1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2020 ARM Ltd.
6 #include <linux/bitops.h>
8 #include <linux/kernel.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>
20 #include <asm/barrier.h>
21 #include <asm/cpufeature.h>
23 #include <asm/ptrace.h>
24 #include <asm/sysreg.h>
26 static DEFINE_PER_CPU_READ_MOSTLY(u64, mte_tcf_preferred);
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);
34 static void mte_sync_page_tags(struct page *page, pte_t old_pte,
35 bool check_swap, bool pte_is_tagged)
37 if (check_swap && is_swap_pte(old_pte)) {
38 swp_entry_t entry = pte_to_swp_entry(old_pte);
40 if (!non_swap_entry(entry) && mte_restore_tags(entry, page))
47 page_kasan_tag_reset(page);
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.
56 mte_clear_page_tags(page_address(page));
59 void mte_sync_tags(pte_t old_pte, pte_t pte)
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);
66 /* Early out if there's nothing to do */
67 if (!check_swap && !pte_is_tagged)
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,
77 /* ensure the tags are visible before the PTE is set */
81 int memcmp_pages(struct page *page1, struct page *page2)
86 addr1 = page_address(page1);
87 addr2 = page_address(page2);
88 ret = memcmp(addr1, addr2, PAGE_SIZE);
90 if (!system_supports_mte() || ret)
94 * If the page content is identical but at least one of the pages is
95 * tagged, return non-zero to avoid KSM merging. If only one of the
96 * pages is tagged, set_pte_at() may zero or change the tags of the
97 * other page via mte_sync_tags().
99 if (test_bit(PG_mte_tagged, &page1->flags) ||
100 test_bit(PG_mte_tagged, &page2->flags))
101 return addr1 != addr2;
106 static inline void __mte_enable_kernel(const char *mode, unsigned long tcf)
108 /* Enable MTE Sync Mode for EL1. */
109 sysreg_clear_set(sctlr_el1, SCTLR_ELx_TCF_MASK, tcf);
112 pr_info_once("MTE: enabled in %s mode at EL1\n", mode);
115 #ifdef CONFIG_KASAN_HW_TAGS
116 void mte_enable_kernel_sync(void)
119 * Make sure we enter this function when no PE has set
120 * async mode previously.
122 WARN_ONCE(system_uses_mte_async_mode(),
123 "MTE async mode enabled system wide!");
125 __mte_enable_kernel("synchronous", SCTLR_ELx_TCF_SYNC);
128 void mte_enable_kernel_async(void)
130 __mte_enable_kernel("asynchronous", SCTLR_ELx_TCF_ASYNC);
133 * MTE async mode is set system wide by the first PE that
134 * executes this function.
136 * Note: If in future KASAN acquires a runtime switching
137 * mode in between sync and async, this strategy needs
140 if (!system_uses_mte_async_mode())
141 static_branch_enable(&mte_async_mode);
145 #ifdef CONFIG_KASAN_HW_TAGS
146 void mte_check_tfsr_el1(void)
148 u64 tfsr_el1 = read_sysreg_s(SYS_TFSR_EL1);
150 if (unlikely(tfsr_el1 & SYS_TFSR_EL1_TF1)) {
152 * Note: isb() is not required after this direct write
153 * because there is no indirect read subsequent to it
154 * (per ARM DDI 0487F.c table D13-1).
156 write_sysreg_s(0, SYS_TFSR_EL1);
158 kasan_report_async();
163 static void mte_update_sctlr_user(struct task_struct *task)
166 * This must be called with preemption disabled and can only be called
167 * on the current or next task since the CPU must match where the thread
168 * is going to run. The caller is responsible for calling
169 * update_sctlr_el1() later in the same preemption disabled block.
