xref: /openbmc/linux/arch/arm64/kernel/mte.c (revision be58f710)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2020 ARM Ltd.
4  */
5 
6 #include <linux/bitops.h>
7 #include <linux/cpu.h>
8 #include <linux/kernel.h>
9 #include <linux/mm.h>
10 #include <linux/prctl.h>
11 #include <linux/sched.h>
12 #include <linux/sched/mm.h>
13 #include <linux/string.h>
14 #include <linux/swap.h>
15 #include <linux/swapops.h>
16 #include <linux/thread_info.h>
17 #include <linux/types.h>
18 #include <linux/uio.h>
19 
20 #include <asm/barrier.h>
21 #include <asm/cpufeature.h>
22 #include <asm/mte.h>
23 #include <asm/ptrace.h>
24 #include <asm/sysreg.h>
25 
26 static DEFINE_PER_CPU_READ_MOSTLY(u64, mte_tcf_preferred);
27 
28 #ifdef CONFIG_KASAN_HW_TAGS
29 /* Whether the MTE asynchronous mode is enabled. */
30 DEFINE_STATIC_KEY_FALSE(mte_async_mode);
31 EXPORT_SYMBOL_GPL(mte_async_mode);
32 #endif
33 
34 static void mte_sync_page_tags(struct page *page, pte_t old_pte,
35 			       bool check_swap, bool pte_is_tagged)
36 {
37 	if (check_swap && is_swap_pte(old_pte)) {
38 		swp_entry_t entry = pte_to_swp_entry(old_pte);
39 
40 		if (!non_swap_entry(entry) && mte_restore_tags(entry, page))
41 			return;
42 	}
43 
44 	if (!pte_is_tagged)
45 		return;
46 
47 	page_kasan_tag_reset(page);
48 	/*
49 	 * We need smp_wmb() in between setting the flags and clearing the
50 	 * tags because if another thread reads page->flags and builds a
51 	 * tagged address out of it, there is an actual dependency to the
52 	 * memory access, but on the current thread we do not guarantee that
53 	 * the new page->flags are visible before the tags were updated.
54 	 */
55 	smp_wmb();
56 	mte_clear_page_tags(page_address(page));
57 }
58 
59 void mte_sync_tags(pte_t old_pte, pte_t pte)
60 {
61 	struct page *page = pte_page(pte);
62 	long i, nr_pages = compound_nr(page);
63 	bool check_swap = nr_pages == 1;
64 	bool pte_is_tagged = pte_tagged(pte);
65 
66 	/* Early out if there's nothing to do */
67 	if (!check_swap && !pte_is_tagged)
68 		return;
69 
70 	/* if PG_mte_tagged is set, tags have already been initialised */
71 	for (i = 0; i < nr_pages; i++, page++) {
72 		if (!test_and_set_bit(PG_mte_tagged, &page->flags))
73 			mte_sync_page_tags(page, old_pte, check_swap,
74 					   pte_is_tagged);
75 	}
76 }
77 
78 int memcmp_pages(struct page *page1, struct page *page2)
79 {
80 	char *addr1, *addr2;
81 	int ret;
82 
83 	addr1 = page_address(page1);
84 	addr2 = page_address(page2);
85 	ret = memcmp(addr1, addr2, PAGE_SIZE);
86 
87 	if (!system_supports_mte() || ret)
88 		return ret;
89 
90 	/*
91 	 * If the page content is identical but at least one of the pages is
92 	 * tagged, return non-zero to avoid KSM merging. If only one of the
93 	 * pages is tagged, set_pte_at() may zero or change the tags of the
94 	 * other page via mte_sync_tags().
95 	 */
96 	if (test_bit(PG_mte_tagged, &page1->flags) ||
97 	    test_bit(PG_mte_tagged, &page2->flags))
98 		return addr1 != addr2;
99 
100 	return ret;
101 }
102 
103 static inline void __mte_enable_kernel(const char *mode, unsigned long tcf)
104 {
105 	/* Enable MTE Sync Mode for EL1. */
106 	sysreg_clear_set(sctlr_el1, SCTLR_ELx_TCF_MASK, tcf);
107 	isb();
108 
109 	pr_info_once("MTE: enabled in %s mode at EL1\n", mode);
110 }
111 
112 #ifdef CONFIG_KASAN_HW_TAGS
113 void mte_enable_kernel_sync(void)
114 {
115 	/*
116 	 * Make sure we enter this function when no PE has set
117 	 * async mode previously.
