]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - drivers/hv/hv.c
vmbus: change to per channel tasklet
[karo-tx-linux.git] / drivers / hv / hv.c
1 /*
2  * Copyright (c) 2009, Microsoft Corporation.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms and conditions of the GNU General Public License,
6  * version 2, as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope it will be useful, but WITHOUT
9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
11  * more details.
12  *
13  * You should have received a copy of the GNU General Public License along with
14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
15  * Place - Suite 330, Boston, MA 02111-1307 USA.
16  *
17  * Authors:
18  *   Haiyang Zhang <haiyangz@microsoft.com>
19  *   Hank Janssen  <hjanssen@microsoft.com>
20  *
21  */
22 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
23
24 #include <linux/kernel.h>
25 #include <linux/mm.h>
26 #include <linux/slab.h>
27 #include <linux/vmalloc.h>
28 #include <linux/hyperv.h>
29 #include <linux/version.h>
30 #include <linux/interrupt.h>
31 #include <linux/clockchips.h>
32 #include <asm/hyperv.h>
33 #include <asm/mshyperv.h>
34 #include "hyperv_vmbus.h"
35
36 /* The one and only */
37 struct hv_context hv_context = {
38         .synic_initialized      = false,
39 };
40
41 #define HV_TIMER_FREQUENCY (10 * 1000 * 1000) /* 100ns period */
42 #define HV_MAX_MAX_DELTA_TICKS 0xffffffff
43 #define HV_MIN_DELTA_TICKS 1
44
45 /*
46  * hv_init - Main initialization routine.
47  *
48  * This routine must be called before any other routines in here are called
49  */
50 int hv_init(void)
51 {
52         if (!hv_is_hypercall_page_setup())
53                 return -ENOTSUPP;
54
55         hv_context.cpu_context = alloc_percpu(struct hv_per_cpu_context);
56         if (!hv_context.cpu_context)
57                 return -ENOMEM;
58
59         return 0;
60 }
61
62 /*
63  * hv_post_message - Post a message using the hypervisor message IPC.
64  *
65  * This involves a hypercall.
66  */
67 int hv_post_message(union hv_connection_id connection_id,
68                   enum hv_message_type message_type,
69                   void *payload, size_t payload_size)
70 {
71         struct hv_input_post_message *aligned_msg;
72         struct hv_per_cpu_context *hv_cpu;
73         u64 status;
74
75         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
76                 return -EMSGSIZE;
77
78         hv_cpu = get_cpu_ptr(hv_context.cpu_context);
79         aligned_msg = hv_cpu->post_msg_page;
80         aligned_msg->connectionid = connection_id;
81         aligned_msg->reserved = 0;
82         aligned_msg->message_type = message_type;
83         aligned_msg->payload_size = payload_size;
84         memcpy((void *)aligned_msg->payload, payload, payload_size);
85         put_cpu_ptr(hv_cpu);
86
87         status = hv_do_hypercall(HVCALL_POST_MESSAGE, aligned_msg, NULL);
88
89         return status & 0xFFFF;
90 }
91
92 static int hv_ce_set_next_event(unsigned long delta,
93                                 struct clock_event_device *evt)
94 {
95         u64 current_tick;
96
97         WARN_ON(!clockevent_state_oneshot(evt));
98
99         hv_get_current_tick(current_tick);
100         current_tick += delta;
101         hv_init_timer(HV_X64_MSR_STIMER0_COUNT, current_tick);
102         return 0;
103 }
104
105 static int hv_ce_shutdown(struct clock_event_device *evt)
106 {
107         hv_init_timer(HV_X64_MSR_STIMER0_COUNT, 0);
108         hv_init_timer_config(HV_X64_MSR_STIMER0_CONFIG, 0);
109
110         return 0;
111 }
112
113 static int hv_ce_set_oneshot(struct clock_event_device *evt)
114 {
115         union hv_timer_config timer_cfg;
116
117         timer_cfg.enable = 1;
118         timer_cfg.auto_enable = 1;
119         timer_cfg.sintx = VMBUS_MESSAGE_SINT;
120         hv_init_timer_config(HV_X64_MSR_STIMER0_CONFIG, timer_cfg.as_uint64);
121
122         return 0;
123 }
124
125 static void hv_init_clockevent_device(struct clock_event_device *dev, int cpu)
126 {
127         dev->name = "Hyper-V clockevent";
128         dev->features = CLOCK_EVT_FEAT_ONESHOT;
129         dev->cpumask = cpumask_of(cpu);
130         dev->rating = 1000;
131         /*
132          * Avoid settint dev->owner = THIS_MODULE deliberately as doing so will
133          * result in clockevents_config_and_register() taking additional
134          * references to the hv_vmbus module making it impossible to unload.
