xref: /freebsd/sys/cam/cam_iosched.c (revision 3aba1d47c841c257036be7ae6f64f212a7c9da31)
1ba6c22ceSWarner Losh /*-
2ba6c22ceSWarner Losh  * CAM IO Scheduler Interface
3ba6c22ceSWarner Losh  *
4f24882ecSPedro F. Giffuni  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
5f24882ecSPedro F. Giffuni  *
6ba6c22ceSWarner Losh  * Copyright (c) 2015 Netflix, Inc.
7ba6c22ceSWarner Losh  *
8ba6c22ceSWarner Losh  * Redistribution and use in source and binary forms, with or without
9ba6c22ceSWarner Losh  * modification, are permitted provided that the following conditions
10ba6c22ceSWarner Losh  * are met:
11ba6c22ceSWarner Losh  * 1. Redistributions of source code must retain the above copyright
12ba6c22ceSWarner Losh  *    notice, this list of conditions, and the following disclaimer,
13ba6c22ceSWarner Losh  *    without modification, immediately at the beginning of the file.
14ba6c22ceSWarner Losh  * 2. The name of the author may not be used to endorse or promote products
15ba6c22ceSWarner Losh  *    derived from this software without specific prior written permission.
16ba6c22ceSWarner Losh  *
17ba6c22ceSWarner Losh  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18ba6c22ceSWarner Losh  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19ba6c22ceSWarner Losh  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20ba6c22ceSWarner Losh  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
21ba6c22ceSWarner Losh  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22ba6c22ceSWarner Losh  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23ba6c22ceSWarner Losh  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24ba6c22ceSWarner Losh  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25ba6c22ceSWarner Losh  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26ba6c22ceSWarner Losh  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27ba6c22ceSWarner Losh  * SUCH DAMAGE.
28ba6c22ceSWarner Losh  *
29ba6c22ceSWarner Losh  * $FreeBSD$
30ba6c22ceSWarner Losh  */
31ba6c22ceSWarner Losh 
32ba6c22ceSWarner Losh #include "opt_cam.h"
33ba6c22ceSWarner Losh #include "opt_ddb.h"
34ba6c22ceSWarner Losh 
35ba6c22ceSWarner Losh #include <sys/cdefs.h>
36ba6c22ceSWarner Losh __FBSDID("$FreeBSD$");
37ba6c22ceSWarner Losh 
38ba6c22ceSWarner Losh #include <sys/param.h>
39ba6c22ceSWarner Losh 
40ba6c22ceSWarner Losh #include <sys/systm.h>
41ba6c22ceSWarner Losh #include <sys/kernel.h>
42ba6c22ceSWarner Losh #include <sys/bio.h>
43ba6c22ceSWarner Losh #include <sys/lock.h>
44ba6c22ceSWarner Losh #include <sys/malloc.h>
45ba6c22ceSWarner Losh #include <sys/mutex.h>
46cf3ec151SWarner Losh #include <sys/sbuf.h>
47ba6c22ceSWarner Losh #include <sys/sysctl.h>
48ba6c22ceSWarner Losh 
49ba6c22ceSWarner Losh #include <cam/cam.h>
50ba6c22ceSWarner Losh #include <cam/cam_ccb.h>
51ba6c22ceSWarner Losh #include <cam/cam_periph.h>
52ba6c22ceSWarner Losh #include <cam/cam_xpt_periph.h>
53cf3ec151SWarner Losh #include <cam/cam_xpt_internal.h>
54ba6c22ceSWarner Losh #include <cam/cam_iosched.h>
55ba6c22ceSWarner Losh 
56ba6c22ceSWarner Losh #include <ddb/ddb.h>
57ba6c22ceSWarner Losh 
58ba6c22ceSWarner Losh static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
59ba6c22ceSWarner Losh     "CAM I/O Scheduler buffers");
60ba6c22ceSWarner Losh 
61ba6c22ceSWarner Losh /*
62ba6c22ceSWarner Losh  * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer
63ba6c22ceSWarner Losh  * over the bioq_* interface, with notions of separate calls for normal I/O and
64ba6c22ceSWarner Losh  * for trims.
65035ec48eSWarner Losh  *
66035ec48eSWarner Losh  * When CAM_IOSCHED_DYNAMIC is defined, the scheduler is enhanced to dynamically
67035ec48eSWarner Losh  * steer the rate of one type of traffic to help other types of traffic (eg
68035ec48eSWarner Losh  * limit writes when read latency deteriorates on SSDs).
69ba6c22ceSWarner Losh  */
70ba6c22ceSWarner Losh 
71df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
72ba6c22ceSWarner Losh 
73035ec48eSWarner Losh static int do_dynamic_iosched = 1;
74035ec48eSWarner Losh TUNABLE_INT("kern.cam.do_dynamic_iosched", &do_dynamic_iosched);
75035ec48eSWarner Losh SYSCTL_INT(_kern_cam, OID_AUTO, do_dynamic_iosched, CTLFLAG_RD,
76035ec48eSWarner Losh     &do_dynamic_iosched, 1,
77035ec48eSWarner Losh     "Enable Dynamic I/O scheduler optimizations.");
78ba6c22ceSWarner Losh 
79cf3ec151SWarner Losh /*
80cf3ec151SWarner Losh  * For an EMA, with an alpha of alpha, we know
81cf3ec151SWarner Losh  * 	alpha = 2 / (N + 1)
82cf3ec151SWarner Losh  * or
83cf3ec151SWarner Losh  * 	N = 1 + (2 / alpha)
84cf3ec151SWarner Losh  * where N is the number of samples that 86% of the current
85cf3ec151SWarner Losh  * EMA is derived from.
86cf3ec151SWarner Losh  *
87cf3ec151SWarner Losh  * So we invent[*] alpha_bits:
88cf3ec151SWarner Losh  *	alpha_bits = -log_2(alpha)
89cf3ec151SWarner Losh  *	alpha = 2^-alpha_bits
90cf3ec151SWarner Losh  * So
91cf3ec151SWarner Losh  *	N = 1 + 2^(alpha_bits + 1)
92cf3ec151SWarner Losh  *
93cf3ec151SWarner Losh  * The default 9 gives a 1025 lookback for 86% of the data.
94cf3ec151SWarner Losh  * For a brief intro: https://en.wikipedia.org/wiki/Moving_average
95cf3ec151SWarner Losh  *
96cf3ec151SWarner Losh  * [*] Steal from the load average code and many other places.
9779d80af2SWarner Losh  * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
98cf3ec151SWarner Losh  */
99ba6c22ceSWarner Losh static int alpha_bits = 9;
100ba6c22ceSWarner Losh TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits);
101ba6c22ceSWarner Losh SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW,
102ba6c22ceSWarner Losh     &alpha_bits, 1,
103ba6c22ceSWarner Losh     "Bits in EMA's alpha.");
104ba6c22ceSWarner Losh 
105ba6c22ceSWarner Losh struct iop_stats;
106ba6c22ceSWarner Losh struct cam_iosched_softc;
107ba6c22ceSWarner Losh 
108ba6c22ceSWarner Losh int iosched_debug = 0;
109ba6c22ceSWarner Losh 
110ba6c22ceSWarner Losh typedef enum {
111ba6c22ceSWarner Losh 	none = 0,				/* No limits */
112ba6c22ceSWarner Losh 	queue_depth,			/* Limit how many ops we queue to SIM */
113ba6c22ceSWarner Losh 	iops,				/* Limit # of IOPS to the drive */
114ba6c22ceSWarner Losh 	bandwidth,			/* Limit bandwidth to the drive */
115ba6c22ceSWarner Losh 	limiter_max
116ba6c22ceSWarner Losh } io_limiter;
117ba6c22ceSWarner Losh 
118ba6c22ceSWarner Losh static const char *cam_iosched_limiter_names[] =
119ba6c22ceSWarner Losh     { "none", "queue_depth", "iops", "bandwidth" };
120ba6c22ceSWarner Losh 
121ba6c22ceSWarner Losh /*
122ba6c22ceSWarner Losh  * Called to initialize the bits of the iop_stats structure relevant to the
123ba6c22ceSWarner Losh  * limiter. Called just after the limiter is set.
124ba6c22ceSWarner Losh  */
125ba6c22ceSWarner Losh typedef int l_init_t(struct iop_stats *);
126ba6c22ceSWarner Losh 
127ba6c22ceSWarner Losh /*
128ba6c22ceSWarner Losh  * Called every tick.
129ba6c22ceSWarner Losh  */
130ba6c22ceSWarner Losh typedef int l_tick_t(struct iop_stats *);
131ba6c22ceSWarner Losh 
132ba6c22ceSWarner Losh /*
133ba6c22ceSWarner Losh  * Called to see if the limiter thinks this IOP can be allowed to
1340b4060b0SEd Maste  * proceed. If so, the limiter assumes that the IOP proceeded
135ba6c22ceSWarner Losh  * and makes any accounting of it that's needed.
136ba6c22ceSWarner Losh  */
137ba6c22ceSWarner Losh typedef int l_iop_t(struct iop_stats *, struct bio *);
138ba6c22ceSWarner Losh 
139ba6c22ceSWarner Losh /*
1400b4060b0SEd Maste  * Called when an I/O completes so the limiter can update its
141ba6c22ceSWarner Losh  * accounting. Pending I/Os may complete in any order (even when
142ba6c22ceSWarner Losh  * sent to the hardware at the same time), so the limiter may not
143ba6c22ceSWarner Losh  * make any assumptions other than this I/O has completed. If it
144ba6c22ceSWarner Losh  * returns 1, then xpt_schedule() needs to be called again.
145ba6c22ceSWarner Losh  */
146ba6c22ceSWarner Losh typedef int l_iodone_t(struct iop_stats *, struct bio *);
147ba6c22ceSWarner Losh 
148ba6c22ceSWarner Losh static l_iop_t cam_iosched_qd_iop;
149ba6c22ceSWarner Losh static l_iop_t cam_iosched_qd_caniop;
150ba6c22ceSWarner Losh static l_iodone_t cam_iosched_qd_iodone;
151ba6c22ceSWarner Losh 
152ba6c22ceSWarner Losh static l_init_t cam_iosched_iops_init;
153ba6c22ceSWarner Losh static l_tick_t cam_iosched_iops_tick;
154ba6c22ceSWarner Losh static l_iop_t cam_iosched_iops_caniop;
155ba6c22ceSWarner Losh static l_iop_t cam_iosched_iops_iop;
156ba6c22ceSWarner Losh 
157ba6c22ceSWarner Losh static l_init_t cam_iosched_bw_init;
158ba6c22ceSWarner Losh static l_tick_t cam_iosched_bw_tick;
159ba6c22ceSWarner Losh static l_iop_t cam_iosched_bw_caniop;
160ba6c22ceSWarner Losh static l_iop_t cam_iosched_bw_iop;
161ba6c22ceSWarner Losh 
162b20c0a07SWarner Losh struct limswitch {
163ba6c22ceSWarner Losh 	l_init_t	*l_init;
164ba6c22ceSWarner Losh 	l_tick_t	*l_tick;
165ba6c22ceSWarner Losh 	l_iop_t		*l_iop;
166ba6c22ceSWarner Losh 	l_iop_t		*l_caniop;
167ba6c22ceSWarner Losh 	l_iodone_t	*l_iodone;
168ba6c22ceSWarner Losh } limsw[] =
169ba6c22ceSWarner Losh {
170ba6c22ceSWarner Losh 	{	/* none */
171ba6c22ceSWarner Losh 		.l_init = NULL,
172ba6c22ceSWarner Losh 		.l_tick = NULL,
173ba6c22ceSWarner Losh 		.l_iop = NULL,
174ba6c22ceSWarner Losh 		.l_iodone= NULL,
175ba6c22ceSWarner Losh 	},
176ba6c22ceSWarner Losh 	{	/* queue_depth */
177ba6c22ceSWarner Losh 		.l_init = NULL,
178ba6c22ceSWarner Losh 		.l_tick = NULL,
179ba6c22ceSWarner Losh 		.l_caniop = cam_iosched_qd_caniop,
180ba6c22ceSWarner Losh 		.l_iop = cam_iosched_qd_iop,
181ba6c22ceSWarner Losh 		.l_iodone= cam_iosched_qd_iodone,
182ba6c22ceSWarner Losh 	},
183ba6c22ceSWarner Losh 	{	/* iops */
184ba6c22ceSWarner Losh 		.l_init = cam_iosched_iops_init,
185ba6c22ceSWarner Losh 		.l_tick = cam_iosched_iops_tick,
186ba6c22ceSWarner Losh 		.l_caniop = cam_iosched_iops_caniop,
187ba6c22ceSWarner Losh 		.l_iop = cam_iosched_iops_iop,
188ba6c22ceSWarner Losh 		.l_iodone= NULL,
189ba6c22ceSWarner Losh 	},
190ba6c22ceSWarner Losh 	{	/* bandwidth */
191ba6c22ceSWarner Losh 		.l_init = cam_iosched_bw_init,
192ba6c22ceSWarner Losh 		.l_tick = cam_iosched_bw_tick,
193ba6c22ceSWarner Losh 		.l_caniop = cam_iosched_bw_caniop,
194ba6c22ceSWarner Losh 		.l_iop = cam_iosched_bw_iop,
195ba6c22ceSWarner Losh 		.l_iodone= NULL,
196ba6c22ceSWarner Losh 	},
197ba6c22ceSWarner Losh };
198ba6c22ceSWarner Losh 
199b20c0a07SWarner Losh struct iop_stats {
200ba6c22ceSWarner Losh 	/*
201ba6c22ceSWarner Losh 	 * sysctl state for this subnode.
202ba6c22ceSWarner Losh 	 */
203ba6c22ceSWarner Losh 	struct sysctl_ctx_list	sysctl_ctx;
204ba6c22ceSWarner Losh 	struct sysctl_oid	*sysctl_tree;
205ba6c22ceSWarner Losh 
206ba6c22ceSWarner Losh 	/*
207ba6c22ceSWarner Losh 	 * Information about the current rate limiters, if any
208ba6c22ceSWarner Losh 	 */
209ba6c22ceSWarner Losh 	io_limiter	limiter;	/* How are I/Os being limited */
210ba6c22ceSWarner Losh 	int		min;		/* Low range of limit */
211ba6c22ceSWarner Losh 	int		max;		/* High range of limit */
212ba6c22ceSWarner Losh 	int		current;	/* Current rate limiter */
213ba6c22ceSWarner Losh 	int		l_value1;	/* per-limiter scratch value 1. */
214ba6c22ceSWarner Losh 	int		l_value2;	/* per-limiter scratch value 2. */
215ba6c22ceSWarner Losh 
216ba6c22ceSWarner Losh 	/*
217ba6c22ceSWarner Losh 	 * Debug information about counts of I/Os that have gone through the
218ba6c22ceSWarner Losh 	 * scheduler.
219ba6c22ceSWarner Losh 	 */
220ba6c22ceSWarner Losh 	int		pending;	/* I/Os pending in the hardware */
221ba6c22ceSWarner Losh 	int		queued;		/* number currently in the queue */
222ba6c22ceSWarner Losh 	int		total;		/* Total for all time -- wraps */
223ba6c22ceSWarner Losh 	int		in;		/* number queued all time -- wraps */
224ba6c22ceSWarner Losh 	int		out;		/* number completed all time -- wraps */
225c4b72d8bSWarner Losh 	int		errs;		/* Number of I/Os completed with error --  wraps */
226ba6c22ceSWarner Losh 
227ba6c22ceSWarner Losh 	/*
228ba6c22ceSWarner Losh 	 * Statistics on different bits of the process.
