The page allocator has been using a locking scheme for its percpu page caches (pcp) based on spin_trylock() with no _irqsave() part. The trick is that if we interrupt the locked section, we fail the trylock and just fallback to the slowpath taking the zone lock. That's more expensive, but rare, so we don't need to pay the irqsave/restore cost all the time in the fastpaths. It's similar to but not exactly local_trylock_t (which is also newer anyway) because in some cases we do lock the pcp of a non-local cpu to drain it, in a way that's cheaper than using IPI or queue_work_on(). The complication of this scheme has been UP non-debug spinlock implementation which assumes spin_trylock() can't fail on UP and has no state to track whether it's locked. It just doesn't anticipate this usage scenario. So to work around that we disable IRQs only on UP, complicating the implementation. Also recently we found years old bug in where we didn't disable IRQs in related paths - see 038a102535eb ("mm/page_alloc: prevent pcp corruption with SMP=n"). We can avoid this UP complication by realizing that we do not need the pcp caching for scalability on UP in the first place. Removing it completely with #ifdefs is not worth the trouble either. Just make pcp_spin_trylock() return NULL unconditionally on CONFIG_SMP=n. This makes the slowpaths unconditional, and we can remove the IRQ save/restore handling in pcp_spin_trylock()/unlock() completely. Suggested-by: David Hildenbrand (Arm) Signed-off-by: Vlastimil Babka (SUSE) --- mm/page_alloc.c | 92 +++++++++++++++++++++------------------------------------ 1 file changed, 34 insertions(+), 58 deletions(-) diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 2d2e9eea077f..65efcaeb8800 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -95,23 +95,6 @@ typedef int __bitwise fpi_t; static DEFINE_MUTEX(pcp_batch_high_lock); #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8) -#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT) -/* - * On SMP, spin_trylock is sufficient protection. - * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP. - * Pass flags to a no-op inline function to typecheck and silence the unused - * variable warning. - */ -static inline void __pcp_trylock_noop(unsigned long *flags) { } -#define pcp_trylock_prepare(flags) __pcp_trylock_noop(&(flags)) -#define pcp_trylock_finish(flags) __pcp_trylock_noop(&(flags)) -#else - -/* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */ -#define pcp_trylock_prepare(flags) local_irq_save(flags) -#define pcp_trylock_finish(flags) local_irq_restore(flags) -#endif - /* * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid * a migration causing the wrong PCP to be locked and remote memory being @@ -150,31 +133,28 @@ static inline void __pcp_trylock_noop(unsigned long *flags) { } pcpu_task_unpin(); \ }) -/* struct per_cpu_pages specific helpers. */ -#define pcp_spin_trylock(ptr, UP_flags) \ -({ \ - struct per_cpu_pages *__ret; \ - pcp_trylock_prepare(UP_flags); \ - __ret = pcpu_spin_trylock(struct per_cpu_pages, lock, ptr); \ - if (!__ret) \ - pcp_trylock_finish(UP_flags); \ - __ret; \ -}) +/* struct per_cpu_pages specific helpers.*/ +#ifdef CONFIG_SMP +#define pcp_spin_trylock(ptr) \ + pcpu_spin_trylock(struct per_cpu_pages, lock, ptr) -#define pcp_spin_unlock(ptr, UP_flags) \ -({ \ - pcpu_spin_unlock(lock, ptr); \ - pcp_trylock_finish(UP_flags); \ -}) +#define pcp_spin_unlock(ptr) \ + pcpu_spin_unlock(lock, ptr) /* - * With the UP spinlock implementation, when we spin_lock(&pcp->lock) (for i.e. - * a potentially remote cpu drain) and get interrupted by an operation that - * attempts pcp_spin_trylock(), we can't rely on the trylock failure due to UP - * spinlock assumptions making the trylock a no-op. So we have to turn that - * spin_lock() to a spin_lock_irqsave(). This works because