171 unsigned long sctlr = task->thread.sctlr_user;
172 unsigned long mte_ctrl = task->thread.mte_ctrl;
173 unsigned long pref, resolved_mte_tcf;
175 pref = __this_cpu_read(mte_tcf_preferred);
176 resolved_mte_tcf = (mte_ctrl & pref) ? pref : mte_ctrl;
177 sctlr &= ~SCTLR_EL1_TCF0_MASK;
178 if (resolved_mte_tcf & MTE_CTRL_TCF_ASYNC)
179 sctlr |= SCTLR_EL1_TCF0_ASYNC;
180 else if (resolved_mte_tcf & MTE_CTRL_TCF_SYNC)
181 sctlr |= SCTLR_EL1_TCF0_SYNC;
182 task->thread.sctlr_user = sctlr;
185 void mte_thread_init_user(void)
187 if (!system_supports_mte())
190 /* clear any pending asynchronous tag fault */
192 write_sysreg_s(0, SYS_TFSRE0_EL1);
193 clear_thread_flag(TIF_MTE_ASYNC_FAULT);
194 /* disable tag checking and reset tag generation mask */
195 set_mte_ctrl(current, 0);
198 void mte_thread_switch(struct task_struct *next)
200 if (!system_supports_mte())
203 mte_update_sctlr_user(next);
206 * Check if an async tag exception occurred at EL1.
208 * Note: On the context switch path we rely on the dsb() present
209 * in __switch_to() to guarantee that the indirect writes to TFSR_EL1
210 * are synchronized before this point.
213 mte_check_tfsr_el1();
216 void mte_suspend_enter(void)
218 if (!system_supports_mte())
222 * The barriers are required to guarantee that the indirect writes
223 * to TFSR_EL1 are synchronized before we report the state.
228 /* Report SYS_TFSR_EL1 before suspend entry */
229 mte_check_tfsr_el1();
232 long set_mte_ctrl(struct task_struct *task, unsigned long arg)
234 u64 mte_ctrl = (~((arg & PR_MTE_TAG_MASK) >> PR_MTE_TAG_SHIFT) &
235 SYS_GCR_EL1_EXCL_MASK) << MTE_CTRL_GCR_USER_EXCL_SHIFT;
237 if (!system_supports_mte())
240 if (arg & PR_MTE_TCF_ASYNC)
241 mte_ctrl |= MTE_CTRL_TCF_ASYNC;
242 if (arg & PR_MTE_TCF_SYNC)
243 mte_ctrl |= MTE_CTRL_TCF_SYNC;
245 task->thread.mte_ctrl = mte_ctrl;
246 if (task == current) {
248 mte_update_sctlr_user(task);
249 update_sctlr_el1(task->thread.sctlr_user);
256 long get_mte_ctrl(struct task_struct *task)
259 u64 mte_ctrl = task->thread.mte_ctrl;
260 u64 incl = (~mte_ctrl >> MTE_CTRL_GCR_USER_EXCL_SHIFT) &
261 SYS_GCR_EL1_EXCL_MASK;
263 if (!system_supports_mte())
266 ret = incl << PR_MTE_TAG_SHIFT;
267 if (mte_ctrl & MTE_CTRL_TCF_ASYNC)
268 ret |= PR_MTE_TCF_ASYNC;
269 if (mte_ctrl & MTE_CTRL_TCF_SYNC)
270 ret |= PR_MTE_TCF_SYNC;
276 * Access MTE tags in another process' address space as given in mm. Update
277 * the number of tags copied. Return 0 if any tags copied, error otherwise.
278 * Inspired by __access_remote_vm().
280 static int __access_remote_tags(struct mm_struct *mm, unsigned long addr,
281 struct iovec *kiov, unsigned int gup_flags)
283 struct vm_area_struct *vma;
284 void __user *buf = kiov->iov_base;
285 size_t len = kiov->iov_len;
287 int write = gup_flags & FOLL_WRITE;
289 if (!access_ok(buf, len))
292 if (mmap_read_lock_killable(mm))
296 unsigned long tags, offset;
298 struct page *page = NULL;
300 ret = get_user_pages_remote(mm, addr, 1, gup_flags, &page,
306 * Only copy tags if the page has been mapped as PROT_MTE
307 * (PG_mte_tagged set). Otherwise the tags are not valid and
308 * not accessible to user. Moreover, an mprotect(PROT_MTE)
309 * would cause the existing tags to be cleared if the page
310 * was never mapped with PROT_MTE.