229ba6c22ceSWarner Losh 	 */
230cf3ec151SWarner Losh 		/* Exp Moving Average, see alpha_bits for more details */
231ba6c22ceSWarner Losh 	sbintime_t      ema;
23279d80af2SWarner Losh 	sbintime_t      emvar;
233ba6c22ceSWarner Losh 	sbintime_t      sd;		/* Last computed sd */
234ba6c22ceSWarner Losh 
235cf3ec151SWarner Losh 	uint32_t	state_flags;
236cf3ec151SWarner Losh #define IOP_RATE_LIMITED		1u
237cf3ec151SWarner Losh 
23808fc2f23SWarner Losh #define LAT_BUCKETS 15			/* < 1ms < 2ms ... < 2^(n-1)ms >= 2^(n-1)ms*/
239cf3ec151SWarner Losh 	uint64_t	latencies[LAT_BUCKETS];
240cf3ec151SWarner Losh 
241ba6c22ceSWarner Losh 	struct cam_iosched_softc *softc;
242ba6c22ceSWarner Losh };
243ba6c22ceSWarner Losh 
244ba6c22ceSWarner Losh 
245ba6c22ceSWarner Losh typedef enum {
246ba6c22ceSWarner Losh 	set_max = 0,			/* current = max */
247ba6c22ceSWarner Losh 	read_latency,			/* Steer read latency by throttling writes */
248ba6c22ceSWarner Losh 	cl_max				/* Keep last */
249ba6c22ceSWarner Losh } control_type;
250ba6c22ceSWarner Losh 
251ba6c22ceSWarner Losh static const char *cam_iosched_control_type_names[] =
252ba6c22ceSWarner Losh     { "set_max", "read_latency" };
253ba6c22ceSWarner Losh 
254b20c0a07SWarner Losh struct control_loop {
255ba6c22ceSWarner Losh 	/*
256ba6c22ceSWarner Losh 	 * sysctl state for this subnode.
257ba6c22ceSWarner Losh 	 */
258ba6c22ceSWarner Losh 	struct sysctl_ctx_list	sysctl_ctx;
259ba6c22ceSWarner Losh 	struct sysctl_oid	*sysctl_tree;
260ba6c22ceSWarner Losh 
261ba6c22ceSWarner Losh 	sbintime_t	next_steer;		/* Time of next steer */
262ba6c22ceSWarner Losh 	sbintime_t	steer_interval;		/* How often do we steer? */
263ba6c22ceSWarner Losh 	sbintime_t	lolat;
264ba6c22ceSWarner Losh 	sbintime_t	hilat;
265ba6c22ceSWarner Losh 	int		alpha;
266ba6c22ceSWarner Losh 	control_type	type;			/* What type of control? */
267ba6c22ceSWarner Losh 	int		last_count;		/* Last I/O count */
268ba6c22ceSWarner Losh 
269ba6c22ceSWarner Losh 	struct cam_iosched_softc *softc;
270ba6c22ceSWarner Losh };
271ba6c22ceSWarner Losh 
272ba6c22ceSWarner Losh #endif
273ba6c22ceSWarner Losh 
274b20c0a07SWarner Losh struct cam_iosched_softc {
275ba6c22ceSWarner Losh 	struct bio_queue_head bio_queue;
276ba6c22ceSWarner Losh 	struct bio_queue_head trim_queue;
277ba6c22ceSWarner Losh 				/* scheduler flags < 16, user flags >= 16 */
278ba6c22ceSWarner Losh 	uint32_t	flags;
279ba6c22ceSWarner Losh 	int		sort_io_queue;
280d900ade5SWarner Losh 	int		trim_goal;		/* # of trims to queue before sending */
281d900ade5SWarner Losh 	int		trim_ticks;		/* Max ticks to hold trims */
282d900ade5SWarner Losh 	int		last_trim_tick;		/* Last 'tick' time ld a trim */
283d900ade5SWarner Losh 	int		queued_trims;		/* Number of trims in the queue */
284df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
285ba6c22ceSWarner Losh 	int		read_bias;		/* Read bias setting */
286ba6c22ceSWarner Losh 	int		current_read_bias;	/* Current read bias state */
287ba6c22ceSWarner Losh 	int		total_ticks;
288cf3ec151SWarner Losh 	int		load;			/* EMA of 'load average' of disk / 2^16 */
289ba6c22ceSWarner Losh 
290ba6c22ceSWarner Losh 	struct bio_queue_head write_queue;
291ba6c22ceSWarner Losh 	struct iop_stats read_stats, write_stats, trim_stats;
292ba6c22ceSWarner Losh 	struct sysctl_ctx_list	sysctl_ctx;
293ba6c22ceSWarner Losh 	struct sysctl_oid	*sysctl_tree;
294ba6c22ceSWarner Losh 
295ba6c22ceSWarner Losh 	int		quanta;			/* Number of quanta per second */
296ba6c22ceSWarner Losh 	struct callout	ticker;			/* Callout for our quota system */
297ba6c22ceSWarner Losh 	struct cam_periph *periph;		/* cam periph associated with this device */
298ba6c22ceSWarner Losh 	uint32_t	this_frac;		/* Fraction of a second (1024ths) for this tick */
299ba6c22ceSWarner Losh 	sbintime_t	last_time;		/* Last time we ticked */
300ba6c22ceSWarner Losh 	struct control_loop cl;
301e5436ab5SWarner Losh 	sbintime_t	max_lat;		/* when != 0, if iop latency > max_lat, call max_lat_fcn */
302e5436ab5SWarner Losh 	cam_iosched_latfcn_t	latfcn;
303e5436ab5SWarner Losh 	void		*latarg;
304ba6c22ceSWarner Losh #endif
305ba6c22ceSWarner Losh };
306ba6c22ceSWarner Losh 
307df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
308ba6c22ceSWarner Losh /*
309ba6c22ceSWarner Losh  * helper functions to call the limsw functions.
310ba6c22ceSWarner Losh  */
311ba6c22ceSWarner Losh static int
312ba6c22ceSWarner Losh cam_iosched_limiter_init(struct iop_stats *ios)
313ba6c22ceSWarner Losh {
314ba6c22ceSWarner Losh 	int lim = ios->limiter;
315ba6c22ceSWarner Losh 
316ba6c22ceSWarner Losh 	/* maybe this should be a kassert */
317ba6c22ceSWarner Losh 	if (lim < none || lim >= limiter_max)
318ba6c22ceSWarner Losh 		return EINVAL;
319ba6c22ceSWarner Losh 
320ba6c22ceSWarner Losh 	if (limsw[lim].l_init)
321ba6c22ceSWarner Losh 		return limsw[lim].l_init(ios);
322ba6c22ceSWarner Losh 
323ba6c22ceSWarner Losh 	return 0;
324ba6c22ceSWarner Losh }
325ba6c22ceSWarner Losh 
326ba6c22ceSWarner Losh static int
327ba6c22ceSWarner Losh cam_iosched_limiter_tick(struct iop_stats *ios)
328ba6c22ceSWarner Losh {
329ba6c22ceSWarner Losh 	int lim = ios->limiter;
330ba6c22ceSWarner Losh 
331ba6c22ceSWarner Losh 	/* maybe this should be a kassert */
332ba6c22ceSWarner Losh 	if (lim < none || lim >= limiter_max)
333ba6c22ceSWarner Losh 		return EINVAL;
334ba6c22ceSWarner Losh 
335ba6c22ceSWarner Losh 	if (limsw[lim].l_tick)
336ba6c22ceSWarner Losh 		return limsw[lim].l_tick(ios);
337ba6c22ceSWarner Losh 
338ba6c22ceSWarner Losh 	return 0;
339ba6c22ceSWarner Losh }
340ba6c22ceSWarner Losh 
341ba6c22ceSWarner Losh static int
342ba6c22ceSWarner Losh cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
343ba6c22ceSWarner Losh {
344ba6c22ceSWarner Losh 	int lim = ios->limiter;
345ba6c22ceSWarner Losh 
346ba6c22ceSWarner Losh 	/* maybe this should be a kassert */
347ba6c22ceSWarner Losh 	if (lim < none || lim >= limiter_max)
348ba6c22ceSWarner Losh 		return EINVAL;
349ba6c22ceSWarner Losh 
350ba6c22ceSWarner Losh 	if (limsw[lim].l_iop)
351ba6c22ceSWarner Losh 		return limsw[lim].l_iop(ios, bp);
352ba6c22ceSWarner Losh 
353ba6c22ceSWarner Losh 	return 0;
354ba6c22ceSWarner Losh }
355ba6c22ceSWarner Losh 
356ba6c22ceSWarner Losh static int
357ba6c22ceSWarner Losh cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
358ba6c22ceSWarner Losh {
359ba6c22ceSWarner Losh 	int lim = ios->limiter;
360ba6c22ceSWarner Losh 
361ba6c22ceSWarner Losh 	/* maybe this should be a kassert */
362ba6c22ceSWarner Losh 	if (lim < none || lim >= limiter_max)
363ba6c22ceSWarner Losh 		return EINVAL;
364ba6c22ceSWarner Losh 
365ba6c22ceSWarner Losh 	if (limsw[lim].l_caniop)
366ba6c22ceSWarner Losh 		return limsw[lim].l_caniop(ios, bp);
367ba6c22ceSWarner Losh 
368ba6c22ceSWarner Losh 	return 0;
369ba6c22ceSWarner Losh }
370ba6c22ceSWarner Losh 
371ba6c22ceSWarner Losh static int
372ba6c22ceSWarner Losh cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
373ba6c22ceSWarner Losh {
374ba6c22ceSWarner Losh 	int lim = ios->limiter;
375ba6c22ceSWarner Losh 
376ba6c22ceSWarner Losh 	/* maybe this should be a kassert */
377ba6c22ceSWarner Losh 	if (lim < none || lim >= limiter_max)
378ba6c22ceSWarner Losh 		return 0;
379ba6c22ceSWarner Losh 
380ba6c22ceSWarner Losh 	if (limsw[lim].l_iodone)
381ba6c22ceSWarner Losh 		return limsw[lim].l_iodone(ios, bp);
382ba6c22ceSWarner Losh 
383ba6c22ceSWarner Losh 	return 0;
384ba6c22ceSWarner Losh }
385ba6c22ceSWarner Losh 
386ba6c22ceSWarner Losh /*
387ba6c22ceSWarner Losh  * Functions to implement the different kinds of limiters
388ba6c22ceSWarner Losh  */
389ba6c22ceSWarner Losh 
390ba6c22ceSWarner Losh static int
391ba6c22ceSWarner Losh cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
392ba6c22ceSWarner Losh {
393ba6c22ceSWarner Losh 
394ba6c22ceSWarner Losh 	if (ios->current <= 0 || ios->pending < ios->current)
395ba6c22ceSWarner Losh 		return 0;
396ba6c22ceSWarner Losh 
397ba6c22ceSWarner Losh 	return EAGAIN;
398ba6c22ceSWarner Losh }
399ba6c22ceSWarner Losh 
400ba6c22ceSWarner Losh static int
401ba6c22ceSWarner Losh cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
402ba6c22ceSWarner Losh {
403ba6c22ceSWarner Losh 
404ba6c22ceSWarner Losh 	if (ios->current <= 0 || ios->pending < ios->current)
405ba6c22ceSWarner Losh 		return 0;
406ba6c22ceSWarner Losh 
407ba6c22ceSWarner Losh 	return EAGAIN;
408ba6c22ceSWarner Losh }
409ba6c22ceSWarner Losh 
410ba6c22ceSWarner Losh static int
411ba6c22ceSWarner Losh cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
412ba6c22ceSWarner Losh {
413ba6c22ceSWarner Losh 
414ba6c22ceSWarner Losh 	if (ios->current <= 0 || ios->pending != ios->current)
415ba6c22ceSWarner Losh 		return 0;
416ba6c22ceSWarner Losh 
417ba6c22ceSWarner Losh 	return 1;
418ba6c22ceSWarner Losh }
419ba6c22ceSWarner Losh 
420ba6c22ceSWarner Losh static int
421ba6c22ceSWarner Losh cam_iosched_iops_init(struct iop_stats *ios)
422ba6c22ceSWarner Losh {
423ba6c22ceSWarner Losh 
424ba6c22ceSWarner Losh 	ios->l_value1 = ios->current / ios->softc->quanta;
425ba6c22ceSWarner Losh 	if (ios->l_value1 <= 0)
426ba6c22ceSWarner Losh 		ios->l_value1 = 1;
427f777123bSWarner Losh 	ios->l_value2 = 0;
428ba6c22ceSWarner Losh 
429ba6c22ceSWarner Losh 	return 0;
430ba6c22ceSWarner Losh }
431ba6c22ceSWarner Losh 
432ba6c22ceSWarner Losh static int
433ba6c22ceSWarner Losh cam_iosched_iops_tick(struct iop_stats *ios)
434ba6c22ceSWarner Losh {
43578ed811eSWarner Losh 	int new_ios;
436ba6c22ceSWarner Losh 
437f777123bSWarner Losh 	/*
438f777123bSWarner Losh 	 * Allow at least one IO per tick until all
439f777123bSWarner Losh 	 * the IOs for this interval have been spent.
440f777123bSWarner Losh 	 */
44178ed811eSWarner Losh 	new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
44278ed811eSWarner Losh 	if (new_ios < 1 && ios->l_value2 < ios->current) {
44378ed811eSWarner Losh 		new_ios = 1;
444f777123bSWarner Losh 		ios->l_value2++;
445f777123bSWarner Losh 	}
446ba6c22ceSWarner Losh 
44778ed811eSWarner Losh 	/*
44878ed811eSWarner Losh 	 * If this a new accounting interval, discard any "unspent" ios
44978ed811eSWarner Losh 	 * granted in the previous interval.  Otherwise add the new ios to
45078ed811eSWarner Losh 	 * the previously granted ones that haven't been spent yet.
45178ed811eSWarner Losh 	 */
45278ed811eSWarner Losh 	if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
45378ed811eSWarner Losh 		ios->l_value1 = new_ios;
45478ed811eSWarner Losh 		ios->l_value2 = 1;
45578ed811eSWarner Losh 	} else {
45678ed811eSWarner Losh 		ios->l_value1 += new_ios;
45778ed811eSWarner Losh 	}
45878ed811eSWarner Losh 
45978ed811eSWarner Losh 
460ba6c22ceSWarner Losh 	return 0;
461ba6c22ceSWarner Losh }
462ba6c22ceSWarner Losh 
463ba6c22ceSWarner Losh static int
464ba6c22ceSWarner Losh cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
465ba6c22ceSWarner Losh {
466ba6c22ceSWarner Losh 
467ba6c22ceSWarner Losh 	/*
468ba6c22ceSWarner Losh 	 * So if we have any more IOPs left, allow it,
4696ca2fb66SWarner Losh 	 * otherwise wait. If current iops is 0, treat that
4706ca2fb66SWarner Losh 	 * as unlimited as a failsafe.
471ba6c22ceSWarner Losh 	 */
4726ca2fb66SWarner Losh 	if (ios->current > 0 && ios->l_value1 <= 0)
473ba6c22ceSWarner Losh 		return EAGAIN;
474ba6c22ceSWarner Losh 	return 0;
475ba6c22ceSWarner Losh }
476ba6c22ceSWarner Losh 
477ba6c22ceSWarner Losh static int
478ba6c22ceSWarner Losh cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
479ba6c22ceSWarner Losh {
480ba6c22ceSWarner Losh 	int rv;
481ba6c22ceSWarner Losh 
482ba6c22ceSWarner Losh 	rv = cam_iosched_limiter_caniop(ios, bp);
483ba6c22ceSWarner Losh 	if (rv == 0)
484ba6c22ceSWarner Losh 		ios->l_value1--;
485ba6c22ceSWarner Losh 
486ba6c22ceSWarner Losh 	return rv;
487ba6c22ceSWarner Losh }
488ba6c22ceSWarner Losh 
489ba6c22ceSWarner Losh static int
490ba6c22ceSWarner Losh cam_iosched_bw_init(struct iop_stats *ios)
491ba6c22ceSWarner Losh {
492ba6c22ceSWarner Losh 
493ba6c22ceSWarner Losh 	/* ios->current is in kB/s, so scale to bytes */
494ba6c22ceSWarner Losh 	ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
495ba6c22ceSWarner Losh 
496ba6c22ceSWarner Losh 	return 0;
497ba6c22ceSWarner Losh }
498ba6c22ceSWarner Losh 
499ba6c22ceSWarner Losh static int
500ba6c22ceSWarner Losh cam_iosched_bw_tick(struct iop_stats *ios)
501ba6c22ceSWarner Losh {
502ba6c22ceSWarner Losh 	int bw;
503ba6c22ceSWarner Losh 
504ba6c22ceSWarner Losh 	/*
505ba6c22ceSWarner Losh 	 * If we're in the hole for available quota from
506ba6c22ceSWarner Losh 	 * the last time, then add the quantum for this.
