xref: /freebsd/sys/amd64/amd64/initcpu.c (revision 1dfcff294e44d4b45813288ef4095c36abb22f0e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) KATO Takenori, 1997, 1998.
5  *
6  * All rights reserved.  Unpublished rights reserved under the copyright
7  * laws of Japan.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  *
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer as
15  *    the first lines of this file unmodified.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
21  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
22  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
23  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
24  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
25  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
26  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
27  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
28  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
29  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_cpu.h"
36 
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/pcpu.h>
40 #include <sys/systm.h>
41 #include <sys/sysctl.h>
42 
43 #include <machine/cputypes.h>
44 #include <machine/md_var.h>
45 #include <machine/psl.h>
46 #include <machine/specialreg.h>
47 
48 #include <vm/vm.h>
49 #include <vm/pmap.h>
50 
51 static int	hw_instruction_sse;
52 SYSCTL_INT(_hw, OID_AUTO, instruction_sse, CTLFLAG_RD,
53     &hw_instruction_sse, 0, "SIMD/MMX2 instructions available in CPU");
54 static int	lower_sharedpage_init;
55 int		hw_lower_amd64_sharedpage;
56 SYSCTL_INT(_hw, OID_AUTO, lower_amd64_sharedpage, CTLFLAG_RDTUN,
57     &hw_lower_amd64_sharedpage, 0,
58    "Lower sharedpage to work around Ryzen issue with executing code near the top of user memory");
59 /*
60  * -1: automatic (default)
61  *  0: keep enable CLFLUSH
62  *  1: force disable CLFLUSH
63  */
64 static int	hw_clflush_disable = -1;
65 
66 static void
67 init_amd(void)
68 {
69 	uint64_t msr;
70 
71 	/*
72 	 * C1E renders the local APIC timer dead, so we disable it by
73 	 * reading the Interrupt Pending Message register and clearing
74 	 * both C1eOnCmpHalt (bit 28) and SmiOnCmpHalt (bit 27).
75 	 *
76 	 * Reference:
77 	 *   "BIOS and Kernel Developer's Guide for AMD NPT Family 0Fh Processors"
78 	 *   #32559 revision 3.00+
79 	 *
80 	 * Detect the presence of C1E capability mostly on latest
81 	 * dual-cores (or future) k8 family.  Affected models range is
82 	 * taken from Linux sources.
83 	 */
84 	if ((CPUID_TO_FAMILY(cpu_id) == 0xf ||
85 	    CPUID_TO_FAMILY(cpu_id) == 0x10) && (cpu_feature2 & CPUID2_HV) == 0)
86 		cpu_amdc1e_bug = 1;
87 
88 	/*
89 	 * Work around Erratum 721 for Family 10h and 12h processors.
90 	 * These processors may incorrectly update the stack pointer
91 	 * after a long series of push and/or near-call instructions,
92 	 * or a long series of pop and/or near-return instructions.
93 	 *
94 	 * http://support.amd.com/us/Processor_TechDocs/41322_10h_Rev_Gd.pdf
95 	 * http://support.amd.com/us/Processor_TechDocs/44739_12h_Rev_Gd.pdf
96 	 *
97 	 * Hypervisors do not provide access to the errata MSR,
98 	 * causing #GP exception on attempt to apply the errata.  The
99 	 * MSR write shall be done on host and persist globally
100 	 * anyway, so do not try to do it when under virtualization.
101 	 */
102 	switch (CPUID_TO_FAMILY(cpu_id)) {
103 	case 0x10:
104 	case 0x12:
105 		if ((cpu_feature2 & CPUID2_HV) == 0)
106 			wrmsr(MSR_DE_CFG, rdmsr(MSR_DE_CFG) | 1);
107 		break;
108 	}
109 
110 	/*
111 	 * BIOS may fail to set InitApicIdCpuIdLo to 1 as it should per BKDG.
112 	 * So, do it here or otherwise some tools could be confused by
113 	 * Initial Local APIC ID reported with CPUID Function 1 in EBX.
114 	 */
115 	if (CPUID_TO_FAMILY(cpu_id) == 0x10) {
116 		if ((cpu_feature2 & CPUID2_HV) == 0) {
117 			msr = rdmsr(MSR_NB_CFG1);
118 			msr |= (uint64_t)1 << 54;
119 			wrmsr(MSR_NB_CFG1, msr);
120 		}
121 	}
122 
123 	/*
124 	 * BIOS may configure Family 10h processors to convert WC+ cache type
125 	 * to CD.  That can hurt performance of guest VMs using nested paging.
126 	 * The relevant MSR bit is not documented in the BKDG,
127 	 * the fix is borrowed from Linux.
