xref: /freebsd/sys/arm64/linux/linux_sysvec.c (revision d4f55cc861b532cc50cf9013cf09afed995396ed) 
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1994-1996 Søren Schmidt
5  * Copyright (c) 2018 Turing Robotic Industries Inc.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/cdefs.h>
35 #include <sys/elf.h>
36 #include <sys/exec.h>
37 #include <sys/imgact.h>
38 #include <sys/imgact_elf.h>
39 #include <sys/kernel.h>
40 #include <sys/lock.h>
41 #include <sys/module.h>
42 #include <sys/mutex.h>
43 #include <sys/proc.h>
44 #include <sys/signalvar.h>
45 #include <sys/sysctl.h>
46 #include <sys/sysent.h>
47 
48 #include <vm/vm_param.h>
49 
50 #include <arm64/linux/linux.h>
51 #include <arm64/linux/linux_proto.h>
52 #include <compat/linux/linux_dtrace.h>
53 #include <compat/linux/linux_emul.h>
54 #include <compat/linux/linux_ioctl.h>
55 #include <compat/linux/linux_mib.h>
56 #include <compat/linux/linux_misc.h>
57 #include <compat/linux/linux_vdso.h>
58 
59 MODULE_VERSION(linux64elf, 1);
60 
61 const char *linux_kplatform;
62 static int linux_szsigcode;
63 static vm_object_t linux_shared_page_obj;
64 static char *linux_shared_page_mapping;
65 extern char _binary_linux_locore_o_start;
66 extern char _binary_linux_locore_o_end;
67 
68 extern struct sysent linux_sysent[LINUX_SYS_MAXSYSCALL];
69 
70 SET_DECLARE(linux_ioctl_handler_set, struct linux_ioctl_handler);
71 
72 static register_t *linux_copyout_strings(struct image_params *imgp);
73 static int	linux_elf_fixup(register_t **stack_base,
74 		    struct image_params *iparams);
75 static bool	linux_trans_osrel(const Elf_Note *note, int32_t *osrel);
76 static void	linux_vdso_install(const void *param);
77 static void	linux_vdso_deinstall(const void *param);
78 static void	linux_set_syscall_retval(struct thread *td, int error);
79 static int	linux_fetch_syscall_args(struct thread *td);
80 static void	linux_exec_setregs(struct thread *td, struct image_params *imgp,
81 		    u_long stack);
82 static int	linux_vsyscall(struct thread *td);
83 
84 /* DTrace init */
85 LIN_SDT_PROVIDER_DECLARE(LINUX_DTRACE);
86 
87 /* DTrace probes */
88 LIN_SDT_PROBE_DEFINE2(sysvec, linux_translate_traps, todo, "int", "int");
89 LIN_SDT_PROBE_DEFINE0(sysvec, linux_exec_setregs, todo);
90 LIN_SDT_PROBE_DEFINE0(sysvec, linux_elf_fixup, todo);
91 LIN_SDT_PROBE_DEFINE0(sysvec, linux_rt_sigreturn, todo);
92 LIN_SDT_PROBE_DEFINE0(sysvec, linux_rt_sendsig, todo);
93 LIN_SDT_PROBE_DEFINE0(sysvec, linux_vsyscall, todo);
94 LIN_SDT_PROBE_DEFINE0(sysvec, linux_vdso_install, todo);
95 LIN_SDT_PROBE_DEFINE0(sysvec, linux_vdso_deinstall, todo);
96 
97 /* LINUXTODO: do we have traps to translate? */
98 static int
99 linux_translate_traps(int signal, int trap_code)
100 {
101 
102 	LIN_SDT_PROBE2(sysvec, linux_translate_traps, todo, signal, trap_code);
103 	return (signal);
104 }
105 
106 LINUX_VDSO_SYM_CHAR(linux_platform);
107 
108 static int
109 linux_fetch_syscall_args(struct thread *td)
110 {
111 	struct proc *p;
112 	struct syscall_args *sa;
113 	register_t *ap;
114 
115 	p = td->td_proc;
116 	ap = td->td_frame->tf_x;
117 	sa = &td->td_sa;
118 
119 	sa->code = td->td_frame->tf_x[8];
120 	/* LINUXTODO: generic syscall? */
121 	if (sa->code >= p->p_sysent->sv_size)
