET_DIM_PARAMS_NUM_PROFILES 5
	280	+#define NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE 256
	281	+#define NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE 128
	282	+#define NET_DIM_DEF_PROFILE_CQE 1
	283	+#define NET_DIM_DEF_PROFILE_EQE 1
	284	+
	285	+#define NET_DIM_RX_EQE_PROFILES { \
	286	+ {1, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
	287	+ {8, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
	288	+ {64, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
	289	+ {128, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
	290	+ {256, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
	291	+}
	292	+
	293	+#define NET_DIM_RX_CQE_PROFILES { \
	294	+ {2, 256}, \
	295	+ {8, 128}, \
	296	+ {16, 64}, \
	297	+ {32, 64}, \
	298	+ {64, 64} \
	299	+}
	300	+
	301	+#define NET_DIM_TX_EQE_PROFILES { \
	302	+ {1, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE}, \
	303	+ {8, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE}, \
	304	+ {32, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE}, \
	305	+ {64, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE}, \
	306	+ {128, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE} \
	307	+}
	308	+
	309	+#define NET_DIM_TX_CQE_PROFILES { \
	310	+ {5, 128}, \
	311	+ {8, 64}, \
	312	+ {16, 32}, \
	313	+ {32, 32}, \
	314	+ {64, 32} \
	315	+}
	316	+
	317	+static const struct dim_cq_moder
	318	+rx_profile[DIM_CQ_PERIOD_NUM_MODES][NET_DIM_PARAMS_NUM_PROFILES] = {
	319	+ NET_DIM_RX_EQE_PROFILES,
	320	+ NET_DIM_RX_CQE_PROFILES,
	321	+};
	322	+
	323	+static const struct dim_cq_moder
	324	+tx_profile[DIM_CQ_PERIOD_NUM_MODES][NET_DIM_PARAMS_NUM_PROFILES] = {
	325	+ NET_DIM_TX_EQE_PROFILES,
	326	+ NET_DIM_TX_CQE_PROFILES,
	327	+};
	328	+
	329	+/**
	330	+ * net_dim_get_rx_moderation - provide a CQ moderation object for the given RX profile
	331	+ * @cq_period_mode: CQ period mode
	332	+ * @ix: Profile index
	333	+ */
	334	+struct dim_cq_moder net_dim_get_rx_moderation(u8 cq_period_mode, int ix);
	335	+
	336	+/**
	337	+ * net_dim_get_def_rx_moderation - provide the default RX moderation
	338	+ * @cq_period_mode: CQ period mode
	339	+ */
	340	+struct dim_cq_moder net_dim_get_def_rx_moderation(u8 cq_period_mode);
	341	+
	342	+/**
	343	+ * net_dim_get_tx_moderation - provide a CQ moderation object for the given TX profile
	344	+ * @cq_period_mode: CQ period mode
	345	+ * @ix: Profile index
	346	+ */
	347	+struct dim_cq_moder net_dim_get_tx_moderation(u8 cq_period_mode, int ix);
	348	+
	349	+/**
	350	+ * net_dim_get_def_tx_moderation - provide the default TX moderation
	351	+ * @cq_period_mode: CQ period mode
	352	+ */
	353	+struct dim_cq_moder net_dim_get_def_tx_moderation(u8 cq_period_mode);
	354	+
	355	+/**
	356	+ * net_dim - main DIM algorithm entry point
	357	+ * @dim: DIM instance information
	358	+ * @end_sample: Current data measurement
	359	+ *
	360	+ * Called by the consumer.
	361	+ * This is the main logic of the algorithm, where data is processed in order to decide on next
	362	+ * required action.
	363	+ */
	364	+void net_dim(struct dim *dim, struct dim_sample end_sample);
	365	+
	366	+#endif /* DIM_H */

include/linux/net_dim.h+0−418

@@ -1,418 +0,0 @@
1		−/*
2		− * Copyright (c) 2016, Mellanox Technologies. All rights reserved.
3		− * Copyright (c) 2017-2018, Broadcom Limited. All rights reserved.