on UP there are no - * remote cpu's so we can only be locking the only existing local one. + * On CONFIG_SMP=n the UP implementation of spin_trylock() never fails and thus + * is not compatible with our locking scheme. However we do not need pcp for + * scalability in the first place, so just make all the trylocks fail and take + * the slow path unconditionally. */ +#else +#define pcp_spin_trylock(ptr) \ + NULL + +#define pcp_spin_unlock(ptr) \ + BUG_ON(1) +#endif + #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT) static inline void __flags_noop(unsigned long *flags) { } #define pcp_spin_lock_maybe_irqsave(ptr, flags) \ @@ -2862,7 +2842,7 @@ static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone, */ static bool free_frozen_page_commit(struct zone *zone, struct per_cpu_pages *pcp, struct page *page, int migratetype, - unsigned int order, fpi_t fpi_flags, unsigned long *UP_flags) + unsigned int order, fpi_t fpi_flags) { int high, batch; int to_free, to_free_batched; @@ -2922,9 +2902,9 @@ static bool free_frozen_page_commit(struct zone *zone, if (to_free == 0 || pcp->count == 0) break; - pcp_spin_unlock(pcp, *UP_flags); + pcp_spin_unlock(pcp); - pcp = pcp_spin_trylock(zone->per_cpu_pageset, *UP_flags); + pcp = pcp_spin_trylock(zone->per_cpu_pageset); if (!pcp) { ret = false; break; @@ -2936,7 +2916,7 @@ static bool free_frozen_page_commit(struct zone *zone, * returned in an unlocked state. */ if (smp_processor_id() != cpu) { - pcp_spin_unlock(pcp, *UP_flags); + pcp_spin_unlock(pcp); ret = false; break; } @@ -2968,7 +2948,6 @@ static bool free_frozen_page_commit(struct zone *zone, static void __free_frozen_pages(struct page *page, unsigned int order, fpi_t fpi_flags) { - unsigned long UP_flags; struct per_cpu_pages *pcp; struct zone *zone; unsigned long pfn = page_to_pfn(page); @@ -3004,12 +2983,12 @@ static void __free_frozen_pages(struct page *page, unsigned int order, add_page_to_zone_llist(zone, page, order); return; } - pcp = pcp_spin_trylock(zone->per_cpu_pageset, UP_flags); + pcp = pcp_spin_trylock(zone->per_cpu_pageset); if (pcp) { if (!free_frozen_page_commit(zone, pcp, page, migratetype, - order, fpi_flags, &UP_flags)) + order, fpi_flags)) return; - pcp_spin_unlock(pcp, UP_flags); + pcp_spin_unlock(pcp); } else { free_one_page(zone, page, pfn, order, fpi_flags); } @@ -3030,7 +3009,6 @@ void free_frozen_pages_nolock(struct page *page, unsigned int order) */ void free_unref_folios(struct folio_batch *folios) { - unsigned long UP_flags; struct per_cpu_pages *pcp = NULL; struct zone *locked_zone = NULL; int i, j; @@ -3073,7 +3051,7 @@ void free_unref_folios(struct folio_batch *folios) if (zone != locked_zone || is_migrate_isolate(migratetype)) { if (pcp) { - pcp_spin_unlock(pcp, UP_flags); + pcp_spin_unlock(pcp); locked_zone = NULL; pcp = NULL; } @@ -3092,7 +3070,7 @@ void free_unref_folios(struct folio_batch *folios) * trylock is necessary as folios may be getting freed * from IRQ or SoftIRQ context after an IO completion. */ - pcp = pcp_spin_trylock(zone->per_cpu_pageset, UP_flags); + pcp = pcp_spin_trylock(zone->per_cpu_pageset); if (unlikely(!pcp)) { free_one_page(zone, &folio->page, pfn, order, FPI_NONE); @@ -3110,14 +3088,14 @@ void free_unref_folios(struct folio_batch *folios) trace_mm_page_free_batched(&folio->page); if (!free_frozen_page_commit(zone, pcp, &folio->page, - migratetype, order, FPI_NONE, &UP_flags)) { + migratetype, order, FPI_NONE)) { pcp = NULL; locked_zone = NULL; } } if (pcp) - pcp_spin_unlock(pcp, UP_flags); + pcp_spin_unlock(pcp); folio_batch_reinit(folios); } @@ -3375,10 +3353,9 @@ static struct page *rmqueue_pcplist(struct zone *preferred_zone, struct per_cpu_pages *pcp; struct list_head *list; struct page *page; - unsigned long UP_flags; /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */ - pcp = pcp_spin_trylock(zone->per_cpu_pageset, UP_flags); + pcp = pcp_spin_trylock(zone->per_cpu_pageset); if (!pcp) return NULL; @@ -3390,7 +3367,7 @@ static struct page *rmqueue_pcplist(struct zone *preferred_zone, pcp->free_count >>= 1; list = &pcp->lists[order_to_pindex(migratetype, order)]; page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list); - pcp_spin_unlock(pcp, UP_flags); + pcp_spin_unlock(pcp); if (page) { __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order); zone_statistics(preferred_zone, zone, 1); @@ -5071,7 +5048,6 @@ unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid, struct page **page_array) { struct page *page; - unsigned long UP_flags; struct zone *zone; struct zoneref *z; struct per_cpu_pages *pcp; @@ -5165,7 +5141,7 @@ unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid, goto failed; /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */ - pcp = pcp_spin_trylock(zone->per_cpu_pageset, UP_flags); + pcp = pcp_spin_trylock(zone->per_cpu_pageset); if (!pcp) goto failed; @@ -5184,7 +5160,7 @@ unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid, if (unlikely(!page)) { /* Try and allocate at least one page */ if (!nr_account) { - pcp_spin_unlock(pcp, UP_flags); + pcp_spin_unlock(pcp); goto failed; } break; @@ -5196,7 +5172,7 @@ unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid, page_array[nr_populated++] = page; } - pcp_spin_unlock(pcp, UP_flags); + pcp_spin_unlock(pcp); __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account); zone_statistics(zonelist_zone(ac.preferred_zoneref), zone, nr_account); -- 2.53.0 Effectively revert commit 038a102535eb ("mm/page_alloc: prevent pcp corruption with SMP=n"). The original problem is now avoided by pcp_spin_trylock() always failing on CONFIG_SMP=n, so we do not need to disable IRQs anymore. It's not a complete revert, because keeping the pcp_spin_(un)lock() wrappers is useful. Rename them from _maybe_irqsave/restore to _nopin. The difference from pcp_spin_trylock()/pcp_spin_unlock() is that the _nopin variants don't perform pcpu_task_pin/unpin(). Signed-off-by: Vlastimil Babka (SUSE) --- mm/page_alloc.c | 46 ++++++++++++++++------------------------------ 1 file changed, 16 insertions(+), 30 deletions(-) diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 65efcaeb8800..8e5b30adfe40 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -155,24 +155,14 @@ static DEFINE_MUTEX(pcp_batch_high_lock); BUG_ON(1) #endif -#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT) -static inline void __flags_noop(unsigned long *flags) { } -#define pcp_spin_lock_maybe_irqsave(ptr, flags) \ -({ \ - __flags_noop(&(flags)); \ - spin_lock(&(ptr)->lock); \ -}) -#define pcp_spin_unlock_maybe_irqrestore(ptr, flags) \ -({ \ - spin_unlock(&(ptr)->lock); \ - __flags_noop(&(flags)); \ -}) -#else -#define pcp_spin_lock_maybe_irqsave(ptr, flags) \ - spin_lock_irqsave(&(ptr)->lock, flags) -#define pcp_spin_unlock_maybe_irqrestore(ptr, flags) \ - spin_unlock_irqrestore(&(ptr)->lock, flags) -#endif +/* + * In some cases we do not need to pin the task to the CPU because we are + * already given a specific cpu's pcp pointer. + */ +#define pcp_spin_lock_nopin(ptr) \ + spin_lock(&(ptr)->lock) +#define pcp_spin_unlock_nopin(ptr) \ + spin_unlock(&(ptr)->lock) #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID DEFINE_PER_CPU(int, numa_node); @@ -2572,7 +2562,6 @@ static int rmqueue_bulk(struct zone *zone, unsigned int order, bool decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp) { int high_min, to_drain, to_drain_batched, batch; - unsigned long UP_flags; bool todo = false; high_min = READ_ONCE(pcp->high_min); @@ -2592,9 +2581,9 @@ bool decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp) to_drain = pcp->count - pcp->high; while (to_drain > 0) { to_drain_batched = min(to_drain, batch); - pcp_spin_lock_maybe_irqsave(pcp, UP_flags); + pcp_spin_lock_nopin(pcp); free_pcppages_bulk(zone, to_drain_batched, pcp, 0); - pcp_spin_unlock_maybe_irqrestore(pcp, UP_flags); + pcp_spin_unlock_nopin(pcp); todo = true; to_drain -= to_drain_batched; @@ -2611,15 +2600,14 @@ bool decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp) */ void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) { - unsigned long UP_flags; int to_drain, batch; batch = READ_ONCE(pcp->batch); to_drain = min(pcp->count, batch); if (to_drain > 0) { - pcp_spin_lock_maybe_irqsave(pcp, UP_flags); + pcp_spin_lock_nopin(pcp); free_pcppages_bulk(zone, to_drain, pcp, 0); - pcp_spin_unlock_maybe_irqrestore(pcp, UP_flags); + pcp_spin_unlock_nopin(pcp); } } #endif @@ -2630,11 +2618,10 @@ void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) static void drain_pages_zone(unsigned int cpu, struct zone *zone) { struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); - unsigned long UP_flags; int count; do { - pcp_spin_lock_maybe_irqsave(pcp, UP_flags); + pcp_spin_lock_nopin(pcp); count = pcp->count; if (count) { int to_drain = min(count, @@ -2643,7 +2630,7 @@ static void drain_pages_zone(unsigned int cpu, struct zone *zone) free_pcppages_bulk(zone, to_drain, pcp, 0); count -= to_drain; } - pcp_spin_unlock_maybe_irqrestore(pcp, UP_flags); + pcp_spin_unlock_nopin(pcp); } while (count); } @@ -6127,7 +6114,6 @@ static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu) { struct per_cpu_pages *pcp; struct cpu_cacheinfo *cci; - unsigned long UP_flags; pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); cci = get_cpu_cacheinfo(cpu); @@ -6138,12 +6124,12 @@ static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu) * This can reduce zone lock contention without hurting * cache-hot pages sharing. */ - pcp_spin_lock_maybe_irqsave(pcp, UP_flags); + pcp_spin_lock_nopin(pcp); if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch) pcp->flags |= PCPF_FREE_HIGH_BATCH; else pcp->flags &= ~PCPF_FREE_HIGH_BATCH; - pcp_spin_unlock_maybe_irqrestore(pcp, UP_flags); + pcp_spin_unlock_nopin(pcp); } void setup_pcp_cacheinfo(unsigned int cpu) -- 2.53.0 We only ever use pcpu_spin_trylock()/unlock() with struct per_cpu_pages so refactor the helpers to remove the generic layer. No functional change intended. Suggested-by: Matthew Wilcox Signed-off-by: Vlastimil Babka (SUSE) --- mm/page_alloc.c | 24 +++++++++--------------- 1 file changed, 9 insertions(+), 15 deletions(-) diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 8e5b30adfe40..b1007afa9492 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -112,35 +112,29 @@ static DEFINE_MUTEX(pcp_batch_high_lock); #endif /* - * Generic helper to lookup and a per-cpu variable with an embedded spinlock. - * Return value should be used with equivalent unlock helper. + * A helper to lookup and trylock pcp with embedded spinlock. + * The return value should be used with the unlock helper. + * NULL return value means the trylock failed. */ -#define pcpu_spin_trylock(type, member, ptr) \ +#ifdef CONFIG_SMP +#define pcp_spin_trylock(ptr) \ ({ \ - type *_ret; \ + struct per_cpu_pages *_ret; \ pcpu_task_pin(); \ _ret = this_cpu_ptr(ptr); \ - if (!spin_trylock(&_ret->member)) { \ + if (!spin_trylock(&_ret->lock)) { \ pcpu_task_unpin(); \ _ret = NULL; \ } \ _ret; \ }) -#define pcpu_spin_unlock(member, ptr) \ +#define pcp_spin_unlock(ptr) \ ({ \ - spin_unlock(&ptr->member); \ + spin_unlock(&ptr->lock); \ pcpu_task_unpin(); \ }) -/* struct per_cpu_pages specific helpers.*/ -#ifdef CONFIG_SMP -#define pcp_spin_trylock(ptr) \ - pcpu_spin_trylock(struct per_cpu_pages, lock, ptr) - -#define pcp_spin_unlock(ptr) \ - pcpu_spin_unlock(lock, ptr) - /* * On CONFIG_SMP=n the UP implementation of spin_trylock() never fails and thus * is not compatible with our locking scheme. However we do not need pcp for -- 2.53.0