xref: /linux/arch/um/os-Linux/skas/process.c (revision 399ead3a6d76cbdd29a716660db5c84a314dab70)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2021 Benjamin Berg <benjamin@sipsolutions.net>
4  * Copyright (C) 2015 Thomas Meyer (thomas@m3y3r.de)
5  * Copyright (C) 2002- 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
6  */
7 
8 #include <stdlib.h>
9 #include <stdbool.h>
10 #include <unistd.h>
11 #include <sched.h>
12 #include <errno.h>
13 #include <string.h>
14 #include <fcntl.h>
15 #include <mem_user.h>
16 #include <sys/mman.h>
17 #include <sys/wait.h>
18 #include <sys/stat.h>
19 #include <sys/socket.h>
20 #include <asm/unistd.h>
21 #include <as-layout.h>
22 #include <init.h>
23 #include <kern_util.h>
24 #include <mem.h>
25 #include <os.h>
26 #include <ptrace_user.h>
27 #include <registers.h>
28 #include <skas.h>
29 #include <sysdep/stub.h>
30 #include <sysdep/mcontext.h>
31 #include <linux/futex.h>
32 #include <linux/threads.h>
33 #include <timetravel.h>
34 #include <asm-generic/rwonce.h>
35 #include "../internal.h"
36 
is_skas_winch(int pid,int fd,void * data)37 int is_skas_winch(int pid, int fd, void *data)
38 {
39 	return pid == getpgrp();
40 }
41 
ptrace_reg_name(int idx)42 static const char *ptrace_reg_name(int idx)
43 {
44 #define R(n) case HOST_##n: return #n
45 
46 	switch (idx) {
47 #ifdef __x86_64__
48 	R(BX);
49 	R(CX);
50 	R(DI);
51 	R(SI);
52 	R(DX);
53 	R(BP);
54 	R(AX);
55 	R(R8);
56 	R(R9);
57 	R(R10);
58 	R(R11);
59 	R(R12);
60 	R(R13);
61 	R(R14);
62 	R(R15);
63 	R(ORIG_AX);
64 	R(CS);
65 	R(SS);
66 	R(EFLAGS);
67 #elif defined(__i386__)
68 	R(IP);
69 	R(SP);
70 	R(EFLAGS);
71 	R(AX);
72 	R(BX);
73 	R(CX);
74 	R(DX);
75 	R(SI);
76 	R(DI);
77 	R(BP);
78 	R(CS);
79 	R(SS);
80 	R(DS);
81 	R(FS);
82 	R(ES);
83 	R(GS);
84 	R(ORIG_AX);
85 #endif
86 	}
87 	return "";
88 }
89 
ptrace_dump_regs(int pid)90 static int ptrace_dump_regs(int pid)
91 {
92 	unsigned long regs[MAX_REG_NR];
93 	int i;
94 
95 	if (ptrace(PTRACE_GETREGS, pid, 0, regs) < 0)
96 		return -errno;
97 
98 	printk(UM_KERN_ERR "Stub registers -\n");
99 	for (i = 0; i < ARRAY_SIZE(regs); i++) {
100 		const char *regname = ptrace_reg_name(i);
101 
102 		printk(UM_KERN_ERR "\t%s\t(%2d): %lx\n", regname, i, regs[i]);
103 	}
104 
105 	return 0;
106 }
107 
108 /*
109  * Signals that are OK to receive in the stub - we'll just continue it.
110  * SIGWINCH will happen when UML is inside a detached screen.