4		− *
5		− * This software is available to you under a choice of one of two
6		− * licenses. You may choose to be licensed under the terms of the GNU
7		− * General Public License (GPL) Version 2, available from the file
8		− * COPYING in the main directory of this source tree, or the
9		− * OpenIB.org BSD license below:
10		− *
11		− * Redistribution and use in source and binary forms, with or
12		− * without modification, are permitted provided that the following
13		− * conditions are met:
14		− *
15		− * - Redistributions of source code must retain the above
16		− * copyright notice, this list of conditions and the following
17		− * disclaimer.
18		− *
19		− * - Redistributions in binary form must reproduce the above
20		− * copyright notice, this list of conditions and the following
21		− * disclaimer in the documentation and/or other materials
22		− * provided with the distribution.
23		− *
24		− * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25		− * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26		− * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27		− * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28		− * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29		− * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30		− * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31		− * SOFTWARE.
32		− */
33		−
34		−#ifndef NET_DIM_H
35		−#define NET_DIM_H
36		−
37		−#include <linux/module.h>
38		−
39		−struct net_dim_cq_moder {
40		− u16 usec;
41		− u16 pkts;
42		− u8 cq_period_mode;
43		−};
44		−
45		−struct net_dim_sample {
46		− ktime_t time;
47		− u32 pkt_ctr;
48		− u32 byte_ctr;
49		− u16 event_ctr;
50		−};
51		−
52		−struct net_dim_stats {
53		− int ppms; /* packets per msec */
54		− int bpms; /* bytes per msec */
55		− int epms; /* events per msec */
56		−};
57		−
58		−struct net_dim { /* Adaptive Moderation */
59		− u8 state;
60		− struct net_dim_stats prev_stats;
61		− struct net_dim_sample start_sample;
62		− struct work_struct work;
63		− u8 profile_ix;
64		− u8 mode;
65		− u8 tune_state;
66		− u8 steps_right;
67		− u8 steps_left;
68		− u8 tired;
69		−};
70		−
71		−enum {
72		− NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE = 0x0,
73		− NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE = 0x1,
74		− NET_DIM_CQ_PERIOD_NUM_MODES
75		−};
76		−
77		−/* Adaptive moderation logic */
78		−enum {
79		− NET_DIM_START_MEASURE,
80		− NET_DIM_MEASURE_IN_PROGRESS,
81		− NET_DIM_APPLY_NEW_PROFILE,
82		−};
83		−
84		−enum {
85		− NET_DIM_PARKING_ON_TOP,
86		− NET_DIM_PARKING_TIRED,
87		− NET_DIM_GOING_RIGHT,
88		− NET_DIM_GOING_LEFT,
89		−};
90		−
91		−enum {
92		− NET_DIM_STATS_WORSE,
93		− NET_DIM_STATS_SAME,
94		− NET_DIM_STATS_BETTER,
95		−};
96		−
97		−enum {
98		− NET_DIM_STEPPED,
99		− NET_DIM_TOO_TIRED,
100		− NET_DIM_ON_EDGE,
101		−};
102		−
103		−#define NET_DIM_PARAMS_NUM_PROFILES 5
104		−/* Adaptive moderation profiles */
105		−#define NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE 256
106		−#define NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE 128
107		−#define NET_DIM_DEF_PROFILE_CQE 1
108		−#define NET_DIM_DEF_PROFILE_EQE 1
109		−
110		−/* All profiles sizes must be NET_PARAMS_DIM_NUM_PROFILES */
111		−#define NET_DIM_RX_EQE_PROFILES { \
112		− {1, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
113		− {8, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
114		− {64, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
115		− {128, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
116		− {256, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
117		−}
118		−
119		−#define NET_DIM_RX_CQE_PROFILES { \
120		− {2, 256}, \
121		− {8, 128}, \
122		− {16, 64}, \
123		− {32, 64}, \
124		− {64, 64} \
125		−}
126		−
127		−#define NET_DIM_TX_EQE_PROFILES { \
128		− {1, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE}, \
129		− {8, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE}, \
130		− {32, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE}, \
131		− {64, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE}, \
132		− {128, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE} \
133		−}
134		−
135		−#define NET_DIM_TX_CQE_PROFILES { \
136		− {5, 128}, \
137		− {8, 64}, \
138		− {16, 32}, \
139		− {32, 32}, \
140		− {64, 32} \
141		−}
142		−
143		−static const struct net_dim_cq_moder
144		−rx_profile[NET_DIM_CQ_PERIOD_NUM_MODES][NET_DIM_PARAMS_NUM_PROFILES] = {
145		− NET_DIM_RX_EQE_PROFILES,
146		− NET_DIM_RX_CQE_PROFILES,
147		−};
148		−
149		−static const struct net_dim_cq_moder
150		−tx_profile[NET_DIM_CQ_PERIOD_NUM_MODES][NET_DIM_PARAMS_NUM_PROFILES] = {
151		− NET_DIM_TX_EQE_PROFILES,
152		− NET_DIM_TX_CQE_PROFILES,
153		−};
154		−
155		−static inline struct net_dim_cq_moder
156		−net_dim_get_rx_moderation(u8 cq_period_mode, int ix)
157		−{
158		− struct net_dim_cq_moder cq_moder = rx_profile[cq_period_mode][ix];
159		−
160		− cq_moder.cq_period_mode = cq_period_mode;
161		− return cq_moder;
162		−}
163		−
164		−static inline struct net_dim_cq_moder
165		−net_dim_get_def_rx_moderation(u8 cq_period_mode)
166		−{
167		− u8 profile_ix = cq_period_mode == NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE ?
