2 * This program is free software; you can redistribute it and/or modify
3 * it under the terms of the GNU General Public License, version 2, as
4 * published by the Free Software Foundation.
6 * Copyright 2012 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 #include <linux/types.h>
10 #include <linux/string.h>
11 #include <linux/kvm.h>
12 #include <linux/kvm_host.h>
13 #include <linux/kernel.h>
16 #include <asm/machdep.h>
17 #include <asm/cputhreads.h>
19 #include <asm/kvm_ppc.h>
21 /* SRR1 bits for machine check on POWER7 */
22 #define SRR1_MC_LDSTERR (1ul << (63-42))
23 #define SRR1_MC_IFETCH_SH (63-45)
24 #define SRR1_MC_IFETCH_MASK 0x7
25 #define SRR1_MC_IFETCH_SLBPAR 2 /* SLB parity error */
26 #define SRR1_MC_IFETCH_SLBMULTI 3 /* SLB multi-hit */
27 #define SRR1_MC_IFETCH_SLBPARMULTI 4 /* SLB parity + multi-hit */
28 #define SRR1_MC_IFETCH_TLBMULTI 5 /* I-TLB multi-hit */
30 /* DSISR bits for machine check on POWER7 */
31 #define DSISR_MC_DERAT_MULTI 0x800 /* D-ERAT multi-hit */
32 #define DSISR_MC_TLB_MULTI 0x400 /* D-TLB multi-hit */
33 #define DSISR_MC_SLB_PARITY 0x100 /* SLB parity error */
34 #define DSISR_MC_SLB_MULTI 0x080 /* SLB multi-hit */
35 #define DSISR_MC_SLB_PARMULTI 0x040 /* SLB parity + multi-hit */
37 /* POWER7 SLB flush and reload */
38 static void reload_slb(struct kvm_vcpu *vcpu)
40 struct slb_shadow *slb;
43 /* First clear out SLB */
44 asm volatile("slbmte %0,%0; slbia" : : "r" (0));
46 /* Do they have an SLB shadow buffer registered? */
47 slb = vcpu->arch.slb_shadow.pinned_addr;
52 n = min_t(u32, be32_to_cpu(slb->persistent), SLB_MIN_SIZE);
53 if ((void *) &slb->save_area[n] > vcpu->arch.slb_shadow.pinned_end)
56 /* Load up the SLB from that */
57 for (i = 0; i < n; ++i) {
58 unsigned long rb = be64_to_cpu(slb->save_area[i].esid);
59 unsigned long rs = be64_to_cpu(slb->save_area[i].vsid);
61 rb = (rb & ~0xFFFul) | i; /* insert entry number */
62 asm volatile("slbmte %0,%1" : : "r" (rs), "r" (rb));
67 * On POWER7, see if we can handle a machine check that occurred inside
68 * the guest in real mode, without switching to the host partition.
70 * Returns: 0 => exit guest, 1 => deliver machine check to guest
72 static long kvmppc_realmode_mc_power7(struct kvm_vcpu *vcpu)
74 unsigned long srr1 = vcpu->arch.shregs.msr;
75 struct machine_check_event mce_evt;
78 if (srr1 & SRR1_MC_LDSTERR) {
79 /* error on load/store */
80 unsigned long dsisr = vcpu->arch.shregs.dsisr;
82 if (dsisr & (DSISR_MC_SLB_PARMULTI | DSISR_MC_SLB_MULTI |
83 DSISR_MC_SLB_PARITY | DSISR_MC_DERAT_MULTI)) {
84 /* flush and reload SLB; flushes D-ERAT too */
86 dsisr &= ~(DSISR_MC_SLB_PARMULTI | DSISR_MC_SLB_MULTI |
87 DSISR_MC_SLB_PARITY | DSISR_MC_DERAT_MULTI);
89 if (dsisr & DSISR_MC_TLB_MULTI) {
90 if (cur_cpu_spec && cur_cpu_spec->flush_tlb)
91 cur_cpu_spec->flush_tlb(TLB_INVAL_SCOPE_LPID);
92 dsisr &= ~DSISR_MC_TLB_MULTI;
94 /* Any other errors we don't understand? */
95 if (dsisr & 0xffffffffUL)
99 switch ((srr1 >> SRR1_MC_IFETCH_SH) & SRR1_MC_IFETCH_MASK) {
102 case SRR1_MC_IFETCH_SLBPAR:
103 case SRR1_MC_IFETCH_SLBMULTI:
104 case SRR1_MC_IFETCH_SLBPARMULTI:
107 case SRR1_MC_IFETCH_TLBMULTI:
108 if (cur_cpu_spec && cur_cpu_spec->flush_tlb)
109 cur_cpu_spec->flush_tlb(TLB_INVAL_SCOPE_LPID);
116 * See if we have already handled the condition in the linux host.
