]> asedeno.scripts.mit.edu Git - linux.git/blob - drivers/hv/hv.c
vfs: do bulk POLL* -> EPOLL* replacement
[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         hv_context.cpu_context = alloc_percpu(struct hv_per_cpu_context);
53         if (!hv_context.cpu_context)
54                 return -ENOMEM;
55
56         return 0;
57 }
58
59 /*
60  * hv_post_message - Post a message using the hypervisor message IPC.
61  *
62  * This involves a hypercall.
63  */
64 int hv_post_message(union hv_connection_id connection_id,
65                   enum hv_message_type message_type,
66                   void *payload, size_t payload_size)
67 {
68         struct hv_input_post_message *aligned_msg;
69         struct hv_per_cpu_context *hv_cpu;
70         u64 status;
71
72         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
73                 return -EMSGSIZE;
74
75         hv_cpu = get_cpu_ptr(hv_context.cpu_context);
76         aligned_msg = hv_cpu->post_msg_page;
77         aligned_msg->connectionid = connection_id;
78         aligned_msg->reserved = 0;
79         aligned_msg->message_type = message_type;
80         aligned_msg->payload_size = payload_size;
81         memcpy((void *)aligned_msg->payload, payload, payload_size);
82
83         status = hv_do_hypercall(HVCALL_POST_MESSAGE, aligned_msg, NULL);
84
85         /* Preemption must remain disabled until after the hypercall
86          * so some other thread can't get scheduled onto this cpu and
87          * corrupt the per-cpu post_msg_page
88          */
89         put_cpu_ptr(hv_cpu);
90
91         return status & 0xFFFF;
92 }
93
94 static int hv_ce_set_next_event(unsigned long delta,
95                                 struct clock_event_device *evt)
96 {
97         u64 current_tick;
98
99         WARN_ON(!clockevent_state_oneshot(evt));
100
101         current_tick = hyperv_cs->read(NULL);
102         current_tick += delta;
103         hv_init_timer(HV_X64_MSR_STIMER0_COUNT, current_tick);
104         return 0;
105 }
106
107 static int hv_ce_shutdown(struct clock_event_device *evt)
108 {
109         hv_init_timer(HV_X64_MSR_STIMER0_COUNT, 0);
110         hv_init_timer_config(HV_X64_MSR_STIMER0_CONFIG, 0);
111
112         return 0;
113 }
114
115 static int hv_ce_set_oneshot(struct clock_event_device *evt)
116 {
117         union hv_timer_config timer_cfg;
118
119         timer_cfg.enable = 1;
120         timer_cfg.auto_enable = 1;
121         timer_cfg.sintx = VMBUS_MESSAGE_SINT;
122         hv_init_timer_config(HV_X64_MSR_STIMER0_CONFIG, timer_cfg.as_uint64);
123
124         return 0;
125 }
126
127 static void hv_init_clockevent_device(struct clock_event_device *dev, int cpu)
128 {
129         dev->name = "Hyper-V clockevent";
130         dev->features = CLOCK_EVT_FEAT_ONESHOT;
131         dev->cpumask = cpumask_of(cpu);
132         dev->rating = 1000;
133         /*
134          * Avoid settint dev->owner = THIS_MODULE deliberately as doing so will
135          * result in clockevents_config_and_register() taking additional
136          * references to the hv_vmbus module making it impossible to unload.