168		− NET_DIM_DEF_PROFILE_CQE : NET_DIM_DEF_PROFILE_EQE;
169		−
170		− return net_dim_get_rx_moderation(cq_period_mode, profile_ix);
171		−}
172		−
173		−static inline struct net_dim_cq_moder
174		−net_dim_get_tx_moderation(u8 cq_period_mode, int ix)
175		−{
176		− struct net_dim_cq_moder cq_moder = tx_profile[cq_period_mode][ix];
177		−
178		− cq_moder.cq_period_mode = cq_period_mode;
179		− return cq_moder;
180		−}
181		−
182		−static inline struct net_dim_cq_moder
183		−net_dim_get_def_tx_moderation(u8 cq_period_mode)
184		−{
185		− u8 profile_ix = cq_period_mode == NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE ?
186		− NET_DIM_DEF_PROFILE_CQE : NET_DIM_DEF_PROFILE_EQE;
187		−
188		− return net_dim_get_tx_moderation(cq_period_mode, profile_ix);
189		−}
190		−
191		−static inline bool net_dim_on_top(struct net_dim *dim)
192		−{
193		− switch (dim->tune_state) {
194		− case NET_DIM_PARKING_ON_TOP:
195		− case NET_DIM_PARKING_TIRED:
196		− return true;
197		− case NET_DIM_GOING_RIGHT:
198		− return (dim->steps_left > 1) && (dim->steps_right == 1);
199		− default: /* NET_DIM_GOING_LEFT */
200		− return (dim->steps_right > 1) && (dim->steps_left == 1);
201		− }
202		−}
203		−
204		−static inline void net_dim_turn(struct net_dim *dim)
205		−{
206		− switch (dim->tune_state) {
207		− case NET_DIM_PARKING_ON_TOP:
208		− case NET_DIM_PARKING_TIRED:
209		− break;
210		− case NET_DIM_GOING_RIGHT:
211		− dim->tune_state = NET_DIM_GOING_LEFT;
212		− dim->steps_left = 0;
213		− break;
214		− case NET_DIM_GOING_LEFT:
215		− dim->tune_state = NET_DIM_GOING_RIGHT;
216		− dim->steps_right = 0;
217		− break;
218		− }
219		−}
220		−
221		−static inline int net_dim_step(struct net_dim *dim)
222		−{
223		− if (dim->tired == (NET_DIM_PARAMS_NUM_PROFILES * 2))
224		− return NET_DIM_TOO_TIRED;
225		−
226		− switch (dim->tune_state) {
227		− case NET_DIM_PARKING_ON_TOP:
228		− case NET_DIM_PARKING_TIRED:
229		− break;
230		− case NET_DIM_GOING_RIGHT:
231		− if (dim->profile_ix == (NET_DIM_PARAMS_NUM_PROFILES - 1))
232		− return NET_DIM_ON_EDGE;
233		− dim->profile_ix++;
234		− dim->steps_right++;
235		− break;
236		− case NET_DIM_GOING_LEFT:
237		− if (dim->profile_ix == 0)
238		− return NET_DIM_ON_EDGE;
239		− dim->profile_ix--;
240		− dim->steps_left++;
241		− break;
242		− }
243		−
244		− dim->tired++;
245		− return NET_DIM_STEPPED;
246		−}
247		−
248		−static inline void net_dim_park_on_top(struct net_dim *dim)
249		−{
250		− dim->steps_right = 0;
251		− dim->steps_left = 0;
252		− dim->tired = 0;
253		− dim->tune_state = NET_DIM_PARKING_ON_TOP;
254		−}
255		−
256		−static inline void net_dim_park_tired(struct net_dim *dim)
257		−{
258		− dim->steps_right = 0;
259		− dim->steps_left = 0;
260		− dim->tune_state = NET_DIM_PARKING_TIRED;
261		−}
262		−
263		−static inline void net_dim_exit_parking(struct net_dim *dim)
264		−{
265		− dim->tune_state = dim->profile_ix ? NET_DIM_GOING_LEFT :
