xref: /openbmc/linux/arch/powerpc/kvm/book3s_hv_ras.c (revision 2cf1c348)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *
4  * Copyright 2012 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
5  */
6 
7 #include <linux/types.h>
8 #include <linux/string.h>
9 #include <linux/kvm.h>
10 #include <linux/kvm_host.h>
11 #include <linux/kernel.h>
12 #include <asm/opal.h>
13 #include <asm/mce.h>
14 #include <asm/machdep.h>
15 #include <asm/cputhreads.h>
16 #include <asm/hmi.h>
17 #include <asm/kvm_ppc.h>
18 
19 /* SRR1 bits for machine check on POWER7 */
20 #define SRR1_MC_LDSTERR		(1ul << (63-42))
21 #define SRR1_MC_IFETCH_SH	(63-45)
22 #define SRR1_MC_IFETCH_MASK	0x7
23 #define SRR1_MC_IFETCH_SLBPAR		2	/* SLB parity error */
24 #define SRR1_MC_IFETCH_SLBMULTI		3	/* SLB multi-hit */
25 #define SRR1_MC_IFETCH_SLBPARMULTI	4	/* SLB parity + multi-hit */
26 #define SRR1_MC_IFETCH_TLBMULTI		5	/* I-TLB multi-hit */
27 
28 /* DSISR bits for machine check on POWER7 */
29 #define DSISR_MC_DERAT_MULTI	0x800		/* D-ERAT multi-hit */
30 #define DSISR_MC_TLB_MULTI	0x400		/* D-TLB multi-hit */
31 #define DSISR_MC_SLB_PARITY	0x100		/* SLB parity error */
32 #define DSISR_MC_SLB_MULTI	0x080		/* SLB multi-hit */
33 #define DSISR_MC_SLB_PARMULTI	0x040		/* SLB parity + multi-hit */
34 
35 /* POWER7 SLB flush and reload */
36 static void reload_slb(struct kvm_vcpu *vcpu)
37 {
38 	struct slb_shadow *slb;
39 	unsigned long i, n;
40 
41 	/* First clear out SLB */
42 	asm volatile("slbmte %0,%0; slbia" : : "r" (0));
43 
44 	/* Do they have an SLB shadow buffer registered? */
45 	slb = vcpu->arch.slb_shadow.pinned_addr;
46 	if (!slb)
47 		return;
48 
49 	/* Sanity check */
50 	n = min_t(u32, be32_to_cpu(slb->persistent), SLB_MIN_SIZE);
51 	if ((void *) &slb->save_area[n] > vcpu->arch.slb_shadow.pinned_end)
52 		return;
53 
54 	/* Load up the SLB from that */
55 	for (i = 0; i < n; ++i) {
56 		unsigned long rb = be64_to_cpu(slb->save_area[i].esid);
57 		unsigned long rs = be64_to_cpu(slb->save_area[i].vsid);
58 
59 		rb = (rb & ~0xFFFul) | i;	/* insert entry number */
60 		asm volatile("slbmte %0,%1" : : "r" (rs), "r" (rb));
61 	}
62 }
63 
64 /*
65  * On POWER7, see if we can handle a machine check that occurred inside
66  * the guest in real mode, without switching to the host partition.