122 		sa->callp = &p->p_sysent->sv_table[0];
123 	else
124 		sa->callp = &p->p_sysent->sv_table[sa->code];
125 
126 	sa->narg = sa->callp->sy_narg;
127 	if (sa->narg > 8)
128 		panic("ARM64TODO: Could we have more than 8 args?");
129 	memcpy(sa->args, ap, 8 * sizeof(register_t));
130 
131 	td->td_retval[0] = 0;
132 	return (0);
133 }
134 
135 static void
136 linux_set_syscall_retval(struct thread *td, int error)
137 {
138 
139 	td->td_retval[1] = td->td_frame->tf_x[1];
140 	cpu_set_syscall_retval(td, error);
141 }
142 
143 static int
144 linux_elf_fixup(register_t **stack_base, struct image_params *imgp)
145 {
146 	Elf_Auxargs *args;
147 	Elf_Auxinfo *argarray, *pos;
148 	Elf_Addr *auxbase, *base;
149 	struct ps_strings *arginfo;
150 	struct proc *p;
151 	int error, issetugid;
152 
153 	LIN_SDT_PROBE0(sysvec, linux_elf_fixup, todo);
154 	p = imgp->proc;
155 	arginfo = (struct ps_strings *)p->p_sysent->sv_psstrings;
156 
157 	KASSERT(curthread->td_proc == imgp->proc,
158 	    ("unsafe linux_elf_fixup(), should be curproc"));
159 	base = (Elf64_Addr *)*stack_base;
160 	args = (Elf64_Auxargs *)imgp->auxargs;
161 	/* Auxargs after argc, and NULL-terminated argv and envv lists. */
162 	auxbase = base + 1 + imgp->args->argc + 1 + imgp->args->envc + 1;
163 	argarray = pos = malloc(LINUX_AT_COUNT * sizeof(*pos), M_TEMP,
164 	    M_WAITOK | M_ZERO);
165 
166 	issetugid = p->p_flag & P_SUGID ? 1 : 0;
167 	AUXARGS_ENTRY(pos, LINUX_AT_SYSINFO_EHDR,
168 	    imgp->proc->p_sysent->sv_shared_page_base);
169 #if 0	/* LINUXTODO: implement arm64 LINUX_AT_HWCAP */
170 	AUXARGS_ENTRY(pos, LINUX_AT_HWCAP, cpu_feature);
171 #endif
172 	AUXARGS_ENTRY(pos, LINUX_AT_CLKTCK, stclohz);
173 	AUXARGS_ENTRY(pos, AT_PHDR, args->phdr);
174 	AUXARGS_ENTRY(pos, AT_PHENT, args->phent);
175 	AUXARGS_ENTRY(pos, AT_PHNUM, args->phnum);
176 	AUXARGS_ENTRY(pos, AT_PAGESZ, args->pagesz);
177 	AUXARGS_ENTRY(pos, AT_BASE, args->base);
178 	AUXARGS_ENTRY(pos, AT_FLAGS, args->flags);
179 	AUXARGS_ENTRY(pos, AT_ENTRY, args->entry);
180 	AUXARGS_ENTRY(pos, AT_UID, imgp->proc->p_ucred->cr_ruid);
181 	AUXARGS_ENTRY(pos, AT_EUID, imgp->proc->p_ucred->cr_svuid);
182 	AUXARGS_ENTRY(pos, AT_GID, imgp->proc->p_ucred->cr_rgid);
183 	AUXARGS_ENTRY(pos, AT_EGID, imgp->proc->p_ucred->cr_svgid);
184 	AUXARGS_ENTRY(pos, LINUX_AT_SECURE, issetugid);
185 #if 0	/* LINUXTODO: implement arm64 LINUX_AT_PLATFORM */
186 	AUXARGS_ENTRY(pos, LINUX_AT_PLATFORM, PTROUT(linux_platform));
187 #endif
188 	AUXARGS_ENTRY(pos, LINUX_AT_RANDOM, imgp->canary);
189 	if (imgp->execpathp != 0)
190 		AUXARGS_ENTRY(pos, LINUX_AT_EXECFN, imgp->execpathp);
191 	if (args->execfd != -1)
192 		AUXARGS_ENTRY(pos, AT_EXECFD, args->execfd);
193 	AUXARGS_ENTRY(pos, AT_NULL, 0);
194 	free(imgp->auxargs, M_TEMP);
195 	imgp->auxargs = NULL;
196 	KASSERT(pos - argarray <= LINUX_AT_COUNT, ("Too many auxargs"));
197 
198 	error = copyout(argarray, auxbase, sizeof(*argarray) * LINUX_AT_COUNT);
199 	free(argarray, M_TEMP);
200 	if (error != 0)
201 		return (error);
202 
203 	return (0);
204 }
205 
206 /*
207  * Copy strings out to the new process address space, constructing new arg
208  * and env vector tables. Return a pointer to the base so that it can be used
209  * as the initial stack pointer.
