xref: /linux/arch/powerpc/kvm/book3s_hv_ras.c (revision 37744feebc086908fd89760650f458ab19071750)
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 void kvmppc_realmode_mc_power7(struct kvm_vcpu *vcpu)
69 {
70 	unsigned long srr1 = vcpu->arch.shregs.msr;
71 	struct machine_check_event mce_evt;
72 	long handled = 1;
73 
74 	if (srr1 & SRR1_MC_LDSTERR) {
75 		/* error on load/store */
76 		unsigned long dsisr = vcpu->arch.shregs.dsisr;
77 
78 		if (dsisr & (DSISR_MC_SLB_PARMULTI | DSISR_MC_SLB_MULTI |
79 			     DSISR_MC_SLB_PARITY | DSISR_MC_DERAT_MULTI)) {
80 			/* flush and reload SLB; flushes D-ERAT too */
81 			reload_slb(vcpu);
82 			dsisr &= ~(DSISR_MC_SLB_PARMULTI | DSISR_MC_SLB_MULTI |
83 				   DSISR_MC_SLB_PARITY | DSISR_MC_DERAT_MULTI);
84 		}
85 		if (dsisr & DSISR_MC_TLB_MULTI) {
86 			tlbiel_all_lpid(vcpu->kvm->arch.radix);
87 			dsisr &= ~DSISR_MC_TLB_MULTI;
88 		}
89 		/* Any other errors we don't understand? */
90 		if (dsisr & 0xffffffffUL)
91 			handled = 0;
92 	}
93 
94 	switch ((srr1 >> SRR1_MC_IFETCH_SH) & SRR1_MC_IFETCH_MASK) {
95 	case 0:
96 		break;
97 	case SRR1_MC_IFETCH_SLBPAR:
98 	case SRR1_MC_IFETCH_SLBMULTI:
99 	case SRR1_MC_IFETCH_SLBPARMULTI:
100 		reload_slb(vcpu);
101 		break;
102 	case SRR1_MC_IFETCH_TLBMULTI:
103 		tlbiel_all_lpid(vcpu->kvm->arch.radix);
104 		break;
105 	default:
106 		handled = 0;
107 	}
108 
109 	/*
110 	 * Now get the event and stash it in the vcpu struct so it can
111 	 * be handled by the primary thread in virtual mode.  We can't
112 	 * call machine_check_queue_event() here if we are running on
113 	 * an offline secondary thread.
114 	 */
115 	if (get_mce_event(&mce_evt, MCE_EVENT_RELEASE)) {
116 		if (handled && mce_evt.version == MCE_V1)
117 			mce_evt.disposition = MCE_DISPOSITION_RECOVERED;
118 	} else {
119 		memset(&mce_evt, 0, sizeof(mce_evt));
120 	}
121 
122 	vcpu->arch.mce_evt = mce_evt;
123 }
124 
125 void kvmppc_realmode_machine_check(struct kvm_vcpu *vcpu)
126 {
127 	kvmppc_realmode_mc_power7(vcpu);
128 }
129 
130 /* Check if dynamic split is in force and return subcore size accordingly. */
131 static inline int kvmppc_cur_subcore_size(void)
132 {
133 	if (local_paca->kvm_hstate.kvm_split_mode)
134 		return local_paca->kvm_hstate.kvm_split_mode->subcore_size;
135 
136 	return threads_per_subcore;
137 }
138 
139 void kvmppc_subcore_enter_guest(void)
140 {
141 	int thread_id, subcore_id;
142 
143 	thread_id = cpu_thread_in_core(local_paca->paca_index);
144 	subcore_id = thread_id / kvmppc_cur_subcore_size();
145 
146 	local_paca->sibling_subcore_state->in_guest[subcore_id] = 1;
147 }
148 EXPORT_SYMBOL_GPL(kvmppc_subcore_enter_guest);
149 
150 void kvmppc_subcore_exit_guest(void)
151 {
152 	int thread_id, subcore_id;
153 
154 	thread_id = cpu_thread_in_core(local_paca->paca_index);
155 	subcore_id = thread_id / kvmppc_cur_subcore_size();
156 
157 	local_paca->sibling_subcore_state->in_guest[subcore_id] = 0;
158 }
159 EXPORT_SYMBOL_GPL(kvmppc_subcore_exit_guest);
160 
161 static bool kvmppc_tb_resync_required(void)
162 {
163 	if (test_and_set_bit(CORE_TB_RESYNC_REQ_BIT,
164 				&local_paca->sibling_subcore_state->flags))
165 		return false;
166 
167 	return true;
168 }
169 
170 static void kvmppc_tb_resync_done(void)
171 {
172 	clear_bit(CORE_TB_RESYNC_REQ_BIT,
173 			&local_paca->sibling_subcore_state->flags);
174 }
175 
176 /*
177  * kvmppc_realmode_hmi_handler() is called only by primary thread during
178  * guest exit path.
179  *
180  * There are multiple reasons why HMI could occur, one of them is
181  * Timebase (TB) error. If this HMI is due to TB error, then TB would
182  * have been in stopped state. The opal hmi handler Will fix it and
183  * restore the TB value with host timebase value. For HMI caused due
184  * to non-TB errors, opal hmi handler will not touch/restore TB register
185  * and hence there won't be any change in TB value.