266		− NET_DIM_GOING_RIGHT;
267		− net_dim_step(dim);
268		−}
269		−
270		−#define IS_SIGNIFICANT_DIFF(val, ref) \
271		− (((100UL * abs((val) - (ref))) / (ref)) > 10) /* more than 10% difference */
272		−
273		−static inline int net_dim_stats_compare(struct net_dim_stats *curr,
274		− struct net_dim_stats *prev)
275		−{
276		− if (!prev->bpms)
277		− return curr->bpms ? NET_DIM_STATS_BETTER :
278		− NET_DIM_STATS_SAME;
279		−
280		− if (IS_SIGNIFICANT_DIFF(curr->bpms, prev->bpms))
281		− return (curr->bpms > prev->bpms) ? NET_DIM_STATS_BETTER :
282		− NET_DIM_STATS_WORSE;
283		−
284		− if (!prev->ppms)
285		− return curr->ppms ? NET_DIM_STATS_BETTER :
286		− NET_DIM_STATS_SAME;
287		−
288		− if (IS_SIGNIFICANT_DIFF(curr->ppms, prev->ppms))
289		− return (curr->ppms > prev->ppms) ? NET_DIM_STATS_BETTER :
290		− NET_DIM_STATS_WORSE;
291		−
292		− if (!prev->epms)
293		− return NET_DIM_STATS_SAME;
294		−
295		− if (IS_SIGNIFICANT_DIFF(curr->epms, prev->epms))
296		− return (curr->epms < prev->epms) ? NET_DIM_STATS_BETTER :
297		− NET_DIM_STATS_WORSE;
298		−
299		− return NET_DIM_STATS_SAME;
300		−}
301		−
302		−static inline bool net_dim_decision(struct net_dim_stats *curr_stats,
303		− struct net_dim *dim)
304		−{
305		− int prev_state = dim->tune_state;
306		− int prev_ix = dim->profile_ix;
307		− int stats_res;
308		− int step_res;
309		−
310		− switch (dim->tune_state) {
311		− case NET_DIM_PARKING_ON_TOP:
312		− stats_res = net_dim_stats_compare(curr_stats, &dim->prev_stats);
313		− if (stats_res != NET_DIM_STATS_SAME)
314		− net_dim_exit_parking(dim);
315		− break;
316		−
317		− case NET_DIM_PARKING_TIRED:
318		− dim->tired--;
319		− if (!dim->tired)
320		− net_dim_exit_parking(dim);
321		− break;
322		−
323		− case NET_DIM_GOING_RIGHT:
324		− case NET_DIM_GOING_LEFT:
325		− stats_res = net_dim_stats_compare(curr_stats, &dim->prev_stats);
326		− if (stats_res != NET_DIM_STATS_BETTER)
327		− net_dim_turn(dim);
328		−
329		− if (net_dim_on_top(dim)) {
330		− net_dim_park_on_top(dim);
331		− break;
332		− }
333		−
334		− step_res = net_dim_step(dim);
335		− switch (step_res) {
336		− case NET_DIM_ON_EDGE:
337		− net_dim_park_on_top(dim);
338		− break;
339		− case NET_DIM_TOO_TIRED:
340		− net_dim_park_tired(dim);
341		− break;
342		− }
343		−
344		− break;
345		− }
346		−
347		− if ((prev_state != NET_DIM_PARKING_ON_TOP) \|\|
348		− (dim->tune_state != NET_DIM_PARKING_ON_TOP))
349		− dim->prev_stats = *curr_stats;
350		−
351		− return dim->profile_ix != prev_ix;
352		−}
353		−
354		−static inline void net_dim_sample(u16 event_ctr,
355		− u64 packets,
356		− u64 bytes,
357		− struct net_dim_sample *s)
358		−{
359		− s->time = ktime_get();
360		− s->pkt_ctr = packets;
361		− s->byte_ctr = bytes;
362		− s->event_ctr = event_ctr;
363		−}
364		−
365		−#define NET_DIM_NEVENTS 64
366		−#define BIT_GAP(bits, end, start) ((((end) - (start)) + BIT_ULL(bits)) & (BIT_ULL(bits) - 1))
367		−