210  * LINUXTODO: deduplicate against other linuxulator archs
211  */
212 static register_t *
213 linux_copyout_strings(struct image_params *imgp)
214 {
215 	char **vectp;
216 	char *stringp, *destp;
217 	register_t *stack_base;
218 	struct ps_strings *arginfo;
219 	char canary[LINUX_AT_RANDOM_LEN];
220 	size_t execpath_len;
221 	struct proc *p;
222 	int argc, envc;
223 
224 	/* Calculate string base and vector table pointers. */
225 	if (imgp->execpath != NULL && imgp->auxargs != NULL)
226 		execpath_len = strlen(imgp->execpath) + 1;
227 	else
228 		execpath_len = 0;
229 
230 	p = imgp->proc;
231 	arginfo = (struct ps_strings *)p->p_sysent->sv_psstrings;
232 	destp = (caddr_t)arginfo - SPARE_USRSPACE -
233 	    roundup(sizeof(canary), sizeof(char *)) -
234 	    roundup(execpath_len, sizeof(char *)) -
235 	    roundup(ARG_MAX - imgp->args->stringspace, sizeof(char *));
236 
237 	if (execpath_len != 0) {
238 		imgp->execpathp = (uintptr_t)arginfo - execpath_len;
239 		copyout(imgp->execpath, (void *)imgp->execpathp, execpath_len);
240 	}
241 
242 	/* Prepare the canary for SSP. */
243 	arc4rand(canary, sizeof(canary), 0);
244 	imgp->canary = (uintptr_t)arginfo -
245 	    roundup(execpath_len, sizeof(char *)) -
246 	    roundup(sizeof(canary), sizeof(char *));
247 	copyout(canary, (void *)imgp->canary, sizeof(canary));
248 
249 	vectp = (char **)destp;
250 	if (imgp->auxargs) {
251 		/*
252 		 * Allocate room on the stack for the ELF auxargs
253 		 * array.  It has up to LINUX_AT_COUNT entries.
254 		 */
255 		vectp -= howmany(LINUX_AT_COUNT * sizeof(Elf64_Auxinfo),
256 		    sizeof(*vectp));
257 	}
258 
259 	/*
260 	 * Allocate room for argc and the argv[] and env vectors including the
261 	 * terminating NULL pointers.
262 	 */
263 	vectp -= 1 + imgp->args->argc + 1 + imgp->args->envc + 1;
264 	vectp = (char **)STACKALIGN(vectp);
265 
266 	/* vectp also becomes our initial stack base. */
267 	stack_base = (register_t *)vectp;
268 
269 	stringp = imgp->args->begin_argv;
270 	argc = imgp->args->argc;
271 	envc = imgp->args->envc;
272 
273 	/* Copy out strings - arguments and environment. */
274 	copyout(stringp, destp, ARG_MAX - imgp->args->stringspace);
275 
276 	/* Fill in "ps_strings" struct for ps, w, etc. */
277 	suword(&arginfo->ps_argvstr, (long)(intptr_t)vectp);
278 	suword(&arginfo->ps_nargvstr, argc);
279 
280 	suword(vectp++, argc);
281 	/* Fill in argument portion of vector table. */
282 	for (; argc > 0; --argc) {
283 		suword(vectp++, (long)(intptr_t)destp);
284 		while (*stringp++ != 0)
285 			destp++;
286 		destp++;
287 	}
288 
289 	/* A null vector table pointer separates the argp's from the envp's. */
290 	suword(vectp++, 0);
291 
292 	suword(&arginfo->ps_envstr, (long)(intptr_t)vectp);
293 	suword(&arginfo->ps_nenvstr, envc);
294 
295 	/* Fill in environment portion of vector table. */
296 	for (; envc > 0; --envc) {
297 		suword(vectp++, (long)(intptr_t)destp);
298 		while (*stringp++ != 0)
299 			destp++;
300 		destp++;
301 	}
302 
303 	/* The end of the vector table is a null pointer. */
304 	suword(vectp, 0);
305 	return (stack_base);
306 }
307 
308 /*
309  * Reset registers to default values on exec.