117 * We assume that if the condition is recovered then linux host
118 * will have generated an error log event that we will pick
120 * Don't release mce event now. We will queue up the event so that
121 * we can log the MCE event info on host console.
123 if (!get_mce_event(&mce_evt, MCE_EVENT_DONTRELEASE))
126 if (mce_evt.version == MCE_V1 &&
127 (mce_evt.severity == MCE_SEV_NO_ERROR ||
128 mce_evt.disposition == MCE_DISPOSITION_RECOVERED))
133 * For guest that supports FWNMI capability, hook the MCE event into
134 * vcpu structure. We are going to exit the guest with KVM_EXIT_NMI
135 * exit reason. On our way to exit we will pull this event from vcpu
136 * structure and print it from thread 0 of the core/subcore.
138 * For guest that does not support FWNMI capability (old QEMU):
139 * We are now going enter guest either through machine check
140 * interrupt (for unhandled errors) or will continue from
141 * current HSRR0 (for handled errors) in guest. Hence
142 * queue up the event so that we can log it from host console later.
144 if (vcpu->kvm->arch.fwnmi_enabled) {
146 * Hook up the mce event on to vcpu structure.
147 * First clear the old event.
149 memset(&vcpu->arch.mce_evt, 0, sizeof(vcpu->arch.mce_evt));
150 if (get_mce_event(&mce_evt, MCE_EVENT_RELEASE)) {
151 vcpu->arch.mce_evt = mce_evt;
154 machine_check_queue_event();
159 long kvmppc_realmode_machine_check(struct kvm_vcpu *vcpu)
161 return kvmppc_realmode_mc_power7(vcpu);
164 /* Check if dynamic split is in force and return subcore size accordingly. */
165 static inline int kvmppc_cur_subcore_size(void)
167 if (local_paca->kvm_hstate.kvm_split_mode)
168 return local_paca->kvm_hstate.kvm_split_mode->subcore_size;
170 return threads_per_subcore;
173 void kvmppc_subcore_enter_guest(void)
175 int thread_id, subcore_id;
177 thread_id = cpu_thread_in_core(local_paca->paca_index);
178 subcore_id = thread_id / kvmppc_cur_subcore_size();
180 local_paca->sibling_subcore_state->in_guest[subcore_id] = 1;
183 void kvmppc_subcore_exit_guest(void)
185 int thread_id, subcore_id;
187 thread_id = cpu_thread_in_core(local_paca->paca_index);
188 subcore_id = thread_id / kvmppc_cur_subcore_size();
190 local_paca->sibling_subcore_state->in_guest[subcore_id] = 0;
193 static bool kvmppc_tb_resync_required(void)
195 if (test_and_set_bit(CORE_TB_RESYNC_REQ_BIT,
196 &local_paca->sibling_subcore_state->flags))
202 static void kvmppc_tb_resync_done(void)
204 clear_bit(CORE_TB_RESYNC_REQ_BIT,
205 &local_paca->sibling_subcore_state->flags);
209 * kvmppc_realmode_hmi_handler() is called only by primary thread during
212 * There are multiple reasons why HMI could occur, one of them is
213 * Timebase (TB) error. If this HMI is due to TB error, then TB would
214 * have been in stopped state. The opal hmi handler Will fix it and
215 * restore the TB value with host timebase value. For HMI caused due
216 * to non-TB errors, opal hmi handler will not touch/restore TB register
217 * and hence there won't be any change in TB value.
219 * Since we are not sure about the cause of this HMI, we can't be sure
220 * about the content of TB register whether it holds guest or host timebase
221 * value. Hence the idea is to resync the TB on every HMI, so that we
222 * know about the exact state of the TB value. Resync TB call will
223 * restore TB to host timebase.