312 if (!(vma->vm_flags & VM_MTE)) {
317 WARN_ON_ONCE(!test_bit(PG_mte_tagged, &page->flags));
319 /* limit access to the end of the page */
320 offset = offset_in_page(addr);
321 tags = min(len, (PAGE_SIZE - offset) / MTE_GRANULE_SIZE);
323 maddr = page_address(page);
325 tags = mte_copy_tags_from_user(maddr + offset, buf, tags);
326 set_page_dirty_lock(page);
328 tags = mte_copy_tags_to_user(buf, maddr + offset, tags);
332 /* error accessing the tracer's buffer */
338 addr += tags * MTE_GRANULE_SIZE;
340 mmap_read_unlock(mm);
342 /* return an error if no tags copied */
343 kiov->iov_len = buf - kiov->iov_base;
344 if (!kiov->iov_len) {
345 /* check for error accessing the tracee's address space */
356 * Copy MTE tags in another process' address space at 'addr' to/from tracer's
357 * iovec buffer. Return 0 on success. Inspired by ptrace_access_vm().
359 static int access_remote_tags(struct task_struct *tsk, unsigned long addr,
360 struct iovec *kiov, unsigned int gup_flags)
362 struct mm_struct *mm;
365 mm = get_task_mm(tsk);
369 if (!tsk->ptrace || (current != tsk->parent) ||
370 ((get_dumpable(mm) != SUID_DUMP_USER) &&
371 !ptracer_capable(tsk, mm->user_ns))) {
376 ret = __access_remote_tags(mm, addr, kiov, gup_flags);
382 int mte_ptrace_copy_tags(struct task_struct *child, long request,
383 unsigned long addr, unsigned long data)
387 struct iovec __user *uiov = (void __user *)data;
388 unsigned int gup_flags = FOLL_FORCE;
390 if (!system_supports_mte())
393 if (get_user(kiov.iov_base, &uiov->iov_base) ||
394 get_user(kiov.iov_len, &uiov->iov_len))
397 if (request == PTRACE_POKEMTETAGS)
398 gup_flags |= FOLL_WRITE;
400 /* align addr to the MTE tag granule */
401 addr &= MTE_GRANULE_MASK;
403 ret = access_remote_tags(child, addr, &kiov, gup_flags);
405 ret = put_user(kiov.iov_len, &uiov->iov_len);
410 static ssize_t mte_tcf_preferred_show(struct device *dev,
411 struct device_attribute *attr, char *buf)
413 switch (per_cpu(mte_tcf_preferred, dev->id)) {
414 case MTE_CTRL_TCF_ASYNC:
415 return sysfs_emit(buf, "async\n");
416 case MTE_CTRL_TCF_SYNC:
417 return sysfs_emit(buf, "sync\n");
419 return sysfs_emit(buf, "???\n");
423 static ssize_t mte_tcf_preferred_store(struct device *dev,
424 struct device_attribute *attr,
425 const char *buf, size_t count)
429 if (sysfs_streq(buf, "async"))
430 tcf = MTE_CTRL_TCF_ASYNC;
431 else if (sysfs_streq(buf, "sync"))
432 tcf = MTE_CTRL_TCF_SYNC;
437 per_cpu(mte_tcf_preferred, dev->id) = tcf;
442 static DEVICE_ATTR_RW(mte_tcf_preferred);
444 static int register_mte_tcf_preferred_sysctl(void)
448 if (!system_supports_mte())
451 for_each_possible_cpu(cpu) {
452 per_cpu(mte_tcf_preferred, cpu) = MTE_CTRL_TCF_ASYNC;
453 device_create_file(get_cpu_device(cpu),
454 &dev_attr_mte_tcf_preferred);
459 subsys_initcall(register_mte_tcf_preferred_sysctl);