111  */
112 #define STUB_SIG_MASK ((1 << SIGALRM) | (1 << SIGWINCH))
113 
114 /* Signals that the stub will finish with - anything else is an error */
115 #define STUB_DONE_MASK (1 << SIGTRAP)
116 
wait_stub_done(int pid)117 void wait_stub_done(int pid)
118 {
119 	int n, status, err;
120 
121 	while (1) {
122 		CATCH_EINTR(n = waitpid(pid, &status, WUNTRACED | __WALL));
123 		if ((n < 0) || !WIFSTOPPED(status))
124 			goto bad_wait;
125 
126 		if (((1 << WSTOPSIG(status)) & STUB_SIG_MASK) == 0)
127 			break;
128 
129 		err = ptrace(PTRACE_CONT, pid, 0, 0);
130 		if (err) {
131 			printk(UM_KERN_ERR "%s : continue failed, errno = %d\n",
132 			       __func__, errno);
133 			fatal_sigsegv();
134 		}
135 	}
136 
137 	if (((1 << WSTOPSIG(status)) & STUB_DONE_MASK) != 0)
138 		return;
139 
140 bad_wait:
141 	err = ptrace_dump_regs(pid);
142 	if (err)
143 		printk(UM_KERN_ERR "Failed to get registers from stub, errno = %d\n",
144 		       -err);
145 	printk(UM_KERN_ERR "%s : failed to wait for SIGTRAP, pid = %d, n = %d, errno = %d, status = 0x%x\n",
146 	       __func__, pid, n, errno, status);
147 	fatal_sigsegv();
148 }
149 
wait_stub_done_seccomp(struct mm_id * mm_idp,int running,int wait_sigsys)150 void wait_stub_done_seccomp(struct mm_id *mm_idp, int running, int wait_sigsys)
151 {
152 	struct stub_data *data = (void *)mm_idp->stack;
153 	int ret;
154 
155 	do {
156 		const char byte = 0;
157 		struct iovec iov = {
158 			.iov_base = (void *)&byte,
159 			.iov_len = sizeof(byte),
160 		};
161 		union {
162 			char data[CMSG_SPACE(sizeof(mm_idp->syscall_fd_map))];
163 			struct cmsghdr align;
164 		} ctrl;
165 		struct msghdr msgh = {
166 			.msg_iov = &iov,
167 			.msg_iovlen = 1,
168 		};
169 
170 		if (!running) {
171 			if (mm_idp->syscall_fd_num) {
172 				unsigned int fds_size =
173 					sizeof(int) * mm_idp->syscall_fd_num;
174 				struct cmsghdr *cmsg;
175 
176 				msgh.msg_control = ctrl.data;
177 				msgh.msg_controllen = CMSG_SPACE(fds_size);
178 				cmsg = CMSG_FIRSTHDR(&msgh);
179 				cmsg->cmsg_level = SOL_SOCKET;
180 				cmsg->cmsg_type = SCM_RIGHTS;
181 				cmsg->cmsg_len = CMSG_LEN(fds_size);
182 				memcpy(CMSG_DATA(cmsg), mm_idp->syscall_fd_map,
183 				       fds_size);
184 
185 				CATCH_EINTR(syscall(__NR_sendmsg, mm_idp->sock,
186 						&msgh, 0));
187 			}
188 
189 			data->signal = 0;
190 			data->futex = FUTEX_IN_CHILD;
191 			CATCH_EINTR(syscall(__NR_futex, &data->futex,
192 					    FUTEX_WAKE, 1, NULL, NULL, 0));
193 		}
194 
195 		do {
196 			/*
197 			 * We need to check whether the child is still alive
198 			 * before and after the FUTEX_WAIT call. Before, in
199 			 * case it just died but we still updated data->futex
200 			 * to FUTEX_IN_CHILD. And after, in case it died while
201 			 * we were waiting (and SIGCHLD woke us up, see the
202 			 * IRQ handler in mmu.c).
203 			 *
204 			 * Either way, if PID is negative, then we have no
205 			 * choice but to kill the task.
206 			 */
207 			if (__READ_ONCE(mm_idp->pid) < 0)
208 				goto out_kill;
209 
210 			ret = syscall(__NR_futex, &data->futex,
211 				      FUTEX_WAIT, FUTEX_IN_CHILD,
212 				      NULL, NULL, 0);
213 			if (ret < 0 && errno != EINTR && errno != EAGAIN) {
214 				printk(UM_KERN_ERR "%s : FUTEX_WAIT failed, errno = %d\n",
215 				       __func__, errno);
216 				goto out_kill;
217 			}
218 		} while (data->futex == FUTEX_IN_CHILD);
219 
220 		if (__READ_ONCE(mm_idp->pid) < 0)
221 			goto out_kill;
222 
223 		running = 0;
224 
225 		/* We may receive a SIGALRM before SIGSYS, iterate again. */
226 	} while (wait_sigsys && data->signal == SIGALRM);
227 
228 	if (data->mctx_offset > sizeof(data->sigstack) - sizeof(mcontext_t)) {
229 		printk(UM_KERN_ERR "%s : invalid mcontext offset", __func__);
230 		goto out_kill;
231 	}
232 
233 	if (wait_sigsys && data->signal != SIGSYS) {
234 		printk(UM_KERN_ERR "%s : expected SIGSYS but got %d",
235 		       __func__, data->signal);
236 		goto out_kill;
237 	}
238 
239 	return;
240 
241 out_kill:
242 	printk(UM_KERN_ERR "%s : failed to wait for stub, pid = %d, errno = %d\n",
243 	       __func__, mm_idp->pid, errno);
244 	/* This is not true inside start_userspace */
245 	if (current_mm_id() == mm_idp)
246 		fatal_sigsegv();
247 }
248 
249 extern unsigned long current_stub_stack(void);
250 
get_skas_faultinfo(int pid,struct faultinfo * fi)251 static void get_skas_faultinfo(int pid, struct faultinfo *fi)
252 {
253 	int err;
254 
255 	err = ptrace(PTRACE_CONT, pid, 0, SIGSEGV);
256 	if (err) {
257 		printk(UM_KERN_ERR "Failed to continue stub, pid = %d, "
258 		       "errno = %d\n", pid, errno);
259 		fatal_sigsegv();
260 	}
261 	wait_stub_done(pid);
262 
263 	/*
264 	 * faultinfo is prepared by the stub_segv_handler at start of
265 	 * the stub stack page. We just have to copy it.
