Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[sfrench/cifs-2.6.git] / arch / arm64 / kvm / hyp / switch.c
1 /*
2  * Copyright (C) 2015 - ARM Ltd
3  * Author: Marc Zyngier <marc.zyngier@arm.com>
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License version 2 as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
16  */
17
18 #include <linux/types.h>
19 #include <linux/jump_label.h>
20
21 #include <asm/kvm_asm.h>
22 #include <asm/kvm_emulate.h>
23 #include <asm/kvm_hyp.h>
24 #include <asm/fpsimd.h>
25 #include <asm/debug-monitors.h>
26
27 static bool __hyp_text __fpsimd_enabled_nvhe(void)
28 {
29         return !(read_sysreg(cptr_el2) & CPTR_EL2_TFP);
30 }
31
32 static bool __hyp_text __fpsimd_enabled_vhe(void)
33 {
34         return !!(read_sysreg(cpacr_el1) & CPACR_EL1_FPEN);
35 }
36
37 static hyp_alternate_select(__fpsimd_is_enabled,
38                             __fpsimd_enabled_nvhe, __fpsimd_enabled_vhe,
39                             ARM64_HAS_VIRT_HOST_EXTN);
40
41 bool __hyp_text __fpsimd_enabled(void)
42 {
43         return __fpsimd_is_enabled()();
44 }
45
46 static void __hyp_text __activate_traps_vhe(void)
47 {
48         u64 val;
49
50         val = read_sysreg(cpacr_el1);
51         val |= CPACR_EL1_TTA;
52         val &= ~(CPACR_EL1_FPEN | CPACR_EL1_ZEN);
53         write_sysreg(val, cpacr_el1);
54
55         write_sysreg(__kvm_hyp_vector, vbar_el1);
56 }
57
58 static void __hyp_text __activate_traps_nvhe(void)
59 {
60         u64 val;
61
62         val = CPTR_EL2_DEFAULT;
63         val |= CPTR_EL2_TTA | CPTR_EL2_TFP | CPTR_EL2_TZ;
64         write_sysreg(val, cptr_el2);
65 }
66
67 static hyp_alternate_select(__activate_traps_arch,
68                             __activate_traps_nvhe, __activate_traps_vhe,
69                             ARM64_HAS_VIRT_HOST_EXTN);
70
71 static void __hyp_text __activate_traps(struct kvm_vcpu *vcpu)
72 {
73         u64 val;
74
75         /*
76          * We are about to set CPTR_EL2.TFP to trap all floating point
77          * register accesses to EL2, however, the ARM ARM clearly states that
78          * traps are only taken to EL2 if the operation would not otherwise
79          * trap to EL1.  Therefore, always make sure that for 32-bit guests,
80          * we set FPEXC.EN to prevent traps to EL1, when setting the TFP bit.
81          * If FP/ASIMD is not implemented, FPEXC is UNDEFINED and any access to
82          * it will cause an exception.
83          */
84         val = vcpu->arch.hcr_el2;
85
86         if (!(val & HCR_RW) && system_supports_fpsimd()) {
87                 write_sysreg(1 << 30, fpexc32_el2);
88                 isb();
89         }
90
91         if (val & HCR_RW) /* for AArch64 only: */
92                 val |= HCR_TID3; /* TID3: trap feature register accesses */
93
94         write_sysreg(val, hcr_el2);
95
96         /* Trap on AArch32 cp15 c15 accesses (EL1 or EL0) */
97         write_sysreg(1 << 15, hstr_el2);
98         /*
99          * Make sure we trap PMU access from EL0 to EL2. Also sanitize
100          * PMSELR_EL0 to make sure it never contains the cycle
101          * counter, which could make a PMXEVCNTR_EL0 access UNDEF at
102          * EL1 instead of being trapped to EL2.