118 	 */
119 	WARN_ONCE(system_uses_mte_async_mode(),
120 			"MTE async mode enabled system wide!");
121 
122 	__mte_enable_kernel("synchronous", SCTLR_ELx_TCF_SYNC);
123 }
124 
125 void mte_enable_kernel_async(void)
126 {
127 	__mte_enable_kernel("asynchronous", SCTLR_ELx_TCF_ASYNC);
128 
129 	/*
130 	 * MTE async mode is set system wide by the first PE that
131 	 * executes this function.
132 	 *
133 	 * Note: If in future KASAN acquires a runtime switching
134 	 * mode in between sync and async, this strategy needs
135 	 * to be reviewed.
136 	 */
137 	if (!system_uses_mte_async_mode())
138 		static_branch_enable(&mte_async_mode);
139 }
140 #endif
141 
142 #ifdef CONFIG_KASAN_HW_TAGS
143 void mte_check_tfsr_el1(void)
144 {
145 	u64 tfsr_el1;
146 
147 	if (!system_supports_mte())
148 		return;
149 
150 	tfsr_el1 = read_sysreg_s(SYS_TFSR_EL1);
151 
152 	if (unlikely(tfsr_el1 & SYS_TFSR_EL1_TF1)) {
153 		/*
154 		 * Note: isb() is not required after this direct write
155 		 * because there is no indirect read subsequent to it
156 		 * (per ARM DDI 0487F.c table D13-1).
157 		 */
158 		write_sysreg_s(0, SYS_TFSR_EL1);
159 
160 		kasan_report_async();
161 	}
162 }
163 #endif
164 
165 static void mte_update_sctlr_user(struct task_struct *task)
166 {
167 	/*
168 	 * This must be called with preemption disabled and can only be called
169 	 * on the current or next task since the CPU must match where the thread
170 	 * is going to run. The caller is responsible for calling
171 	 * update_sctlr_el1() later in the same preemption disabled block.
172 	 */
173 	unsigned long sctlr = task->thread.sctlr_user;
174 	unsigned long mte_ctrl = task->thread.mte_ctrl;
175 	unsigned long pref, resolved_mte_tcf;
176 
177 	pref = __this_cpu_read(mte_tcf_preferred);
178 	resolved_mte_tcf = (mte_ctrl & pref) ? pref : mte_ctrl;
179 	sctlr &= ~SCTLR_EL1_TCF0_MASK;
180 	if (resolved_mte_tcf & MTE_CTRL_TCF_ASYNC)
181 		sctlr |= SCTLR_EL1_TCF0_ASYNC;
182 	else if (resolved_mte_tcf & MTE_CTRL_TCF_SYNC)
183 		sctlr |= SCTLR_EL1_TCF0_SYNC;
184 	task->thread.sctlr_user = sctlr;
185 }
186 
187 void mte_thread_init_user(void)
188 {
189 	if (!system_supports_mte())
190 		return;
191 
192 	/* clear any pending asynchronous tag fault */
193 	dsb(ish);
194 	write_sysreg_s(0, SYS_TFSRE0_EL1);
195 	clear_thread_flag(TIF_MTE_ASYNC_FAULT);
196 	/* disable tag checking and reset tag generation mask */
197 	set_mte_ctrl(current, 0);
198 }
199 
200 void mte_thread_switch(struct task_struct *next)
201 {
202 	mte_update_sctlr_user(next);
203 
204 	/*
205 	 * Check if an async tag exception occurred at EL1.
206 	 *
207 	 * Note: On the context switch path we rely on the dsb() present
208 	 * in __switch_to() to guarantee that the indirect writes to TFSR_EL1
209 	 * are synchronized before this point.
210 	 */
211 	isb();
212 	mte_check_tfsr_el1();
213 }
214 
215 void mte_suspend_enter(void)
216 {
217 	if (!system_supports_mte())
218 		return;
219 
220 	/*
221 	 * The barriers are required to guarantee that the indirect writes
222 	 * to TFSR_EL1 are synchronized before we report the state.