266 	 */
267 	memcpy(fi, (void *)current_stub_stack(), sizeof(*fi));
268 }
269 
handle_trap(struct uml_pt_regs * regs)270 static void handle_trap(struct uml_pt_regs *regs)
271 {
272 	if ((UPT_IP(regs) >= STUB_START) && (UPT_IP(regs) < STUB_END))
273 		fatal_sigsegv();
274 
275 	handle_syscall(regs);
276 }
277 
278 extern char __syscall_stub_start[];
279 
280 static int stub_exe_fd;
281 
282 struct tramp_data {
283 	struct stub_data *stub_data;
284 	/* 0 is inherited, 1 is the kernel side */
285 	int sockpair[2];
286 };
287 
288 #ifndef CLOSE_RANGE_CLOEXEC
289 #define CLOSE_RANGE_CLOEXEC	(1U << 2)
290 #endif
291 
userspace_tramp(void * data)292 static int userspace_tramp(void *data)
293 {
294 	struct tramp_data *tramp_data = data;
295 	char *const argv[] = { "uml-userspace", NULL };
296 	unsigned long long offset;
297 	struct stub_init_data init_data = {
298 		.seccomp = using_seccomp,
299 		.stub_start = STUB_START,
300 	};
301 	int ret;
302 
303 	if (using_seccomp) {
304 		init_data.signal_handler = STUB_CODE +
305 					   (unsigned long) stub_signal_interrupt -
306 					   (unsigned long) __syscall_stub_start;
307 		init_data.signal_restorer = STUB_CODE +
308 					   (unsigned long) stub_signal_restorer -
309 					   (unsigned long) __syscall_stub_start;
310 	} else {
311 		init_data.signal_handler = STUB_CODE +
312 					   (unsigned long) stub_segv_handler -
313 					   (unsigned long) __syscall_stub_start;
314 		init_data.signal_restorer = 0;
315 	}
316 
317 	init_data.stub_code_fd = phys_mapping(uml_to_phys(__syscall_stub_start),
318 					      &offset);
319 	init_data.stub_code_offset = MMAP_OFFSET(offset);
320 
321 	init_data.stub_data_fd = phys_mapping(uml_to_phys(tramp_data->stub_data),
322 					      &offset);
323 	init_data.stub_data_offset = MMAP_OFFSET(offset);
324 
325 	/*
326 	 * Avoid leaking unneeded FDs to the stub by setting CLOEXEC on all FDs
327 	 * and then unsetting it on all memory related FDs.
328 	 * This is not strictly necessary from a safety perspective.