137          */
138
139         dev->set_state_shutdown = hv_ce_shutdown;
140         dev->set_state_oneshot = hv_ce_set_oneshot;
141         dev->set_next_event = hv_ce_set_next_event;
142 }
143
144
145 int hv_synic_alloc(void)
146 {
147         int cpu;
148
149         hv_context.hv_numa_map = kzalloc(sizeof(struct cpumask) * nr_node_ids,
150                                          GFP_ATOMIC);
151         if (hv_context.hv_numa_map == NULL) {
152                 pr_err("Unable to allocate NUMA map\n");
153                 goto err;
154         }
155
156         for_each_present_cpu(cpu) {
157                 struct hv_per_cpu_context *hv_cpu
158                         = per_cpu_ptr(hv_context.cpu_context, cpu);
159
160                 memset(hv_cpu, 0, sizeof(*hv_cpu));
161                 tasklet_init(&hv_cpu->msg_dpc,
162                              vmbus_on_msg_dpc, (unsigned long) hv_cpu);
163
164                 hv_cpu->clk_evt = kzalloc(sizeof(struct clock_event_device),
165                                           GFP_KERNEL);
166                 if (hv_cpu->clk_evt == NULL) {
167                         pr_err("Unable to allocate clock event device\n");
168                         goto err;
169                 }
170                 hv_init_clockevent_device(hv_cpu->clk_evt, cpu);
171
172                 hv_cpu->synic_message_page =
173                         (void *)get_zeroed_page(GFP_ATOMIC);
174                 if (hv_cpu->synic_message_page == NULL) {
175                         pr_err("Unable to allocate SYNIC message page\n");
176                         goto err;
177                 }
178
179                 hv_cpu->synic_event_page = (void *)get_zeroed_page(GFP_ATOMIC);
180                 if (hv_cpu->synic_event_page == NULL) {
181                         pr_err("Unable to allocate SYNIC event page\n");
182                         goto err;
183                 }
184
185                 hv_cpu->post_msg_page = (void *)get_zeroed_page(GFP_ATOMIC);
186                 if (hv_cpu->post_msg_page == NULL) {
187                         pr_err("Unable to allocate post msg page\n");
188                         goto err;
189                 }
190
191                 INIT_LIST_HEAD(&hv_cpu->chan_list);
192         }
193
194         return 0;
195 err:
196         return -ENOMEM;
197 }
198
199
200 void hv_synic_free(void)
201 {
202         int cpu;
203
204         for_each_present_cpu(cpu) {
205                 struct hv_per_cpu_context *hv_cpu
206                         = per_cpu_ptr(hv_context.cpu_context, cpu);
207
208                 if (hv_cpu->synic_event_page)
209                         free_page((unsigned long)hv_cpu->synic_event_page);
210                 if (hv_cpu->synic_message_page)
211                         free_page((unsigned long)hv_cpu->synic_message_page);
212                 if (hv_cpu->post_msg_page)
213                         free_page((unsigned long)hv_cpu->post_msg_page);
214         }
215
216         kfree(hv_context.hv_numa_map);
217 }
218
219 /*
220  * hv_synic_init - Initialize the Synthethic Interrupt Controller.
221  *
222  * If it is already initialized by another entity (ie x2v shim), we need to
223  * retrieve the initialized message and event pages.  Otherwise, we create and
224  * initialize the message and event pages.
225  */
226 int hv_synic_init(unsigned int cpu)
227 {
228         struct hv_per_cpu_context *hv_cpu
229                 = per_cpu_ptr(hv_context.cpu_context, cpu);
230         union hv_synic_simp simp;
231         union hv_synic_siefp siefp;
232         union hv_synic_sint shared_sint;
233         union hv_synic_scontrol sctrl;
234
235         /* Setup the Synic's message page */
236         hv_get_simp(simp.as_uint64);
237         simp.simp_enabled = 1;
238         simp.base_simp_gpa = virt_to_phys(hv_cpu->synic_message_page)
239                 >> PAGE_SHIFT;
240
241         hv_set_simp(simp.as_uint64);
242
243         /* Setup the Synic's event page */
244         hv_get_siefp(siefp.as_uint64);
245         siefp.siefp_enabled = 1;
246         siefp.base_siefp_gpa = virt_to_phys(hv_cpu->synic_event_page)
247                 >> PAGE_SHIFT;
248
249         hv_set_siefp(siefp.as_uint64);
250
251         /* Setup the shared SINT. */
252         hv_get_synint_state(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT,
253                             shared_sint.as_uint64);
254
255         shared_sint.as_uint64 = 0;
256         shared_sint.vector = HYPERVISOR_CALLBACK_VECTOR;
257         shared_sint.masked = false;
258         if (ms_hyperv.hints & HV_X64_DEPRECATING_AEOI_RECOMMENDED)
259                 shared_sint.auto_eoi = false;
260         else
261                 shared_sint.auto_eoi = true;
262
263         hv_set_synint_state(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT,
264                             shared_sint.as_uint64);
265
266         /* Enable the global synic bit */
267         hv_get_synic_state(sctrl.as_uint64);
268         sctrl.enable = 1;
269
270         hv_set_synic_state(sctrl.as_uint64);
271
272         hv_context.synic_initialized = true;
273
274         /*
275          * Register the per-cpu clockevent source.