368		−static inline void net_dim_calc_stats(struct net_dim_sample *start,
369		− struct net_dim_sample *end,
370		− struct net_dim_stats *curr_stats)
371		−{
372		− /* u32 holds up to 71 minutes, should be enough */
373		− u32 delta_us = ktime_us_delta(end->time, start->time);
374		− u32 npkts = BIT_GAP(BITS_PER_TYPE(u32), end->pkt_ctr, start->pkt_ctr);
375		− u32 nbytes = BIT_GAP(BITS_PER_TYPE(u32), end->byte_ctr,
376		− start->byte_ctr);
377		−
378		− if (!delta_us)
379		− return;
380		−
381		− curr_stats->ppms = DIV_ROUND_UP(npkts * USEC_PER_MSEC, delta_us);
382		− curr_stats->bpms = DIV_ROUND_UP(nbytes * USEC_PER_MSEC, delta_us);
383		− curr_stats->epms = DIV_ROUND_UP(NET_DIM_NEVENTS * USEC_PER_MSEC,
384		− delta_us);
385		−}
386		−
387		−static inline void net_dim(struct net_dim *dim,
388		− struct net_dim_sample end_sample)
389		−{
390		− struct net_dim_stats curr_stats;
391		− u16 nevents;
392		−
393		− switch (dim->state) {
394		− case NET_DIM_MEASURE_IN_PROGRESS:
395		− nevents = BIT_GAP(BITS_PER_TYPE(u16),
396		− end_sample.event_ctr,
397		− dim->start_sample.event_ctr);
398		− if (nevents < NET_DIM_NEVENTS)
399		− break;
400		− net_dim_calc_stats(&dim->start_sample, &end_sample,
401		− &curr_stats);
402		− if (net_dim_decision(&curr_stats, dim)) {
403		− dim->state = NET_DIM_APPLY_NEW_PROFILE;
404		− schedule_work(&dim->work);
405		− break;
406		− }
407		− /* fall through */
408		− case NET_DIM_START_MEASURE:
409		− net_dim_sample(end_sample.event_ctr, end_sample.pkt_ctr, end_sample.byte_ctr,
410		− &dim->start_sample);
411		− dim->state = NET_DIM_MEASURE_IN_PROGRESS;
412		− break;
413		− case NET_DIM_APPLY_NEW_PROFILE:
414		− break;
415		− }
416		−}
417		−
418		−#endif /* NET_DIM_H */

lib/Kconfig+8−0

@@ -562,6 +562,14 @@ config SIGNATURE · 561 unmodified lines
562	562	Digital signature verification. Currently only RSA is supported.
563	563	Implementation is done using GnuPG MPI library
564	564
	565	+config DIMLIB
	566	+ bool "DIM library"
	567	+ default y
	568	+ help
	569	+ Dynamic Interrupt Moderation library.
	570	+ Implements an algorithm for dynamically change CQ modertion values
	571	+ according to run time performance.
	572	+
565	573	#
566	574	# libfdt files, only selected if needed.
567	575	#

lib/Makefile+1−0

@@ -202,6 +202,7 @@ obj-$(CONFIG_GLOB) += glob.o · 201 unmodified lines
202	202	obj-$(CONFIG_GLOB_SELFTEST) += globtest.o
203	203
204	204	obj-$(CONFIG_MPILIB) += mpi/
	205	+obj-$(CONFIG_DIMLIB) += dim/
205	206	obj-$(CONFIG_SIGNATURE) += digsig.o
206	207
207	208	lib-$(CONFIG_CLZ_TAB) += clz_tab.o

lib/dim/Makefile+9−0

@@ -0,0 +1,9 @@
	1	+#
	2	+# DIM Dynamic Interrupt Moderation library
	3	+#
	4	+
	5	+obj-$(CONFIG_DIMLIB) = net_dim.o
	6	+
	7	+net_dim-y = \
	8	+ dim.o \
	9	+ net_dim.o

lib/dim/dim.c+83−0

@@ -0,0 +1,83 @@
	1	+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
	2	+/*
	3	+ * Copyright (c) 2019, Mellanox Technologies inc. All rights reserved.