223 	 */
224 	dsb(nsh);
225 	isb();
226 
227 	/* Report SYS_TFSR_EL1 before suspend entry */
228 	mte_check_tfsr_el1();
229 }
230 
231 long set_mte_ctrl(struct task_struct *task, unsigned long arg)
232 {
233 	u64 mte_ctrl = (~((arg & PR_MTE_TAG_MASK) >> PR_MTE_TAG_SHIFT) &
234 			SYS_GCR_EL1_EXCL_MASK) << MTE_CTRL_GCR_USER_EXCL_SHIFT;
235 
236 	if (!system_supports_mte())
237 		return 0;
238 
239 	if (arg & PR_MTE_TCF_ASYNC)
240 		mte_ctrl |= MTE_CTRL_TCF_ASYNC;
241 	if (arg & PR_MTE_TCF_SYNC)
242 		mte_ctrl |= MTE_CTRL_TCF_SYNC;
243 
244 	task->thread.mte_ctrl = mte_ctrl;
245 	if (task == current) {
246 		preempt_disable();
247 		mte_update_sctlr_user(task);
248 		update_sctlr_el1(task->thread.sctlr_user);
249 		preempt_enable();
250 	}
251 
252 	return 0;
253 }
254 
255 long get_mte_ctrl(struct task_struct *task)
256 {
257 	unsigned long ret;
258 	u64 mte_ctrl = task->thread.mte_ctrl;
259 	u64 incl = (~mte_ctrl >> MTE_CTRL_GCR_USER_EXCL_SHIFT) &
260 		   SYS_GCR_EL1_EXCL_MASK;
261 
262 	if (!system_supports_mte())
263 		return 0;
264 
265 	ret = incl << PR_MTE_TAG_SHIFT;
266 	if (mte_ctrl & MTE_CTRL_TCF_ASYNC)
267 		ret |= PR_MTE_TCF_ASYNC;
268 	if (mte_ctrl & MTE_CTRL_TCF_SYNC)
269 		ret |= PR_MTE_TCF_SYNC;
270 
271 	return ret;
272 }
273 
274 /*
275  * Access MTE tags in another process' address space as given in mm. Update
276  * the number of tags copied. Return 0 if any tags copied, error otherwise.
277  * Inspired by __access_remote_vm().
278  */
279 static int __access_remote_tags(struct mm_struct *mm, unsigned long addr,
280 				struct iovec *kiov, unsigned int gup_flags)
281 {
282 	struct vm_area_struct *vma;
283 	void __user *buf = kiov->iov_base;
284 	size_t len = kiov->iov_len;
285 	int ret;
286 	int write = gup_flags & FOLL_WRITE;
287 
288 	if (!access_ok(buf, len))
289 		return -EFAULT;
290 
291 	if (mmap_read_lock_killable(mm))
292 		return -EIO;
293 
294 	while (len) {
295 		unsigned long tags, offset;
296 		void *maddr;
297 		struct page *page = NULL;
298 
299 		ret = get_user_pages_remote(mm, addr, 1, gup_flags, &page,
300 					    &vma, NULL);
301 		if (ret <= 0)
302 			break;
303 
304 		/*
305 		 * Only copy tags if the page has been mapped as PROT_MTE
306 		 * (PG_mte_tagged set). Otherwise the tags are not valid and
307 		 * not accessible to user. Moreover, an mprotect(PROT_MTE)
308 		 * would cause the existing tags to be cleared if the page
309 		 * was never mapped with PROT_MTE.
310 		 */
311 		if (!(vma->vm_flags & VM_MTE)) {
312 			ret = -EOPNOTSUPP;
313 			put_page(page);
314 			break;
315 		}
316 		WARN_ON_ONCE(!test_bit(PG_mte_tagged, &page->flags));
317 
318 		/* limit access to the end of the page */
319 		offset = offset_in_page(addr);
320 		tags = min(len, (PAGE_SIZE - offset) / MTE_GRANULE_SIZE);
321 
322 		maddr = page_address(page);
323 		if (write) {
324 			tags = mte_copy_tags_from_user(maddr + offset, buf, tags);
325 			set_page_dirty_lock(page);
326 		} else {
327 			tags = mte_copy_tags_to_user(buf, maddr + offset, tags);
328 		}
329 		put_page(page);
330 
331 		/* error accessing the tracer's buffer */
332 		if (!tags)
333 			break;
334 
335 		len -= tags;
336 		buf += tags;
337 		addr += tags * MTE_GRANULE_SIZE;
338 	}
339 	mmap_read_unlock(mm);
340 
341 	/* return an error if no tags copied */
342 	kiov->iov_len = buf - kiov->iov_base;
343 	if (!kiov->iov_len) {
344 		/* check for error accessing the tracee's address space */
345 		if (ret <= 0)
346 			return -EIO;
347 		else
348 			return -EFAULT;
349 	}
350 
351 	return 0;
352 }
353 
354 /*
355  * Copy MTE tags in another process' address space at 'addr' to/from tracer's
356  * iovec buffer. Return 0 on success. Inspired by ptrace_access_vm().