310  */
311 static void
312 linux_exec_setregs(struct thread *td, struct image_params *imgp, u_long stack)
313 {
314 	struct trapframe *regs = td->td_frame;
315 
316 	/* LINUXTODO: validate */
317 	LIN_SDT_PROBE0(sysvec, linux_exec_setregs, todo);
318 
319 	memset(regs, 0, sizeof(*regs));
320 	/* glibc start.S registers function pointer in x0 with atexit. */
321         regs->tf_sp = stack;
322 #if 0	/* LINUXTODO: See if this is used. */
323 	regs->tf_lr = imgp->entry_addr;
324 #else
325         regs->tf_lr = 0xffffffffffffffff;
326 #endif
327         regs->tf_elr = imgp->entry_addr;
328 }
329 
330 int
331 linux_rt_sigreturn(struct thread *td, struct linux_rt_sigreturn_args *args)
332 {
333 
334 	/* LINUXTODO: implement */
335 	LIN_SDT_PROBE0(sysvec, linux_rt_sigreturn, todo);
336 	return (EDOOFUS);
337 }
338 
339 static void
340 linux_rt_sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
341 {
342 
343 	/* LINUXTODO: implement */
344 	LIN_SDT_PROBE0(sysvec, linux_rt_sendsig, todo);
345 }
346 
347 static int
348 linux_vsyscall(struct thread *td)
349 {
350 
351 	/* LINUXTODO: implement */
352 	LIN_SDT_PROBE0(sysvec, linux_vsyscall, todo);
353 	return (EDOOFUS);
354 }
355 
356 struct sysentvec elf_linux_sysvec = {
357 	.sv_size	= LINUX_SYS_MAXSYSCALL,
358 	.sv_table	= linux_sysent,
359 	.sv_errsize	= ELAST + 1,
360 	.sv_errtbl	= linux_errtbl,
361 	.sv_transtrap	= linux_translate_traps,
362 	.sv_fixup	= linux_elf_fixup,
363 	.sv_sendsig	= linux_rt_sendsig,
364 	.sv_sigcode	= &_binary_linux_locore_o_start,
365 	.sv_szsigcode	= &linux_szsigcode,
366 	.sv_name	= "Linux ELF64",
367 	.sv_coredump	= elf64_coredump,
368 	.sv_imgact_try	= linux_exec_imgact_try,
369 	.sv_minsigstksz	= LINUX_MINSIGSTKSZ,
370 	.sv_minuser	= VM_MIN_ADDRESS,
371 	.sv_maxuser	= VM_MAXUSER_ADDRESS,
372 	.sv_usrstack	= USRSTACK,
373 	.sv_psstrings	= PS_STRINGS, /* XXX */
374 	.sv_stackprot	= VM_PROT_READ | VM_PROT_WRITE,
375 	.sv_copyout_strings = linux_copyout_strings,
376 	.sv_setregs	= linux_exec_setregs,
377 	.sv_fixlimit	= NULL,
378 	.sv_maxssiz	= NULL,
379 	.sv_flags	= SV_ABI_LINUX | SV_LP64 | SV_SHP,
380 	.sv_set_syscall_retval = linux_set_syscall_retval,
381 	.sv_fetch_syscall_args = linux_fetch_syscall_args,
382 	.sv_syscallnames = NULL,
383 	.sv_shared_page_base = SHAREDPAGE,
384 	.sv_shared_page_len = PAGE_SIZE,
385 	.sv_schedtail	= linux_schedtail,
386 	.sv_thread_detach = linux_thread_detach,
387 	.sv_trap	= linux_vsyscall,
388 };
389 
390 static void
391 linux_vdso_install(const void *param)
392 {
393 
394 	linux_szsigcode = (&_binary_linux_locore_o_end -
395 	    &_binary_linux_locore_o_start);
396 
397 	if (linux_szsigcode > elf_linux_sysvec.sv_shared_page_len)
398 		panic("invalid Linux VDSO size\n");
399 
400 	__elfN(linux_vdso_fixup)(&elf_linux_sysvec);
401 
402 	linux_shared_page_obj = __elfN(linux_shared_page_init)
403 	    (&linux_shared_page_mapping);
404 
405 	__elfN(linux_vdso_reloc)(&elf_linux_sysvec);
406 
407 	memcpy(linux_shared_page_mapping, elf_linux_sysvec.sv_sigcode,
408 	    linux_szsigcode);
409 	elf_linux_sysvec.sv_shared_page_obj = linux_shared_page_obj;
410 
411 	printf("LINUXTODO: %s: fix linux_kplatform\n", __func__);
412 #if 0
413 	linux_kplatform = linux_shared_page_mapping +
414 	    (linux_platform - (caddr_t)elf_linux_sysvec.sv_shared_page_base);
415 #else
416 	linux_kplatform = "arm64";
417 #endif