507ba6c22ceSWarner Losh 	 * If we have any left over from last quantum,
508ba6c22ceSWarner Losh 	 * then too bad, that's lost. Also, ios->current
509ba6c22ceSWarner Losh 	 * is in kB/s, so scale.
510ba6c22ceSWarner Losh 	 *
511ba6c22ceSWarner Losh 	 * We also allow up to 4 quanta of credits to
512ba6c22ceSWarner Losh 	 * accumulate to deal with burstiness. 4 is extremely
513ba6c22ceSWarner Losh 	 * arbitrary.
514ba6c22ceSWarner Losh 	 */
515ba6c22ceSWarner Losh 	bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
516ba6c22ceSWarner Losh 	if (ios->l_value1 < bw * 4)
517ba6c22ceSWarner Losh 		ios->l_value1 += bw;
518ba6c22ceSWarner Losh 
519ba6c22ceSWarner Losh 	return 0;
520ba6c22ceSWarner Losh }
521ba6c22ceSWarner Losh 
522ba6c22ceSWarner Losh static int
523ba6c22ceSWarner Losh cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
524ba6c22ceSWarner Losh {
525ba6c22ceSWarner Losh 	/*
526ba6c22ceSWarner Losh 	 * So if we have any more bw quota left, allow it,
5270b4060b0SEd Maste 	 * otherwise wait. Note, we'll go negative and that's
5280b4060b0SEd Maste 	 * OK. We'll just get a little less next quota.
529ba6c22ceSWarner Losh 	 *
530ba6c22ceSWarner Losh 	 * Note on going negative: that allows us to process
531ba6c22ceSWarner Losh 	 * requests in order better, since we won't allow
532ba6c22ceSWarner Losh 	 * shorter reads to get around the long one that we
533ba6c22ceSWarner Losh 	 * don't have the quota to do just yet. It also prevents
534ba6c22ceSWarner Losh 	 * starvation by being a little more permissive about
535ba6c22ceSWarner Losh 	 * what we let through this quantum (to prevent the
536ba6c22ceSWarner Losh 	 * starvation), at the cost of getting a little less
537ba6c22ceSWarner Losh 	 * next quantum.
5386ca2fb66SWarner Losh 	 *
5396ca2fb66SWarner Losh 	 * Also note that if the current limit is <= 0,
5406ca2fb66SWarner Losh 	 * we treat it as unlimited as a failsafe.
541ba6c22ceSWarner Losh 	 */
5426ca2fb66SWarner Losh 	if (ios->current > 0 && ios->l_value1 <= 0)
543ba6c22ceSWarner Losh 		return EAGAIN;
544ba6c22ceSWarner Losh 
545ba6c22ceSWarner Losh 
546ba6c22ceSWarner Losh 	return 0;
547ba6c22ceSWarner Losh }
548ba6c22ceSWarner Losh 
549ba6c22ceSWarner Losh static int
550ba6c22ceSWarner Losh cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
551ba6c22ceSWarner Losh {
552ba6c22ceSWarner Losh 	int rv;
553ba6c22ceSWarner Losh 
554ba6c22ceSWarner Losh 	rv = cam_iosched_limiter_caniop(ios, bp);
555ba6c22ceSWarner Losh 	if (rv == 0)
556ba6c22ceSWarner Losh 		ios->l_value1 -= bp->bio_length;
557ba6c22ceSWarner Losh 
558ba6c22ceSWarner Losh 	return rv;
559ba6c22ceSWarner Losh }
560ba6c22ceSWarner Losh 
561ba6c22ceSWarner Losh static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
562ba6c22ceSWarner Losh 
563ba6c22ceSWarner Losh static void
564ba6c22ceSWarner Losh cam_iosched_ticker(void *arg)
565ba6c22ceSWarner Losh {
566ba6c22ceSWarner Losh 	struct cam_iosched_softc *isc = arg;
567ba6c22ceSWarner Losh 	sbintime_t now, delta;
568cf3ec151SWarner Losh 	int pending;
569ba6c22ceSWarner Losh 
5703028dd8dSWarner Losh 	callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
571ba6c22ceSWarner Losh 
572ba6c22ceSWarner Losh 	now = sbinuptime();
573ba6c22ceSWarner Losh 	delta = now - isc->last_time;
574ba6c22ceSWarner Losh 	isc->this_frac = (uint32_t)delta >> 16;		/* Note: discards seconds -- should be 0 harmless if not */
575ba6c22ceSWarner Losh 	isc->last_time = now;
576ba6c22ceSWarner Losh 
577ba6c22ceSWarner Losh 	cam_iosched_cl_maybe_steer(&isc->cl);
578ba6c22ceSWarner Losh 
579ba6c22ceSWarner Losh 	cam_iosched_limiter_tick(&isc->read_stats);
580ba6c22ceSWarner Losh 	cam_iosched_limiter_tick(&isc->write_stats);
581ba6c22ceSWarner Losh 	cam_iosched_limiter_tick(&isc->trim_stats);
582ba6c22ceSWarner Losh 
583ba6c22ceSWarner Losh 	cam_iosched_schedule(isc, isc->periph);
584ba6c22ceSWarner Losh 
585cf3ec151SWarner Losh 	/*
586cf3ec151SWarner Losh 	 * isc->load is an EMA of the pending I/Os at each tick. The number of
587cf3ec151SWarner Losh 	 * pending I/Os is the sum of the I/Os queued to the hardware, and those
588cf3ec151SWarner Losh 	 * in the software queue that could be queued to the hardware if there
589cf3ec151SWarner Losh 	 * were slots.
590cf3ec151SWarner Losh 	 *
591cf3ec151SWarner Losh 	 * ios_stats.pending is a count of requests in the SIM right now for
592cf3ec151SWarner Losh 	 * each of these types of I/O. So the total pending count is the sum of
593cf3ec151SWarner Losh 	 * these I/Os and the sum of the queued I/Os still in the software queue
594cf3ec151SWarner Losh 	 * for those operations that aren't being rate limited at the moment.
595cf3ec151SWarner Losh 	 *
596cf3ec151SWarner Losh 	 * The reason for the rate limiting bit is because those I/Os
597cf3ec151SWarner Losh 	 * aren't part of the software queued load (since we could
598cf3ec151SWarner Losh 	 * give them to hardware, but choose not to).
599cf3ec151SWarner Losh 	 *
600cf3ec151SWarner Losh 	 * Note: due to a bug in counting pending TRIM in the device, we
601cf3ec151SWarner Losh 	 * don't include them in this count. We count each BIO_DELETE in
602cf3ec151SWarner Losh 	 * the pending count, but the periph drivers collapse them down
603cf3ec151SWarner Losh 	 * into one TRIM command. That one trim command gets the completion
604cf3ec151SWarner Losh 	 * so the counts get off.
605cf3ec151SWarner Losh 	 */
606cf3ec151SWarner Losh 	pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
607cf3ec151SWarner Losh 	pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
608cf3ec151SWarner Losh 	    !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
609cf3ec151SWarner Losh 	    !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
610cf3ec151SWarner Losh 	pending <<= 16;
611cf3ec151SWarner Losh 	pending /= isc->periph->path->device->ccbq.total_openings;
612cf3ec151SWarner Losh 
613cf3ec151SWarner Losh 	isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
614cf3ec151SWarner Losh 
615ba6c22ceSWarner Losh 	isc->total_ticks++;
616ba6c22ceSWarner Losh }
617ba6c22ceSWarner Losh 
618ba6c22ceSWarner Losh 
619ba6c22ceSWarner Losh static void
620ba6c22ceSWarner Losh cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
621ba6c22ceSWarner Losh {
622ba6c22ceSWarner Losh 
623ba6c22ceSWarner Losh 	clp->next_steer = sbinuptime();
624ba6c22ceSWarner Losh 	clp->softc = isc;
625ba6c22ceSWarner Losh 	clp->steer_interval = SBT_1S * 5;	/* Let's start out steering every 5s */
626ba6c22ceSWarner Losh 	clp->lolat = 5 * SBT_1MS;
627ba6c22ceSWarner Losh 	clp->hilat = 15 * SBT_1MS;
628ba6c22ceSWarner Losh 	clp->alpha = 20;			/* Alpha == gain. 20 = .2 */
629ba6c22ceSWarner Losh 	clp->type = set_max;
630ba6c22ceSWarner Losh }
631ba6c22ceSWarner Losh 
632ba6c22ceSWarner Losh static void
633ba6c22ceSWarner Losh cam_iosched_cl_maybe_steer(struct control_loop *clp)
634ba6c22ceSWarner Losh {
635ba6c22ceSWarner Losh 	struct cam_iosched_softc *isc;
636ba6c22ceSWarner Losh 	sbintime_t now, lat;
637ba6c22ceSWarner Losh 	int old;
638ba6c22ceSWarner Losh 
639ba6c22ceSWarner Losh 	isc = clp->softc;
640ba6c22ceSWarner Losh 	now = isc->last_time;
641ba6c22ceSWarner Losh 	if (now < clp->next_steer)
642ba6c22ceSWarner Losh 		return;
643ba6c22ceSWarner Losh 
644ba6c22ceSWarner Losh 	clp->next_steer = now + clp->steer_interval;
645ba6c22ceSWarner Losh 	switch (clp->type) {
646ba6c22ceSWarner Losh 	case set_max:
647ba6c22ceSWarner Losh 		if (isc->write_stats.current != isc->write_stats.max)
648ba6c22ceSWarner Losh 			printf("Steering write from %d kBps to %d kBps\n",
649ba6c22ceSWarner Losh 			    isc->write_stats.current, isc->write_stats.max);
650ba6c22ceSWarner Losh 		isc->read_stats.current = isc->read_stats.max;
651ba6c22ceSWarner Losh 		isc->write_stats.current = isc->write_stats.max;
652ba6c22ceSWarner Losh 		isc->trim_stats.current = isc->trim_stats.max;
653ba6c22ceSWarner Losh 		break;
654ba6c22ceSWarner Losh 	case read_latency:
655ba6c22ceSWarner Losh 		old = isc->write_stats.current;
656ba6c22ceSWarner Losh 		lat = isc->read_stats.ema;
657ba6c22ceSWarner Losh 		/*
658ba6c22ceSWarner Losh 		 * Simple PLL-like engine. Since we're steering to a range for
659ba6c22ceSWarner Losh 		 * the SP (set point) that makes things a little more
660ba6c22ceSWarner Losh 		 * complicated. In addition, we're not directly controlling our
661ba6c22ceSWarner Losh 		 * PV (process variable), the read latency, but instead are
662ba6c22ceSWarner Losh 		 * manipulating the write bandwidth limit for our MV
663ba6c22ceSWarner Losh 		 * (manipulation variable), analysis of this code gets a bit
664ba6c22ceSWarner Losh 		 * messy. Also, the MV is a very noisy control surface for read
665ba6c22ceSWarner Losh 		 * latency since it is affected by many hidden processes inside
666ba6c22ceSWarner Losh 		 * the device which change how responsive read latency will be
667ba6c22ceSWarner Losh 		 * in reaction to changes in write bandwidth. Unlike the classic
668ba6c22ceSWarner Losh 		 * boiler control PLL. this may result in over-steering while
669ba6c22ceSWarner Losh 		 * the SSD takes its time to react to the new, lower load. This
670ba6c22ceSWarner Losh 		 * is why we use a relatively low alpha of between .1 and .25 to
671ba6c22ceSWarner Losh 		 * compensate for this effect. At .1, it takes ~22 steering
672ba6c22ceSWarner Losh 		 * intervals to back off by a factor of 10. At .2 it only takes
673ba6c22ceSWarner Losh 		 * ~10. At .25 it only takes ~8. However some preliminary data
674ba6c22ceSWarner Losh 		 * from the SSD drives suggests a reasponse time in 10's of
675ba6c22ceSWarner Losh 		 * seconds before latency drops regardless of the new write
6760b4060b0SEd Maste 		 * rate. Careful observation will be required to tune this
677ba6c22ceSWarner Losh 		 * effectively.
678ba6c22ceSWarner Losh 		 *
679ba6c22ceSWarner Losh 		 * Also, when there's no read traffic, we jack up the write
680ba6c22ceSWarner Losh 		 * limit too regardless of the last read latency.  10 is
681ba6c22ceSWarner Losh 		 * somewhat arbitrary.
682ba6c22ceSWarner Losh 		 */
683ba6c22ceSWarner Losh 		if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
684ba6c22ceSWarner Losh 			isc->write_stats.current = isc->write_stats.current *
685ba6c22ceSWarner Losh 			    (100 + clp->alpha) / 100;	/* Scale up */
686ba6c22ceSWarner Losh 		else if (lat > clp->hilat)
687ba6c22ceSWarner Losh 			isc->write_stats.current = isc->write_stats.current *
688ba6c22ceSWarner Losh 			    (100 - clp->alpha) / 100;	/* Scale down */
689ba6c22ceSWarner Losh 		clp->last_count = isc->read_stats.total;
690ba6c22ceSWarner Losh 
691ba6c22ceSWarner Losh 		/*
692ba6c22ceSWarner Losh 		 * Even if we don't steer, per se, enforce the min/max limits as
693ba6c22ceSWarner Losh 		 * those may have changed.
694ba6c22ceSWarner Losh 		 */
695ba6c22ceSWarner Losh 		if (isc->write_stats.current < isc->write_stats.min)
696ba6c22ceSWarner Losh 			isc->write_stats.current = isc->write_stats.min;
697ba6c22ceSWarner Losh 		if (isc->write_stats.current > isc->write_stats.max)
698ba6c22ceSWarner Losh 			isc->write_stats.current = isc->write_stats.max;
6992b5c19f1SWarner Losh 		if (old != isc->write_stats.current && 	iosched_debug)
700cf3ec151SWarner Losh 			printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
701ba6c22ceSWarner Losh 			    old, isc->write_stats.current,
7022b5c19f1SWarner Losh 			    (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
703ba6c22ceSWarner Losh 		break;
704ba6c22ceSWarner Losh 	case cl_max:
705ba6c22ceSWarner Losh 		break;
706ba6c22ceSWarner Losh 	}
707ba6c22ceSWarner Losh }
708ba6c22ceSWarner Losh #endif
709ba6c22ceSWarner Losh 
71007e5967aSWarner Losh /*
71107e5967aSWarner Losh  * Trim or similar currently pending completion. Should only be set for
71207e5967aSWarner Losh  * those drivers wishing only one Trim active at a time.
