From 2ca8cf738907180e7fbda90f25f32b86feda609f Mon Sep 17 00:00:00 2001 From: Benjamin Herrenschmidt Date: Thu, 18 Dec 2008 19:13:29 +0000 Subject: powerpc/mm: Rework context management for CPUs with no hash table This reworks the context management code used by 4xx,8xx and freescale BookE. It adds support for SMP by implementing a concept of stale context map to lazily flush the TLB on processors where a context may have been invalidated. This also contains the ground work for generalizing such lazy TLB flushing by just picking up a new PID and marking the old one stale. This will be implemented later. This is a first implementation that uses a global spinlock. Ideally, we should try to get at least the fast path (context ID already assigned) lockless or limited to a per context lock, but for now this will do. I tried to keep the UP case reasonably simple to avoid adding too much overhead to 8xx which does a lot of context stealing since it effectively has only 16 PIDs available. Signed-off-by: Benjamin Herrenschmidt Acked-by: Kumar Gala Signed-off-by: Paul Mackerras --- arch/powerpc/include/asm/mmu-40x.h | 5 +- arch/powerpc/include/asm/mmu-44x.h | 5 +- arch/powerpc/include/asm/mmu-8xx.h | 3 +- arch/powerpc/include/asm/mmu-fsl-booke.h | 5 +- arch/powerpc/include/asm/tlbflush.h | 2 + arch/powerpc/mm/mmu_context_nohash.c | 268 +++++++++++++++++++++++++------ 6 files changed, 234 insertions(+), 54 deletions(-) (limited to 'arch') diff --git a/arch/powerpc/include/asm/mmu-40x.h b/arch/powerpc/include/asm/mmu-40x.h index 3d108676584c..776f415a36aa 100644 --- a/arch/powerpc/include/asm/mmu-40x.h +++ b/arch/powerpc/include/asm/mmu-40x.h @@ -54,8 +54,9 @@ #ifndef __ASSEMBLY__ typedef struct { - unsigned long id; - unsigned long vdso_base; + unsigned int id; + unsigned int active; + unsigned long vdso_base; } mm_context_t; #endif /* !__ASSEMBLY__ */ diff --git a/arch/powerpc/include/asm/mmu-44x.h b/arch/powerpc/include/asm/mmu-44x.h index a825524c981a..b21af32ac6d6 100644 --- a/arch/powerpc/include/asm/mmu-44x.h +++ b/arch/powerpc/include/asm/mmu-44x.h @@ -56,8 +56,9 @@ extern unsigned int tlb_44x_hwater; typedef struct { - unsigned long id; - unsigned long vdso_base; + unsigned int id; + unsigned int active; + unsigned long vdso_base; } mm_context_t; #endif /* !__ASSEMBLY__ */ diff --git a/arch/powerpc/include/asm/mmu-8xx.h b/arch/powerpc/include/asm/mmu-8xx.h index 9db877eb88db..07865a357848 100644 --- a/arch/powerpc/include/asm/mmu-8xx.h +++ b/arch/powerpc/include/asm/mmu-8xx.h @@ -137,7 +137,8 @@ #ifndef __ASSEMBLY__ typedef struct { - unsigned long id; + unsigned int id; + unsigned int active; unsigned long vdso_base; } mm_context_t; #endif /* !__ASSEMBLY__ */ diff --git a/arch/powerpc/include/asm/mmu-fsl-booke.h b/arch/powerpc/include/asm/mmu-fsl-booke.h index 5588a41f439c..3f941c0f7e8e 100644 --- a/arch/powerpc/include/asm/mmu-fsl-booke.h +++ b/arch/powerpc/include/asm/mmu-fsl-booke.h @@ -76,8 +76,9 @@ #ifndef __ASSEMBLY__ typedef struct { - unsigned long id; - unsigned long vdso_base; + unsigned int id; + unsigned int active; + unsigned long vdso_base; } mm_context_t; #endif /* !