329 	 */
330 	syscall(__NR_close_range, 0, ~0U, CLOSE_RANGE_CLOEXEC);
331 
332 	fcntl(init_data.stub_data_fd, F_SETFD, 0);
333 
334 	/* dup2 signaling FD/socket to STDIN */
335 	if (dup2(tramp_data->sockpair[0], 0) < 0)
336 		exit(3);
337 	close(tramp_data->sockpair[0]);
338 
339 	/* Write init_data and close write side */
340 	ret = write(tramp_data->sockpair[1], &init_data, sizeof(init_data));
341 	close(tramp_data->sockpair[1]);
342 
343 	if (ret != sizeof(init_data))
344 		exit(4);
345 
346 	/* Raw execveat for compatibility with older libc versions */
347 	syscall(__NR_execveat, stub_exe_fd, (unsigned long)"",
348 		(unsigned long)argv, NULL, AT_EMPTY_PATH);
349 
350 	exit(5);
351 }
352 
353 extern char stub_exe_start[];
354 extern char stub_exe_end[];
355 
356 extern char *tempdir;
357 
358 #define STUB_EXE_NAME_TEMPLATE "/uml-userspace-XXXXXX"
359 
360 #ifndef MFD_EXEC
361 #define MFD_EXEC 0x0010U
362 #endif
363 
init_stub_exe_fd(void)364 static int __init init_stub_exe_fd(void)
365 {
366 	size_t written = 0;
367 	char *tmpfile = NULL;
368 
369 	stub_exe_fd = memfd_create("uml-userspace",
370 				   MFD_EXEC | MFD_CLOEXEC | MFD_ALLOW_SEALING);
371 
372 	if (stub_exe_fd < 0) {
373 		printk(UM_KERN_INFO "Could not create executable memfd, using temporary file!");
374 
375 		tmpfile = malloc(strlen(tempdir) +
376 				  strlen(STUB_EXE_NAME_TEMPLATE) + 1);
377 		if (tmpfile == NULL)
378 			panic("Failed to allocate memory for stub binary name");
379 
380 		strcpy(tmpfile, tempdir);
381 		strcat(tmpfile, STUB_EXE_NAME_TEMPLATE);
382 
383 		stub_exe_fd = mkstemp(tmpfile);
384 		if (stub_exe_fd < 0)
385 			panic("Could not create temporary file for stub binary: %d",
386 			      -errno);
387 	}
388 
389 	while (written < stub_exe_end - stub_exe_start) {
390 		ssize_t res = write(stub_exe_fd, stub_exe_start + written,
391 				    stub_exe_end - stub_exe_start - written);
392 		if (res < 0) {
393 			if (errno == EINTR)
394 				continue;
395 
396 			if (tmpfile)
397 				unlink(tmpfile);
398 			panic("Failed write stub binary: %d", -errno);
399 		}
400 
401 		written += res;
402 	}
403 
404 	if (!tmpfile) {
405 		fcntl(stub_exe_fd, F_ADD_SEALS,
406 		      F_SEAL_WRITE | F_SEAL_SHRINK | F_SEAL_GROW | F_SEAL_SEAL);
407 	} else {
408 		if (fchmod(stub_exe_fd, 00500) < 0) {
409 			unlink(tmpfile);
410 			panic("Could not make stub binary executable: %d",
411 			      -errno);
412 		}
413 
414 		close(stub_exe_fd);
415 		stub_exe_fd = open(tmpfile, O_RDONLY | O_CLOEXEC | O_NOFOLLOW);
416 		if (stub_exe_fd < 0) {
417 			unlink(tmpfile);
418 			panic("Could not reopen stub binary: %d", -errno);
419 		}
420 
421 		unlink(tmpfile);
422 		free(tmpfile);
423 	}
424 
425 	return 0;
426 }
427 __initcall(init_stub_exe_fd);
428 
429 int using_seccomp;
430 
431 /**
432  * start_userspace() - prepare a new userspace process
433  * @mm_id: The corresponding struct mm_id
434  *
435  * Setups a new temporary stack page that is used while userspace_tramp() runs
436  * Clones the kernel process into a new userspace process, with FDs only.
437  *
438  * Return: When positive: the process id of the new userspace process,
439  *         when negative: an error number.
440  * FIXME: can PIDs become negative?!