71307e5967aSWarner Losh  */
71407e5967aSWarner Losh #define CAM_IOSCHED_FLAG_TRIM_ACTIVE	(1ul << 0)
71507e5967aSWarner Losh 			/* Callout active, and needs to be torn down */
71607e5967aSWarner Losh #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
71707e5967aSWarner Losh 
71807e5967aSWarner Losh 			/* Periph drivers set these flags to indicate work */
71907e5967aSWarner Losh #define CAM_IOSCHED_FLAG_WORK_FLAGS	((0xffffu) << 16)
72007e5967aSWarner Losh 
721df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
722ba6c22ceSWarner Losh static void
723ba6c22ceSWarner Losh cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
724ba6c22ceSWarner Losh     sbintime_t sim_latency, int cmd, size_t size);
7255ede5b8cSWarner Losh #endif
726ba6c22ceSWarner Losh 
7272d87718fSWarner Losh static inline bool
728ba6c22ceSWarner Losh cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
729ba6c22ceSWarner Losh {
730ba6c22ceSWarner Losh 	return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
731ba6c22ceSWarner Losh }
732ba6c22ceSWarner Losh 
7332d87718fSWarner Losh static inline bool
734ba6c22ceSWarner Losh cam_iosched_has_io(struct cam_iosched_softc *isc)
735ba6c22ceSWarner Losh {
736df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
737035ec48eSWarner Losh 	if (do_dynamic_iosched) {
738ba6c22ceSWarner Losh 		struct bio *rbp = bioq_first(&isc->bio_queue);
739ba6c22ceSWarner Losh 		struct bio *wbp = bioq_first(&isc->write_queue);
7402d87718fSWarner Losh 		bool can_write = wbp != NULL &&
741ba6c22ceSWarner Losh 		    cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
7422d87718fSWarner Losh 		bool can_read = rbp != NULL &&
743ba6c22ceSWarner Losh 		    cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
744ba6c22ceSWarner Losh 		if (iosched_debug > 2) {
745ba6c22ceSWarner Losh 			printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
746ba6c22ceSWarner Losh 			printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
747ba6c22ceSWarner Losh 			printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
748ba6c22ceSWarner Losh 		}
749ba6c22ceSWarner Losh 		return can_read || can_write;
750ba6c22ceSWarner Losh 	}
751ba6c22ceSWarner Losh #endif
752ba6c22ceSWarner Losh 	return bioq_first(&isc->bio_queue) != NULL;
753ba6c22ceSWarner Losh }
754ba6c22ceSWarner Losh 
7552d87718fSWarner Losh static inline bool
756ba6c22ceSWarner Losh cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
757ba6c22ceSWarner Losh {
758c6171b44SWarner Losh 	struct bio *bp;
759c6171b44SWarner Losh 
760c6171b44SWarner Losh 	bp = bioq_first(&isc->trim_queue);
761c6171b44SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
762c6171b44SWarner Losh 	if (do_dynamic_iosched) {
763c6171b44SWarner Losh 		/*
764c6171b44SWarner Losh 		 * If we're limiting trims, then defer action on trims
765c6171b44SWarner Losh 		 * for a bit.
766c6171b44SWarner Losh 		 */
767c6171b44SWarner Losh 		if (bp == NULL || cam_iosched_limiter_caniop(&isc->trim_stats, bp) != 0)
768c6171b44SWarner Losh 			return false;
769c6171b44SWarner Losh 	}
770c6171b44SWarner Losh #endif
771d900ade5SWarner Losh 
772d900ade5SWarner Losh 	/*
773d900ade5SWarner Losh 	 * If we've set a trim_goal, then if we exceed that allow trims
774d900ade5SWarner Losh 	 * to be passed back to the driver. If we've also set a tick timeout
775d900ade5SWarner Losh 	 * allow trims back to the driver. Otherwise, don't allow trims yet.
776d900ade5SWarner Losh 	 */
777d900ade5SWarner Losh 	if (isc->trim_goal > 0) {
778d900ade5SWarner Losh 		if (isc->queued_trims >= isc->trim_goal)
779d900ade5SWarner Losh 			return true;
780d900ade5SWarner Losh 		if (isc->queued_trims > 0 &&
781d900ade5SWarner Losh 		    isc->trim_ticks > 0 &&
782d900ade5SWarner Losh 		    ticks - isc->last_trim_tick > isc->trim_ticks)
783d900ade5SWarner Losh 			return true;
784d900ade5SWarner Losh 		return false;
785d900ade5SWarner Losh 	}
786d900ade5SWarner Losh 
787c6171b44SWarner Losh 	/* NB: Should perhaps have a max trim active independent of I/O limiters */
788c6171b44SWarner Losh 	return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) && bp != NULL;
789ba6c22ceSWarner Losh }
790ba6c22ceSWarner Losh 
791ba6c22ceSWarner Losh #define cam_iosched_sort_queue(isc)	((isc)->sort_io_queue >= 0 ?	\
792ba6c22ceSWarner Losh     (isc)->sort_io_queue : cam_sort_io_queues)
793ba6c22ceSWarner Losh 
794ba6c22ceSWarner Losh 
7952d87718fSWarner Losh static inline bool
796ba6c22ceSWarner Losh cam_iosched_has_work(struct cam_iosched_softc *isc)
797ba6c22ceSWarner Losh {
798df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
799ba6c22ceSWarner Losh 	if (iosched_debug > 2)
800ba6c22ceSWarner Losh 		printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
801ba6c22ceSWarner Losh 		    cam_iosched_has_more_trim(isc),
802ba6c22ceSWarner Losh 		    cam_iosched_has_flagged_work(isc));
803ba6c22ceSWarner Losh #endif
804ba6c22ceSWarner Losh 
805ba6c22ceSWarner Losh 	return cam_iosched_has_io(isc) ||
806ba6c22ceSWarner Losh 		cam_iosched_has_more_trim(isc) ||
807ba6c22ceSWarner Losh 		cam_iosched_has_flagged_work(isc);
808ba6c22ceSWarner Losh }
809ba6c22ceSWarner Losh 
810df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
811ba6c22ceSWarner Losh static void
812ba6c22ceSWarner Losh cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
813ba6c22ceSWarner Losh {
814ba6c22ceSWarner Losh 
815ba6c22ceSWarner Losh 	ios->limiter = none;
816ba6c22ceSWarner Losh 	ios->in = 0;
81789d26636SWarner Losh 	ios->max = ios->current = 300000;
818ba6c22ceSWarner Losh 	ios->min = 1;
819ba6c22ceSWarner Losh 	ios->out = 0;
820c4b72d8bSWarner Losh 	ios->errs = 0;
821ba6c22ceSWarner Losh 	ios->pending = 0;
822ba6c22ceSWarner Losh 	ios->queued = 0;
823ba6c22ceSWarner Losh 	ios->total = 0;
824ba6c22ceSWarner Losh 	ios->ema = 0;
82579d80af2SWarner Losh 	ios->emvar = 0;
826ba6c22ceSWarner Losh 	ios->softc = isc;
82789d26636SWarner Losh 	cam_iosched_limiter_init(ios);
828ba6c22ceSWarner Losh }
829ba6c22ceSWarner Losh 
830ba6c22ceSWarner Losh static int
831ba6c22ceSWarner Losh cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
832ba6c22ceSWarner Losh {
833ba6c22ceSWarner Losh 	char buf[16];
834ba6c22ceSWarner Losh 	struct iop_stats *ios;
835ba6c22ceSWarner Losh 	struct cam_iosched_softc *isc;
836cf3ec151SWarner Losh 	int value, i, error;
837ba6c22ceSWarner Losh 	const char *p;
838ba6c22ceSWarner Losh 
839ba6c22ceSWarner Losh 	ios = arg1;
840ba6c22ceSWarner Losh 	isc = ios->softc;
841ba6c22ceSWarner Losh 	value = ios->limiter;
842ba6c22ceSWarner Losh 	if (value < none || value >= limiter_max)
843ba6c22ceSWarner Losh 		p = "UNKNOWN";
844ba6c22ceSWarner Losh 	else
845ba6c22ceSWarner Losh 		p = cam_iosched_limiter_names[value];
846ba6c22ceSWarner Losh 
847ba6c22ceSWarner Losh 	strlcpy(buf, p, sizeof(buf));
848ba6c22ceSWarner Losh 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
849ba6c22ceSWarner Losh 	if (error != 0 || req->newptr == NULL)
850ba6c22ceSWarner Losh 		return error;
851ba6c22ceSWarner Losh 
852ba6c22ceSWarner Losh 	cam_periph_lock(isc->periph);
853ba6c22ceSWarner Losh 
854ba6c22ceSWarner Losh 	for (i = none; i < limiter_max; i++) {
855ba6c22ceSWarner Losh 		if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
856ba6c22ceSWarner Losh 			continue;
857ba6c22ceSWarner Losh 		ios->limiter = i;
858ba6c22ceSWarner Losh 		error = cam_iosched_limiter_init(ios);
859ba6c22ceSWarner Losh 		if (error != 0) {
860ba6c22ceSWarner Losh 			ios->limiter = value;
861ba6c22ceSWarner Losh 			cam_periph_unlock(isc->periph);
862ba6c22ceSWarner Losh 			return error;
863ba6c22ceSWarner Losh 		}
864cf3ec151SWarner Losh 		/* Note: disk load averate requires ticker to be always running */
8653028dd8dSWarner Losh 		callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
866ba6c22ceSWarner Losh 		isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
867ba6c22ceSWarner Losh 
868ba6c22ceSWarner Losh 		cam_periph_unlock(isc->periph);
869ba6c22ceSWarner Losh 		return 0;
870ba6c22ceSWarner Losh 	}
871ba6c22ceSWarner Losh 
872ba6c22ceSWarner Losh 	cam_periph_unlock(isc->periph);
873ba6c22ceSWarner Losh 	return EINVAL;
874ba6c22ceSWarner Losh }
875ba6c22ceSWarner Losh 
876ba6c22ceSWarner Losh static int
877ba6c22ceSWarner Losh cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
878ba6c22ceSWarner Losh {
879ba6c22ceSWarner Losh 	char buf[16];
880ba6c22ceSWarner Losh 	struct control_loop *clp;
881ba6c22ceSWarner Losh 	struct cam_iosched_softc *isc;
882ba6c22ceSWarner Losh 	int value, i, error;
883ba6c22ceSWarner Losh 	const char *p;
884ba6c22ceSWarner Losh 
885ba6c22ceSWarner Losh 	clp = arg1;
886ba6c22ceSWarner Losh 	isc = clp->softc;
887ba6c22ceSWarner Losh 	value = clp->type;
888ba6c22ceSWarner Losh 	if (value < none || value >= cl_max)
889ba6c22ceSWarner Losh 		p = "UNKNOWN";
890ba6c22ceSWarner Losh 	else
891ba6c22ceSWarner Losh 		p = cam_iosched_control_type_names[value];
892ba6c22ceSWarner Losh 
893ba6c22ceSWarner Losh 	strlcpy(buf, p, sizeof(buf));
894ba6c22ceSWarner Losh 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
895ba6c22ceSWarner Losh 	if (error != 0 || req->newptr == NULL)
896ba6c22ceSWarner Losh 		return error;
897ba6c22ceSWarner Losh 
898ba6c22ceSWarner Losh 	for (i = set_max; i < cl_max; i++) {
899ba6c22ceSWarner Losh 		if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
900ba6c22ceSWarner Losh 			continue;
901ba6c22ceSWarner Losh 		cam_periph_lock(isc->periph);
902ba6c22ceSWarner Losh 		clp->type = i;
903ba6c22ceSWarner Losh 		cam_periph_unlock(isc->periph);
904ba6c22ceSWarner Losh 		return 0;
905ba6c22ceSWarner Losh 	}
906ba6c22ceSWarner Losh 
907ba6c22ceSWarner Losh 	return EINVAL;
908ba6c22ceSWarner Losh }
909ba6c22ceSWarner Losh 
910ba6c22ceSWarner Losh static int
911ba6c22ceSWarner Losh cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
912ba6c22ceSWarner Losh {
913ba6c22ceSWarner Losh 	char buf[16];
914ba6c22ceSWarner Losh 	sbintime_t value;
915ba6c22ceSWarner Losh 	int error;
916ba6c22ceSWarner Losh 	uint64_t us;
917ba6c22ceSWarner Losh 
918ba6c22ceSWarner Losh 	value = *(sbintime_t *)arg1;
919ba6c22ceSWarner Losh 	us = (uint64_t)value / SBT_1US;
920ba6c22ceSWarner Losh 	snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
921ba6c22ceSWarner Losh 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
922ba6c22ceSWarner Losh 	if (error != 0 || req->newptr == NULL)
923ba6c22ceSWarner Losh 		return error;
924ba6c22ceSWarner Losh 	us = strtoul(buf, NULL, 10);
925ba6c22ceSWarner Losh 	if (us == 0)
926ba6c22ceSWarner Losh 		return EINVAL;
927ba6c22ceSWarner Losh 	*(sbintime_t *)arg1 = us * SBT_1US;
928ba6c22ceSWarner Losh 	return 0;
929ba6c22ceSWarner Losh }
930ba6c22ceSWarner Losh 
931cf3ec151SWarner Losh static int
932cf3ec151SWarner Losh cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
933cf3ec151SWarner Losh {
934cf3ec151SWarner Losh 	int i, error;
935cf3ec151SWarner Losh 	struct sbuf sb;
936cf3ec151SWarner Losh 	uint64_t *latencies;
937cf3ec151SWarner Losh 
938cf3ec151SWarner Losh 	latencies = arg1;
939cf3ec151SWarner Losh 	sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
940cf3ec151SWarner Losh 
941cf3ec151SWarner Losh 	for (i = 0; i < LAT_BUCKETS - 1; i++)
942cf3ec151SWarner Losh 		sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
943cf3ec151SWarner Losh 	sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
944cf3ec151SWarner Losh 	error = sbuf_finish(&sb);
945cf3ec151SWarner Losh 	sbuf_delete(&sb);
946cf3ec151SWarner Losh 
947cf3ec151SWarner Losh 	return (error);
948cf3ec151SWarner Losh }
949cf3ec151SWarner Losh 
9502d22619aSWarner Losh static int
9512d22619aSWarner Losh cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
9522d22619aSWarner Losh {
9532d22619aSWarner Losh 	int *quanta;
9542d22619aSWarner Losh 	int error, value;
9552d22619aSWarner Losh 
9562d22619aSWarner Losh 	quanta = (unsigned *)arg1;
9572d22619aSWarner Losh 	value = *quanta;
9582d22619aSWarner Losh 
9592d22619aSWarner Losh 	error = sysctl_handle_int(oidp, (int *)&value, 0, req);
9602d22619aSWarner Losh 	if ((error != 0) || (req->newptr == NULL))
9612d22619aSWarner Losh 		return (error);
9622d22619aSWarner Losh 
9632d22619aSWarner Losh 	if (value < 1 || value > hz)
9642d22619aSWarner Losh 		return (EINVAL);
9652d22619aSWarner Losh 
9662d22619aSWarner Losh 	*quanta = value;
9672d22619aSWarner Losh 
9682d22619aSWarner Losh 	return (0);
9692d22619aSWarner Losh }
9702d22619aSWarner Losh 
971ba6c22ceSWarner Losh static void
972ba6c22ceSWarner Losh cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
973ba6c22ceSWarner Losh {
974ba6c22ceSWarner Losh 	struct sysctl_oid_list *n;
975ba6c22ceSWarner Losh 	struct sysctl_ctx_list *ctx;
976ba6c22ceSWarner Losh 
977ba6c22ceSWarner Losh 	ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
978ba6c22ceSWarner Losh 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
979ba6c22ceSWarner Losh 	    CTLFLAG_RD, 0, name);
980ba6c22ceSWarner Losh 	n = SYSCTL_CHILDREN(ios->sysctl_tree);
981ba6c22ceSWarner Losh 	ctx = &ios->sysctl_ctx;
982ba6c22ceSWarner Losh 
983ba6c22ceSWarner Losh 	SYSCTL_ADD_UQUAD(ctx, n,
984ba6c22ceSWarner Losh 	    OID_AUTO, "ema", CTLFLAG_RD,
985ba6c22ceSWarner Losh 	    &ios->ema,
986ba6c22ceSWarner Losh 	    "Fast Exponentially Weighted Moving Average");
987ba6c22ceSWarner Losh 	SYSCTL_ADD_UQUAD(ctx, n,
98879d80af2SWarner Losh 	    OID_AUTO, "emvar", CTLFLAG_RD,
98979d80af2SWarner Losh 	    &ios->emvar,
99079d80af2SWarner Losh 	    "Fast Exponentially Weighted Moving Variance");
991ba6c22ceSWarner Losh 
992ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
993ba6c22ceSWarner Losh 	    OID_AUTO, "pending", CTLFLAG_RD,
994ba6c22ceSWarner Losh 	    &ios->pending, 0,
995ba6c22ceSWarner Losh 	    "Instantaneous # of pending transactions");
996ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
997ba6c22ceSWarner Losh 	    OID_AUTO, "count", CTLFLAG_RD,
998ba6c22ceSWarner Losh 	    &ios->total, 0,
999ba6c22ceSWarner Losh 	    "# of transactions submitted to hardware");
1000ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1001ba6c22ceSWarner Losh 	    OID_AUTO, "queued", CTLFLAG_RD,