135          */
136
137         dev->set_state_shutdown = hv_ce_shutdown;
138         dev->set_state_oneshot = hv_ce_set_oneshot;
139         dev->set_next_event = hv_ce_set_next_event;
140 }
141
142
143 int hv_synic_alloc(void)
144 {
145         int cpu;
146
147         hv_context.hv_numa_map = kzalloc(sizeof(struct cpumask) * nr_node_ids,
148                                          GFP_ATOMIC);
149         if (hv_context.hv_numa_map == NULL) {
150                 pr_err("Unable to allocate NUMA map\n");
151                 goto err;
152         }
153
154         for_each_present_cpu(cpu) {
155                 struct hv_per_cpu_context *hv_cpu
156                         = per_cpu_ptr(hv_context.cpu_context, cpu);
157
158                 memset(hv_cpu, 0, sizeof(*hv_cpu));
159                 tasklet_init(&hv_cpu->msg_dpc,
160                              vmbus_on_msg_dpc, (unsigned long) hv_cpu);
161
162                 hv_cpu->clk_evt = kzalloc(sizeof(struct clock_event_device),
163                                           GFP_KERNEL);
164                 if (hv_cpu->clk_evt == NULL) {
165                         pr_err("Unable to allocate clock event device\n");
166                         goto err;
167                 }
168                 hv_init_clockevent_device(hv_cpu->clk_evt, cpu);
169
170                 hv_cpu->synic_message_page =
171                         (void *)get_zeroed_page(GFP_ATOMIC);
172                 if (hv_cpu->synic_message_page == NULL) {
173                         pr_err("Unable to allocate SYNIC message page\n");
174                         goto err;
175                 }
176
177                 hv_cpu->synic_event_page = (void *)get_zeroed_page(GFP_ATOMIC);
178                 if (hv_cpu->synic_event_page == NULL) {
179                         pr_err("Unable to allocate SYNIC event page\n");
180                         goto err;
181                 }
182
183                 hv_cpu->post_msg_page = (void *)get_zeroed_page(GFP_ATOMIC);
184                 if (hv_cpu->post_msg_page == NULL) {
185                         pr_err("Unable to allocate post msg page\n");
186                         goto err;
187                 }
188
189                 INIT_LIST_HEAD(&hv_cpu->chan_list);
190         }
191
192         return 0;
193 err:
194         return -ENOMEM;
195 }
196
197
198 void hv_synic_free(void)
199 {
200         int cpu;
201
202         for_each_present_cpu(cpu) {
203                 struct hv_per_cpu_context *hv_cpu
204                         = per_cpu_ptr(hv_context.cpu_context, cpu);
205
206                 if (hv_cpu->synic_event_page)
207                         free_page((unsigned long)hv_cpu->synic_event_page);
208                 if (hv_cpu->synic_message_page)
209                         free_page((unsigned long)hv_cpu->synic_message_page);
210                 if (hv_cpu->post_msg_page)
211                         free_page((unsigned long)hv_cpu->post_msg_page);
212         }
213
214         kfree(hv_context.hv_numa_map);
215 }
216
217 /*
218  * hv_synic_init - Initialize the Synthethic Interrupt Controller.
219  *
220  * If it is already initialized by another entity (ie x2v shim), we need to
221  * retrieve the initialized message and event pages.  Otherwise, we create and
222  * initialize the message and event pages.
223  */
224 int hv_synic_init(unsigned int cpu)
225 {
226         struct hv_per_cpu_context *hv_cpu
227                 = per_cpu_ptr(hv_context.cpu_context, cpu);
228         union hv_synic_simp simp;
229         union hv_synic_siefp siefp;
230         union hv_synic_sint shared_sint;
231         union hv_synic_scontrol sctrl;
232         u64 vp_index;
233
234         /* Setup the Synic's message page */
235         hv_get_simp(simp.as_uint64);
236         simp.simp_enabled = 1;
237         simp.base_simp_gpa = virt_to_phys(hv_cpu->synic_message_page)
238                 >> PAGE_SHIFT;
239
240         hv_set_simp(simp.as_uint64);
241
242         /* Setup the Synic's event page */
243         hv_get_siefp(siefp.as_uint64);
244         siefp.siefp_enabled = 1;
245         siefp.base_siefp_gpa = virt_to_phys(hv_cpu->synic_event_page)
246                 >> PAGE_SHIFT;
247
248         hv_set_siefp(siefp.as_uint64);
249
250         /* Setup the shared SINT. */
251         hv_get_synint_state(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT,
252                             shared_sint.as_uint64);
253
254         shared_sint.as_uint64 = 0;
255         shared_sint.vector = HYPERVISOR_CALLBACK_VECTOR;
256         shared_sint.masked = false;
257         shared_sint.auto_eoi = true;
258
259         hv_set_synint_state(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT,
260                             shared_sint.as_uint64);
261
262         /* Enable the global synic bit */
263         hv_get_synic_state(sctrl.as_uint64);
264         sctrl.enable = 1;
265
266         hv_set_synic_state(sctrl.as_uint64);
267
268         hv_context.synic_initialized = true;
269
270         /*
271          * Setup the mapping between Hyper-V's notion
272          * of cpuid and Linux' notion of cpuid.
