1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2012 Regents of the University of California
4 * Copyright (C) 2017 SiFive
5 * Copyright (C) 2021 Western Digital Corporation or its affiliates.
8 #include <linux/bitops.h>
9 #include <linux/cpumask.h>
11 #include <linux/percpu.h>
12 #include <linux/slab.h>
13 #include <linux/spinlock.h>
14 #include <linux/static_key.h>
15 #include <asm/tlbflush.h>
16 #include <asm/cacheflush.h>
17 #include <asm/mmu_context.h>
21 DEFINE_STATIC_KEY_FALSE(use_asid_allocator);
23 static unsigned long asid_bits;
24 static unsigned long num_asids;
25 unsigned long asid_mask;
27 static atomic_long_t current_version;
29 static DEFINE_RAW_SPINLOCK(context_lock);
30 static cpumask_t context_tlb_flush_pending;
31 static unsigned long *context_asid_map;
33 static DEFINE_PER_CPU(atomic_long_t, active_context);
34 static DEFINE_PER_CPU(unsigned long, reserved_context);
36 static bool check_update_reserved_context(unsigned long cntx,
37 unsigned long newcntx)
43 * Iterate over the set of reserved CONTEXT looking for a match.
44 * If we find one, then we can update our mm to use new CONTEXT
45 * (i.e. the same CONTEXT in the current_version) but we can't
46 * exit the loop early, since we need to ensure that all copies
47 * of the old CONTEXT are updated to reflect the mm. Failure to do
48 * so could result in us missing the reserved CONTEXT in a future
51 for_each_possible_cpu(cpu) {
52 if (per_cpu(reserved_context, cpu) == cntx) {
54 per_cpu(reserved_context, cpu) = newcntx;
61 static void __flush_context(void)
66 /* Must be called with context_lock held */
67 lockdep_assert_held(&context_lock);
69 /* Update the list of reserved ASIDs and the ASID bitmap. */
70 bitmap_zero(context_asid_map, num_asids);
72 /* Mark already active ASIDs as used */
73 for_each_possible_cpu(i) {
74 cntx = atomic_long_xchg_relaxed(&per_cpu(active_context, i), 0);
76 * If this CPU has already been through a rollover, but
77 * hasn't run another task in the meantime, we must preserve
78 * its reserved CONTEXT, as this is the only trace we have of
79 * the process it is still running.
82 cntx = per_cpu(reserved_context, i);
84 __set_bit(cntx & asid_mask, context_asid_map);
85 per_cpu(reserved_context, i) = cntx;
88 /* Mark ASID #0 as used because it is used at boot-time */
89 __set_bit(0, context_asid_map);
91 /* Queue a TLB invalidation for each CPU on next context-switch */
92 cpumask_setall(&context_tlb_flush_pending);
95 static unsigned long __new_context(struct mm_struct *mm)
97 static u32 cur_idx = 1;
98 unsigned long cntx = atomic_long_read(&mm->context.id);
99 unsigned long asid, ver = atomic_long_read(¤t_version);
101 /* Must be called with context_lock held */
102 lockdep_assert_held(&context_lock);
105 unsigned long newcntx = ver | (cntx & asid_mask);
108 * If our current CONTEXT was active during a rollover, we
109 * can continue to use it and this was just a false alarm.
111 if (check_update_reserved_context(cntx, newcntx))
115 * We had a valid CONTEXT in a previous life, so try to
116 * re-use it if possible.
118 if (!__test_and_set_bit(cntx & asid_mask, context_asid_map))
123 * Allocate a free ASID. If we can't find one then increment
124 * current_version and flush all ASIDs.
126 asid = find_next_zero_bit(context_asid_map, num_asids, cur_idx);
127 if (asid != num_asids)
130 /* We're out of ASIDs, so increment current_version */
131 ver = atomic_long_add_return_relaxed(num_asids, ¤t_version);
133 /* Flush everything */
136 /* We have more ASIDs than CPUs, so this will always succeed */
137 asid = find_next_zero_bit(context_asid_map, num_asids, 1);
140 __set_bit(asid, context_asid_map);
145 static void set_mm_asid(struct mm_struct *mm, unsigned int cpu)
148 bool need_flush_tlb = false;
149 unsigned long cntx, old_active_cntx;
151 cntx = atomic_long_read(&mm->context.id);
154 * If our active_context is non-zero and the context matches the
155 * current_version, then we update the active_context entry with a
158 * Following is how we handle racing with a concurrent rollover:
160 * - We get a zero back from the cmpxchg and end up waiting on the
161 * lock. Taking the lock synchronises with the rollover and so
162 * we are forced to see the updated verion.
164 * - We get a valid context back from the cmpxchg then we continue
165 * using old ASID because __flush_context() would have marked ASID
166 * of active_context as used and next context switch we will
167 * allocate new context.