1002ba6c22ceSWarner Losh 	    &ios->queued, 0,
1003ba6c22ceSWarner Losh 	    "# of transactions in the queue");
1004ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1005ba6c22ceSWarner Losh 	    OID_AUTO, "in", CTLFLAG_RD,
1006ba6c22ceSWarner Losh 	    &ios->in, 0,
1007ba6c22ceSWarner Losh 	    "# of transactions queued to driver");
1008ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1009ba6c22ceSWarner Losh 	    OID_AUTO, "out", CTLFLAG_RD,
1010ba6c22ceSWarner Losh 	    &ios->out, 0,
1011c4b72d8bSWarner Losh 	    "# of transactions completed (including with error)");
1012c4b72d8bSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1013c4b72d8bSWarner Losh 	    OID_AUTO, "errs", CTLFLAG_RD,
1014c4b72d8bSWarner Losh 	    &ios->errs, 0,
1015c4b72d8bSWarner Losh 	    "# of transactions completed with an error");
1016ba6c22ceSWarner Losh 
1017ba6c22ceSWarner Losh 	SYSCTL_ADD_PROC(ctx, n,
1018ba6c22ceSWarner Losh 	    OID_AUTO, "limiter", CTLTYPE_STRING | CTLFLAG_RW,
1019ba6c22ceSWarner Losh 	    ios, 0, cam_iosched_limiter_sysctl, "A",
1020ba6c22ceSWarner Losh 	    "Current limiting type.");
1021ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1022ba6c22ceSWarner Losh 	    OID_AUTO, "min", CTLFLAG_RW,
1023ba6c22ceSWarner Losh 	    &ios->min, 0,
1024ba6c22ceSWarner Losh 	    "min resource");
1025ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1026ba6c22ceSWarner Losh 	    OID_AUTO, "max", CTLFLAG_RW,
1027ba6c22ceSWarner Losh 	    &ios->max, 0,
1028ba6c22ceSWarner Losh 	    "max resource");
1029ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1030ba6c22ceSWarner Losh 	    OID_AUTO, "current", CTLFLAG_RW,
1031ba6c22ceSWarner Losh 	    &ios->current, 0,
1032ba6c22ceSWarner Losh 	    "current resource");
1033ba6c22ceSWarner Losh 
1034cf3ec151SWarner Losh 	SYSCTL_ADD_PROC(ctx, n,
1035cf3ec151SWarner Losh 	    OID_AUTO, "latencies", CTLTYPE_STRING | CTLFLAG_RD,
1036cf3ec151SWarner Losh 	    &ios->latencies, 0,
1037cf3ec151SWarner Losh 	    cam_iosched_sysctl_latencies, "A",
1038cf3ec151SWarner Losh 	    "Array of power of 2 latency from 1ms to 1.024s");
1039ba6c22ceSWarner Losh }
1040ba6c22ceSWarner Losh 
1041ba6c22ceSWarner Losh static void
1042ba6c22ceSWarner Losh cam_iosched_iop_stats_fini(struct iop_stats *ios)
1043ba6c22ceSWarner Losh {
1044ba6c22ceSWarner Losh 	if (ios->sysctl_tree)
1045ba6c22ceSWarner Losh 		if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1046ba6c22ceSWarner Losh 			printf("can't remove iosched sysctl stats context\n");
1047ba6c22ceSWarner Losh }
1048ba6c22ceSWarner Losh 
1049ba6c22ceSWarner Losh static void
1050ba6c22ceSWarner Losh cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1051ba6c22ceSWarner Losh {
1052ba6c22ceSWarner Losh 	struct sysctl_oid_list *n;
1053ba6c22ceSWarner Losh 	struct sysctl_ctx_list *ctx;
1054ba6c22ceSWarner Losh 	struct control_loop *clp;
1055ba6c22ceSWarner Losh 
1056ba6c22ceSWarner Losh 	clp = &isc->cl;
1057ba6c22ceSWarner Losh 	clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1058ba6c22ceSWarner Losh 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1059ba6c22ceSWarner Losh 	    CTLFLAG_RD, 0, "Control loop info");
1060ba6c22ceSWarner Losh 	n = SYSCTL_CHILDREN(clp->sysctl_tree);
1061ba6c22ceSWarner Losh 	ctx = &clp->sysctl_ctx;
1062ba6c22ceSWarner Losh 
1063ba6c22ceSWarner Losh 	SYSCTL_ADD_PROC(ctx, n,
1064ba6c22ceSWarner Losh 	    OID_AUTO, "type", CTLTYPE_STRING | CTLFLAG_RW,
1065ba6c22ceSWarner Losh 	    clp, 0, cam_iosched_control_type_sysctl, "A",
1066ba6c22ceSWarner Losh 	    "Control loop algorithm");
1067ba6c22ceSWarner Losh 	SYSCTL_ADD_PROC(ctx, n,
1068ba6c22ceSWarner Losh 	    OID_AUTO, "steer_interval", CTLTYPE_STRING | CTLFLAG_RW,
1069ba6c22ceSWarner Losh 	    &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1070ba6c22ceSWarner Losh 	    "How often to steer (in us)");
1071ba6c22ceSWarner Losh 	SYSCTL_ADD_PROC(ctx, n,
1072ba6c22ceSWarner Losh 	    OID_AUTO, "lolat", CTLTYPE_STRING | CTLFLAG_RW,
1073ba6c22ceSWarner Losh 	    &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1074ba6c22ceSWarner Losh 	    "Low water mark for Latency (in us)");
1075ba6c22ceSWarner Losh 	SYSCTL_ADD_PROC(ctx, n,
1076ba6c22ceSWarner Losh 	    OID_AUTO, "hilat", CTLTYPE_STRING | CTLFLAG_RW,
1077ba6c22ceSWarner Losh 	    &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1078ba6c22ceSWarner Losh 	    "Hi water mark for Latency (in us)");
1079ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1080ba6c22ceSWarner Losh 	    OID_AUTO, "alpha", CTLFLAG_RW,
1081ba6c22ceSWarner Losh 	    &clp->alpha, 0,
1082ba6c22ceSWarner Losh 	    "Alpha for PLL (x100) aka gain");
1083ba6c22ceSWarner Losh }
1084ba6c22ceSWarner Losh 
1085ba6c22ceSWarner Losh static void
1086ba6c22ceSWarner Losh cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1087ba6c22ceSWarner Losh {
1088ba6c22ceSWarner Losh 	if (clp->sysctl_tree)
1089ba6c22ceSWarner Losh 		if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1090ba6c22ceSWarner Losh 			printf("can't remove iosched sysctl control loop context\n");
1091ba6c22ceSWarner Losh }
1092ba6c22ceSWarner Losh #endif
1093ba6c22ceSWarner Losh 
1094ba6c22ceSWarner Losh /*
1095ba6c22ceSWarner Losh  * Allocate the iosched structure. This also insulates callers from knowing
1096ba6c22ceSWarner Losh  * sizeof struct cam_iosched_softc.
1097ba6c22ceSWarner Losh  */
1098ba6c22ceSWarner Losh int
10990028abe6SWarner Losh cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1100ba6c22ceSWarner Losh {
1101ba6c22ceSWarner Losh 
1102ba6c22ceSWarner Losh 	*iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1103ba6c22ceSWarner Losh 	if (*iscp == NULL)
1104ba6c22ceSWarner Losh 		return ENOMEM;
1105df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1106ba6c22ceSWarner Losh 	if (iosched_debug)
1107ba6c22ceSWarner Losh 		printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1108ba6c22ceSWarner Losh #endif
1109ba6c22ceSWarner Losh 	(*iscp)->sort_io_queue = -1;
1110ba6c22ceSWarner Losh 	bioq_init(&(*iscp)->bio_queue);
1111ba6c22ceSWarner Losh 	bioq_init(&(*iscp)->trim_queue);
1112df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1113035ec48eSWarner Losh 	if (do_dynamic_iosched) {
1114ba6c22ceSWarner Losh 		bioq_init(&(*iscp)->write_queue);
1115ba6c22ceSWarner Losh 		(*iscp)->read_bias = 100;
1116ba6c22ceSWarner Losh 		(*iscp)->current_read_bias = 100;
1117d7fa1ab0SWarner Losh 		(*iscp)->quanta = min(hz, 200);
1118ba6c22ceSWarner Losh 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1119ba6c22ceSWarner Losh 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1120ba6c22ceSWarner Losh 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1121ba6c22ceSWarner Losh 		(*iscp)->trim_stats.max = 1;	/* Trims are special: one at a time for now */
1122ba6c22ceSWarner Losh 		(*iscp)->last_time = sbinuptime();
1123ba6c22ceSWarner Losh 		callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1124ba6c22ceSWarner Losh 		(*iscp)->periph = periph;
1125ba6c22ceSWarner Losh 		cam_iosched_cl_init(&(*iscp)->cl, *iscp);
11263028dd8dSWarner Losh 		callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1127ba6c22ceSWarner Losh 		(*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1128ba6c22ceSWarner Losh 	}
1129ba6c22ceSWarner Losh #endif
1130ba6c22ceSWarner Losh 
1131ba6c22ceSWarner Losh 	return 0;
1132ba6c22ceSWarner Losh }
1133ba6c22ceSWarner Losh 
1134ba6c22ceSWarner Losh /*
1135ba6c22ceSWarner Losh  * Reclaim all used resources. This assumes that other folks have
1136ba6c22ceSWarner Losh  * drained the requests in the hardware. Maybe an unwise assumption.
1137ba6c22ceSWarner Losh  */
1138ba6c22ceSWarner Losh void
1139ba6c22ceSWarner Losh cam_iosched_fini(struct cam_iosched_softc *isc)
1140ba6c22ceSWarner Losh {
1141ba6c22ceSWarner Losh 	if (isc) {
1142ba6c22ceSWarner Losh 		cam_iosched_flush(isc, NULL, ENXIO);
1143df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1144ba6c22ceSWarner Losh 		cam_iosched_iop_stats_fini(&isc->read_stats);
1145ba6c22ceSWarner Losh 		cam_iosched_iop_stats_fini(&isc->write_stats);
1146ba6c22ceSWarner Losh 		cam_iosched_iop_stats_fini(&isc->trim_stats);
1147ba6c22ceSWarner Losh 		cam_iosched_cl_sysctl_fini(&isc->cl);
1148ba6c22ceSWarner Losh 		if (isc->sysctl_tree)
1149ba6c22ceSWarner Losh 			if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1150ba6c22ceSWarner Losh 				printf("can't remove iosched sysctl stats context\n");
1151ba6c22ceSWarner Losh 		if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1152ba6c22ceSWarner Losh 			callout_drain(&isc->ticker);
1153ba6c22ceSWarner Losh 			isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1154ba6c22ceSWarner Losh 		}
1155ba6c22ceSWarner Losh #endif
1156ba6c22ceSWarner Losh 		free(isc, M_CAMSCHED);
1157ba6c22ceSWarner Losh 	}
1158ba6c22ceSWarner Losh }
1159ba6c22ceSWarner Losh 
1160ba6c22ceSWarner Losh /*
1161ba6c22ceSWarner Losh  * After we're sure we're attaching a device, go ahead and add
1162ba6c22ceSWarner Losh  * hooks for any sysctl we may wish to honor.
1163ba6c22ceSWarner Losh  */
1164ba6c22ceSWarner Losh void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1165ba6c22ceSWarner Losh     struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1166ba6c22ceSWarner Losh {
1167ba6c22ceSWarner Losh 	struct sysctl_oid_list *n;
1168ba6c22ceSWarner Losh 
1169d900ade5SWarner Losh 	n = SYSCTL_CHILDREN(node);
1170d900ade5SWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1171ba6c22ceSWarner Losh 		OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1172ba6c22ceSWarner Losh 		&isc->sort_io_queue, 0,
1173ba6c22ceSWarner Losh 		"Sort IO queue to try and optimise disk access patterns");
1174d900ade5SWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1175d900ade5SWarner Losh 	    OID_AUTO, "trim_goal", CTLFLAG_RW,
1176d900ade5SWarner Losh 	    &isc->trim_goal, 0,
1177d900ade5SWarner Losh 	    "Number of trims to try to accumulate before sending to hardware");
1178d900ade5SWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1179d900ade5SWarner Losh 	    OID_AUTO, "trim_ticks", CTLFLAG_RW,
1180d900ade5SWarner Losh 	    &isc->trim_goal, 0,
1181d900ade5SWarner Losh 	    "IO Schedul qaunta to hold back trims for when accumulating");
1182ba6c22ceSWarner Losh 
1183df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1184035ec48eSWarner Losh 	if (!do_dynamic_iosched)
1185ba6c22ceSWarner Losh 		return;
1186ba6c22ceSWarner Losh 
1187ba6c22ceSWarner Losh 	isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1188ba6c22ceSWarner Losh 	    SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1189ba6c22ceSWarner Losh 	    CTLFLAG_RD, 0, "I/O scheduler statistics");
1190ba6c22ceSWarner Losh 	n = SYSCTL_CHILDREN(isc->sysctl_tree);
1191ba6c22ceSWarner Losh 	ctx = &isc->sysctl_ctx;
1192ba6c22ceSWarner Losh 
1193ba6c22ceSWarner Losh 	cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1194ba6c22ceSWarner Losh 	cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1195ba6c22ceSWarner Losh 	cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1196ba6c22ceSWarner Losh 	cam_iosched_cl_sysctl_init(isc);
1197ba6c22ceSWarner Losh 
1198ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1199ba6c22ceSWarner Losh 	    OID_AUTO, "read_bias", CTLFLAG_RW,
1200ba6c22ceSWarner Losh 	    &isc->read_bias, 100,
1201ba6c22ceSWarner Losh 	    "How biased towards read should we be independent of limits");
1202ba6c22ceSWarner Losh 
12032d22619aSWarner Losh 	SYSCTL_ADD_PROC(ctx, n,
12042d22619aSWarner Losh 	    OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW,
12052d22619aSWarner Losh 	    &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1206ba6c22ceSWarner Losh 	    "How many quanta per second do we slice the I/O up into");
1207ba6c22ceSWarner Losh 
1208ba6c22ceSWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1209ba6c22ceSWarner Losh 	    OID_AUTO, "total_ticks", CTLFLAG_RD,
1210ba6c22ceSWarner Losh 	    &isc->total_ticks, 0,
1211ba6c22ceSWarner Losh 	    "Total number of ticks we've done");
1212cf3ec151SWarner Losh 
1213cf3ec151SWarner Losh 	SYSCTL_ADD_INT(ctx, n,
1214cf3ec151SWarner Losh 	    OID_AUTO, "load", CTLFLAG_RD,
1215cf3ec151SWarner Losh 	    &isc->load, 0,
1216cf3ec151SWarner Losh 	    "scaled load average / 100");
1217e5436ab5SWarner Losh 
1218e5436ab5SWarner Losh 	SYSCTL_ADD_U64(ctx, n,
1219e5436ab5SWarner Losh 	    OID_AUTO, "latency_trigger", CTLFLAG_RW,
1220e5436ab5SWarner Losh 	    &isc->max_lat, 0,
1221e5436ab5SWarner Losh 	    "Latency treshold to trigger callbacks");
1222e5436ab5SWarner Losh #endif
1223e5436ab5SWarner Losh }
1224e5436ab5SWarner Losh 
1225e5436ab5SWarner Losh void
1226e5436ab5SWarner Losh cam_iosched_set_latfcn(struct cam_iosched_softc *isc,
1227e5436ab5SWarner Losh     cam_iosched_latfcn_t fnp, void *argp)
1228e5436ab5SWarner Losh {
1229e5436ab5SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1230e5436ab5SWarner Losh 	isc->latfcn = fnp;
1231e5436ab5SWarner Losh 	isc->latarg = argp;
1232ba6c22ceSWarner Losh #endif
1233ba6c22ceSWarner Losh }
1234ba6c22ceSWarner Losh 
1235ba6c22ceSWarner Losh /*
1236d900ade5SWarner Losh  * Client drivers can set two parameters. "goal" is the number of BIO_DELETEs
1237d900ade5SWarner Losh  * that will be queued up before iosched will "release" the trims to the client
1238d900ade5SWarner Losh  * driver to wo with what they will (usually combine as many as possible). If we
1239d900ade5SWarner Losh  * don't get this many, after trim_ticks we'll submit the I/O anyway with
1240d900ade5SWarner Losh  * whatever we have.  We do need an I/O of some kind of to clock the deferred
1241d900ade5SWarner Losh  * trims out to disk. Since we will eventually get a write for the super block
1242d900ade5SWarner Losh  * or something before we shutdown, the trims will complete. To be safe, when a
1243d900ade5SWarner Losh  * BIO_FLUSH is presented to the iosched work queue, we set the ticks time far
1244d900ade5SWarner Losh  * enough in the past so we'll present the BIO_DELETEs to the client driver.