186  *
187  * Since we are not sure about the cause of this HMI, we can't be sure
188  * about the content of TB register whether it holds guest or host timebase
189  * value. Hence the idea is to resync the TB on every HMI, so that we
190  * know about the exact state of the TB value. Resync TB call will
191  * restore TB to host timebase.
192  *
193  * Things to consider:
194  * - On TB error, HMI interrupt is reported on all the threads of the core
195  *   that has encountered TB error irrespective of split-core mode.
196  * - The very first thread on the core that get chance to fix TB error
197  *   would rsync the TB with local chipTOD value.
198  * - The resync TB is a core level action i.e. it will sync all the TBs
199  *   in that core independent of split-core mode. This means if we trigger
200  *   TB sync from a thread from one subcore, it would affect TB values of
201  *   sibling subcores of the same core.
202  *
203  * All threads need to co-ordinate before making opal hmi handler.
204  * All threads will use sibling_subcore_state->in_guest[] (shared by all
205  * threads in the core) in paca which holds information about whether
206  * sibling subcores are in Guest mode or host mode. The in_guest[] array
207  * is of size MAX_SUBCORE_PER_CORE=4, indexed using subcore id to set/unset
208  * subcore status. Only primary threads from each subcore is responsible
209  * to set/unset its designated array element while entering/exiting the
210  * guset.
211  *
212  * After invoking opal hmi handler call, one of the thread (of entire core)
213  * will need to resync the TB. Bit 63 from subcore state bitmap flags
214  * (sibling_subcore_state->flags) will be used to co-ordinate between
215  * primary threads to decide who takes up the responsibility.
216  *
217  * This is what we do:
218  * - Primary thread from each subcore tries to set resync required bit[63]
219  *   of paca->sibling_subcore_state->flags.
220  * - The first primary thread that is able to set the flag takes the
221  *   responsibility of TB resync. (Let us call it as thread leader)
222  * - All other threads which are in host will call
223  *   wait_for_subcore_guest_exit() and wait for in_guest[0-3] from
224  *   paca->sibling_subcore_state to get cleared.
225  * - All the primary thread will clear its subcore status from subcore
226  *   state in_guest[] array respectively.
227  * - Once all primary threads clear in_guest[0-3], all of them will invoke
228  *   opal hmi handler.
229  * - Now all threads will wait for TB resync to complete by invoking
230  *   wait_for_tb_resync() except the thread leader.
231  * - Thread leader will do a TB resync by invoking opal_resync_timebase()
232  *   call and the it will clear the resync required bit.
233  * - All other threads will now come out of resync wait loop and proceed
234  *   with individual execution.
235  * - On return of this function, primary thread will signal all
236  *   secondary threads to proceed.
237  * - All secondary threads will eventually call opal hmi handler on
238  *   their exit path.
239  *
240  * Returns 1 if the timebase offset should be applied, 0 if not.
241  */
242 
243 long kvmppc_realmode_hmi_handler(void)
244 {
245 	bool resync_req;
246 
247 	__this_cpu_inc(irq_stat.hmi_exceptions);
248 
249 	if (hmi_handle_debugtrig(NULL) >= 0)
250 		return 1;
251 
252 	/*
253 	 * By now primary thread has already completed guest->host
254 	 * partition switch but haven't signaled secondaries yet.
255 	 * All the secondary threads on this subcore is waiting
256 	 * for primary thread to signal them to go ahead.
257 	 *
258 	 * For threads from subcore which isn't in guest, they all will
259 	 * wait until all other subcores on this core exit the guest.
260 	 *
261 	 * Now set the resync required bit. If you are the first to
262 	 * set this bit then kvmppc_tb_resync_required() function will
263 	 * return true. For rest all other subcores
264 	 * kvmppc_tb_resync_required() will return false.
265 	 *
266 	 * If resync_req == true, then this thread is responsible to
267 	 * initiate TB resync after hmi handler has completed.
268 	 * All other threads on this core will wait until this thread
269 	 * clears the resync required bit flag.
270 	 */
271 	resync_req = kvmppc_tb_resync_required();
272 
273 	/* Reset the subcore status to indicate it has exited guest */
274 	kvmppc_subcore_exit_guest();
275 
276 	/*
277 	 * Wait for other subcores on this core to exit the guest.
278 	 * All the primary threads and threads from subcore that are
279 	 * not in guest will wait here until all subcores are out
280 	 * of guest context.
281 	 */
282 	wait_for_subcore_guest_exit();
283 
284 	/*
285 	 * At this point we are sure that primary threads from each
286 	 * subcore on this core have completed guest->host partition
287 	 * switch. Now it is safe to call HMI handler.
288 	 */
289 	if (ppc_md.hmi_exception_early)
290 		ppc_md.hmi_exception_early(NULL);
291 
292 	/*
293 	 * Check if this thread is responsible to resync TB.
294 	 * All other threads will wait until this thread completes the
295 	 * TB resync.
296 	 */
297 	if (resync_req) {
298 		opal_resync_timebase();
299 		/* Reset TB resync req bit */
300 		kvmppc_tb_resync_done();
301 	} else {
302 		wait_for_tb_resync();
303 	}
304 
305 	/*
306 	 * Reset tb_offset_applied so the guest exit code won't try
307 	 * to subtract the previous timebase offset from the timebase.
308 	 */
309 	if (local_paca->kvm_hstate.kvm_vcore)
310 		local_paca->kvm_hstate.kvm_vcore->tb_offset_applied = 0;
311 
312 	return 0;
313 }
314