	4	+ */
	5	+
	6	+#include <linux/dim.h>
	7	+
	8	+bool dim_on_top(struct dim *dim)
	9	+{
	10	+ switch (dim->tune_state) {
	11	+ case DIM_PARKING_ON_TOP:
	12	+ case DIM_PARKING_TIRED:
	13	+ return true;
	14	+ case DIM_GOING_RIGHT:
	15	+ return (dim->steps_left > 1) && (dim->steps_right == 1);
	16	+ default: /* DIM_GOING_LEFT */
	17	+ return (dim->steps_right > 1) && (dim->steps_left == 1);
	18	+ }
	19	+}
	20	+EXPORT_SYMBOL(dim_on_top);
	21	+
	22	+void dim_turn(struct dim *dim)
	23	+{
	24	+ switch (dim->tune_state) {
	25	+ case DIM_PARKING_ON_TOP:
	26	+ case DIM_PARKING_TIRED:
	27	+ break;
	28	+ case DIM_GOING_RIGHT:
	29	+ dim->tune_state = DIM_GOING_LEFT;
	30	+ dim->steps_left = 0;
	31	+ break;
	32	+ case DIM_GOING_LEFT:
	33	+ dim->tune_state = DIM_GOING_RIGHT;
	34	+ dim->steps_right = 0;
	35	+ break;
	36	+ }
	37	+}
	38	+EXPORT_SYMBOL(dim_turn);
	39	+
	40	+void dim_park_on_top(struct dim *dim)
	41	+{
	42	+ dim->steps_right = 0;
	43	+ dim->steps_left = 0;
	44	+ dim->tired = 0;
	45	+ dim->tune_state = DIM_PARKING_ON_TOP;
	46	+}
	47	+EXPORT_SYMBOL(dim_park_on_top);
	48	+
	49	+void dim_park_tired(struct dim *dim)
	50	+{
	51	+ dim->steps_right = 0;
	52	+ dim->steps_left = 0;
	53	+ dim->tune_state = DIM_PARKING_TIRED;
	54	+}
	55	+EXPORT_SYMBOL(dim_park_tired);
	56	+
	57	+void dim_calc_stats(struct dim_sample start, struct dim_sample end,
	58	+ struct dim_stats *curr_stats)
	59	+{
	60	+ /* u32 holds up to 71 minutes, should be enough */
	61	+ u32 delta_us = ktime_us_delta(end->time, start->time);
	62	+ u32 npkts = BIT_GAP(BITS_PER_TYPE(u32), end->pkt_ctr, start->pkt_ctr);
	63	+ u32 nbytes = BIT_GAP(BITS_PER_TYPE(u32), end->byte_ctr,
	64	+ start->byte_ctr);
	65	+ u32 ncomps = BIT_GAP(BITS_PER_TYPE(u32), end->comp_ctr,
	66	+ start->comp_ctr);
	67	+
	68	+ if (!delta_us)
	69	+ return;
	70	+
	71	+ curr_stats->ppms = DIV_ROUND_UP(npkts * USEC_PER_MSEC, delta_us);
	72	+ curr_stats->bpms = DIV_ROUND_UP(nbytes * USEC_PER_MSEC, delta_us);
	73	+ curr_stats->epms = DIV_ROUND_UP(DIM_NEVENTS * USEC_PER_MSEC,
	74	+ delta_us);
	75	+ curr_stats->cpms = DIV_ROUND_UP(ncomps * USEC_PER_MSEC, delta_us);
	76	+ if (curr_stats->epms != 0)
	77	+ curr_stats->cpe_ratio =
	78	+ (curr_stats->cpms * 100) / curr_stats->epms;
	79	+ else
	80	+ curr_stats->cpe_ratio = 0;
	81	+
	82	+}
	83	+EXPORT_SYMBOL(dim_calc_stats);

lib/dim/net_dim.c+190−0

@@ -0,0 +1,190 @@
	1	+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
	2	+/*
	3	+ * Copyright (c) 2018, Mellanox Technologies inc. All rights reserved.
