xref: /illumos-gate/usr/src/uts/sun4v/os/mach_mp_states.c (revision 560f878bce5cdf0661659001415019ca5c8a01b4)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/cpuvar.h>
30 #include <sys/cpu_module.h>
31 #include <sys/machsystm.h>
32 #include <sys/archsystm.h>
33 #include <sys/prom_plat.h>
34 #include <sys/hypervisor_api.h>
35 #include <sys/hsvc.h>
36 
37 extern uint64_t xc_tick_limit;
38 extern uint64_t xc_tick_jump_limit;
39 
40 extern void cpu_intrq_unregister_powerdown(uint64_t doneflag_va);
41 
42 /*
43  * set_idle_cpu is called from idle() when a CPU becomes idle.
44  */
45 /*ARGSUSED*/
46 void
47 set_idle_cpu(int cpun)
48 {
49 }
50 
51 /*
52  * unset_idle_cpu is called from idle() when a CPU is no longer idle.
53  */
54 /*ARGSUSED*/
55 void
56 unset_idle_cpu(int cpun)
57 {
58 }
59 
60 /*
61  * Stop a CPU based on its cpuid, using the cpu_stop hypervisor call.
62  * Since this requires that the hypervisor force a remote CPU to stop,
63  * the assumption is made that this should take roughly the same amount
64  * of time as a CPU mondo. Consequently, the mondo timeout is used to
65  * determine when to give up waiting for the CPU to stop.
66  *
67  * Attempts to stop a CPU already in the stopped or error state will
68  * silently succeed. Zero is returned on success and a non-negative
69  * errno value is returned on failure.
70  */
71 int
72 stopcpu_bycpuid(int cpuid)
73 {
74 	uint64_t	loop_cnt;
75 	uint64_t	state;
76 	uint64_t	rv;
77 	uint64_t	major = 0;
78 	uint64_t	minor = 0;
79 	uint64_t	cpu_stop_time_limit;
80 	extern uint64_t	xc_mondo_time_limit;
81 
82 	ASSERT(MUTEX_HELD(&cpu_lock));
83 
84 	/*
85 	 * Check the state of the CPU up front to see if an
86 	 * attempt to stop it is even necessary.
87 	 */
88 	if (hv_cpu_state(cpuid, &state) != H_EOK)
89 		return (EINVAL);
90 
91 	/* treat stopped and error state the same */
92 	if (state != CPU_STATE_RUNNING) {
93 		/* nothing to do */
94 		return (0);
95 	}
96 
97 	/*
98 	 * The HV API to stop a CPU is only supported in
99 	 * version 1.1 and later of the core group. If an
100 	 * older version of the HV is in use, return not
101 	 * supported.
102 	 */
103 	if (hsvc_version(HSVC_GROUP_CORE, &major, &minor) != 0)
104 		return (EINVAL);
105 
106 	ASSERT(major != 0);
107 
108 	if ((major == 1) && (minor < 1))
109 		return (ENOTSUP);
110 
111 	/* use the mondo timeout if it has been initialized */
112 	cpu_stop_time_limit = xc_mondo_time_limit;
113 
114 	/*
115 	 * If called early in boot before the mondo time limit
116 	 * is set, use a reasonable timeout based on the the
117 	 * clock frequency of the current CPU.
118 	 */
119 	if (cpu_stop_time_limit == 0)
120 		cpu_stop_time_limit = cpunodes[CPU->cpu_id].clock_freq;
121 
122 	/* should only fail if called too early in boot */
123 	ASSERT(cpu_stop_time_limit > 0);
124 
125 	loop_cnt = 0;
126 
127 	/*
128 	 * Attempt to stop the CPU, retrying if it is busy.
129 	 */
130 	while (loop_cnt++ < cpu_stop_time_limit) {
131 
132 		if ((rv = hv_cpu_stop(cpuid)) != H_EWOULDBLOCK)
133 			break;
134 	}
135 
136 	if (loop_cnt == cpu_stop_time_limit)
137 		return (ETIMEDOUT);
138 
139 	if (rv != H_EOK)
140 		return (EINVAL);
141 
142 	/*
143 	 * Verify that the CPU has reached the stopped state.