357  */
358 static int access_remote_tags(struct task_struct *tsk, unsigned long addr,
359 			      struct iovec *kiov, unsigned int gup_flags)
360 {
361 	struct mm_struct *mm;
362 	int ret;
363 
364 	mm = get_task_mm(tsk);
365 	if (!mm)
366 		return -EPERM;
367 
368 	if (!tsk->ptrace || (current != tsk->parent) ||
369 	    ((get_dumpable(mm) != SUID_DUMP_USER) &&
370 	     !ptracer_capable(tsk, mm->user_ns))) {
371 		mmput(mm);
372 		return -EPERM;
373 	}
374 
375 	ret = __access_remote_tags(mm, addr, kiov, gup_flags);
376 	mmput(mm);
377 
378 	return ret;
379 }
380 
381 int mte_ptrace_copy_tags(struct task_struct *child, long request,
382 			 unsigned long addr, unsigned long data)
383 {
384 	int ret;
385 	struct iovec kiov;
386 	struct iovec __user *uiov = (void __user *)data;
387 	unsigned int gup_flags = FOLL_FORCE;
388 
389 	if (!system_supports_mte())
390 		return -EIO;
391 
392 	if (get_user(kiov.iov_base, &uiov->iov_base) ||
393 	    get_user(kiov.iov_len, &uiov->iov_len))
394 		return -EFAULT;
395 
396 	if (request == PTRACE_POKEMTETAGS)
397 		gup_flags |= FOLL_WRITE;
398 
399 	/* align addr to the MTE tag granule */
400 	addr &= MTE_GRANULE_MASK;
401 
402 	ret = access_remote_tags(child, addr, &kiov, gup_flags);
403 	if (!ret)
404 		ret = put_user(kiov.iov_len, &uiov->iov_len);
405 
406 	return ret;
407 }
408 
409 static ssize_t mte_tcf_preferred_show(struct device *dev,
410 				      struct device_attribute *attr, char *buf)
411 {
412 	switch (per_cpu(mte_tcf_preferred, dev->id)) {
413 	case MTE_CTRL_TCF_ASYNC:
414 		return sysfs_emit(buf, "async\n");
415 	case MTE_CTRL_TCF_SYNC:
416 		return sysfs_emit(buf, "sync\n");
417 	default:
418 		return sysfs_emit(buf, "???\n");
419 	}
420 }
421 
422 static ssize_t mte_tcf_preferred_store(struct device *dev,
423 				       struct device_attribute *attr,
424 				       const char *buf, size_t count)
425 {
426 	u64 tcf;
427 
428 	if (sysfs_streq(buf, "async"))
429 		tcf = MTE_CTRL_TCF_ASYNC;
430 	else if (sysfs_streq(buf, "sync"))
431 		tcf = MTE_CTRL_TCF_SYNC;
432 	else
433 		return -EINVAL;
434 
435 	device_lock(dev);
436 	per_cpu(mte_tcf_preferred, dev->id) = tcf;
437 	device_unlock(dev);
438 
439 	return count;
440 }
441 static DEVICE_ATTR_RW(mte_tcf_preferred);
442 
443 static int register_mte_tcf_preferred_sysctl(void)
444 {
445 	unsigned int cpu;
446 
447 	if (!system_supports_mte())
448 		return 0;
449 
450 	for_each_possible_cpu(cpu) {
451 		per_cpu(mte_tcf_preferred, cpu) = MTE_CTRL_TCF_ASYNC;
452 		device_create_file(get_cpu_device(cpu),
453 				   &dev_attr_mte_tcf_preferred);
454 	}
455 
456 	return 0;
457 }
458 subsys_initcall(register_mte_tcf_preferred_sysctl);
459