103          */
104         write_sysreg(0, pmselr_el0);
105         write_sysreg(ARMV8_PMU_USERENR_MASK, pmuserenr_el0);
106         write_sysreg(vcpu->arch.mdcr_el2, mdcr_el2);
107         __activate_traps_arch()();
108 }
109
110 static void __hyp_text __deactivate_traps_vhe(void)
111 {
112         extern char vectors[];  /* kernel exception vectors */
113         u64 mdcr_el2 = read_sysreg(mdcr_el2);
114
115         mdcr_el2 &= MDCR_EL2_HPMN_MASK |
116                     MDCR_EL2_E2PB_MASK << MDCR_EL2_E2PB_SHIFT |
117                     MDCR_EL2_TPMS;
118
119         write_sysreg(mdcr_el2, mdcr_el2);
120         write_sysreg(HCR_HOST_VHE_FLAGS, hcr_el2);
121         write_sysreg(CPACR_EL1_DEFAULT, cpacr_el1);
122         write_sysreg(vectors, vbar_el1);
123 }
124
125 static void __hyp_text __deactivate_traps_nvhe(void)
126 {
127         u64 mdcr_el2 = read_sysreg(mdcr_el2);
128
129         mdcr_el2 &= MDCR_EL2_HPMN_MASK;
130         mdcr_el2 |= MDCR_EL2_E2PB_MASK << MDCR_EL2_E2PB_SHIFT;
131
132         write_sysreg(mdcr_el2, mdcr_el2);
133         write_sysreg(HCR_RW, hcr_el2);
134         write_sysreg(CPTR_EL2_DEFAULT, cptr_el2);
135 }
136
137 static hyp_alternate_select(__deactivate_traps_arch,
138                             __deactivate_traps_nvhe, __deactivate_traps_vhe,
139                             ARM64_HAS_VIRT_HOST_EXTN);
140
141 static void __hyp_text __deactivate_traps(struct kvm_vcpu *vcpu)
142 {
143         /*
144          * If we pended a virtual abort, preserve it until it gets
145          * cleared. See D1.14.3 (Virtual Interrupts) for details, but
146          * the crucial bit is "On taking a vSError interrupt,
147          * HCR_EL2.VSE is cleared to 0."
148          */
149         if (vcpu->arch.hcr_el2 & HCR_VSE)
150                 vcpu->arch.hcr_el2 = read_sysreg(hcr_el2);
151
152         __deactivate_traps_arch()();
153         write_sysreg(0, hstr_el2);
154         write_sysreg(0, pmuserenr_el0);
155 }
156
157 static void __hyp_text __activate_vm(struct kvm_vcpu *vcpu)
158 {
159         struct kvm *kvm = kern_hyp_va(vcpu->kvm);
160         write_sysreg(kvm->arch.vttbr, vttbr_el2);
161 }
162
163 static void __hyp_text __deactivate_vm(struct kvm_vcpu *vcpu)
164 {
165         write_sysreg(0, vttbr_el2);
166 }
167
168 static void __hyp_text __vgic_save_state(struct kvm_vcpu *vcpu)
169 {
170         if (static_branch_unlikely(&kvm_vgic_global_state.gicv3_cpuif))
171                 __vgic_v3_save_state(vcpu);
172         else
173                 __vgic_v2_save_state(vcpu);
174
175         write_sysreg(read_sysreg(hcr_el2) & ~HCR_INT_OVERRIDE, hcr_el2);
176 }
177
178 static void __hyp_text __vgic_restore_state(struct kvm_vcpu *vcpu)
179 {
180         u64 val;
181
182         val = read_sysreg(hcr_el2);
183         val |=  HCR_INT_OVERRIDE;
184         val |= vcpu->arch.irq_lines;
185         write_sysreg(val, hcr_el2);
186
187         if (static_branch_unlikely(&kvm_vgic_global_state.gicv3_cpuif))
188                 __vgic_v3_restore_state(vcpu);
189         else
190                 __vgic_v2_restore_state(vcpu);
191 }
192
193 static bool __hyp_text __true_value(void)
194 {
195         return true;
196 }
197
198 static bool __hyp_text __false_value(void)
199 {
200         return false;
201 }
202
203 static hyp_alternate_select(__check_arm_834220,
204                             __false_value, __true_value,
205                             ARM64_WORKAROUND_834220);
206
207 static bool __hyp_text __translate_far_to_hpfar(u64 far, u64 *hpfar)
208 {
209         u64 par, tmp;
210
211         /*
212          * Resolve the IPA the hard way using the guest VA.
213          *
214          * Stage-1 translation already validated the memory access
215          * rights. As such, we can use the EL1 translation regime, and
216          * don't have to distinguish between EL0 and EL1 access.
217          *
218          * We do need to save/restore PAR_EL1 though, as we haven't
219          * saved the guest context yet, and we may return early...
220          */
221         par = read_sysreg(par_el1);
222         asm volatile("at s1e1r, %0" : : "r" (far));
223         isb();
224
225         tmp = read_sysreg(par_el1);
226         write_sysreg(par, par_el1);
227
228         if (unlikely(tmp & 1))
229                 return false; /* Translation failed, back to guest */
230
231         /* Convert PAR to HPFAR format */
232         *hpfar = ((tmp >> 12) & ((1UL << 36) - 1)) << 4;
233         return true;
234 }
235
236 static bool __hyp_text __populate_fault_info(struct kvm_vcpu *vcpu)
237 {
238         u64 esr = read_sysreg_el2(esr);
239         u8 ec = ESR_ELx_EC(esr);
240         u64 hpfar, far;
241
242         vcpu->arch.fault.esr_el2 = esr;
243
244         if (ec != ESR_ELx_EC_DABT_LOW && ec != ESR_ELx_EC_IABT_LOW)
245                 return true;
246
247         far = read_sysreg_el2(far);
248
249         /*
250          * The HPFAR can be invalid if the stage 2 fault did not
251          * happen during a stage 1 page table walk (the ESR_EL2.S1PTW
252          * bit is clear) and one of the two following cases are true:
253          *   1. The fault was due to a permission fault
254          *   2. The processor carries errata 834220
255          *
256          * Therefore, for all non S1PTW faults where we either have a
257          * permission fault or the errata workaround is enabled, we
258          * resolve the IPA using the AT instruction.