1245d900ade5SWarner Losh  * There might be a race if no BIO_DELETESs were queued, a BIO_FLUSH comes in
1246d900ade5SWarner Losh  * and then a BIO_DELETE is sent down. No know client does this, and there's
1247d900ade5SWarner Losh  * already a race between an ordered BIO_FLUSH and any BIO_DELETEs in flight,
1248d900ade5SWarner Losh  * but no client depends on the ordering being honored.
1249d900ade5SWarner Losh  *
1250d900ade5SWarner Losh  * XXX I'm not sure what the interaction between UFS direct BIOs and the BUF
1251d900ade5SWarner Losh  * flushing on shutdown. I think there's bufs that would be dependent on the BIO
1252d900ade5SWarner Losh  * finishing to write out at least metadata, so we'll be fine. To be safe, keep
1253d900ade5SWarner Losh  * the number of ticks low (less than maybe 10s) to avoid shutdown races.
1254d900ade5SWarner Losh  */
1255d900ade5SWarner Losh 
1256d900ade5SWarner Losh void
1257d900ade5SWarner Losh cam_iosched_set_trim_goal(struct cam_iosched_softc *isc, int goal)
1258d900ade5SWarner Losh {
1259d900ade5SWarner Losh 
1260d900ade5SWarner Losh 	isc->trim_goal = goal;
1261d900ade5SWarner Losh }
1262d900ade5SWarner Losh 
1263d900ade5SWarner Losh void
1264d900ade5SWarner Losh cam_iosched_set_trim_ticks(struct cam_iosched_softc *isc, int trim_ticks)
1265d900ade5SWarner Losh {
1266d900ade5SWarner Losh 
1267d900ade5SWarner Losh 	isc->trim_ticks = trim_ticks;
1268d900ade5SWarner Losh }
1269d900ade5SWarner Losh 
1270d900ade5SWarner Losh /*
1271ba6c22ceSWarner Losh  * Flush outstanding I/O. Consumers of this library don't know all the
1272ba6c22ceSWarner Losh  * queues we may keep, so this allows all I/O to be flushed in one
1273ba6c22ceSWarner Losh  * convenient call.
1274ba6c22ceSWarner Losh  */
1275ba6c22ceSWarner Losh void
1276ba6c22ceSWarner Losh cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1277ba6c22ceSWarner Losh {
1278ba6c22ceSWarner Losh 	bioq_flush(&isc->bio_queue, stp, err);
1279ba6c22ceSWarner Losh 	bioq_flush(&isc->trim_queue, stp, err);
1280df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1281035ec48eSWarner Losh 	if (do_dynamic_iosched)
1282ba6c22ceSWarner Losh 		bioq_flush(&isc->write_queue, stp, err);
1283ba6c22ceSWarner Losh #endif
1284ba6c22ceSWarner Losh }
1285ba6c22ceSWarner Losh 
1286df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1287ba6c22ceSWarner Losh static struct bio *
12880028abe6SWarner Losh cam_iosched_get_write(struct cam_iosched_softc *isc)
1289ba6c22ceSWarner Losh {
1290ba6c22ceSWarner Losh 	struct bio *bp;
1291ba6c22ceSWarner Losh 
1292ba6c22ceSWarner Losh 	/*
1293ba6c22ceSWarner Losh 	 * We control the write rate by controlling how many requests we send
1294ba6c22ceSWarner Losh 	 * down to the drive at any one time. Fewer requests limits the
1295ba6c22ceSWarner Losh 	 * effects of both starvation when the requests take a while and write
1296ba6c22ceSWarner Losh 	 * amplification when each request is causing more than one write to
1297ba6c22ceSWarner Losh 	 * the NAND media. Limiting the queue depth like this will also limit
1298ba6c22ceSWarner Losh 	 * the write throughput and give and reads that want to compete to
1299ba6c22ceSWarner Losh 	 * compete unfairly.
1300ba6c22ceSWarner Losh 	 */
1301ba6c22ceSWarner Losh 	bp = bioq_first(&isc->write_queue);
1302ba6c22ceSWarner Losh 	if (bp == NULL) {
1303ba6c22ceSWarner Losh 		if (iosched_debug > 3)
1304ba6c22ceSWarner Losh 			printf("No writes present in write_queue\n");
1305ba6c22ceSWarner Losh 		return NULL;
1306ba6c22ceSWarner Losh 	}
1307ba6c22ceSWarner Losh 
1308ba6c22ceSWarner Losh 	/*
1309ba6c22ceSWarner Losh 	 * If pending read, prefer that based on current read bias
1310ba6c22ceSWarner Losh 	 * setting.
1311ba6c22ceSWarner Losh 	 */
1312ba6c22ceSWarner Losh 	if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1313ba6c22ceSWarner Losh 		if (iosched_debug)
1314f2b98850SWarner Losh 			printf(
1315f2b98850SWarner Losh 			    "Reads present and current_read_bias is %d queued "
1316f2b98850SWarner Losh 			    "writes %d queued reads %d\n",
1317f2b98850SWarner Losh 			    isc->current_read_bias, isc->write_stats.queued,
1318f2b98850SWarner Losh 			    isc->read_stats.queued);
1319ba6c22ceSWarner Losh 		isc->current_read_bias--;
1320cf3ec151SWarner Losh 		/* We're not limiting writes, per se, just doing reads first */
1321ba6c22ceSWarner Losh 		return NULL;
1322ba6c22ceSWarner Losh 	}
1323ba6c22ceSWarner Losh 
1324ba6c22ceSWarner Losh 	/*
1325ba6c22ceSWarner Losh 	 * See if our current limiter allows this I/O.
1326ba6c22ceSWarner Losh 	 */
1327ba6c22ceSWarner Losh 	if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1328ba6c22ceSWarner Losh 		if (iosched_debug)
1329ba6c22ceSWarner Losh 			printf("Can't write because limiter says no.\n");
1330cf3ec151SWarner Losh 		isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1331ba6c22ceSWarner Losh 		return NULL;
1332ba6c22ceSWarner Losh 	}
1333ba6c22ceSWarner Losh 
1334ba6c22ceSWarner Losh 	/*
1335ba6c22ceSWarner Losh 	 * Let's do this: We've passed all the gates and we're a go
1336ba6c22ceSWarner Losh 	 * to schedule the I/O in the SIM.
1337ba6c22ceSWarner Losh 	 */
1338ba6c22ceSWarner Losh 	isc->current_read_bias = isc->read_bias;
1339ba6c22ceSWarner Losh 	bioq_remove(&isc->write_queue, bp);
1340ba6c22ceSWarner Losh 	if (bp->bio_cmd == BIO_WRITE) {
1341ba6c22ceSWarner Losh 		isc->write_stats.queued--;
1342ba6c22ceSWarner Losh 		isc->write_stats.total++;
1343ba6c22ceSWarner Losh 		isc->write_stats.pending++;
1344ba6c22ceSWarner Losh 	}
1345ba6c22ceSWarner Losh 	if (iosched_debug > 9)
1346ba6c22ceSWarner Losh 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1347cf3ec151SWarner Losh 	isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1348ba6c22ceSWarner Losh 	return bp;
1349ba6c22ceSWarner Losh }
1350ba6c22ceSWarner Losh #endif
1351ba6c22ceSWarner Losh 
1352ba6c22ceSWarner Losh /*
1353ba6c22ceSWarner Losh  * Put back a trim that you weren't able to actually schedule this time.
1354ba6c22ceSWarner Losh  */
1355ba6c22ceSWarner Losh void
1356ba6c22ceSWarner Losh cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1357ba6c22ceSWarner Losh {
1358ba6c22ceSWarner Losh 	bioq_insert_head(&isc->trim_queue, bp);
1359d900ade5SWarner Losh 	if (isc->queued_trims == 0)
1360d900ade5SWarner Losh 		isc->last_trim_tick = ticks;
1361d900ade5SWarner Losh 	isc->queued_trims++;
1362df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1363ba6c22ceSWarner Losh 	isc->trim_stats.queued++;
1364ba6c22ceSWarner Losh 	isc->trim_stats.total--;		/* since we put it back, don't double count */
1365ba6c22ceSWarner Losh 	isc->trim_stats.pending--;
1366ba6c22ceSWarner Losh #endif
1367ba6c22ceSWarner Losh }
1368ba6c22ceSWarner Losh 
1369ba6c22ceSWarner Losh /*
1370ba6c22ceSWarner Losh  * gets the next trim from the trim queue.
1371ba6c22ceSWarner Losh  *
1372ba6c22ceSWarner Losh  * Assumes we're called with the periph lock held.  It removes this
13730b4060b0SEd Maste  * trim from the queue and the device must explicitly reinsert it
1374ba6c22ceSWarner Losh  * should the need arise.
1375ba6c22ceSWarner Losh  */
1376ba6c22ceSWarner Losh struct bio *
1377ba6c22ceSWarner Losh cam_iosched_next_trim(struct cam_iosched_softc *isc)
1378ba6c22ceSWarner Losh {
1379ba6c22ceSWarner Losh 	struct bio *bp;
1380ba6c22ceSWarner Losh 
1381ba6c22ceSWarner Losh 	bp  = bioq_first(&isc->trim_queue);
1382ba6c22ceSWarner Losh 	if (bp == NULL)
1383ba6c22ceSWarner Losh 		return NULL;
1384ba6c22ceSWarner Losh 	bioq_remove(&isc->trim_queue, bp);
1385d900ade5SWarner Losh 	isc->queued_trims--;
1386d900ade5SWarner Losh 	isc->last_trim_tick = ticks;	/* Reset the tick timer when we take trims */
1387df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1388ba6c22ceSWarner Losh 	isc->trim_stats.queued--;
1389ba6c22ceSWarner Losh 	isc->trim_stats.total++;
1390ba6c22ceSWarner Losh 	isc->trim_stats.pending++;
1391ba6c22ceSWarner Losh #endif
1392ba6c22ceSWarner Losh 	return bp;
1393ba6c22ceSWarner Losh }
1394ba6c22ceSWarner Losh 
1395ba6c22ceSWarner Losh /*
13960b4060b0SEd Maste  * gets an available trim from the trim queue, if there's no trim
1397ba6c22ceSWarner Losh  * already pending. It removes this trim from the queue and the device
13980b4060b0SEd Maste  * must explicitly reinsert it should the need arise.
1399ba6c22ceSWarner Losh  *
1400ba6c22ceSWarner Losh  * Assumes we're called with the periph lock held.
1401ba6c22ceSWarner Losh  */
14020028abe6SWarner Losh struct bio *
14030028abe6SWarner Losh cam_iosched_get_trim(struct cam_iosched_softc *isc)
1404ba6c22ceSWarner Losh {
1405c6171b44SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1406c6171b44SWarner Losh 	struct bio *bp;
1407c6171b44SWarner Losh #endif
1408ba6c22ceSWarner Losh 
14090028abe6SWarner Losh 	if (!cam_iosched_has_more_trim(isc))
1410ba6c22ceSWarner Losh 		return NULL;
141162c94a05SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1412c6171b44SWarner Losh 	bp  = bioq_first(&isc->trim_queue);
1413c6171b44SWarner Losh 	if (bp == NULL)
1414c6171b44SWarner Losh 		return NULL;
1415c6171b44SWarner Losh 
141662c94a05SWarner Losh 	/*
14171759fd77SWarner Losh 	 * If pending read, prefer that based on current read bias setting. The
14181759fd77SWarner Losh 	 * read bias is shared for both writes and TRIMs, but on TRIMs the bias
14191759fd77SWarner Losh 	 * is for a combined TRIM not a single TRIM request that's come in.
142062c94a05SWarner Losh 	 */
14211759fd77SWarner Losh 	if (do_dynamic_iosched) {
142262c94a05SWarner Losh 		if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
14231759fd77SWarner Losh 			if (iosched_debug)
14241759fd77SWarner Losh 				printf("Reads present and current_read_bias is %d"
14251759fd77SWarner Losh 				    " queued trims %d queued reads %d\n",
14261759fd77SWarner Losh 				    isc->current_read_bias, isc->trim_stats.queued,
14271759fd77SWarner Losh 				    isc->read_stats.queued);
142862c94a05SWarner Losh 			isc->current_read_bias--;
142962c94a05SWarner Losh 			/* We're not limiting TRIMS, per se, just doing reads first */
143062c94a05SWarner Losh 			return NULL;
143162c94a05SWarner Losh 		}
143262c94a05SWarner Losh 		/*
143362c94a05SWarner Losh 		 * We're going to do a trim, so reset the bias.
143462c94a05SWarner Losh 		 */
143562c94a05SWarner Losh 		isc->current_read_bias = isc->read_bias;
143662c94a05SWarner Losh 	}
1437c6171b44SWarner Losh 
1438c6171b44SWarner Losh 	/*
1439c6171b44SWarner Losh 	 * See if our current limiter allows this I/O. Because we only call this
1440c6171b44SWarner Losh 	 * here, and not in next_trim, the 'bandwidth' limits for trims won't
1441c6171b44SWarner Losh 	 * work, while the iops or max queued limits will work. It's tricky
1442c6171b44SWarner Losh 	 * because we want the limits to be from the perspective of the
1443c6171b44SWarner Losh 	 * "commands sent to the device." To make iops work, we need to check
1444c6171b44SWarner Losh 	 * only here (since we want all the ops we combine to count as one). To
1445c6171b44SWarner Losh 	 * make bw limits work, we'd need to check in next_trim, but that would
1446c6171b44SWarner Losh 	 * have the effect of limiting the iops as seen from the upper layers.
1447c6171b44SWarner Losh 	 */
1448c6171b44SWarner Losh 	if (cam_iosched_limiter_iop(&isc->trim_stats, bp) != 0) {
1449c6171b44SWarner Losh 		if (iosched_debug)
1450c6171b44SWarner Losh 			printf("Can't trim because limiter says no.\n");
1451c6171b44SWarner Losh 		isc->trim_stats.state_flags |= IOP_RATE_LIMITED;
1452c6171b44SWarner Losh 		return NULL;
1453c6171b44SWarner Losh 	}
1454c6171b44SWarner Losh 	isc->current_read_bias = isc->read_bias;
1455c6171b44SWarner Losh 	isc->trim_stats.state_flags &= ~IOP_RATE_LIMITED;
1456c6171b44SWarner Losh 	/* cam_iosched_next_trim below keeps proper book */
145762c94a05SWarner Losh #endif
1458ba6c22ceSWarner Losh 	return cam_iosched_next_trim(isc);
1459ba6c22ceSWarner Losh }
1460ba6c22ceSWarner Losh 
1461ba6c22ceSWarner Losh /*
1462ba6c22ceSWarner Losh  * Determine what the next bit of work to do is for the periph. The
1463ba6c22ceSWarner Losh  * default implementation looks to see if we have trims to do, but no
1464ba6c22ceSWarner Losh  * trims outstanding. If so, we do that. Otherwise we see if we have
1465ba6c22ceSWarner Losh  * other work. If we do, then we do that. Otherwise why were we called?