225 * Things to consider:
226 * - On TB error, HMI interrupt is reported on all the threads of the core
227 * that has encountered TB error irrespective of split-core mode.
228 * - The very first thread on the core that get chance to fix TB error
229 * would rsync the TB with local chipTOD value.
230 * - The resync TB is a core level action i.e. it will sync all the TBs
231 * in that core independent of split-core mode. This means if we trigger
232 * TB sync from a thread from one subcore, it would affect TB values of
233 * sibling subcores of the same core.
235 * All threads need to co-ordinate before making opal hmi handler.
236 * All threads will use sibling_subcore_state->in_guest[] (shared by all
237 * threads in the core) in paca which holds information about whether
238 * sibling subcores are in Guest mode or host mode. The in_guest[] array
239 * is of size MAX_SUBCORE_PER_CORE=4, indexed using subcore id to set/unset
240 * subcore status. Only primary threads from each subcore is responsible
241 * to set/unset its designated array element while entering/exiting the
244 * After invoking opal hmi handler call, one of the thread (of entire core)
245 * will need to resync the TB. Bit 63 from subcore state bitmap flags
246 * (sibling_subcore_state->flags) will be used to co-ordinate between
247 * primary threads to decide who takes up the responsibility.
249 * This is what we do:
250 * - Primary thread from each subcore tries to set resync required bit[63]
251 * of paca->sibling_subcore_state->flags.
252 * - The first primary thread that is able to set the flag takes the
253 * responsibility of TB resync. (Let us call it as thread leader)
254 * - All other threads which are in host will call
255 * wait_for_subcore_guest_exit() and wait for in_guest[0-3] from
256 * paca->sibling_subcore_state to get cleared.
257 * - All the primary thread will clear its subcore status from subcore
258 * state in_guest[] array respectively.
259 * - Once all primary threads clear in_guest[0-3], all of them will invoke
261 * - Now all threads will wait for TB resync to complete by invoking
262 * wait_for_tb_resync() except the thread leader.
263 * - Thread leader will do a TB resync by invoking opal_resync_timebase()
264 * call and the it will clear the resync required bit.
265 * - All other threads will now come out of resync wait loop and proceed
266 * with individual execution.
267 * - On return of this function, primary thread will signal all
268 * secondary threads to proceed.
269 * - All secondary threads will eventually call opal hmi handler on
273 long kvmppc_realmode_hmi_handler(void)
275 int ptid = local_paca->kvm_hstate.ptid;
278 /* This is only called on primary thread. */
280 __this_cpu_inc(irq_stat.hmi_exceptions);
283 * By now primary thread has already completed guest->host
284 * partition switch but haven't signaled secondaries yet.
285 * All the secondary threads on this subcore is waiting
286 * for primary thread to signal them to go ahead.
288 * For threads from subcore which isn't in guest, they all will
289 * wait until all other subcores on this core exit the guest.
291 * Now set the resync required bit. If you are the first to
292 * set this bit then kvmppc_tb_resync_required() function will
293 * return true. For rest all other subcores
294 * kvmppc_tb_resync_required() will return false.
296 * If resync_req == true, then this thread is responsible to
297 * initiate TB resync after hmi handler has completed.
298 * All other threads on this core will wait until this thread
299 * clears the resync required bit flag.
301 resync_req = kvmppc_tb_resync_required();
303 /* Reset the subcore status to indicate it has exited guest */
304 kvmppc_subcore_exit_guest();
307 * Wait for other subcores on this core to exit the guest.
308 * All the primary threads and threads from subcore that are
309 * not in guest will wait here until all subcores are out
312 wait_for_subcore_guest_exit();
315 * At this point we are sure that primary threads from each
316 * subcore on this core have completed guest->host partition
317 * switch. Now it is safe to call HMI handler.
319 if (ppc_md.hmi_exception_early)
320 ppc_md.hmi_exception_early(NULL);
323 * Check if this thread is responsible to resync TB.
324 * All other threads will wait until this thread completes the
328 opal_resync_timebase();
329 /* Reset TB resync req bit */
330 kvmppc_tb_resync_done();
332 wait_for_tb_resync();