	4	+ */
	5	+
	6	+#include <linux/dim.h>
	7	+
	8	+struct dim_cq_moder
	9	+net_dim_get_rx_moderation(u8 cq_period_mode, int ix)
	10	+{
	11	+ struct dim_cq_moder cq_moder = rx_profile[cq_period_mode][ix];
	12	+
	13	+ cq_moder.cq_period_mode = cq_period_mode;
	14	+ return cq_moder;
	15	+}
	16	+EXPORT_SYMBOL(net_dim_get_rx_moderation);
	17	+
	18	+struct dim_cq_moder
	19	+net_dim_get_def_rx_moderation(u8 cq_period_mode)
	20	+{
	21	+ u8 profile_ix = cq_period_mode == DIM_CQ_PERIOD_MODE_START_FROM_CQE ?
	22	+ NET_DIM_DEF_PROFILE_CQE : NET_DIM_DEF_PROFILE_EQE;
	23	+
	24	+ return net_dim_get_rx_moderation(cq_period_mode, profile_ix);
	25	+}
	26	+EXPORT_SYMBOL(net_dim_get_def_rx_moderation);
	27	+
	28	+struct dim_cq_moder
	29	+net_dim_get_tx_moderation(u8 cq_period_mode, int ix)
	30	+{
	31	+ struct dim_cq_moder cq_moder = tx_profile[cq_period_mode][ix];
	32	+
	33	+ cq_moder.cq_period_mode = cq_period_mode;
	34	+ return cq_moder;
	35	+}
	36	+EXPORT_SYMBOL(net_dim_get_tx_moderation);
	37	+
	38	+struct dim_cq_moder
	39	+net_dim_get_def_tx_moderation(u8 cq_period_mode)
	40	+{
	41	+ u8 profile_ix = cq_period_mode == DIM_CQ_PERIOD_MODE_START_FROM_CQE ?
	42	+ NET_DIM_DEF_PROFILE_CQE : NET_DIM_DEF_PROFILE_EQE;
	43	+
	44	+ return net_dim_get_tx_moderation(cq_period_mode, profile_ix);
	45	+}
	46	+EXPORT_SYMBOL(net_dim_get_def_tx_moderation);
	47	+
	48	+static int net_dim_step(struct dim *dim)
	49	+{
	50	+ if (dim->tired == (NET_DIM_PARAMS_NUM_PROFILES * 2))
	51	+ return DIM_TOO_TIRED;
	52	+
	53	+ switch (dim->tune_state) {
	54	+ case DIM_PARKING_ON_TOP:
	55	+ case DIM_PARKING_TIRED:
	56	+ break;
	57	+ case DIM_GOING_RIGHT:
	58	+ if (dim->profile_ix == (NET_DIM_PARAMS_NUM_PROFILES - 1))
	59	+ return DIM_ON_EDGE;
	60	+ dim->profile_ix++;
	61	+ dim->steps_right++;
	62	+ break;
	63	+ case DIM_GOING_LEFT:
	64	+ if (dim->profile_ix == 0)
	65	+ return DIM_ON_EDGE;
	66	+ dim->profile_ix--;
	67	+ dim->steps_left++;
	68	+ break;
	69	+ }
	70	+
	71	+ dim->tired++;
	72	+ return DIM_STEPPED;
	73	+}
	74	+
	75	+static void net_dim_exit_parking(struct dim *dim)
	76	+{
	77	+ dim->tune_state = dim->profile_ix ? DIM_GOING_LEFT : DIM_GOING_RIGHT;
	78	+ net_dim_step(dim);
	79	+}
	80	+
	81	+static int net_dim_stats_compare(struct dim_stats *curr,
	82	+ struct dim_stats *prev)
	83	+{
	84	+ if (!prev->bpms)
	85	+ return curr->bpms ? DIM_STATS_BETTER : DIM_STATS_SAME;
	86	+
	87	+ if (IS_SIGNIFICANT_DIFF(curr->bpms, prev->bpms))
	88	+ return (curr->bpms > prev->bpms) ? DIM_STATS_BETTER :
	89	+ DIM_STATS_WORSE;
	90	+
	91	+ if (!prev->ppms)
	92	+ return curr->ppms ? DIM_STATS_BETTER :
	93	+ DIM_STATS_SAME;
	94	+
	95	+ if (IS_SIGNIFICANT_DIFF(curr->ppms, prev->ppms))
