1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Block rq-qos policy for assigning an I/O priority class to requests.
  4 *
  5 * Using an rq-qos policy for assigning I/O priority class has two advantages
  6 * over using the ioprio_set() system call:
  7 *
  8 * - This policy is cgroup based so it has all the advantages of cgroups.
  9 * - While ioprio_set() does not affect page cache writeback I/O, this rq-qos
 10 *   controller affects page cache writeback I/O for filesystems that support
 11 *   assiociating a cgroup with writeback I/O. See also
 12 *   Documentation/admin-guide/cgroup-v2.rst.
 13 */
 14
 15#include <linux/blk-mq.h>
 16#include <linux/blk_types.h>
 17#include <linux/kernel.h>
 18#include <linux/module.h>
 19#include "blk-cgroup.h"
 20#include "blk-ioprio.h"
 21#include "blk-rq-qos.h"
 22
 23/**
 24 * enum prio_policy - I/O priority class policy.
 25 * @POLICY_NO_CHANGE: (default) do not modify the I/O priority class.
 26 * @POLICY_PROMOTE_TO_RT: modify no-IOPRIO_CLASS_RT to IOPRIO_CLASS_RT.
 27 * @POLICY_RESTRICT_TO_BE: modify IOPRIO_CLASS_NONE and IOPRIO_CLASS_RT into
 28 *		IOPRIO_CLASS_BE.
 29 * @POLICY_ALL_TO_IDLE: change the I/O priority class into IOPRIO_CLASS_IDLE.
 30 * @POLICY_NONE_TO_RT: an alias for POLICY_PROMOTE_TO_RT.
 31 *
 32 * See also <linux/ioprio.h>.
 33 */
 34enum prio_policy {
 35	POLICY_NO_CHANGE	= 0,
 36	POLICY_PROMOTE_TO_RT	= 1,
 37	POLICY_RESTRICT_TO_BE	= 2,
 38	POLICY_ALL_TO_IDLE	= 3,
 39	POLICY_NONE_TO_RT	= 4,
 40};
 41
 42static const char *policy_name[] = {
 43	[POLICY_NO_CHANGE]	= "no-change",
 44	[POLICY_PROMOTE_TO_RT]	= "promote-to-rt",
 45	[POLICY_RESTRICT_TO_BE]	= "restrict-to-be",
 46	[POLICY_ALL_TO_IDLE]	= "idle",
 47	[POLICY_NONE_TO_RT]	= "none-to-rt",
 48};
 49
 50static struct blkcg_policy ioprio_policy;
 51
 52/**
 53 * struct ioprio_blkcg - Per cgroup data.
 54 * @cpd: blkcg_policy_data structure.
 55 * @prio_policy: One of the IOPRIO_CLASS_* values. See also <linux/ioprio.h>.