259          */
260         if (!(esr & ESR_ELx_S1PTW) &&
261             (__check_arm_834220()() || (esr & ESR_ELx_FSC_TYPE) == FSC_PERM)) {
262                 if (!__translate_far_to_hpfar(far, &hpfar))
263                         return false;
264         } else {
265                 hpfar = read_sysreg(hpfar_el2);
266         }
267
268         vcpu->arch.fault.far_el2 = far;
269         vcpu->arch.fault.hpfar_el2 = hpfar;
270         return true;
271 }
272
273 /* Skip an instruction which has been emulated. Returns true if
274  * execution can continue or false if we need to exit hyp mode because
275  * single-step was in effect.
276  */
277 static bool __hyp_text __skip_instr(struct kvm_vcpu *vcpu)
278 {
279         *vcpu_pc(vcpu) = read_sysreg_el2(elr);
280
281         if (vcpu_mode_is_32bit(vcpu)) {
282                 vcpu->arch.ctxt.gp_regs.regs.pstate = read_sysreg_el2(spsr);
283                 kvm_skip_instr32(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
284                 write_sysreg_el2(vcpu->arch.ctxt.gp_regs.regs.pstate, spsr);
285         } else {
286                 *vcpu_pc(vcpu) += 4;
287         }
288
289         write_sysreg_el2(*vcpu_pc(vcpu), elr);
290
291         if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP) {
292                 vcpu->arch.fault.esr_el2 =
293                         (ESR_ELx_EC_SOFTSTP_LOW << ESR_ELx_EC_SHIFT) | 0x22;
294                 return false;
295         } else {
296                 return true;
297         }
298 }
299
300 int __hyp_text __kvm_vcpu_run(struct kvm_vcpu *vcpu)
301 {
302         struct kvm_cpu_context *host_ctxt;
303         struct kvm_cpu_context *guest_ctxt;
304         bool fp_enabled;
305         u64 exit_code;
306
307         vcpu = kern_hyp_va(vcpu);
308         write_sysreg(vcpu, tpidr_el2);
309
310         host_ctxt = kern_hyp_va(vcpu->arch.host_cpu_context);
311         guest_ctxt = &vcpu->arch.ctxt;
312
313         __sysreg_save_host_state(host_ctxt);
314         __debug_cond_save_host_state(vcpu);
315
316         __activate_traps(vcpu);
317         __activate_vm(vcpu);
318
319         __vgic_restore_state(vcpu);
320         __timer_enable_traps(vcpu);
321
322         /*
323          * We must restore the 32-bit state before the sysregs, thanks
324          * to erratum #852523 (Cortex-A57) or #853709 (Cortex-A72).
325          */
326         __sysreg32_restore_state(vcpu);
327         __sysreg_restore_guest_state(guest_ctxt);
328         __debug_restore_state(vcpu, kern_hyp_va(vcpu->arch.debug_ptr), guest_ctxt);
329
330         /* Jump in the fire! */
331 again:
332         exit_code = __guest_enter(vcpu, host_ctxt);
333         /* And we're baaack! */
334
335         /*
336          * We're using the raw exception code in order to only process
337          * the trap if no SError is pending. We will come back to the
338          * same PC once the SError has been injected, and replay the
339          * trapping instruction.
340          */
341         if (exit_code == ARM_EXCEPTION_TRAP && !__populate_fault_info(vcpu))
342                 goto again;
343
344         if (static_branch_unlikely(&vgic_v2_cpuif_trap) &&
345             exit_code == ARM_EXCEPTION_TRAP) {
346                 bool valid;
347
348                 valid = kvm_vcpu_trap_get_class(vcpu) == ESR_ELx_EC_DABT_LOW &&
349                         kvm_vcpu_trap_get_fault_type(vcpu) == FSC_FAULT &&
350                         kvm_vcpu_dabt_isvalid(vcpu) &&
351                         !kvm_vcpu_dabt_isextabt(vcpu) &&
352                         !kvm_vcpu_dabt_iss1tw(vcpu);
353
354                 if (valid) {
355                         int ret = __vgic_v2_perform_cpuif_access(vcpu);
356
357                         if (ret == 1) {
358                                 if (__skip_instr(vcpu))
359                                         goto again;
360                                 else
361                                         exit_code = ARM_EXCEPTION_TRAP;
362                         }
363
364                         if (ret == -1) {
365                                 /* Promote an illegal access to an
366                                  * SError. If we would be returning
367                                  * due to single-step clear the SS
368                                  * bit so handle_exit knows what to
369                                  * do after dealing with the error.