441  */
start_userspace(struct mm_id * mm_id)442 int start_userspace(struct mm_id *mm_id)
443 {
444 	struct stub_data *proc_data = (void *)mm_id->stack;
445 	struct tramp_data tramp_data = {
446 		.stub_data = proc_data,
447 	};
448 	void *stack;
449 	unsigned long sp;
450 	int status, n, err;
451 
452 	/* setup a temporary stack page */
453 	stack = mmap(NULL, UM_KERN_PAGE_SIZE,
454 		     PROT_READ | PROT_WRITE | PROT_EXEC,
455 		     MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
456 	if (stack == MAP_FAILED) {
457 		err = -errno;
458 		printk(UM_KERN_ERR "%s : mmap failed, errno = %d\n",
459 		       __func__, errno);
460 		return err;
461 	}
462 
463 	/* set stack pointer to the end of the stack page, so it can grow downwards */
464 	sp = (unsigned long)stack + UM_KERN_PAGE_SIZE;
465 
466 	/* socket pair for init data and SECCOMP FD passing (no CLOEXEC here) */
467 	if (socketpair(AF_UNIX, SOCK_STREAM, 0, tramp_data.sockpair)) {
468 		err = -errno;
469 		printk(UM_KERN_ERR "%s : socketpair failed, errno = %d\n",
470 		       __func__, errno);
471 		return err;
472 	}
473 
474 	if (using_seccomp)
475 		proc_data->futex = FUTEX_IN_CHILD;
476 
477 	mm_id->pid = clone(userspace_tramp, (void *) sp,
478 		    CLONE_VFORK | CLONE_VM | SIGCHLD,
479 		    (void *)&tramp_data);
480 	if (mm_id->pid < 0) {
481 		err = -errno;
482 		printk(UM_KERN_ERR "%s : clone failed, errno = %d\n",
483 		       __func__, errno);
484 		goto out_close;
485 	}
486 
487 	if (using_seccomp) {
488 		wait_stub_done_seccomp(mm_id, 1, 1);
489 	} else {
490 		do {
491 			CATCH_EINTR(n = waitpid(mm_id->pid, &status,
492 						WUNTRACED | __WALL));
493 			if (n < 0) {
494 				err = -errno;
495 				printk(UM_KERN_ERR "%s : wait failed, errno = %d\n",
496 				       __func__, errno);
497 				goto out_kill;
498 			}
499 		} while (WIFSTOPPED(status) && (WSTOPSIG(status) == SIGALRM));
500 
501 		if (!WIFSTOPPED(status) || (WSTOPSIG(status) != SIGSTOP)) {
502 			err = -EINVAL;
503 			printk(UM_KERN_ERR "%s : expected SIGSTOP, got status = %d\n",
504 			       __func__, status);
505 			goto out_kill;
506 		}
507 
508 		if (ptrace(PTRACE_SETOPTIONS, mm_id->pid, NULL,
509 			   (void *) PTRACE_O_TRACESYSGOOD) < 0) {
510 			err = -errno;
511 			printk(UM_KERN_ERR "%s : PTRACE_SETOPTIONS failed, errno = %d\n",
512 			       __func__, errno);
513 			goto out_kill;
514 		}
515 	}
516 
517 	if (munmap(stack, UM_KERN_PAGE_SIZE) < 0) {
518 		err = -errno;
519 		printk(UM_KERN_ERR "%s : munmap failed, errno = %d\n",
520 		       __func__, errno);
521 		goto out_kill;
522 	}
523 
524 	close(tramp_data.sockpair[0]);
525 	if (using_seccomp)
526 		mm_id->sock = tramp_data.sockpair[1];
527 	else
528 		close(tramp_data.sockpair[1]);
529 
530 	return 0;
531 
532 out_kill:
533 	os_kill_ptraced_process(mm_id->pid, 1);
534 out_close:
535 	close(tramp_data.sockpair[0]);
536 	close(tramp_data.sockpair[1]);
537 
538 	mm_id->pid = -1;
539 
540 	return err;
541 }
542 
543 static int unscheduled_userspace_iterations;
544 extern unsigned long tt_extra_sched_jiffies;
545 
userspace(struct uml_pt_regs * regs)546 void userspace(struct uml_pt_regs *regs)
547 {
548 	int err, status, op;
549 	siginfo_t si_local;
550 	siginfo_t *si;
551 	int sig;
552 
553 	/* Handle any immediate reschedules or signals */
554 	interrupt_end();
555 
556 	while (1) {
557 		struct mm_id *mm_id = current_mm_id();
558 
559 		/*
560 		 * At any given time, only one CPU thread can enter the
561 		 * turnstile to operate on the same stub process, including
562 		 * executing stub system calls (mmap and munmap).
563 		 */
564 		enter_turnstile(mm_id);
565 
566 		/*
567 		 * When we are in time-travel mode, userspace can theoretically
568 		 * do a *lot* of work without being scheduled. The problem with
569 		 * this is that it will prevent kernel bookkeeping (primarily
570 		 * the RCU) from running and this can for example cause OOM
571 		 * situations.