 56 */
 57struct ioprio_blkcg {
 58	struct blkcg_policy_data cpd;
 59	enum prio_policy	 prio_policy;
 60};
 61
 62static struct ioprio_blkcg *blkcg_to_ioprio_blkcg(struct blkcg *blkcg)
 63{
 64	return container_of(blkcg_to_cpd(blkcg, &ioprio_policy),
 65			    struct ioprio_blkcg, cpd);
 66}
 67
 68static struct ioprio_blkcg *
 69ioprio_blkcg_from_css(struct cgroup_subsys_state *css)
 70{
 71	return blkcg_to_ioprio_blkcg(css_to_blkcg(css));
 72}
 73
 74static int ioprio_show_prio_policy(struct seq_file *sf, void *v)
 75{
 76	struct ioprio_blkcg *blkcg = ioprio_blkcg_from_css(seq_css(sf));
 77
 78	seq_printf(sf, "%s\n", policy_name[blkcg->prio_policy]);
 79	return 0;
 80}
 81
 82static ssize_t ioprio_set_prio_policy(struct kernfs_open_file *of, char *buf,
 83				      size_t nbytes, loff_t off)
 84{
 85	struct ioprio_blkcg *blkcg = ioprio_blkcg_from_css(of_css(of));
 86	int ret;
 87
 88	if (off != 0)
 89		return -EIO;
 90	/* kernfs_fop_write_iter() terminates 'buf' with '\0'. */
 91	ret = sysfs_match_string(policy_name, buf);
 92	if (ret < 0)
 93		return ret;
 94	blkcg->prio_policy = ret;
 95	return nbytes;
 96}
 97
 98static struct blkcg_policy_data *ioprio_alloc_cpd(gfp_t gfp)
 99{
100	struct ioprio_blkcg *blkcg;
101
102	blkcg = kzalloc(sizeof(*blkcg), gfp);
103	if (!blkcg)
104		return NULL;
105	blkcg->prio_policy = POLICY_NO_CHANGE;
106	return &blkcg->cpd;
107}
108
109static void ioprio_free_cpd(struct blkcg_policy_data *cpd)
110{
111	struct ioprio_blkcg *blkcg = container_of(cpd, typeof(*blkcg), cpd);
112
113	kfree(blkcg);
114}
115
116#define IOPRIO_ATTRS						\
117	{							\
118		.name		= "prio.class",			\
119		.seq_show	= ioprio_show_prio_policy,	\
120		.write		= ioprio_set_prio_policy,	\
121	},							\
122	{ } /* sentinel */
123
124/* cgroup v2 attributes */
125static struct cftype ioprio_files[] = {
126	IOPRIO_ATTRS
127};
128
129/* cgroup v1 attributes */
130static struct cftype ioprio_legacy_files[] = {
131	IOPRIO_ATTRS
132};
133
134static struct blkcg_policy ioprio_policy = {
135	.dfl_cftypes	= ioprio_files,
136	.legacy_cftypes = ioprio_legacy_files,
137
138	.cpd_alloc_fn	= ioprio_alloc_cpd,
139	.cpd_free_fn	= ioprio_free_cpd,
140};
141
142void blkcg_set_ioprio(struct bio *bio)
143{
144	struct ioprio_blkcg *blkcg = blkcg_to_ioprio_blkcg(bio->bi_blkg->blkcg);
145	u16 prio;
146
147	if (!blkcg || blkcg->prio_policy == POLICY_NO_CHANGE)
148		return;
149
150	if (blkcg->prio_policy == POLICY_PROMOTE_TO_RT ||
151	    blkcg->prio_policy == POLICY_NONE_TO_RT) {
152		/*
153		 * For RT threads, the default priority level is 4 because
154		 * task_nice is 0. By promoting non-RT io-priority to RT-class
155		 * and default level 4, those requests that are already
156		 * RT-class but need a higher io-priority can use ioprio_set()
157		 * to achieve this.
158		 */
159		if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) != IOPRIO_CLASS_RT)
160			bio->bi_ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_RT, 4);
161		return;
162	}
163
164	/*
165	 * Except for IOPRIO_CLASS_NONE, higher I/O priority numbers
166	 * correspond to a lower priority. Hence, the max_t() below selects
167	 * the lower priority of bi_ioprio and the cgroup I/O priority class.
168	 * If the bio I/O priority equals IOPRIO_CLASS_NONE, the cgroup I/O
169	 * priority is assigned to the bio.
170	 */
171	prio = max_t(u16, bio->bi_ioprio,
172			IOPRIO_PRIO_VALUE(blkcg->prio_policy, 0));
173	if (prio > bio->bi_ioprio)
174		bio->bi_ioprio = prio;
175}
176
177static int __init ioprio_init(void)
178{
179	return blkcg_policy_register(&ioprio_policy);
180}
181
182static void __exit ioprio_exit(void)
183{
184	blkcg_policy_unregister(&ioprio_policy);
185}
186
187module_init(ioprio_init);
188module_exit(ioprio_exit);