	96	+ return (curr->ppms > prev->ppms) ? DIM_STATS_BETTER :
	97	+ DIM_STATS_WORSE;
	98	+
	99	+ if (!prev->epms)
	100	+ return DIM_STATS_SAME;
	101	+
	102	+ if (IS_SIGNIFICANT_DIFF(curr->epms, prev->epms))
	103	+ return (curr->epms < prev->epms) ? DIM_STATS_BETTER :
	104	+ DIM_STATS_WORSE;
	105	+
	106	+ return DIM_STATS_SAME;
	107	+}
	108	+
	109	+static bool net_dim_decision(struct dim_stats curr_stats, struct dim dim)
	110	+{
	111	+ int prev_state = dim->tune_state;
	112	+ int prev_ix = dim->profile_ix;
	113	+ int stats_res;
	114	+ int step_res;
	115	+
	116	+ switch (dim->tune_state) {
	117	+ case DIM_PARKING_ON_TOP:
	118	+ stats_res = net_dim_stats_compare(curr_stats,
	119	+ &dim->prev_stats);
	120	+ if (stats_res != DIM_STATS_SAME)
	121	+ net_dim_exit_parking(dim);
	122	+ break;
	123	+
	124	+ case DIM_PARKING_TIRED:
	125	+ dim->tired--;
	126	+ if (!dim->tired)
	127	+ net_dim_exit_parking(dim);
	128	+ break;
	129	+
	130	+ case DIM_GOING_RIGHT:
	131	+ case DIM_GOING_LEFT:
	132	+ stats_res = net_dim_stats_compare(curr_stats,
	133	+ &dim->prev_stats);
	134	+ if (stats_res != DIM_STATS_BETTER)
	135	+ dim_turn(dim);
	136	+
	137	+ if (dim_on_top(dim)) {
	138	+ dim_park_on_top(dim);
	139	+ break;
	140	+ }
	141	+
	142	+ step_res = net_dim_step(dim);
	143	+ switch (step_res) {
	144	+ case DIM_ON_EDGE:
	145	+ dim_park_on_top(dim);
	146	+ break;
	147	+ case DIM_TOO_TIRED:
	148	+ dim_park_tired(dim);
	149	+ break;
	150	+ }
	151	+
	152	+ break;
	153	+ }
	154	+
	155	+ if (prev_state != DIM_PARKING_ON_TOP \|\|
	156	+ dim->tune_state != DIM_PARKING_ON_TOP)
	157	+ dim->prev_stats = *curr_stats;
	158	+
	159	+ return dim->profile_ix != prev_ix;
	160	+}
	161	+
	162	+void net_dim(struct dim *dim, struct dim_sample end_sample)
	163	+{
	164	+ struct dim_stats curr_stats;
	165	+ u16 nevents;
	166	+
	167	+ switch (dim->state) {
	168	+ case DIM_MEASURE_IN_PROGRESS:
	169	+ nevents = BIT_GAP(BITS_PER_TYPE(u16),
	170	+ end_sample.event_ctr,
	171	+ dim->start_sample.event_ctr);
	172	+ if (nevents < DIM_NEVENTS)
	173	+ break;
	174	+ dim_calc_stats(&dim->start_sample, &end_sample, &curr_stats);
	175	+ if (net_dim_decision(&curr_stats, dim)) {
	176	+ dim->state = DIM_APPLY_NEW_PROFILE;
	177	+ schedule_work(&dim->work);
	178	+ break;
	179	+ }
	180	+ /* fall through */
	181	+ case DIM_START_MEASURE:
	182	+ dim_update_sample(end_sample.event_ctr, end_sample.pkt_ctr,
	183	+ end_sample.byte_ctr, &dim->start_sample);
	184	+ dim->state = DIM_MEASURE_IN_PROGRESS;
	185	+ break;
	186	+ case DIM_APPLY_NEW_PROFILE:
	187	+ break;
	188	+ }
	189	+}
	190	+EXPORT_SYMBOL(net_dim);

Merge tag 'blk-dim-v2' into rdma.git for-next

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code