572 		 *
573 		 * This code accounts a jiffie against the scheduling clock
574 		 * after the defined userspace iterations in the same thread.
575 		 * By doing so the situation is effectively prevented.
576 		 */
577 		if (time_travel_mode == TT_MODE_INFCPU ||
578 		    time_travel_mode == TT_MODE_EXTERNAL) {
579 #ifdef CONFIG_UML_MAX_USERSPACE_ITERATIONS
580 			if (CONFIG_UML_MAX_USERSPACE_ITERATIONS &&
581 			    unscheduled_userspace_iterations++ >
582 			    CONFIG_UML_MAX_USERSPACE_ITERATIONS) {
583 				tt_extra_sched_jiffies += 1;
584 				unscheduled_userspace_iterations = 0;
585 			}
586 #endif
587 		}
588 
589 		time_travel_print_bc_msg();
590 
591 		current_mm_sync();
592 
593 		if (using_seccomp) {
594 			struct stub_data *proc_data = (void *) mm_id->stack;
595 
596 			err = set_stub_state(regs, proc_data, singlestepping());
597 			if (err) {
598 				printk(UM_KERN_ERR "%s - failed to set regs: %d",
599 				       __func__, err);
600 				fatal_sigsegv();
601 			}
602 
603 			/* Must have been reset by the syscall caller */
604 			if (proc_data->restart_wait != 0)
605 				panic("Programming error: Flag to only run syscalls in child was not cleared!");
606 
607 			/* Mark pending syscalls for flushing */
608 			proc_data->syscall_data_len = mm_id->syscall_data_len;
609 
610 			wait_stub_done_seccomp(mm_id, 0, 0);
611 
612 			sig = proc_data->signal;
613 
614 			if (sig == SIGTRAP && proc_data->err != 0) {
615 				printk(UM_KERN_ERR "%s - Error flushing stub syscalls",
616 				       __func__);
617 				syscall_stub_dump_error(mm_id);
618 				mm_id->syscall_data_len = proc_data->err;
619 				fatal_sigsegv();
620 			}
621 
622 			mm_id->syscall_data_len = 0;
623 			mm_id->syscall_fd_num = 0;
624 
625 			err = get_stub_state(regs, proc_data, NULL);
626 			if (err) {
627 				printk(UM_KERN_ERR "%s - failed to get regs: %d",
628 				       __func__, err);
629 				fatal_sigsegv();
630 			}
631 
632 			if (proc_data->si_offset > sizeof(proc_data->sigstack) - sizeof(*si))
633 				panic("%s - Invalid siginfo offset from child", __func__);
634 
635 			si = &si_local;
636 			memcpy(si, &proc_data->sigstack[proc_data->si_offset], sizeof(*si));
637 
638 			regs->is_user = 1;
639 
640 			/* Fill in ORIG_RAX and extract fault information */
641 			PT_SYSCALL_NR(regs->gp) = si->si_syscall;
642 			if (sig == SIGSEGV) {
643 				mcontext_t *mcontext = (void *)&proc_data->sigstack[proc_data->mctx_offset];
644 
645 				GET_FAULTINFO_FROM_MC(regs->faultinfo, mcontext);
646 			}
647 		} else {
648 			int pid = mm_id->pid;
649 
650 			/* Flush out any pending syscalls */
651 			err = syscall_stub_flush(mm_id);
652 			if (err) {
653 				if (err == -ENOMEM)
654 					report_enomem();
655 
656 				printk(UM_KERN_ERR "%s - Error flushing stub syscalls: %d",
657 					__func__, -err);
658 				fatal_sigsegv();
659 			}
660 
661 			/*
662 			 * This can legitimately fail if the process loads a
663 			 * bogus value into a segment register.  It will
664 			 * segfault and PTRACE_GETREGS will read that value
665 			 * out of the process.  However, PTRACE_SETREGS will
666 			 * fail.  In this case, there is nothing to do but
667 			 * just kill the process.