1466ba6c22ceSWarner Losh  */
1467ba6c22ceSWarner Losh struct bio *
1468ba6c22ceSWarner Losh cam_iosched_next_bio(struct cam_iosched_softc *isc)
1469ba6c22ceSWarner Losh {
1470ba6c22ceSWarner Losh 	struct bio *bp;
1471ba6c22ceSWarner Losh 
1472ba6c22ceSWarner Losh 	/*
1473ba6c22ceSWarner Losh 	 * See if we have a trim that can be scheduled. We can only send one
14740028abe6SWarner Losh 	 * at a time down, so this takes that into account.
14750028abe6SWarner Losh 	 *
14760028abe6SWarner Losh 	 * XXX newer TRIM commands are queueable. Revisit this when we
14770028abe6SWarner Losh 	 * implement them.
1478ba6c22ceSWarner Losh 	 */
14790028abe6SWarner Losh 	if ((bp = cam_iosched_get_trim(isc)) != NULL)
1480ba6c22ceSWarner Losh 		return bp;
1481ba6c22ceSWarner Losh 
1482df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1483ba6c22ceSWarner Losh 	/*
1484ba6c22ceSWarner Losh 	 * See if we have any pending writes, and room in the queue for them,
1485ba6c22ceSWarner Losh 	 * and if so, those are next.
1486ba6c22ceSWarner Losh 	 */
1487035ec48eSWarner Losh 	if (do_dynamic_iosched) {
14880028abe6SWarner Losh 		if ((bp = cam_iosched_get_write(isc)) != NULL)
1489ba6c22ceSWarner Losh 			return bp;
1490ba6c22ceSWarner Losh 	}
1491ba6c22ceSWarner Losh #endif
1492ba6c22ceSWarner Losh 
1493ba6c22ceSWarner Losh 	/*
1494ba6c22ceSWarner Losh 	 * next, see if there's other, normal I/O waiting. If so return that.
1495ba6c22ceSWarner Losh 	 */
1496ba6c22ceSWarner Losh 	if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1497ba6c22ceSWarner Losh 		return NULL;
1498ba6c22ceSWarner Losh 
1499df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1500ba6c22ceSWarner Losh 	/*
1501cf3ec151SWarner Losh 	 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1502ba6c22ceSWarner Losh 	 * the limits here. Enforce only for reads.
1503ba6c22ceSWarner Losh 	 */
1504035ec48eSWarner Losh 	if (do_dynamic_iosched) {
1505ba6c22ceSWarner Losh 		if (bp->bio_cmd == BIO_READ &&
1506cf3ec151SWarner Losh 		    cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1507cf3ec151SWarner Losh 			isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1508ba6c22ceSWarner Losh 			return NULL;
1509ba6c22ceSWarner Losh 		}
1510cf3ec151SWarner Losh 	}
1511cf3ec151SWarner Losh 	isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1512ba6c22ceSWarner Losh #endif
1513ba6c22ceSWarner Losh 	bioq_remove(&isc->bio_queue, bp);
1514df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1515035ec48eSWarner Losh 	if (do_dynamic_iosched) {
1516ba6c22ceSWarner Losh 		if (bp->bio_cmd == BIO_READ) {
1517ba6c22ceSWarner Losh 			isc->read_stats.queued--;
1518ba6c22ceSWarner Losh 			isc->read_stats.total++;
1519ba6c22ceSWarner Losh 			isc->read_stats.pending++;
1520ba6c22ceSWarner Losh 		} else
1521ba6c22ceSWarner Losh 			printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1522ba6c22ceSWarner Losh 	}
1523ba6c22ceSWarner Losh 	if (iosched_debug > 9)
1524ba6c22ceSWarner Losh 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1525ba6c22ceSWarner Losh #endif
1526ba6c22ceSWarner Losh 	return bp;
1527ba6c22ceSWarner Losh }
1528ba6c22ceSWarner Losh 
1529ba6c22ceSWarner Losh /*
1530ba6c22ceSWarner Losh  * Driver has been given some work to do by the block layer. Tell the
1531ba6c22ceSWarner Losh  * scheduler about it and have it queue the work up. The scheduler module
1532ba6c22ceSWarner Losh  * will then return the currently most useful bit of work later, possibly
1533ba6c22ceSWarner Losh  * deferring work for various reasons.
1534ba6c22ceSWarner Losh  */
1535ba6c22ceSWarner Losh void
1536ba6c22ceSWarner Losh cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1537ba6c22ceSWarner Losh {
1538ba6c22ceSWarner Losh 
1539ba6c22ceSWarner Losh 	/*
1540d900ade5SWarner Losh 	 * If we get a BIO_FLUSH, and we're doing delayed BIO_DELETEs then we
1541d900ade5SWarner Losh 	 * set the last tick time to one less than the current ticks minus the
1542d900ade5SWarner Losh 	 * delay to force the BIO_DELETEs to be presented to the client driver.
1543d900ade5SWarner Losh 	 */
1544d900ade5SWarner Losh 	if (bp->bio_cmd == BIO_FLUSH && isc->trim_ticks > 0)
1545d900ade5SWarner Losh 		isc->last_trim_tick = ticks - isc->trim_ticks - 1;
1546d900ade5SWarner Losh 
1547d900ade5SWarner Losh 	/*
1548d900ade5SWarner Losh 	 * Put all trims on the trim queue. Otherwise put the work on the bio
1549d900ade5SWarner Losh 	 * queue.
1550ba6c22ceSWarner Losh 	 */
1551ba6c22ceSWarner Losh 	if (bp->bio_cmd == BIO_DELETE) {
155297f8aa05SWarner Losh 		bioq_insert_tail(&isc->trim_queue, bp);
1553d900ade5SWarner Losh 		if (isc->queued_trims == 0)
1554d900ade5SWarner Losh 			isc->last_trim_tick = ticks;
1555d900ade5SWarner Losh 		isc->queued_trims++;
1556df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1557ba6c22ceSWarner Losh 		isc->trim_stats.in++;
1558ba6c22ceSWarner Losh 		isc->trim_stats.queued++;
1559ba6c22ceSWarner Losh #endif
1560ba6c22ceSWarner Losh 	}
1561df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
156284c12dcdSWarner Losh 	else if (do_dynamic_iosched && (bp->bio_cmd != BIO_READ)) {
1563ba6c22ceSWarner Losh 		if (cam_iosched_sort_queue(isc))
1564ba6c22ceSWarner Losh 			bioq_disksort(&isc->write_queue, bp);
1565ba6c22ceSWarner Losh 		else
1566ba6c22ceSWarner Losh 			bioq_insert_tail(&isc->write_queue, bp);
1567ba6c22ceSWarner Losh 		if (iosched_debug > 9)
1568ba6c22ceSWarner Losh 			printf("Qw  : %p %#x\n", bp, bp->bio_cmd);
1569ba6c22ceSWarner Losh 		if (bp->bio_cmd == BIO_WRITE) {
1570ba6c22ceSWarner Losh 			isc->write_stats.in++;
1571ba6c22ceSWarner Losh 			isc->write_stats.queued++;
1572ba6c22ceSWarner Losh 		}
1573ba6c22ceSWarner Losh 	}
1574ba6c22ceSWarner Losh #endif
1575ba6c22ceSWarner Losh 	else {
1576ba6c22ceSWarner Losh 		if (cam_iosched_sort_queue(isc))
1577ba6c22ceSWarner Losh 			bioq_disksort(&isc->bio_queue, bp);
1578ba6c22ceSWarner Losh 		else
1579ba6c22ceSWarner Losh 			bioq_insert_tail(&isc->bio_queue, bp);
1580df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1581ba6c22ceSWarner Losh 		if (iosched_debug > 9)
1582ba6c22ceSWarner Losh 			printf("Qr  : %p %#x\n", bp, bp->bio_cmd);
1583ba6c22ceSWarner Losh 		if (bp->bio_cmd == BIO_READ) {
1584ba6c22ceSWarner Losh 			isc->read_stats.in++;
1585ba6c22ceSWarner Losh 			isc->read_stats.queued++;
1586ba6c22ceSWarner Losh 		} else if (bp->bio_cmd == BIO_WRITE) {
1587ba6c22ceSWarner Losh 			isc->write_stats.in++;
1588ba6c22ceSWarner Losh 			isc->write_stats.queued++;
1589ba6c22ceSWarner Losh 		}
1590ba6c22ceSWarner Losh #endif
1591ba6c22ceSWarner Losh 	}
1592ba6c22ceSWarner Losh }
1593ba6c22ceSWarner Losh 
1594ba6c22ceSWarner Losh /*
1595ba6c22ceSWarner Losh  * If we have work, get it scheduled. Called with the periph lock held.
1596ba6c22ceSWarner Losh  */
1597ba6c22ceSWarner Losh void
1598ba6c22ceSWarner Losh cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1599ba6c22ceSWarner Losh {
1600ba6c22ceSWarner Losh 
1601ba6c22ceSWarner Losh 	if (cam_iosched_has_work(isc))
1602ba6c22ceSWarner Losh 		xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1603ba6c22ceSWarner Losh }
1604ba6c22ceSWarner Losh 
1605ba6c22ceSWarner Losh /*
160655c770b4SWarner Losh  * Complete a trim request. Mark that we no longer have one in flight.
1607ba6c22ceSWarner Losh  */
1608ba6c22ceSWarner Losh void
1609ba6c22ceSWarner Losh cam_iosched_trim_done(struct cam_iosched_softc *isc)
1610ba6c22ceSWarner Losh {
1611ba6c22ceSWarner Losh 
1612ba6c22ceSWarner Losh 	isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1613ba6c22ceSWarner Losh }
1614ba6c22ceSWarner Losh 
1615ba6c22ceSWarner Losh /*
1616ba6c22ceSWarner Losh  * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1617ba6c22ceSWarner Losh  * might use notes in the ccb for statistics.
1618ba6c22ceSWarner Losh  */
1619ba6c22ceSWarner Losh int
1620ba6c22ceSWarner Losh cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1621ba6c22ceSWarner Losh     union ccb *done_ccb)
1622ba6c22ceSWarner Losh {
1623ba6c22ceSWarner Losh 	int retval = 0;
1624df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1625035ec48eSWarner Losh 	if (!do_dynamic_iosched)
1626ba6c22ceSWarner Losh 		return retval;
1627ba6c22ceSWarner Losh 
1628ba6c22ceSWarner Losh 	if (iosched_debug > 10)
1629ba6c22ceSWarner Losh 		printf("done: %p %#x\n", bp, bp->bio_cmd);
1630ba6c22ceSWarner Losh 	if (bp->bio_cmd == BIO_WRITE) {
1631ba6c22ceSWarner Losh 		retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1632157cb465SWarner Losh 		if ((bp->bio_flags & BIO_ERROR) != 0)
1633c4b72d8bSWarner Losh 			isc->write_stats.errs++;
1634ba6c22ceSWarner Losh 		isc->write_stats.out++;
1635ba6c22ceSWarner Losh 		isc->write_stats.pending--;
1636ba6c22ceSWarner Losh 	} else if (bp->bio_cmd == BIO_READ) {
1637ba6c22ceSWarner Losh 		retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1638157cb465SWarner Losh 		if ((bp->bio_flags & BIO_ERROR) != 0)
1639c4b72d8bSWarner Losh 			isc->read_stats.errs++;
1640ba6c22ceSWarner Losh 		isc->read_stats.out++;
1641ba6c22ceSWarner Losh 		isc->read_stats.pending--;
1642ba6c22ceSWarner Losh 	} else if (bp->bio_cmd == BIO_DELETE) {
1643157cb465SWarner Losh 		if ((bp->bio_flags & BIO_ERROR) != 0)
1644c4b72d8bSWarner Losh 			isc->trim_stats.errs++;
1645ba6c22ceSWarner Losh 		isc->trim_stats.out++;
1646ba6c22ceSWarner Losh 		isc->trim_stats.pending--;
1647ba6c22ceSWarner Losh 	} else if (bp->bio_cmd != BIO_FLUSH) {
1648ba6c22ceSWarner Losh 		if (iosched_debug)
1649ba6c22ceSWarner Losh 			printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1650ba6c22ceSWarner Losh 	}
1651ba6c22ceSWarner Losh 
1652e5436ab5SWarner Losh 	if (!(bp->bio_flags & BIO_ERROR) && done_ccb != NULL) {
1653e5436ab5SWarner Losh 		sbintime_t sim_latency;
1654e5436ab5SWarner Losh 
1655e5436ab5SWarner Losh 		sim_latency = cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data);
1656e5436ab5SWarner Losh 
1657e5436ab5SWarner Losh 		cam_iosched_io_metric_update(isc, sim_latency,
1658ba6c22ceSWarner Losh 		    bp->bio_cmd, bp->bio_bcount);
1659e5436ab5SWarner Losh 		/*
1660e5436ab5SWarner Losh 		 * Debugging code: allow callbacks to the periph driver when latency max
1661e5436ab5SWarner Losh 		 * is exceeded. This can be useful for triggering external debugging actions.
1662e5436ab5SWarner Losh 		 */
1663e5436ab5SWarner Losh 		if (isc->latfcn && isc->max_lat != 0 && sim_latency > isc->max_lat)
1664e5436ab5SWarner Losh 			isc->latfcn(isc->latarg, sim_latency, bp);
1665e5436ab5SWarner Losh 	}
1666e5436ab5SWarner Losh 
1667ba6c22ceSWarner Losh #endif
1668ba6c22ceSWarner Losh 	return retval;
1669ba6c22ceSWarner Losh }
1670ba6c22ceSWarner Losh 
1671ba6c22ceSWarner Losh /*
1672ba6c22ceSWarner Losh  * Tell the io scheduler that you've pushed a trim down into the sim.
167355c770b4SWarner Losh  * This also tells the I/O scheduler not to push any more trims down, so
167455c770b4SWarner Losh  * some periphs do not call it if they can cope with multiple trims in flight.
1675ba6c22ceSWarner Losh  */
1676ba6c22ceSWarner Losh void
1677ba6c22ceSWarner Losh cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1678ba6c22ceSWarner Losh {
1679ba6c22ceSWarner Losh 
1680ba6c22ceSWarner Losh 	isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1681ba6c22ceSWarner Losh }
1682ba6c22ceSWarner Losh 
1683ba6c22ceSWarner Losh /*
1684ba6c22ceSWarner Losh  * Change the sorting policy hint for I/O transactions for this device.
1685ba6c22ceSWarner Losh  */
1686ba6c22ceSWarner Losh void
1687ba6c22ceSWarner Losh cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1688ba6c22ceSWarner Losh {
1689ba6c22ceSWarner Losh 
1690ba6c22ceSWarner Losh 	isc->sort_io_queue = val;
1691ba6c22ceSWarner Losh }
1692ba6c22ceSWarner Losh 
1693ba6c22ceSWarner Losh int
1694ba6c22ceSWarner Losh cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1695ba6c22ceSWarner Losh {
1696ba6c22ceSWarner Losh 	return isc->flags & flags;
1697ba6c22ceSWarner Losh }
1698ba6c22ceSWarner Losh 
1699ba6c22ceSWarner Losh void
1700ba6c22ceSWarner Losh cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1701ba6c22ceSWarner Losh {
1702ba6c22ceSWarner Losh 	isc->flags |= flags;
1703ba6c22ceSWarner Losh }
1704ba6c22ceSWarner Losh 
1705ba6c22ceSWarner Losh void
1706ba6c22ceSWarner Losh cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1707ba6c22ceSWarner Losh {
1708ba6c22ceSWarner Losh 	isc->flags &= ~flags;
1709ba6c22ceSWarner Losh }
1710ba6c22ceSWarner Losh 
1711df236247SWarner Losh #ifdef CAM_IOSCHED_DYNAMIC
1712ba6c22ceSWarner Losh /*
1713ba6c22ceSWarner Losh  * After the method presented in Jack Crenshaw's 1998 article "Integer
17140b4060b0SEd Maste  * Square Roots," reprinted at
1715ba6c22ceSWarner Losh  * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1716ba6c22ceSWarner Losh  * and well worth the read. Briefly, we find the power of 4 that's the
1717ba6c22ceSWarner Losh  * largest smaller than val. We then check each smaller power of 4 to
1718ba6c22ceSWarner Losh  * see if val is still bigger. The right shifts at each step divide
1719ba6c22ceSWarner Losh  * the result by 2 which after successive application winds up
1720ba6c22ceSWarner Losh  * accumulating the right answer. It could also have been accumulated
1721ba6c22ceSWarner Losh  * using a separate root counter, but this code is smaller and faster
1722ba6c22ceSWarner Losh  * than that method. This method is also integer size invariant.