169 old_active_cntx = atomic_long_read(&per_cpu(active_context, cpu));
170 if (old_active_cntx &&
171 ((cntx & ~asid_mask) == atomic_long_read(¤t_version)) &&
172 atomic_long_cmpxchg_relaxed(&per_cpu(active_context, cpu),
173 old_active_cntx, cntx))
176 raw_spin_lock_irqsave(&context_lock, flags);
178 /* Check that our ASID belongs to the current_version. */
179 cntx = atomic_long_read(&mm->context.id);
180 if ((cntx & ~asid_mask) != atomic_long_read(¤t_version)) {
181 cntx = __new_context(mm);
182 atomic_long_set(&mm->context.id, cntx);
185 if (cpumask_test_and_clear_cpu(cpu, &context_tlb_flush_pending))
186 need_flush_tlb = true;
188 atomic_long_set(&per_cpu(active_context, cpu), cntx);
190 raw_spin_unlock_irqrestore(&context_lock, flags);
193 csr_write(CSR_SATP, virt_to_pfn(mm->pgd) |
194 ((cntx & asid_mask) << SATP_ASID_SHIFT) |
198 local_flush_tlb_all();
201 static void set_mm_noasid(struct mm_struct *mm)
203 /* Switch the page table and blindly nuke entire local TLB */
204 csr_write(CSR_SATP, virt_to_pfn(mm->pgd) | satp_mode);
205 local_flush_tlb_all();
208 static inline void set_mm(struct mm_struct *prev,
209 struct mm_struct *next, unsigned int cpu)
212 * The mm_cpumask indicates which harts' TLBs contain the virtual
213 * address mapping of the mm. Compared to noasid, using asid
214 * can't guarantee that stale TLB entries are invalidated because
215 * the asid mechanism wouldn't flush TLB for every switch_mm for
216 * performance. So when using asid, keep all CPUs footmarks in
217 * cpumask() until mm reset.
219 cpumask_set_cpu(cpu, mm_cpumask(next));
220 if (static_branch_unlikely(&use_asid_allocator)) {
221 set_mm_asid(next, cpu);
223 cpumask_clear_cpu(cpu, mm_cpumask(prev));
228 static int __init asids_init(void)
232 /* Figure-out number of ASID bits in HW */
233 old = csr_read(CSR_SATP);
234 asid_bits = old | (SATP_ASID_MASK << SATP_ASID_SHIFT);
235 csr_write(CSR_SATP, asid_bits);
236 asid_bits = (csr_read(CSR_SATP) >> SATP_ASID_SHIFT) & SATP_ASID_MASK;
237 asid_bits = fls_long(asid_bits);
238 csr_write(CSR_SATP, old);
241 * In the process of determining number of ASID bits (above)
242 * we polluted the TLB of current HART so let's do TLB flushed
243 * to remove unwanted TLB enteries.
245 local_flush_tlb_all();
247 /* Pre-compute ASID details */
249 num_asids = 1 << asid_bits;
250 asid_mask = num_asids - 1;
254 * Use ASID allocator only if number of HW ASIDs are
255 * at-least twice more than CPUs
257 if (num_asids > (2 * num_possible_cpus())) {
258 atomic_long_set(¤t_version, num_asids);
260 context_asid_map = bitmap_zalloc(num_asids, GFP_KERNEL);
261 if (!context_asid_map)
262 panic("Failed to allocate bitmap for %lu ASIDs\n",
265 __set_bit(0, context_asid_map);
267 static_branch_enable(&use_asid_allocator);
269 pr_info("ASID allocator using %lu bits (%lu entries)\n",
270 asid_bits, num_asids);
272 pr_info("ASID allocator disabled (%lu bits)\n", asid_bits);
277 early_initcall(asids_init);
279 static inline void set_mm(struct mm_struct *prev,
280 struct mm_struct *next, unsigned int cpu)
282 /* Nothing to do here when there is no MMU */
287 * When necessary, performs a deferred icache flush for the given MM context,
288 * on the local CPU. RISC-V has no direct mechanism for instruction cache
289 * shoot downs, so instead we send an IPI that informs the remote harts they
290 * need to flush their local instruction caches. To avoid pathologically slow
291 * behavior in a common case (a bunch of single-hart processes on a many-hart
292 * machine, ie 'make -j') we avoid the IPIs for harts that are not currently
293 * executing a MM context and instead schedule a deferred local instruction
294 * cache flush to be performed before execution resumes on each hart. This
295 * actually performs that local instruction cache flush, which implicitly only
296 * refers to the current hart.
298 * The "cpu" argument must be the current local CPU number.
300 static inline void flush_icache_deferred(struct mm_struct *mm, unsigned int cpu)
303 cpumask_t *mask = &mm->context.icache_stale_mask;
305 if (cpumask_test_cpu(cpu, mask)) {
306 cpumask_clear_cpu(cpu, mask);
308 * Ensure the remote hart's writes are visible to this hart.
309 * This pairs with a barrier in flush_icache_mm.
312 local_flush_icache_all();
318 void switch_mm(struct mm_struct *prev, struct mm_struct *next,
319 struct task_struct *task)
323 if (unlikely(prev == next))
327 * Mark the current MM context as inactive, and the next as
328 * active. This is at least used by the icache flushing
329 * routines in order to determine who should be flushed.
331 cpu = smp_processor_id();
333 set_mm(prev, next, cpu);
335 flush_icache_deferred(next, cpu);