273          * This array will be indexed using Linux cpuid.
274          */
275         hv_get_vp_index(vp_index);
276         hv_context.vp_index[cpu] = (u32)vp_index;
277
278         /*
279          * Register the per-cpu clockevent source.
280          */
281         if (ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE)
282                 clockevents_config_and_register(hv_cpu->clk_evt,
283                                                 HV_TIMER_FREQUENCY,
284                                                 HV_MIN_DELTA_TICKS,
285                                                 HV_MAX_MAX_DELTA_TICKS);
286         return 0;
287 }
288
289 /*
290  * hv_synic_clockevents_cleanup - Cleanup clockevent devices
291  */
292 void hv_synic_clockevents_cleanup(void)
293 {
294         int cpu;
295
296         if (!(ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE))
297                 return;
298
299         for_each_present_cpu(cpu) {
300                 struct hv_per_cpu_context *hv_cpu
301                         = per_cpu_ptr(hv_context.cpu_context, cpu);
302
303                 clockevents_unbind_device(hv_cpu->clk_evt, cpu);
304         }
305 }
306
307 /*
308  * hv_synic_cleanup - Cleanup routine for hv_synic_init().
309  */
310 int hv_synic_cleanup(unsigned int cpu)
311 {
312         union hv_synic_sint shared_sint;
313         union hv_synic_simp simp;
314         union hv_synic_siefp siefp;
315         union hv_synic_scontrol sctrl;
316         struct vmbus_channel *channel, *sc;
317         bool channel_found = false;
318         unsigned long flags;
319
320         if (!hv_context.synic_initialized)
321                 return -EFAULT;
322
323         /*
324          * Search for channels which are bound to the CPU we're about to
325          * cleanup. In case we find one and vmbus is still connected we need to
326          * fail, this will effectively prevent CPU offlining. There is no way
327          * we can re-bind channels to different CPUs for now.
328          */
329         mutex_lock(&vmbus_connection.channel_mutex);
330         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
331                 if (channel->target_cpu == cpu) {
332                         channel_found = true;
333                         break;
334                 }
335                 spin_lock_irqsave(&channel->lock, flags);
336                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
337                         if (sc->target_cpu == cpu) {
338                                 channel_found = true;
339                                 break;
340                         }
341                 }
342                 spin_unlock_irqrestore(&channel->lock, flags);
343                 if (channel_found)
344                         break;
345         }
346         mutex_unlock(&vmbus_connection.channel_mutex);
347
348         if (channel_found && vmbus_connection.conn_state == CONNECTED)
349                 return -EBUSY;
350
351         /* Turn off clockevent device */
352         if (ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE) {
353                 struct hv_per_cpu_context *hv_cpu
354                         = this_cpu_ptr(hv_context.cpu_context);
355
356                 clockevents_unbind_device(hv_cpu->clk_evt, cpu);
357                 hv_ce_shutdown(hv_cpu->clk_evt);
358                 put_cpu_ptr(hv_cpu);
359         }
360
361         hv_get_synint_state(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT,
362                             shared_sint.as_uint64);
363
364         shared_sint.masked = 1;
365
366         /* Need to correctly cleanup in the case of SMP!!! */
367         /* Disable the interrupt */
368         hv_set_synint_state(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT,
369                             shared_sint.as_uint64);
370
371         hv_get_simp(simp.as_uint64);
372         simp.simp_enabled = 0;
373         simp.base_simp_gpa = 0;
374
375         hv_set_simp(simp.as_uint64);
376
377         hv_get_siefp(siefp.as_uint64);
378         siefp.siefp_enabled = 0;
379         siefp.base_siefp_gpa = 0;
380
381         hv_set_siefp(siefp.as_uint64);
382
383         /* Disable the global synic bit */
384         hv_get_synic_state(sctrl.as_uint64);
385         sctrl.enable = 0;
386         hv_set_synic_state(sctrl.as_uint64);
387
388         return 0;
389 }