67  */
68 static long kvmppc_realmode_mc_power7(struct kvm_vcpu *vcpu)
69 {
70 	unsigned long srr1 = vcpu->arch.shregs.msr;
71 	long handled = 1;
72 
73 	if (srr1 & SRR1_MC_LDSTERR) {
74 		/* error on load/store */
75 		unsigned long dsisr = vcpu->arch.shregs.dsisr;
76 
77 		if (dsisr & (DSISR_MC_SLB_PARMULTI | DSISR_MC_SLB_MULTI |
78 			     DSISR_MC_SLB_PARITY | DSISR_MC_DERAT_MULTI)) {
79 			/* flush and reload SLB; flushes D-ERAT too */
80 			reload_slb(vcpu);
81 			dsisr &= ~(DSISR_MC_SLB_PARMULTI | DSISR_MC_SLB_MULTI |
82 				   DSISR_MC_SLB_PARITY | DSISR_MC_DERAT_MULTI);
83 		}
84 		if (dsisr & DSISR_MC_TLB_MULTI) {
85 			tlbiel_all_lpid(vcpu->kvm->arch.radix);
86 			dsisr &= ~DSISR_MC_TLB_MULTI;
87 		}
88 		/* Any other errors we don't understand? */
89 		if (dsisr & 0xffffffffUL)
90 			handled = 0;
91 	}
92 
93 	switch ((srr1 >> SRR1_MC_IFETCH_SH) & SRR1_MC_IFETCH_MASK) {
94 	case 0:
95 		break;
96 	case SRR1_MC_IFETCH_SLBPAR:
97 	case SRR1_MC_IFETCH_SLBMULTI:
98 	case SRR1_MC_IFETCH_SLBPARMULTI:
99 		reload_slb(vcpu);
100 		break;
101 	case SRR1_MC_IFETCH_TLBMULTI:
102 		tlbiel_all_lpid(vcpu->kvm->arch.radix);
103 		break;
104 	default:
105 		handled = 0;
106 	}
107 
108 	return handled;
109 }
110 
111 void kvmppc_realmode_machine_check(struct kvm_vcpu *vcpu)
112 {
113 	struct machine_check_event mce_evt;
114 	long handled;
115 
116 	if (vcpu->kvm->arch.fwnmi_enabled) {
117 		/* FWNMI guests handle their own recovery */
118 		handled = 0;
119 	} else {
120 		handled = kvmppc_realmode_mc_power7(vcpu);
121 	}
122 
123 	/*
124 	 * Now get the event and stash it in the vcpu struct so it can
125 	 * be handled by the primary thread in virtual mode.  We can't
126 	 * call machine_check_queue_event() here if we are running on
127 	 * an offline secondary thread.
128 	 */
129 	if (get_mce_event(&mce_evt, MCE_EVENT_RELEASE)) {
130 		if (handled && mce_evt.version == MCE_V1)
131 			mce_evt.disposition = MCE_DISPOSITION_RECOVERED;
132 	} else {
133 		memset(&mce_evt, 0, sizeof(mce_evt));
134 	}
135 
136 	vcpu->arch.mce_evt = mce_evt;
137 }
138 
139 
140 long kvmppc_p9_realmode_hmi_handler(struct kvm_vcpu *vcpu)
141 {
142 	struct kvmppc_vcore *vc = vcpu->arch.vcore;
143 	long ret = 0;
144 
145 	/*
146 	 * Unapply and clear the offset first. That way, if the TB was not
147 	 * resynced then it will remain in host-offset, and if it was resynced
148 	 * then it is brought into host-offset. Then the tb offset is
149 	 * re-applied before continuing with the KVM exit.
150 	 *
151 	 * This way, we don't need to actually know whether not OPAL resynced
152 	 * the timebase or do any of the complicated dance that the P7/8
153 	 * path requires.
154 	 */
155 	if (vc->tb_offset_applied) {
156 		u64 new_tb = mftb() - vc->tb_offset_applied;
157 		mtspr(SPRN_TBU40, new_tb);
158 		if ((mftb() & 0xffffff) < (new_tb & 0xffffff)) {
159 			new_tb += 0x1000000;
160 			mtspr(SPRN_TBU40, new_tb);
161 		}
162 		vc->tb_offset_applied = 0;
163 	}
164 
165 	local_paca->hmi_irqs++;
166 
167 	if (hmi_handle_debugtrig(NULL) >= 0) {
168 		ret = 1;
169 		goto out;
170 	}
171 
172 	if (ppc_md.hmi_exception_early)
173 		ppc_md.hmi_exception_early(NULL);
174 
175 out:
176 	if (vc->tb_offset) {
177 		u64 new_tb = mftb() + vc->tb_offset;
178 		mtspr(SPRN_TBU40, new_tb);
179 		if ((mftb() & 0xffffff) < (new_tb & 0xffffff)) {
180 			new_tb += 0x1000000;
181 			mtspr(SPRN_TBU40, new_tb);
182 		}
183 		vc->tb_offset_applied = vc->tb_offset;
184 	}
185 
186 	return ret;
187 }
188 
189 /*
190  * The following subcore HMI handling is all only for pre-POWER9 CPUs.