370                                  */
371                                 if (!__skip_instr(vcpu))
372                                         *vcpu_cpsr(vcpu) &= ~DBG_SPSR_SS;
373                                 exit_code = ARM_EXCEPTION_EL1_SERROR;
374                         }
375
376                         /* 0 falls through to be handler out of EL2 */
377                 }
378         }
379
380         if (static_branch_unlikely(&vgic_v3_cpuif_trap) &&
381             exit_code == ARM_EXCEPTION_TRAP &&
382             (kvm_vcpu_trap_get_class(vcpu) == ESR_ELx_EC_SYS64 ||
383              kvm_vcpu_trap_get_class(vcpu) == ESR_ELx_EC_CP15_32)) {
384                 int ret = __vgic_v3_perform_cpuif_access(vcpu);
385
386                 if (ret == 1) {
387                         if (__skip_instr(vcpu))
388                                 goto again;
389                         else
390                                 exit_code = ARM_EXCEPTION_TRAP;
391                 }
392
393                 /* 0 falls through to be handled out of EL2 */
394         }
395
396         fp_enabled = __fpsimd_enabled();
397
398         __sysreg_save_guest_state(guest_ctxt);
399         __sysreg32_save_state(vcpu);
400         __timer_disable_traps(vcpu);
401         __vgic_save_state(vcpu);
402
403         __deactivate_traps(vcpu);
404         __deactivate_vm(vcpu);
405
406         __sysreg_restore_host_state(host_ctxt);
407
408         if (fp_enabled) {
409                 __fpsimd_save_state(&guest_ctxt->gp_regs.fp_regs);
410                 __fpsimd_restore_state(&host_ctxt->gp_regs.fp_regs);
411         }
412
413         __debug_save_state(vcpu, kern_hyp_va(vcpu->arch.debug_ptr), guest_ctxt);
414         /*
415          * This must come after restoring the host sysregs, since a non-VHE
416          * system may enable SPE here and make use of the TTBRs.
417          */
418         __debug_cond_restore_host_state(vcpu);
419
420         return exit_code;
421 }
422
423 static const char __hyp_panic_string[] = "HYP panic:\nPS:%08llx PC:%016llx ESR:%08llx\nFAR:%016llx HPFAR:%016llx PAR:%016llx\nVCPU:%p\n";
424
425 static void __hyp_text __hyp_call_panic_nvhe(u64 spsr, u64 elr, u64 par)
426 {
427         unsigned long str_va;
428
429         /*
430          * Force the panic string to be loaded from the literal pool,
431          * making sure it is a kernel address and not a PC-relative
432          * reference.
433          */
434         asm volatile("ldr %0, =__hyp_panic_string" : "=r" (str_va));
435
436         __hyp_do_panic(str_va,
437                        spsr,  elr,
438                        read_sysreg(esr_el2),   read_sysreg_el2(far),
439                        read_sysreg(hpfar_el2), par,
440                        (void *)read_sysreg(tpidr_el2));
441 }
442
443 static void __hyp_text __hyp_call_panic_vhe(u64 spsr, u64 elr, u64 par)
444 {
445         panic(__hyp_panic_string,
446               spsr,  elr,
447               read_sysreg_el2(esr),   read_sysreg_el2(far),
448               read_sysreg(hpfar_el2), par,
449               (void *)read_sysreg(tpidr_el2));
450 }
451
452 static hyp_alternate_select(__hyp_call_panic,
453                             __hyp_call_panic_nvhe, __hyp_call_panic_vhe,
454                             ARM64_HAS_VIRT_HOST_EXTN);
455
456 void __hyp_text __noreturn __hyp_panic(void)
457 {
458         u64 spsr = read_sysreg_el2(spsr);
459         u64 elr = read_sysreg_el2(elr);
460         u64 par = read_sysreg(par_el1);
461
462         if (read_sysreg(vttbr_el2)) {
463                 struct kvm_vcpu *vcpu;
464                 struct kvm_cpu_context *host_ctxt;
465
466                 vcpu = (struct kvm_vcpu *)read_sysreg(tpidr_el2);
467                 host_ctxt = kern_hyp_va(vcpu->arch.host_cpu_context);
468                 __timer_disable_traps(vcpu);
469                 __deactivate_traps(vcpu);
470                 __deactivate_vm(vcpu);
471                 __sysreg_restore_host_state(host_ctxt);
472         }
473
474         /* Call panic for real */
475         __hyp_call_panic()(spsr, elr, par);
476
477         unreachable();
478 }