17230b4060b0SEd Maste  * It returns floor(sqrt((float)val)), or the largest integer less than
1724ba6c22ceSWarner Losh  * or equal to the square root.
1725ba6c22ceSWarner Losh  */
1726ba6c22ceSWarner Losh static uint64_t
1727ba6c22ceSWarner Losh isqrt64(uint64_t val)
1728ba6c22ceSWarner Losh {
1729ba6c22ceSWarner Losh 	uint64_t res = 0;
1730ba6c22ceSWarner Losh 	uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1731ba6c22ceSWarner Losh 
1732ba6c22ceSWarner Losh 	/*
1733ba6c22ceSWarner Losh 	 * Find the largest power of 4 smaller than val.
1734ba6c22ceSWarner Losh 	 */
1735ba6c22ceSWarner Losh 	while (bit > val)
1736ba6c22ceSWarner Losh 		bit >>= 2;
1737ba6c22ceSWarner Losh 
1738ba6c22ceSWarner Losh 	/*
1739ba6c22ceSWarner Losh 	 * Accumulate the answer, one bit at a time (we keep moving
1740ba6c22ceSWarner Losh 	 * them over since 2 is the square root of 4 and we test
1741ba6c22ceSWarner Losh 	 * powers of 4). We accumulate where we find the bit, but
1742ba6c22ceSWarner Losh 	 * the successive shifts land the bit in the right place
1743ba6c22ceSWarner Losh 	 * by the end.
1744ba6c22ceSWarner Losh 	 */
1745ba6c22ceSWarner Losh 	while (bit != 0) {
1746ba6c22ceSWarner Losh 		if (val >= res + bit) {
1747ba6c22ceSWarner Losh 			val -= res + bit;
1748ba6c22ceSWarner Losh 			res = (res >> 1) + bit;
1749ba6c22ceSWarner Losh 		} else
1750ba6c22ceSWarner Losh 			res >>= 1;
1751ba6c22ceSWarner Losh 		bit >>= 2;
1752ba6c22ceSWarner Losh 	}
1753ba6c22ceSWarner Losh 
1754ba6c22ceSWarner Losh 	return res;
1755ba6c22ceSWarner Losh }
1756ba6c22ceSWarner Losh 
175708fc2f23SWarner Losh static sbintime_t latencies[LAT_BUCKETS - 1] = {
1758cf3ec151SWarner Losh 	SBT_1MS <<  0,
1759cf3ec151SWarner Losh 	SBT_1MS <<  1,
1760cf3ec151SWarner Losh 	SBT_1MS <<  2,
1761cf3ec151SWarner Losh 	SBT_1MS <<  3,
1762cf3ec151SWarner Losh 	SBT_1MS <<  4,
1763cf3ec151SWarner Losh 	SBT_1MS <<  5,
1764cf3ec151SWarner Losh 	SBT_1MS <<  6,
1765cf3ec151SWarner Losh 	SBT_1MS <<  7,
1766cf3ec151SWarner Losh 	SBT_1MS <<  8,
1767cf3ec151SWarner Losh 	SBT_1MS <<  9,
176808fc2f23SWarner Losh 	SBT_1MS << 10,
176908fc2f23SWarner Losh 	SBT_1MS << 11,
177008fc2f23SWarner Losh 	SBT_1MS << 12,
177108fc2f23SWarner Losh 	SBT_1MS << 13		/* 8.192s */
1772cf3ec151SWarner Losh };
1773cf3ec151SWarner Losh 
1774ba6c22ceSWarner Losh static void
1775ba6c22ceSWarner Losh cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1776ba6c22ceSWarner Losh {
177779d80af2SWarner Losh 	sbintime_t y, deltasq, delta;
1778cf3ec151SWarner Losh 	int i;
1779cf3ec151SWarner Losh 
1780cf3ec151SWarner Losh 	/*
1781cf3ec151SWarner Losh 	 * Keep counts for latency. We do it by power of two buckets.
1782cf3ec151SWarner Losh 	 * This helps us spot outlier behavior obscured by averages.
1783cf3ec151SWarner Losh 	 */
1784cf3ec151SWarner Losh 	for (i = 0; i < LAT_BUCKETS - 1; i++) {
1785cf3ec151SWarner Losh 		if (sim_latency < latencies[i]) {
1786cf3ec151SWarner Losh 			iop->latencies[i]++;
1787cf3ec151SWarner Losh 			break;
1788cf3ec151SWarner Losh 		}
1789cf3ec151SWarner Losh 	}
1790cf3ec151SWarner Losh 	if (i == LAT_BUCKETS - 1)
1791cf3ec151SWarner Losh 		iop->latencies[i]++; 	 /* Put all > 1024ms values into the last bucket. */
1792ba6c22ceSWarner Losh 
1793ba6c22ceSWarner Losh 	/*
17940b4060b0SEd Maste 	 * Classic exponentially decaying average with a tiny alpha
1795ba6c22ceSWarner Losh 	 * (2 ^ -alpha_bits). For more info see the NIST statistical
1796ba6c22ceSWarner Losh 	 * handbook.
1797ba6c22ceSWarner Losh 	 *
179879d80af2SWarner Losh 	 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha)		[nist]
179979d80af2SWarner Losh 	 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
180079d80af2SWarner Losh 	 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1801ba6c22ceSWarner Losh 	 * alpha = 1 / (1 << alpha_bits)
180279d80af2SWarner Losh 	 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
180379d80af2SWarner Losh 	 *	= y_t/b - e/b + be/b
180479d80af2SWarner Losh 	 *      = (y_t - e + be) / b
180579d80af2SWarner Losh 	 *	= (e + d) / b
1806ba6c22ceSWarner Losh 	 *
1807ba6c22ceSWarner Losh 	 * Since alpha is a power of two, we can compute this w/o any mult or
1808ba6c22ceSWarner Losh 	 * division.
180979d80af2SWarner Losh 	 *
181079d80af2SWarner Losh 	 * Variance can also be computed. Usually, it would be expressed as follows:
181179d80af2SWarner Losh 	 *	diff_t = y_t - ema_t-1
181279d80af2SWarner Losh 	 *	emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
181379d80af2SWarner Losh 	 *	  = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
181479d80af2SWarner Losh 	 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
181579d80af2SWarner Losh 	 *	  = e - e/b + dd/b + dd/bb
181679d80af2SWarner Losh 	 *	  = (bbe - be + bdd + dd) / bb
181779d80af2SWarner Losh 	 *	  = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
181879d80af2SWarner Losh 	 */
181979d80af2SWarner Losh 	/*
182079d80af2SWarner Losh 	 * XXX possible numeric issues
182179d80af2SWarner Losh 	 *	o We assume right shifted integers do the right thing, since that's
182279d80af2SWarner Losh 	 *	  implementation defined. You can change the right shifts to / (1LL << alpha).
182379d80af2SWarner Losh 	 *	o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
182479d80af2SWarner Losh 	 *	  for emvar. This puts a ceiling of 13 bits on alpha since we need a
182579d80af2SWarner Losh 	 *	  few tens of seconds of representation.
182679d80af2SWarner Losh 	 *	o We mitigate alpha issues by never setting it too high.
1827ba6c22ceSWarner Losh 	 */
1828ba6c22ceSWarner Losh 	y = sim_latency;
182979d80af2SWarner Losh 	delta = (y - iop->ema);					/* d */
183079d80af2SWarner Losh 	iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1831ba6c22ceSWarner Losh 
1832ba6c22ceSWarner Losh 	/*
183379d80af2SWarner Losh 	 * Were we to naively plow ahead at this point, we wind up with many numerical
183479d80af2SWarner Losh 	 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
183579d80af2SWarner Losh 	 * us with microsecond level precision in the input, so the same in the
183679d80af2SWarner Losh 	 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
183779d80af2SWarner Losh 	 * also means that emvar can be up 46 bits 40 of which are fraction, which
183879d80af2SWarner Losh 	 * gives us a way to measure up to ~8s in the SD before the computation goes
183979d80af2SWarner Losh 	 * unstable. Even the worst hard disk rarely has > 1s service time in the
184079d80af2SWarner Losh 	 * drive. It does mean we have to shift left 12 bits after taking the
184179d80af2SWarner Losh 	 * square root to compute the actual standard deviation estimate. This loss of
184279d80af2SWarner Losh 	 * precision is preferable to needing int128 types to work. The above numbers
184379d80af2SWarner Losh 	 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
184479d80af2SWarner Losh 	 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1845ba6c22ceSWarner Losh 	 */
184679d80af2SWarner Losh 	delta >>= 12;
184779d80af2SWarner Losh 	deltasq = delta * delta;				/* dd */
184879d80af2SWarner Losh 	iop->emvar = ((iop->emvar << (2 * alpha_bits)) +	/* bbe */
184979d80af2SWarner Losh 	    ((deltasq - iop->emvar) << alpha_bits) +		/* b(dd-e) */
185079d80af2SWarner Losh 	    deltasq)						/* dd */
185179d80af2SWarner Losh 	    >> (2 * alpha_bits);				/* div bb */
185279d80af2SWarner Losh 	iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1853ba6c22ceSWarner Losh }
1854ba6c22ceSWarner Losh 
1855ba6c22ceSWarner Losh static void
1856ba6c22ceSWarner Losh cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1857ba6c22ceSWarner Losh     sbintime_t sim_latency, int cmd, size_t size)
1858ba6c22ceSWarner Losh {
1859ba6c22ceSWarner Losh 	/* xxx Do we need to scale based on the size of the I/O ? */
1860ba6c22ceSWarner Losh 	switch (cmd) {
1861ba6c22ceSWarner Losh 	case BIO_READ:
1862ba6c22ceSWarner Losh 		cam_iosched_update(&isc->read_stats, sim_latency);
1863ba6c22ceSWarner Losh 		break;
1864ba6c22ceSWarner Losh 	case BIO_WRITE:
1865ba6c22ceSWarner Losh 		cam_iosched_update(&isc->write_stats, sim_latency);
1866ba6c22ceSWarner Losh 		break;
1867ba6c22ceSWarner Losh 	case BIO_DELETE:
1868ba6c22ceSWarner Losh 		cam_iosched_update(&isc->trim_stats, sim_latency);
1869ba6c22ceSWarner Losh 		break;
1870ba6c22ceSWarner Losh 	default:
1871ba6c22ceSWarner Losh 		break;
1872ba6c22ceSWarner Losh 	}
1873ba6c22ceSWarner Losh }
1874ba6c22ceSWarner Losh 
1875ba6c22ceSWarner Losh #ifdef DDB
1876ba6c22ceSWarner Losh static int biolen(struct bio_queue_head *bq)
1877ba6c22ceSWarner Losh {
1878ba6c22ceSWarner Losh 	int i = 0;
1879ba6c22ceSWarner Losh 	struct bio *bp;
1880ba6c22ceSWarner Losh 
1881ba6c22ceSWarner Losh 	TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1882ba6c22ceSWarner Losh 		i++;
1883ba6c22ceSWarner Losh 	}
1884ba6c22ceSWarner Losh 	return i;
1885ba6c22ceSWarner Losh }
1886ba6c22ceSWarner Losh 
1887ba6c22ceSWarner Losh /*
1888ba6c22ceSWarner Losh  * Show the internal state of the I/O scheduler.
1889ba6c22ceSWarner Losh  */
1890ba6c22ceSWarner Losh DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1891ba6c22ceSWarner Losh {
1892ba6c22ceSWarner Losh 	struct cam_iosched_softc *isc;
1893ba6c22ceSWarner Losh 
1894ba6c22ceSWarner Losh 	if (!have_addr) {
1895ba6c22ceSWarner Losh 		db_printf("Need addr\n");
1896ba6c22ceSWarner Losh 		return;
1897ba6c22ceSWarner Losh 	}
1898ba6c22ceSWarner Losh 	isc = (struct cam_iosched_softc *)addr;
1899ba6c22ceSWarner Losh 	db_printf("pending_reads:     %d\n", isc->read_stats.pending);
1900ba6c22ceSWarner Losh 	db_printf("min_reads:         %d\n", isc->read_stats.min);
1901ba6c22ceSWarner Losh 	db_printf("max_reads:         %d\n", isc->read_stats.max);
1902ba6c22ceSWarner Losh 	db_printf("reads:             %d\n", isc->read_stats.total);
1903ba6c22ceSWarner Losh 	db_printf("in_reads:          %d\n", isc->read_stats.in);
1904ba6c22ceSWarner Losh 	db_printf("out_reads:         %d\n", isc->read_stats.out);
1905ba6c22ceSWarner Losh 	db_printf("queued_reads:      %d\n", isc->read_stats.queued);
1906*3aba1d47SWarner Losh 	db_printf("Read Q len         %d\n", biolen(&isc->bio_queue));
1907ba6c22ceSWarner Losh 	db_printf("pending_writes:    %d\n", isc->write_stats.pending);
1908ba6c22ceSWarner Losh 	db_printf("min_writes:        %d\n", isc->write_stats.min);
1909ba6c22ceSWarner Losh 	db_printf("max_writes:        %d\n", isc->write_stats.max);
1910ba6c22ceSWarner Losh 	db_printf("writes:            %d\n", isc->write_stats.total);
1911ba6c22ceSWarner Losh 	db_printf("in_writes:         %d\n", isc->write_stats.in);
1912ba6c22ceSWarner Losh 	db_printf("out_writes:        %d\n", isc->write_stats.out);
1913ba6c22ceSWarner Losh 	db_printf("queued_writes:     %d\n", isc->write_stats.queued);
1914*3aba1d47SWarner Losh 	db_printf("Write Q len        %d\n", biolen(&isc->write_queue));
1915ba6c22ceSWarner Losh 	db_printf("pending_trims:     %d\n", isc->trim_stats.pending);
1916ba6c22ceSWarner Losh 	db_printf("min_trims:         %d\n", isc->trim_stats.min);
1917ba6c22ceSWarner Losh 	db_printf("max_trims:         %d\n", isc->trim_stats.max);
1918ba6c22ceSWarner Losh 	db_printf("trims:             %d\n", isc->trim_stats.total);
1919ba6c22ceSWarner Losh 	db_printf("in_trims:          %d\n", isc->trim_stats.in);
1920ba6c22ceSWarner Losh 	db_printf("out_trims:         %d\n", isc->trim_stats.out);
1921ba6c22ceSWarner Losh 	db_printf("queued_trims:      %d\n", isc->trim_stats.queued);
1922*3aba1d47SWarner Losh 	db_printf("Trim Q len         %d\n", biolen(&isc->trim_queue));
1923ba6c22ceSWarner Losh 	db_printf("read_bias:         %d\n", isc->read_bias);
1924ba6c22ceSWarner Losh 	db_printf("current_read_bias: %d\n", isc->current_read_bias);
1925ba6c22ceSWarner Losh 	db_printf("Trim active?       %s\n",
1926ba6c22ceSWarner Losh 	    (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");
1927ba6c22ceSWarner Losh }
1928ba6c22ceSWarner Losh #endif
1929ba6c22ceSWarner Losh #endif
1930