__ASSEMBLY__ */ diff --git a/arch/powerpc/include/asm/tlbflush.h b/arch/powerpc/include/asm/tlbflush.h index 333c24b54379..9ed363d3de44 100644 --- a/arch/powerpc/include/asm/tlbflush.h +++ b/arch/powerpc/include/asm/tlbflush.h @@ -29,6 +29,8 @@ #include +#define MMU_NO_CONTEXT ((unsigned int)-1) + extern void _tlbie(unsigned long address, unsigned int pid); extern void _tlbil_all(void); extern void _tlbil_pid(unsigned int pid); diff --git a/arch/powerpc/mm/mmu_context_nohash.c b/arch/powerpc/mm/mmu_context_nohash.c index 00e02150abef..8b5de52de0ad 100644 --- a/arch/powerpc/mm/mmu_context_nohash.c +++ b/arch/powerpc/mm/mmu_context_nohash.c @@ -14,13 +14,28 @@ * as published by the Free Software Foundation; either version * 2 of the License, or (at your option) any later version. * + * TODO: + * + * - The global context lock will not scale very well + * - The maps should be dynamically allocated to allow for processors + * that support more PID bits at runtime + * - Implement flush_tlb_mm() by making the context stale and picking + * a new one + * - More aggressively clear stale map bits and maybe find some way to + * also clear mm->cpu_vm_mask bits when processes are migrated */ +#undef DEBUG +#define DEBUG_STEAL_ONLY +#undef DEBUG_MAP_CONSISTENCY + +#include #include #include #include #include +#include /* * The MPC8xx has only 16 contexts. We rotate through them on each @@ -40,17 +55,14 @@ */ #ifdef CONFIG_8xx -#define NO_CONTEXT 16 #define LAST_CONTEXT 15 #define FIRST_CONTEXT 0 #elif defined(CONFIG_4xx) -#define NO_CONTEXT 256 #define LAST_CONTEXT 255 #define FIRST_CONTEXT 1 #elif defined(CONFIG_E200) || defined(CONFIG_E500) -#define NO_CONTEXT 256 #define LAST_CONTEXT 255 #define FIRST_CONTEXT 1 @@ -58,66 +70,208 @@ #error Unsupported processor type #endif -static unsigned long next_mmu_context; +static unsigned int next_context, nr_free_contexts; static unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1]; -static atomic_t nr_free_contexts; +static unsigned long stale_map[NR_CPUS][LAST_CONTEXT / BITS_PER_LONG + 1]; static struct mm_struct *context_mm[LAST_CONTEXT+1]; -static void steal_context(void); +static spinlock_t context_lock = SPIN_LOCK_UNLOCKED; /* Steal a context from a task that has one at the moment. - * This is only used on 8xx and 4xx and we presently assume that - * they don't do SMP. If they do then this will have to check - * whether the MM we steal is in use. - * We also assume that this is only used on systems that don't - * use an MMU hash table - this is true for 8xx and 4xx. + * + * This is used when we are running out of available PID numbers + * on the processors. + * * This isn't an LRU system, it just frees up each context in * turn (sort-of pseudo-random replacement :). This would be the * place to implement an LRU scheme if anyone was motivated to do it. * -- paulus + * + * For context stealing, we use a slightly different approach for + * SMP and UP. Basically, the UP one is simpler and doesn't use + * the stale map as we can just flush the local CPU + * -- benh */ -static void steal_context(void) +#ifdef CONFIG_SMP +static unsigned int steal_context_smp(unsigned int id) { struct mm_struct *mm; + unsigned int cpu, max; - /* free up context `next_mmu_context' */ - /* if we shouldn't free context 0, don't... */ - if (next_mmu_context < FIRST_CONTEXT) - next_mmu_context = FIRST_CONTEXT; - mm = context_mm[next_mmu_context]; - flush_tlb_mm(mm); - destroy_context(mm); -} + again: + max = LAST_CONTEXT - FIRST_CONTEXT; + /* Attempt to free next_context first and then loop until we manage */ + while (max--) { + /* Pick up the victim mm */ + mm = context_mm[id]; -/* - * Get a new mmu context for the address space described by `mm'. + /* We have a candidate victim, check if it's active, on SMP + * we cannot steal active contexts + */ + if (mm->context.active) { + id++; + if (id > LAST_CONTEXT) + id = FIRST_CONTEXT; + continue; + } + pr_debug("[%d] steal context %d from mm @%p\n", + smp_processor_id(), id, mm); + + /* Mark this mm has having no context anymore */ + mm->context.id = MMU_NO_CONTEXT; + + /* Mark it stale on all CPUs that used this mm */ + for_each_cpu_mask_nr(cpu, mm->cpu_vm_mask) + __set_bit(id, stale_map[cpu]); + return id; + } + + /* This will happen if you have more CPUs than available contexts, + * all we can do here is wait a bit and try again + */ + spin_unlock(&context_lock); + cpu_relax(); + spin_lock(&context_lock); + goto again; +} +#endif /* CONFIG_SMP */ + +/* Note that this will also be called on SMP if all other CPUs are + * offlined, which means that it may be called for cpu != 0. For + * this to work, we somewhat assume that CPUs that are onlined + * come up with a fully clean TLB (or are cleaned when offlined) */ -static inline void get_mmu_context(struct mm_struct *mm) +static unsigned int steal_context_up(unsigned int id) { - unsigned long ctx; + struct mm_struct *mm; + int cpu = smp_processor_id(); - if (mm->context.id != NO_CONTEXT) - return; + /* Pick up the victim mm */ + mm = context_mm[id]; + + pr_debug("[%d] steal context %d from mm @%p\n", cpu, id, mm); - while (atomic_dec_if_positive(&nr_free_contexts) < 0) - steal_context(); + /* Mark this mm has having no context anymore */ + mm->context.id = MMU_NO_CONTEXT; - ctx = next_mmu_context; - while (test_and_set_bit(ctx, context_map)) { - ctx = find_next_zero_bit(context_map, LAST_CONTEXT+1, ctx); - if (ctx > LAST_CONTEXT) - ctx = 0; + /* Flush the TLB for that context */ + local_flush_tlb_mm(mm); + + /* XXX This clear should ultimately be part of local_flush_tlb_mm */ + __clear_bit(id, stale_map[cpu]); + + return id; +} + +#ifdef DEBUG_MAP_CONSISTENCY +static void context_check_map(void) +{ + unsigned int id, nrf, nact; + + nrf = nact = 0; + for (id = FIRST_CONTEXT; id <= LAST_CONTEXT; id++) { + int used = test_bit(id, context_map); + if (!used) + nrf++; + if (used != (context_mm[id] != NULL)) + pr_err("MMU: Context %d is %s and MM is %p !\n", + id, used ? "used" : "free", context_mm[id]); + if (context_mm[id] != NULL) + nact += context_mm[id]->context.active; } - next_mmu_context = (ctx + 1) & LAST_CONTEXT; - mm->context.id = ctx; - context_mm[ctx] = mm; + if (nrf != nr_free_contexts) { + pr_err("MMU: Free context count out of sync ! (%d vs %d)\n", + nr_free_contexts, nrf); + nr_free_contexts = nrf; + } + if (nact > num_online_cpus()) + pr_err("MMU: More active contexts than CPUs ! (%d vs %d)\n", + nact, num_online_cpus()); } +#else +static void context_check_map(void) { } +#endif void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next) { - get_mmu_context(next); + unsigned int id, cpu = smp_processor_id(); + unsigned long *map; - set_context(next->context.id, next->pgd); + /* No lockless fast path .. yet */ + spin_lock(&context_lock); + +#ifndef DEBUG_STEAL_ONLY + pr_debug("[%d] activating context for mm @%p, active=%d, id=%d\n", + cpu, next, next->context.active, next->context.id); +#endif + +#ifdef CONFIG_SMP + /* Mark us active and the previous one not anymore */ + next->context.active++; + if (prev) { + WARN_ON(prev->context.active < 1); + prev->context.active--; + } +#endif /* CONFIG_SMP */ + + /* If we already have a valid assigned context, skip all that */ + id = next->context.id; + if (likely(id != MMU_NO_CONTEXT)) + goto ctxt_ok; + + /* We really don't have a context, let's try to acquire one */ + id = next_context; + if (id > LAST_CONTEXT) + id = FIRST_CONTEXT; + map = context_map; + + /* No more free contexts, let's try to steal one */ + if (nr_free_contexts == 0) { +#ifdef CONFIG_SMP + if (num_online_cpus() > 1) { + id = steal_context_smp(id); + goto stolen; + } +#endif /* CONFIG_SMP */ + id = steal_context_up(id); + goto stolen; + } + nr_free_contexts--; + + /* We know there's at least one free context, try to find it */ + while (__test_and_set_bit(id, map)) { + id = find_next_zero_bit(map, LAST_CONTEXT+1, id); + if (id > LAST_CONTEXT) + id = FIRST_CONTEXT; + } + stolen: + next_context = id + 1; + context_mm[id] = next; + next->context.id = id; + +#ifndef DEBUG_STEAL_ONLY + pr_debug("[%d] picked up new id %d, nrf is now %d\n", + cpu, id, nr_free_contexts); +#endif + + context_check_map(); + ctxt_ok: + + /* If that context got marked stale on this CPU, then flush the + * local TLB for it and unmark it before we use it + */ + if (test_bit(id, stale_map[cpu])) { + pr_debug("[%d] flushing stale context %d for mm @%p !\n", + cpu, id, next); + local_flush_tlb_mm(next); + + /* XXX This clear should ultimately be part of local_flush_tlb_mm */ + __clear_bit(id, stale_map[cpu]); + } + + /* Flick the MMU and release lock */ + set_context(id, next->pgd); + spin_unlock(&context_lock); } /* @@ -125,7 +279,9 @@ void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next) */ int init_new_context(struct task_struct *t, struct mm_struct *mm) { - mm->context.id = NO_CONTEXT; + mm->context.id = MMU_NO_CONTEXT; + mm->context.active = 0; + return 0; } @@ -134,13 +290,25 @@ int init_new_context(struct task_struct *t, struct mm_struct *mm) */ void destroy_context(struct mm_struct *mm) { - preempt_disable(); - if (mm->context.id != NO_CONTEXT) { - clear_bit(mm->context.id, context_map); - mm->context.id = NO_CONTEXT; - atomic_inc(&nr_free_contexts); + unsigned int id; + + if (mm->context.id == MMU_NO_CONTEXT) + return; + + WARN_ON(mm->context.active != 0); + + spin_lock(&context_lock); + id = mm->context.id; + if (id != MMU_NO_CONTEXT) { + __clear_bit(id, context_map); + mm->context.id = MMU_NO_CONTEXT; +#ifdef DEBUG_MAP_CONSISTENCY + mm->context.active = 0; + context_mm[id] = NULL; +#endif + nr_free_contexts++; } - preempt_enable(); + spin_unlock(&context_lock); } @@ -149,6 +317,12 @@ void destroy_context(struct mm_struct *mm) */ void __init mmu_context_init(void) { + /* Mark init_mm as being active on all possible CPUs since + * we'll get called with prev == init_mm the first time + * we schedule on a given CPU + */ + init_mm.context.active = NR_CPUS; + /* * Some processors have too few contexts to reserve one for * init_mm, and require using context 0 for a normal task. @@ -156,7 +330,7 @@ void __init mmu_context_init(void) * This code assumes FIRST_CONTEXT < 32. */ context_map[0] = (1 << FIRST_CONTEXT) - 1; - next_mmu_context = FIRST_CONTEXT; - atomic_set(&nr_free_contexts, LAST_CONTEXT - FIRST_CONTEXT + 1); + next_context = FIRST_CONTEXT; + nr_free_contexts = LAST_CONTEXT - FIRST_CONTEXT + 1; } -- cgit v1.2.3