144 	 */
145 	while (loop_cnt++ < cpu_stop_time_limit) {
146 
147 		if (hv_cpu_state(cpuid, &state) != H_EOK)
148 			return (EINVAL);
149 
150 		/* treat stopped and error state the same */
151 		if (state != CPU_STATE_RUNNING)
152 			break;
153 	}
154 
155 	return ((loop_cnt == cpu_stop_time_limit) ? ETIMEDOUT : 0);
156 }
157 
158 /*
159  * X-trap to the target to unregister its interrupt and error queues
160  * and put it in a safe place just before the CPU is stopped. After
161  * unregistering its queues, the target CPU must not return from the
162  * trap to priv or user context. Ensure that the interrupt CPU unregister
163  * succeeded.
164  */
165 void
166 xt_cpu_unreg_powerdown(struct cpu *cpup)
167 {
168 	uint8_t volatile not_done;
169 	uint64_t starttick, endtick, tick, lasttick;
170 	processorid_t cpuid = cpup->cpu_id;
171 
172 	kpreempt_disable();
173 
174 	/*
175 	 * Sun4v uses a queue for receiving mondos. Successful
176 	 * transmission of a mondo only indicates that the mondo
177 	 * has been written into the queue.
178 	 *
179 	 * Set the not_done flag to 1 before sending the cross
180 	 * trap and wait until the other cpu resets it to 0.
181 	 */
182 
183 	not_done = 1;
184 
185 	xt_one_unchecked(cpuid, (xcfunc_t *)cpu_intrq_unregister_powerdown,
186 	    (uint64_t)&not_done, 0);
187 
188 	starttick = lasttick = gettick();
189 	endtick = starttick + xc_tick_limit;
190 
191 	while (not_done) {
192 
193 		tick = gettick();
194 
195 		/*
196 		 * If there is a big jump between the current tick
197 		 * count and lasttick, we have probably hit a break
198 		 * point. Adjust endtick accordingly to avoid panic.
199 		 */
200 		if (tick > (lasttick + xc_tick_jump_limit)) {
201 			endtick += (tick - lasttick);
202 		}
203 
204 		lasttick = tick;
205 		if (tick > endtick) {
206 			cmn_err(CE_CONT, "Cross trap timeout at cpu id %x\n",
207 			    cpuid);
208 			cmn_err(CE_WARN, "xt_intrq_unreg_powerdown: timeout");
209 		}
210 	}
211 
212 	kpreempt_enable();
213 }
214 
215 int
216 plat_cpu_poweroff(struct cpu *cp)
217 {
218 	int		rv = 0;
219 	int		status;
220 	processorid_t	cpuid = cp->cpu_id;
221 
222 	ASSERT(MUTEX_HELD(&cpu_lock));
223 
224 	/*
225 	 * Capture all CPUs (except for detaching proc) to prevent
226 	 * crosscalls to the detaching proc until it has cleared its
227 	 * bit in cpu_ready_set.
228 	 *
229 	 * The CPU's remain paused and the prom_mutex is known to be free.
230 	 * This prevents the x-trap victim from blocking when doing prom
231 	 * IEEE-1275 calls at a high PIL level.
232 	 */
233 	promsafe_pause_cpus();
234 
235 	/*
236 	 * Quiesce interrupts on the target CPU. We do this by setting
237 	 * the CPU 'not ready'- (i.e. removing the CPU from cpu_ready_set)
238 	 * to prevent it from receiving cross calls and cross traps. This
239 	 * prevents the processor from receiving any new soft interrupts.
240 	 */
241 	mp_cpu_quiesce(cp);
242 
243 	/*
244 	 * Send a cross trap to the cpu to unregister its interrupt
245 	 * error queues.
246 	 */
247 	xt_cpu_unreg_powerdown(cp);
248 
249 	cp->cpu_flags = CPU_OFFLINE | CPU_QUIESCED | CPU_POWEROFF;
250 
251 	/* call into the Hypervisor to stop the CPU */
252 	if ((status = stopcpu_bycpuid(cpuid)) != 0) {
253 		rv = -1;
254 	}
255 
256 	start_cpus();
257 
258 	if (rv != 0) {
259 		cmn_err(CE_WARN, "failed to stop cpu %d (%d)", cpuid, status);
260 		/* mark the CPU faulted so that it cannot be onlined */
261 		cp->cpu_flags = CPU_OFFLINE | CPU_QUIESCED | CPU_FAULTED;
262 	}
263 
264 	return (rv);
265 }
266 
267 int
268 plat_cpu_poweron(struct cpu *cp)
269 {
270 	extern void	restart_other_cpu(int);
271 
272 	ASSERT(MUTEX_HELD(&cpu_lock));
273 
274 	cp->cpu_flags &= ~CPU_POWEROFF;
275 
276 	restart_other_cpu(cp->cpu_id);
277 
278 	return (0);
279 }
280