668 			 */
669 			if (ptrace(PTRACE_SETREGS, pid, 0, regs->gp)) {
670 				printk(UM_KERN_ERR "%s - ptrace set regs failed, errno = %d\n",
671 				       __func__, errno);
672 				fatal_sigsegv();
673 			}
674 
675 			if (put_fp_registers(pid, regs->fp)) {
676 				printk(UM_KERN_ERR "%s - ptrace set fp regs failed, errno = %d\n",
677 				       __func__, errno);
678 				fatal_sigsegv();
679 			}
680 
681 			if (singlestepping())
682 				op = PTRACE_SYSEMU_SINGLESTEP;
683 			else
684 				op = PTRACE_SYSEMU;
685 
686 			if (ptrace(op, pid, 0, 0)) {
687 				printk(UM_KERN_ERR "%s - ptrace continue failed, op = %d, errno = %d\n",
688 				       __func__, op, errno);
689 				fatal_sigsegv();
690 			}
691 
692 			CATCH_EINTR(err = waitpid(pid, &status, WUNTRACED | __WALL));
693 			if (err < 0) {
694 				printk(UM_KERN_ERR "%s - wait failed, errno = %d\n",
695 				       __func__, errno);
696 				fatal_sigsegv();
697 			}
698 
699 			regs->is_user = 1;
700 			if (ptrace(PTRACE_GETREGS, pid, 0, regs->gp)) {
701 				printk(UM_KERN_ERR "%s - PTRACE_GETREGS failed, errno = %d\n",
702 				       __func__, errno);
703 				fatal_sigsegv();
704 			}
705 
706 			if (get_fp_registers(pid, regs->fp)) {
707 				printk(UM_KERN_ERR "%s -  get_fp_registers failed, errno = %d\n",
708 				       __func__, errno);
709 				fatal_sigsegv();
710 			}
711 
712 			if (WIFSTOPPED(status)) {
713 				sig = WSTOPSIG(status);
714 
715 				/*
716 				 * These signal handlers need the si argument
717 				 * and SIGSEGV needs the faultinfo.
718 				 * The SIGIO and SIGALARM handlers which constitute
719 				 * the majority of invocations, do not use it.
720 				 */
721 				switch (sig) {
722 				case SIGSEGV:
723 					get_skas_faultinfo(pid,
724 							   &regs->faultinfo);
725 					fallthrough;
726 				case SIGTRAP:
727 				case SIGILL:
728 				case SIGBUS:
729 				case SIGFPE:
730 				case SIGWINCH:
731 					ptrace(PTRACE_GETSIGINFO, pid, 0,
732 					       (struct siginfo *)&si_local);
733 					si = &si_local;
734 					break;
735 				default:
736 					si = NULL;
737 					break;
738 				}
739 			} else {
740 				sig = 0;
741 			}
742 		}
743 
744 		exit_turnstile(mm_id);
745 
746 		UPT_SYSCALL_NR(regs) = -1; /* Assume: It's not a syscall */
747 
748 		if (sig) {
749 			switch (sig) {
750 			case SIGSEGV:
751 				if (using_seccomp || PTRACE_FULL_FAULTINFO)
752 					(*sig_info[SIGSEGV])(SIGSEGV,
753 							     (struct siginfo *)si,
754 							     regs, NULL);
755 				else
756 					segv(regs->faultinfo, 0, 1, NULL, NULL);
757 
758 				break;
759 			case SIGSYS:
760 				handle_syscall(regs);
761 				break;
762 			case SIGTRAP + 0x80:
763 				handle_trap(regs);
764 				break;
765 			case SIGTRAP:
766 				relay_signal(SIGTRAP, (struct siginfo *)si, regs, NULL);
767 				break;
768 			case SIGALRM:
769 				break;
770 			case SIGIO:
771 			case SIGILL:
772 			case SIGBUS:
773 			case SIGFPE:
774 			case SIGWINCH:
775 				block_signals_trace();
776 				(*sig_info[sig])(sig, (struct siginfo *)si, regs, NULL);
777 				unblock_signals_trace();
778 				break;
779 			default:
780 				printk(UM_KERN_ERR "%s - child stopped with signal %d\n",
781 				       __func__, sig);
782 				fatal_sigsegv();
783 			}
784 			interrupt_end();
785 
786 			/* Avoid -ERESTARTSYS handling in host */
787 			if (PT_SYSCALL_NR_OFFSET != PT_SYSCALL_RET_OFFSET)
788 				PT_SYSCALL_NR(regs->gp) = -1;