128 	 */
129 	if (CPUID_TO_FAMILY(cpu_id) == 0x10) {
130 		if ((cpu_feature2 & CPUID2_HV) == 0) {
131 			msr = rdmsr(0xc001102a);
132 			msr &= ~((uint64_t)1 << 24);
133 			wrmsr(0xc001102a, msr);
134 		}
135 	}
136 
137 	/*
138 	 * Work around Erratum 793: Specific Combination of Writes to Write
139 	 * Combined Memory Types and Locked Instructions May Cause Core Hang.
140 	 * See Revision Guide for AMD Family 16h Models 00h-0Fh Processors,
141 	 * revision 3.04 or later, publication 51810.
142 	 */
143 	if (CPUID_TO_FAMILY(cpu_id) == 0x16 && CPUID_TO_MODEL(cpu_id) <= 0xf) {
144 		if ((cpu_feature2 & CPUID2_HV) == 0) {
145 			msr = rdmsr(MSR_LS_CFG);
146 			msr |= (uint64_t)1 << 15;
147 			wrmsr(MSR_LS_CFG, msr);
148 		}
149 	}
150 
151 	/* Ryzen erratas. */
152 	if (CPUID_TO_FAMILY(cpu_id) == 0x17 && CPUID_TO_MODEL(cpu_id) == 0x1 &&
153 	    (cpu_feature2 & CPUID2_HV) == 0) {
154 		/* 1021 */
155 		msr = rdmsr(MSR_DE_CFG);
156 		msr |= 0x2000;
157 		wrmsr(MSR_DE_CFG, msr);
158 
159 		/* 1033 */
160 		msr = rdmsr(MSR_LS_CFG);
161 		msr |= 0x10;
162 		wrmsr(MSR_LS_CFG, msr);
163 
164 		/* 1049 */
165 		msr = rdmsr(0xc0011028);
166 		msr |= 0x10;
167 		wrmsr(0xc0011028, msr);
168 
169 		/* 1095 */
170 		msr = rdmsr(MSR_LS_CFG);
171 		msr |= 0x200000000000000;
172 		wrmsr(MSR_LS_CFG, msr);
173 	}
174 
175 	/*
176 	 * Work around a problem on Ryzen that is triggered by executing
177 	 * code near the top of user memory, in our case the signal
178 	 * trampoline code in the shared page on amd64.
179 	 *
180 	 * This function is executed once for the BSP before tunables take
181 	 * effect so the value determined here can be overridden by the
182 	 * tunable.  This function is then executed again for each AP and
183 	 * also on resume.  Set a flag the first time so that value set by
184 	 * the tunable is not overwritten.
185 	 *
186 	 * The stepping and/or microcode versions should be checked after
187 	 * this issue is fixed by AMD so that we don't use this mode if not
188 	 * needed.
189 	 */
190 	if (lower_sharedpage_init == 0) {
191 		lower_sharedpage_init = 1;
192 		if (CPUID_TO_FAMILY(cpu_id) == 0x17 ||
193 		    CPUID_TO_FAMILY(cpu_id) == 0x18) {
194 			hw_lower_amd64_sharedpage = 1;
195 		}
196 	}
197 }
198 
199 /*
200  * Initialize special VIA features
201  */
202 static void
203 init_via(void)
204 {
205 	u_int regs[4], val;
206 
207 	/*
208 	 * Check extended CPUID for PadLock features.
209 	 *
210 	 * http://www.via.com.tw/en/downloads/whitepapers/initiatives/padlock/programming_guide.pdf
211 	 */
212 	do_cpuid(0xc0000000, regs);
213 	if (regs[0] >= 0xc0000001) {
214 		do_cpuid(0xc0000001, regs);
215 		val = regs[3];
216 	} else
217 		return;
218 
219 	/* Enable RNG if present. */
220 	if ((val & VIA_CPUID_HAS_RNG) != 0) {
221 		via_feature_rng = VIA_HAS_RNG;
222 		wrmsr(0x110B, rdmsr(0x110B) | VIA_CPUID_DO_RNG);
223 	}
224 
225 	/* Enable PadLock if present. */
226 	if ((val & VIA_CPUID_HAS_ACE) != 0)
227 		via_feature_xcrypt |= VIA_HAS_AES;
228 	if ((val & VIA_CPUID_HAS_ACE2) != 0)
229 		via_feature_xcrypt |= VIA_HAS_AESCTR;
230 	if ((val & VIA_CPUID_HAS_PHE) != 0)
231 		via_feature_xcrypt |= VIA_HAS_SHA;
232 	if ((val & VIA_CPUID_HAS_PMM) != 0)
233 		via_feature_xcrypt |= VIA_HAS_MM;
234 	if (via_feature_xcrypt != 0)
235 		wrmsr(0x1107, rdmsr(0x1107) | (1 << 28));
236 }
237 
238 /*
239  * The value for the TSC_AUX MSR and rdtscp/rdpid on the invoking CPU.
240  *
241  * Caller should prevent CPU migration.