418 }
419 SYSINIT(elf_linux_vdso_init, SI_SUB_EXEC, SI_ORDER_ANY,
420     linux_vdso_install, NULL);
421 
422 static void
423 linux_vdso_deinstall(const void *param)
424 {
425 
426 	LIN_SDT_PROBE0(sysvec, linux_vdso_deinstall, todo);
427 	__elfN(linux_shared_page_fini)(linux_shared_page_obj);
428 }
429 SYSUNINIT(elf_linux_vdso_uninit, SI_SUB_EXEC, SI_ORDER_FIRST,
430     linux_vdso_deinstall, NULL);
431 
432 static char GNU_ABI_VENDOR[] = "GNU";
433 static int GNU_ABI_LINUX = 0;
434 
435 /* LINUXTODO: deduplicate */
436 static bool
437 linux_trans_osrel(const Elf_Note *note, int32_t *osrel)
438 {
439 	const Elf32_Word *desc;
440 	uintptr_t p;
441 
442 	p = (uintptr_t)(note + 1);
443 	p += roundup2(note->n_namesz, sizeof(Elf32_Addr));
444 
445 	desc = (const Elf32_Word *)p;
446 	if (desc[0] != GNU_ABI_LINUX)
447 		return (false);
448 
449 	*osrel = LINUX_KERNVER(desc[1], desc[2], desc[3]);
450 	return (true);
451 }
452 
453 static Elf_Brandnote linux64_brandnote = {
454 	.hdr.n_namesz	= sizeof(GNU_ABI_VENDOR),
455 	.hdr.n_descsz	= 16,
456 	.hdr.n_type	= 1,
457 	.vendor		= GNU_ABI_VENDOR,
458 	.flags		= BN_TRANSLATE_OSREL,
459 	.trans_osrel	= linux_trans_osrel
460 };
461 
462 static Elf64_Brandinfo linux_glibc2brand = {
463 	.brand		= ELFOSABI_LINUX,
464 	.machine	= EM_AARCH64,
465 	.compat_3_brand	= "Linux",
466 	.emul_path	= "/compat/linux",
467 	.interp_path	= "/lib64/ld-linux-x86-64.so.2",
468 	.sysvec		= &elf_linux_sysvec,
469 	.interp_newpath	= NULL,
470 	.brand_note	= &linux64_brandnote,
471 	.flags		= BI_CAN_EXEC_DYN | BI_BRAND_NOTE
472 };
473 
474 Elf64_Brandinfo *linux_brandlist[] = {
475 	&linux_glibc2brand,
476 	NULL
477 };
478 
479 static int
480 linux64_elf_modevent(module_t mod, int type, void *data)
481 {
482 	Elf64_Brandinfo **brandinfo;
483 	struct linux_ioctl_handler**lihp;
484 	int error;
485 
486 	error = 0;
487 	switch(type) {
488 	case MOD_LOAD:
489 		for (brandinfo = &linux_brandlist[0]; *brandinfo != NULL;
490 		    ++brandinfo)
491 			if (elf64_insert_brand_entry(*brandinfo) < 0)
492 				error = EINVAL;
493 		if (error == 0) {
494 			SET_FOREACH(lihp, linux_ioctl_handler_set)
495 				linux_ioctl_register_handler(*lihp);
496 			stclohz = (stathz ? stathz : hz);
497 			if (bootverbose)
498 				printf("Linux arm64 ELF exec handler installed\n");
499 		}
500 		break;
501 	case MOD_UNLOAD:
502 		for (brandinfo = &linux_brandlist[0]; *brandinfo != NULL;
503 		    ++brandinfo)
504 			if (elf64_brand_inuse(*brandinfo))
505 				error = EBUSY;
506 		if (error == 0) {
507 			for (brandinfo = &linux_brandlist[0];
508 			    *brandinfo != NULL; ++brandinfo)
509 				if (elf64_remove_brand_entry(*brandinfo) < 0)
510 					error = EINVAL;
511 		}
512 		if (error == 0) {
513 			SET_FOREACH(lihp, linux_ioctl_handler_set)
514 				linux_ioctl_unregister_handler(*lihp);
515 			if (bootverbose)
516 				printf("Linux ELF exec handler removed\n");
517 		} else
518 			printf("Could not deinstall ELF interpreter entry\n");
519 		break;
520 	default:
521 		return (EOPNOTSUPP);
522 	}
523 	return (error);
524 }
525 
526 static moduledata_t linux64_elf_mod = {
527 	"linux64elf",
528 	linux64_elf_modevent,
529 	0
530 };
531 
532 DECLARE_MODULE_TIED(linux64elf, linux64_elf_mod, SI_SUB_EXEC, SI_ORDER_ANY);
533 MODULE_DEPEND(linux64elf, linux_common, 1, 1, 1);
534 FEATURE(linux64, "AArch64 Linux 64bit support");
535