789 		}
790 	}
791 }
792 
new_thread(void * stack,jmp_buf * buf,void (* handler)(void))793 void new_thread(void *stack, jmp_buf *buf, void (*handler)(void))
794 {
795 	(*buf)[0].JB_IP = (unsigned long) handler;
796 	(*buf)[0].JB_SP = (unsigned long) stack + UM_THREAD_SIZE -
797 		sizeof(void *);
798 }
799 
800 #define INIT_JMP_NEW_THREAD 0
801 #define INIT_JMP_CALLBACK 1
802 #define INIT_JMP_HALT 2
803 #define INIT_JMP_REBOOT 3
804 
switch_threads(jmp_buf * me,jmp_buf * you)805 void switch_threads(jmp_buf *me, jmp_buf *you)
806 {
807 	unscheduled_userspace_iterations = 0;
808 
809 	if (UML_SETJMP(me) == 0)
810 		UML_LONGJMP(you, 1);
811 }
812 
813 static jmp_buf initial_jmpbuf;
814 
815 static __thread void (*cb_proc)(void *arg);
816 static __thread void *cb_arg;
817 static __thread jmp_buf *cb_back;
818 
start_idle_thread(void * stack,jmp_buf * switch_buf)819 int start_idle_thread(void *stack, jmp_buf *switch_buf)
820 {
821 	int n;
822 
823 	set_handler(SIGWINCH);
824 
825 	/*
826 	 * Can't use UML_SETJMP or UML_LONGJMP here because they save
827 	 * and restore signals, with the possible side-effect of
828 	 * trying to handle any signals which came when they were
829 	 * blocked, which can't be done on this stack.
830 	 * Signals must be blocked when jumping back here and restored
831 	 * after returning to the jumper.
832 	 */
833 	n = setjmp(initial_jmpbuf);
834 	switch (n) {
835 	case INIT_JMP_NEW_THREAD:
836 		(*switch_buf)[0].JB_IP = (unsigned long) uml_finishsetup;
837 		(*switch_buf)[0].JB_SP = (unsigned long) stack +
838 			UM_THREAD_SIZE - sizeof(void *);
839 		break;
840 	case INIT_JMP_CALLBACK:
841 		(*cb_proc)(cb_arg);
842 		longjmp(*cb_back, 1);
843 		break;
844 	case INIT_JMP_HALT:
845 		kmalloc_ok = 0;
846 		return 0;
847 	case INIT_JMP_REBOOT:
848 		kmalloc_ok = 0;
849 		return 1;
850 	default:
851 		printk(UM_KERN_ERR "Bad sigsetjmp return in %s - %d\n",
852 		       __func__, n);
853 		fatal_sigsegv();
854 	}
855 	longjmp(*switch_buf, 1);
856 
857 	/* unreachable */
858 	printk(UM_KERN_ERR "impossible long jump!");
859 	fatal_sigsegv();
860 	return 0;
861 }
862 
initial_thread_cb_skas(void (* proc)(void *),void * arg)863 void initial_thread_cb_skas(void (*proc)(void *), void *arg)
864 {
865 	jmp_buf here;
866 
867 	cb_proc = proc;
868 	cb_arg = arg;
869 	cb_back = &here;
870 
871 	initial_jmpbuf_lock();
872 	if (UML_SETJMP(&here) == 0)
873 		UML_LONGJMP(&initial_jmpbuf, INIT_JMP_CALLBACK);
874 	initial_jmpbuf_unlock();
875 
876 	cb_proc = NULL;
877 	cb_arg = NULL;
878 	cb_back = NULL;
879 }
880 
halt_skas(void)881 void halt_skas(void)
882 {
883 	initial_jmpbuf_lock();
884 	UML_LONGJMP(&initial_jmpbuf, INIT_JMP_HALT);
885 	/* unreachable */
886 }
887 
888 static bool noreboot;
889 
noreboot_cmd_param(char * str,int * add)890 static int __init noreboot_cmd_param(char *str, int *add)
891 {
892 	*add = 0;
893 	noreboot = true;
894 	return 0;
895 }
896 
897 __uml_setup("noreboot", noreboot_cmd_param,
898 "noreboot\n"
899 "    Rather than rebooting, exit always, akin to QEMU's -no-reboot option.\n"
900 "    This is useful if you're using CONFIG_PANIC_TIMEOUT in order to catch\n"
901 "    crashes in CI\n\n");
902 
reboot_skas(void)903 void reboot_skas(void)
904 {
905 	initial_jmpbuf_lock();
906 	UML_LONGJMP(&initial_jmpbuf, noreboot ? INIT_JMP_HALT : INIT_JMP_REBOOT);
907 	/* unreachable */
908 }
909