242  */
243 u_int
244 cpu_auxmsr(void)
245 {
246 	KASSERT((read_rflags() & PSL_I) == 0, ("context switch possible"));
247 	return (PCPU_GET(cpuid));
248 }
249 
250 /*
251  * Initialize CPU control registers
252  */
253 void
254 initializecpu(void)
255 {
256 	uint64_t msr;
257 	uint32_t cr4;
258 
259 	cr4 = rcr4();
260 	if ((cpu_feature & CPUID_XMM) && (cpu_feature & CPUID_FXSR)) {
261 		cr4 |= CR4_FXSR | CR4_XMM;
262 		cpu_fxsr = hw_instruction_sse = 1;
263 	}
264 	if (cpu_stdext_feature & CPUID_STDEXT_FSGSBASE)
265 		cr4 |= CR4_FSGSBASE;
266 
267 	if (cpu_stdext_feature2 & CPUID_STDEXT2_PKU)
268 		cr4 |= CR4_PKE;
269 
270 	/*
271 	 * If SMEP is present, we only need to flush RSB (by default)
272 	 * on context switches, to prevent cross-process ret2spec
273 	 * attacks.  Do it automatically if ibrs_disable is set, to
274 	 * complete the mitigation.
275 	 *
276 	 * Postpone enabling the SMEP on the boot CPU until the page
277 	 * tables are switched from the boot loader identity mapping
278 	 * to the kernel tables.  The boot loader enables the U bit in
279 	 * its tables.
280 	 */
281 	if (IS_BSP()) {
282 		if (cpu_stdext_feature & CPUID_STDEXT_SMEP &&
283 		    !TUNABLE_INT_FETCH(
284 		    "machdep.mitigations.cpu_flush_rsb_ctxsw",
285 		    &cpu_flush_rsb_ctxsw) &&
286 		    hw_ibrs_disable)
287 			cpu_flush_rsb_ctxsw = 1;
288 	} else {
289 		if (cpu_stdext_feature & CPUID_STDEXT_SMEP)
290 			cr4 |= CR4_SMEP;
291 		if (cpu_stdext_feature & CPUID_STDEXT_SMAP)
292 			cr4 |= CR4_SMAP;
293 	}
294 	load_cr4(cr4);
295 	/* Reload cpu ext features to reflect cr4 changes */
296 	if (IS_BSP())
297 		identify_cpu_ext_features();
298 	if (IS_BSP() && (amd_feature & AMDID_NX) != 0) {
299 		msr = rdmsr(MSR_EFER) | EFER_NXE;
300 		wrmsr(MSR_EFER, msr);
301 		pg_nx = PG_NX;
302 	}
303 	hw_ibrs_recalculate(false);
304 	hw_ssb_recalculate(false);
305 	amd64_syscall_ret_flush_l1d_recalc();
306 	x86_rngds_mitg_recalculate(false);
307 	switch (cpu_vendor_id) {
308 	case CPU_VENDOR_AMD:
309 	case CPU_VENDOR_HYGON:
310 		init_amd();
311 		break;
312 	case CPU_VENDOR_CENTAUR:
313 		init_via();
314 		break;
315 	}
316 
317 	if ((amd_feature & AMDID_RDTSCP) != 0 ||
318 	    (cpu_stdext_feature2 & CPUID_STDEXT2_RDPID) != 0)
319 		wrmsr(MSR_TSC_AUX, cpu_auxmsr());
320 }
321 
322 void
323 initializecpucache(void)
324 {
325 
326 	/*
327 	 * CPUID with %eax = 1, %ebx returns
328 	 * Bits 15-8: CLFLUSH line size
329 	 * 	(Value * 8 = cache line size in bytes)
330 	 */
331 	if ((cpu_feature & CPUID_CLFSH) != 0)
332 		cpu_clflush_line_size = ((cpu_procinfo >> 8) & 0xff) * 8;
333 	/*
334 	 * XXXKIB: (temporary) hack to work around traps generated
335 	 * when CLFLUSHing APIC register window under virtualization
336 	 * environments.  These environments tend to disable the
337 	 * CPUID_SS feature even though the native CPU supports it.
338 	 */
339 	TUNABLE_INT_FETCH("hw.clflush_disable", &hw_clflush_disable);
340 	if (vm_guest != VM_GUEST_NO && hw_clflush_disable == -1) {
341 		cpu_feature &= ~CPUID_CLFSH;
342 		cpu_stdext_feature &= ~CPUID_STDEXT_CLFLUSHOPT;
343 	}
344 
345 	/*
346 	 * The kernel's use of CLFLUSH{,OPT} can be disabled manually
347 	 * by setting the hw.clflush_disable tunable.
348 	 */
349 	if (hw_clflush_disable == 1) {
350 		cpu_feature &= ~CPUID_CLFSH;
351 		cpu_stdext_feature &= ~CPUID_STDEXT_CLFLUSHOPT;
352 	}
353 }
354