From 0d6526f82c3cdefcca47f73f5fc08dc6f335eac6 Mon Sep 17 00:00:00 2001 From: Tim Chen Date: Mon, 21 Sep 2026 17:37:22 -0700 Subject: sched/cache: Keep nr_pref_llc_running in the runnable domain, to fix LLC mis-scheduling bug alb_break_llc() decides whether to break LLC preference during active load balance. It does so by testing that every runnable fair task on the source rq prefers its LLC: env->src_rq->nr_pref_llc_running == env->src_rq->cfs.h_nr_runnable But the two counters cover different sets. nr_pref_llc_running is updated in account_llc_enqueue()/account_llc_dequeue(), next to cfs_rq->nr_queued, so it follows queued tasks. h_nr_runnable is updated in set_delayed()/ clear_delayed() and drops delay-dequeued tasks. So under DELAY_DEQUEUE, a preferring task that goes to sleep stays counted in nr_pref_llc_running while h_nr_runnable falls. The equality then breaks, alb_break_llc() returns false, and active balance is free to pull a task off its preferred LLC. Active balance only moves runnable tasks, and this is the only LLC check it consults: once the stopper runs, LBF_ACTIVE_LB skips the per-task test in can_migrate_task(). The runnable set is the one we want. Fix it on the counter side. A task should be counted in nr_pref_llc_running exactly while it is both queued on its preferred LLC (pref_llc_queued) and runnable (!sched_delayed). Define that membership once in task_pref_llc_runnable(), and adjust the counter only through pref_llc_running_inc()/pref_llc_running_dec() from the four sites that change either input: account_llc_enqueue(), account_llc_dequeue(), set_delayed() and clear_delayed(). Gating every update on the same predicate keeps the delay, wake and dequeue paths from double-counting or underflowing; see the comments at those sites for the ordering. nr_llc_running and sd->llc_counts are not touched and stay on queued semantics. Fixes: 714059f79ff0 ("sched/cache: Handle moving single tasks to/from their preferred LLC") Closes: https://lore.kernel.org/lkml/20260827135000.735138-1-zhanxusheng@xiaomi.com/ Reported-by: Zhan Xusheng Suggested-by: Chen Yu Signed-off-by: Tim Chen Signed-off-by: Peter Zijlstra (Intel) Signed-off-by: Ingo Molnar Reviewed-by: Kayra Cizmeci Cc: # v7.2.x Link: https://patch.msgid.link/06af61afedac32e6477f57feb4d658f6c411c3af.1790035273.git.tim.c.chen@linux.intel.com --- mm/kmsan/init.c | 237 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 237 insertions(+) create mode 100644 mm/kmsan/init.c (limited to 'mm/kmsan/init.c') diff --git a/mm/kmsan/init.c b/mm/kmsan/init.c new file mode 100644 index 000000000..4983b6e9f --- /dev/null +++ b/mm/kmsan/init.c @@ -0,0 +1,237 @@ +// SPDX-License-Identifier: GPL-2.0 +/* + * KMSAN initialization routines. + * + * Copyright (C) 2017-2021 Google LLC + * Author: Alexander Potapenko + * + */ + +#include "kmsan.h" + +#include +#include +#include + +#include "../page_alloc.h" + +#define NUM_FUTURE_RANGES 128 +struct start_end_pair { + u64 start, end; +}; + +static struct start_end_pair start_end_pairs[NUM_FUTURE_RANGES] __initdata; +static int future_index __initdata; + +/* + * Record a range of memory for which the metadata pages will be created once + * the page allocator becomes available. + */ +static void __init kmsan_record_future_shadow_range(void *start, void *end) +{ + u64 nstart = (u64)start, nend = (u64)end, cstart, cend; + bool merged = false; + + KMSAN_WARN_ON(future_index == NUM_FUTURE_RANGES); + KMSAN_WARN_ON((nstart >= nend) || + /* Virtual address 0 is valid on s390. */ + (!IS_ENABLED(CONFIG_S390) && !nstart) || !nend); + nstart = ALIGN_DOWN(nstart, PAGE_SIZE); + nend = ALIGN(nend, PAGE_SIZE); + + /* + * Scan the existing ranges to see if any of them overlaps with + * [start, end). In that case, merge the two ranges instead of + * creating a new one. + * The number of ranges is less than 20, so there is no need to organize + * them into a more intelligent data structure. + */ + for (int i = 0; i < future_index; i++) { + cstart = start_end_pairs[i].start; + cend = start_end_pairs[i].end; + if ((cstart < nstart && cend < nstart) || + (cstart > nend && cend > nend)) + /* ranges are disjoint - do not merge */ + continue; + start_end_pairs[i].start = min(nstart, cstart); + start_end_pairs[i].end = max(nend, cend); + merged = true; + break; + } + if (merged) + return; + start_end_pairs[future_index].start = nstart; + start_end_pairs[future_index].end = nend; + future_index++; +} + +/* + * Initialize the shadow for existing mappings during kernel initialization. + * These include kernel text/data sections, NODE_DATA and future ranges + * registered while creating other data (e.g. percpu). + * + * Allocations via memblock can be only done before slab is initialized. + */ +void __init kmsan_init_shadow(void) +{ + const size_t nd_size = sizeof(pg_data_t); + phys_addr_t p_start, p_end; + u64 loop; + int nid; + + for_each_reserved_mem_range(loop, &p_start, &p_end) + kmsan_record_future_shadow_range(phys_to_virt(p_start), + phys_to_virt(p_end)); + /* Allocate shadow for .data */ + kmsan_record_future_shadow_range(_sdata, _edata); + + for_each_online_node(nid) + kmsan_record_future_shadow_range( + NODE_DATA(nid), (char *)NODE_DATA(nid) + nd_size); + + for (int i = 0; i < future_index; i++) + kmsan_init_alloc_meta_for_range( + (void *)start_end_pairs[i].start, + (void *)start_end_pairs[i].end); +} + +struct metadata_page_pair { + struct page *shadow, *origin; +}; +static struct metadata_page_pair held_back[NR_PAGE_ORDERS] __initdata; + +/* + * Eager metadata allocation. When the memblock allocator is freeing pages to + * pagealloc, we use 2/3 of them as metadata for the remaining 1/3. + * We store the pointers to the returned blocks of pages in held_back[] grouped + * by their order: when kmsan_memblock_free_pages() is called for the first + * time with a certain order, it is reserved as a shadow block, for the second + * time - as an origin block. On the third time the incoming block receives its + * shadow and origin ranges from the previously saved shadow and origin blocks, + * after which held_back[order] can be used again. + * + * At the very end there may be leftover blocks in held_back[]. They are + * collected later by kmsan_memblock_discard(). + */ +bool kmsan_memblock_free_pages(struct page *page, unsigned int order) +{ + struct page *shadow, *origin; + + if (!held_back[order].shadow) { + held_back[order].shadow = page; + return false; + } + if (!held_back[order].origin) { + held_back[order].origin = page; + return false; + } + shadow = held_back[order].shadow; + origin = held_back[order].origin; + kmsan_setup_meta(page, shadow, origin, order); + + held_back[order].shadow = NULL; + held_back[order].origin = NULL; + return true; +} + +#define MAX_BLOCKS 8 +struct smallstack { + struct page *items[MAX_BLOCKS]; + int index; + int order; +}; + +static struct smallstack collect = { + .index = 0, + .order = MAX_PAGE_ORDER, +}; + +static void smallstack_push(struct smallstack *stack, struct page *pages) +{ + KMSAN_WARN_ON(stack->index == MAX_BLOCKS); + stack->items[stack->index] = pages; + stack->index++; +} +#undef MAX_BLOCKS + +static struct page *smallstack_pop(struct smallstack *stack) +{ + struct page *ret; + + KMSAN_WARN_ON(stack->index == 0); + stack->index--; + ret = stack->items[stack->index]; + stack->items[stack->index] = NULL; + return ret; +} + +static void do_collection(void) +{ + struct page *page, *shadow, *origin; + + while (collect.index >= 3) { + page = smallstack_pop(&collect); + shadow = smallstack_pop(&collect); + origin = smallstack_pop(&collect); + kmsan_setup_meta(page, shadow, origin, collect.order); + __free_pages_core(page, collect.order, MEMINIT_EARLY); + } +} + +static void collect_split(void) +{ + struct smallstack tmp = { + .order = collect.order - 1, + .index = 0, + }; + struct page *page; + + if (!collect.order) + return; + while (collect.index) { + page = smallstack_pop(&collect); + smallstack_push(&tmp, &page[0]); + smallstack_push(&tmp, &page[1 << tmp.order]); + } + __memcpy(&collect, &tmp, sizeof(tmp)); +} + +/* + * Memblock is about to go away. Split the page blocks left over in held_back[] + * and return 1/3 of that memory to the system. + */ +static void kmsan_memblock_discard(void) +{ + /* + * For each order=N: + * - push held_back[N].shadow and .origin to @collect; + * - while there are >= 3 elements in @collect, do garbage collection: + * - pop 3 ranges from @collect; + * - use two of them as shadow and origin for the third one; + * - repeat; + * - split each remaining element from @collect into 2 ranges of + * order=N-1, + * - repeat. + */ + collect.order = MAX_PAGE_ORDER; + for (int i = MAX_PAGE_ORDER; i >= 0; i--) { + if (held_back[i].shadow) + smallstack_push(&collect, held_back[i].shadow); + if (held_back[i].origin) + smallstack_push(&collect, held_back[i].origin); + held_back[i].shadow = NULL; + held_back[i].origin = NULL; + do_collection(); + collect_split(); + } +} + +void __init kmsan_init_runtime(void) +{ + /* Assuming current is init_task */ + kmsan_internal_task_create(current); + kmsan_memblock_discard(); + pr_info("Starting KernelMemorySanitizer\n"); + pr_info("ATTENTION: KMSAN is a debugging tool! Do not use it on production machines!\n"); + kmsan_enabled = true; +} -- cgit v1.3.1