276          */
277         if (ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE)
278                 clockevents_config_and_register(hv_cpu->clk_evt,
279                                                 HV_TIMER_FREQUENCY,
280                                                 HV_MIN_DELTA_TICKS,
281                                                 HV_MAX_MAX_DELTA_TICKS);
282         return 0;
283 }
284
285 /*
286  * hv_synic_clockevents_cleanup - Cleanup clockevent devices
287  */
288 void hv_synic_clockevents_cleanup(void)
289 {
290         int cpu;
291
292         if (!(ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE))
293                 return;
294
295         for_each_present_cpu(cpu) {
296                 struct hv_per_cpu_context *hv_cpu
297                         = per_cpu_ptr(hv_context.cpu_context, cpu);
298
299                 clockevents_unbind_device(hv_cpu->clk_evt, cpu);
300         }
301 }
302
303 /*
304  * hv_synic_cleanup - Cleanup routine for hv_synic_init().
305  */
306 int hv_synic_cleanup(unsigned int cpu)
307 {
308         union hv_synic_sint shared_sint;
309         union hv_synic_simp simp;
310         union hv_synic_siefp siefp;
311         union hv_synic_scontrol sctrl;
312         struct vmbus_channel *channel, *sc;
313         bool channel_found = false;
314         unsigned long flags;
315
316         if (!hv_context.synic_initialized)
317                 return -EFAULT;
318
319         /*
320          * Search for channels which are bound to the CPU we're about to
321          * cleanup. In case we find one and vmbus is still connected we need to
322          * fail, this will effectively prevent CPU offlining. There is no way
323          * we can re-bind channels to different CPUs for now.
324          */
325         mutex_lock(&vmbus_connection.channel_mutex);
326         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
327                 if (channel->target_cpu == cpu) {
328                         channel_found = true;
329                         break;
330                 }
331                 spin_lock_irqsave(&channel->lock, flags);
332                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
333                         if (sc->target_cpu == cpu) {
334                                 channel_found = true;
335                                 break;
336                         }
337                 }
338                 spin_unlock_irqrestore(&channel->lock, flags);
339                 if (channel_found)
340                         break;
341         }
342         mutex_unlock(&vmbus_connection.channel_mutex);
343
344         if (channel_found && vmbus_connection.conn_state == CONNECTED)
345                 return -EBUSY;
346
347         /* Turn off clockevent device */
348         if (ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE) {
349                 struct hv_per_cpu_context *hv_cpu
350                         = this_cpu_ptr(hv_context.cpu_context);
351
352                 clockevents_unbind_device(hv_cpu->clk_evt, cpu);
353                 hv_ce_shutdown(hv_cpu->clk_evt);
354                 put_cpu_ptr(hv_cpu);
355         }
356
357         hv_get_synint_state(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT,
358                             shared_sint.as_uint64);
359
360         shared_sint.masked = 1;
361
362         /* Need to correctly cleanup in the case of SMP!!! */
363         /* Disable the interrupt */
364         hv_set_synint_state(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT,
365                             shared_sint.as_uint64);
366
367         hv_get_simp(simp.as_uint64);
368         simp.simp_enabled = 0;
369         simp.base_simp_gpa = 0;
370
371         hv_set_simp(simp.as_uint64);
372
373         hv_get_siefp(siefp.as_uint64);
374         siefp.siefp_enabled = 0;
375         siefp.base_siefp_gpa = 0;
376
377         hv_set_siefp(siefp.as_uint64);
378
379         /* Disable the global synic bit */
380         hv_get_synic_state(sctrl.as_uint64);
381         sctrl.enable = 0;
382         hv_set_synic_state(sctrl.as_uint64);
383
384         return 0;
385 }