191  */
192 
193 /* Check if dynamic split is in force and return subcore size accordingly. */
194 static inline int kvmppc_cur_subcore_size(void)
195 {
196 	if (local_paca->kvm_hstate.kvm_split_mode)
197 		return local_paca->kvm_hstate.kvm_split_mode->subcore_size;
198 
199 	return threads_per_subcore;
200 }
201 
202 void kvmppc_subcore_enter_guest(void)
203 {
204 	int thread_id, subcore_id;
205 
206 	thread_id = cpu_thread_in_core(local_paca->paca_index);
207 	subcore_id = thread_id / kvmppc_cur_subcore_size();
208 
209 	local_paca->sibling_subcore_state->in_guest[subcore_id] = 1;
210 }
211 EXPORT_SYMBOL_GPL(kvmppc_subcore_enter_guest);
212 
213 void kvmppc_subcore_exit_guest(void)
214 {
215 	int thread_id, subcore_id;
216 
217 	thread_id = cpu_thread_in_core(local_paca->paca_index);
218 	subcore_id = thread_id / kvmppc_cur_subcore_size();
219 
220 	local_paca->sibling_subcore_state->in_guest[subcore_id] = 0;
221 }
222 EXPORT_SYMBOL_GPL(kvmppc_subcore_exit_guest);
223 
224 static bool kvmppc_tb_resync_required(void)
225 {
226 	if (test_and_set_bit(CORE_TB_RESYNC_REQ_BIT,
227 				&local_paca->sibling_subcore_state->flags))
228 		return false;
229 
230 	return true;
231 }
232 
233 static void kvmppc_tb_resync_done(void)
234 {
235 	clear_bit(CORE_TB_RESYNC_REQ_BIT,
236 			&local_paca->sibling_subcore_state->flags);
237 }
238 
239 /*
240  * kvmppc_realmode_hmi_handler() is called only by primary thread during
241  * guest exit path.
242  *
243  * There are multiple reasons why HMI could occur, one of them is
244  * Timebase (TB) error. If this HMI is due to TB error, then TB would
245  * have been in stopped state. The opal hmi handler Will fix it and
246  * restore the TB value with host timebase value. For HMI caused due
247  * to non-TB errors, opal hmi handler will not touch/restore TB register
248  * and hence there won't be any change in TB value.
249  *
250  * Since we are not sure about the cause of this HMI, we can't be sure
251  * about the content of TB register whether it holds guest or host timebase
252  * value. Hence the idea is to resync the TB on every HMI, so that we
253  * know about the exact state of the TB value. Resync TB call will
254  * restore TB to host timebase.
255  *
256  * Things to consider:
257  * - On TB error, HMI interrupt is reported on all the threads of the core
258  *   that has encountered TB error irrespective of split-core mode.
259  * - The very first thread on the core that get chance to fix TB error
260  *   would rsync the TB with local chipTOD value.
261  * - The resync TB is a core level action i.e. it will sync all the TBs
262  *   in that core independent of split-core mode. This means if we trigger
263  *   TB sync from a thread from one subcore, it would affect TB values of
264  *   sibling subcores of the same core.
265  *
266  * All threads need to co-ordinate before making opal hmi handler.
267  * All threads will use sibling_subcore_state->in_guest[] (shared by all
268  * threads in the core) in paca which holds information about whether
269  * sibling subcores are in Guest mode or host mode. The in_guest[] array
270  * is of size MAX_SUBCORE_PER_CORE=4, indexed using subcore id to set/unset
271  * subcore status. Only primary threads from each subcore is responsible
272  * to set/unset its designated array element while entering/exiting the
273  * guset.
274  *
275  * After invoking opal hmi handler call, one of the thread (of entire core)
276  * will need to resync the TB. Bit 63 from subcore state bitmap flags
277  * (sibling_subcore_state->flags) will be used to co-ordinate between
278  * primary threads to decide who takes up the responsibility.
279  *
280  * This is what we do:
281  * - Primary thread from each subcore tries to set resync required bit[63]
282  *   of paca->sibling_subcore_state->flags.
283  * - The first primary thread that is able to set the flag takes the
284  *   responsibility of TB resync. (Let us call it as thread leader)
285  * - All other threads which are in host will call
286  *   wait_for_subcore_guest_exit() and wait for in_guest[0-3] from
287  *   paca->sibling_subcore_state to get cleared.
288  * - All the primary thread will clear its subcore status from subcore
289  *   state in_guest[] array respectively.
290  * - Once all primary threads clear in_guest[0-3], all of them will invoke
291  *   opal hmi handler.
292  * - Now all threads will wait for TB resync to complete by invoking
293  *   wait_for_tb_resync() except the thread leader.
294  * - Thread leader will do a TB resync by invoking opal_resync_timebase()
295  *   call and the it will clear the resync required bit.
296  * - All other threads will now come out of resync wait loop and proceed
297  *   with individual execution.
298  * - On return of this function, primary thread will signal all
299  *   secondary threads to proceed.
300  * - All secondary threads will eventually call opal hmi handler on
301  *   their exit path.
302  *
303  * Returns 1 if the timebase offset should be applied, 0 if not.
304  */
305 
306 long kvmppc_realmode_hmi_handler(void)
307 {
308 	bool resync_req;
309 
310 	local_paca->hmi_irqs++;
311 
312 	if (hmi_handle_debugtrig(NULL) >= 0)
313 		return 1;
314 
315 	/*
316 	 * By now primary thread has already completed guest->host
317 	 * partition switch but haven't signaled secondaries yet.
318 	 * All the secondary threads on this subcore is waiting
319 	 * for primary thread to signal them to go ahead.
320 	 *
321 	 * For threads from subcore which isn't in guest, they all will
322 	 * wait until all other subcores on this core exit the guest.
323 	 *
324 	 * Now set the resync required bit. If you are the first to
325 	 * set this bit then kvmppc_tb_resync_required() function will
326 	 * return true. For rest all other subcores
327 	 * kvmppc_tb_resync_required() will return false.
328 	 *
329 	 * If resync_req == true, then this thread is responsible to
330 	 * initiate TB resync after hmi handler has completed.
331 	 * All other threads on this core will wait until this thread
332 	 * clears the resync required bit flag.
333 	 */
334 	resync_req = kvmppc_tb_resync_required();
335 
336 	/* Reset the subcore status to indicate it has exited guest */
337 	kvmppc_subcore_exit_guest();
338 
339 	/*
340 	 * Wait for other subcores on this core to exit the guest.
341 	 * All the primary threads and threads from subcore that are
342 	 * not in guest will wait here until all subcores are out
343 	 * of guest context.
344 	 */
345 	wait_for_subcore_guest_exit();
346 
347 	/*
348 	 * At this point we are sure that primary threads from each
349 	 * subcore on this core have completed guest->host partition
350 	 * switch. Now it is safe to call HMI handler.
351 	 */
352 	if (ppc_md.hmi_exception_early)
353 		ppc_md.hmi_exception_early(NULL);
354 
355 	/*
356 	 * Check if this thread is responsible to resync TB.
357 	 * All other threads will wait until this thread completes the
358 	 * TB resync.
359 	 */
360 	if (resync_req) {
361 		opal_resync_timebase();
362 		/* Reset TB resync req bit */
363 		kvmppc_tb_resync_done();
364 	} else {
365 		wait_for_tb_resync();
366 	}
367 
368 	/*
369 	 * Reset tb_offset_applied so the guest exit code won't try
370 	 * to subtract the previous timebase offset from the timebase.
371 	 */
372 	if (local_paca->kvm_hstate.kvm_vcore)
373 		local_paca->kvm_hstate.kvm_vcore->tb_offset_applied = 0;
374 
375 	return 0;
376 }
377