sched/cache: Keep nr_pref_llc_running in the runnable domain, to fix LLC mis-scheduling buggrafted

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 <zhanxusheng@xiaomi.com>
Suggested-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kayra Cizmeci <kayracizmeci@gmail.com>
Cc: <stable@kernel.org> # v7.2.x
Link: https://patch.msgid.link/06af61afedac32e6477f57feb4d658f6c411c3af.1790035273.git.tim.c.chen@linux.intel.com

1 files changed, 296 insertions, 0 deletions

diff --git a/drivers/misc/ocxl/context.c b/drivers/misc/ocxl/context.c
new file mode 100644
index 000000000..33f25be46
--- /dev/null
+++ b/drivers/misc/ocxl/context.c
@@ -0,0 +1,296 @@
+// SPDX-License-Identifier: GPL-2.0+
+// Copyright 2017 IBM Corp.
+#include <linux/sched/mm.h>
+#include "trace.h"
+#include "ocxl_internal.h"
+
+int ocxl_context_alloc(struct ocxl_context **context, struct ocxl_afu *afu,
+		struct address_space *mapping)
+{
+	int pasid;
+	struct ocxl_context *ctx;
+
+	ctx = kzalloc_obj(*ctx);
+	if (!ctx)
+		return -ENOMEM;
+
+	ctx->afu = afu;
+	mutex_lock(&afu->contexts_lock);
+	pasid = idr_alloc(&afu->contexts_idr, ctx, afu->pasid_base,
+			afu->pasid_base + afu->pasid_max, GFP_KERNEL);
+	if (pasid < 0) {
+		mutex_unlock(&afu->contexts_lock);
+		kfree(ctx);
+		return pasid;
+	}
+	afu->pasid_count++;
+	mutex_unlock(&afu->contexts_lock);
+
+	ctx->pasid = pasid;
+	ctx->status = OPENED;
+	mutex_init(&ctx->status_mutex);
+	ctx->mapping = mapping;
+	mutex_init(&ctx->mapping_lock);
+	init_waitqueue_head(&ctx->events_wq);
+	mutex_init(&ctx->xsl_error_lock);
+	mutex_init(&ctx->irq_lock);
+	idr_init(&ctx->irq_idr);
+	ctx->tidr = 0;
+
+	/*
+	 * Keep a reference on the AFU to make sure it's valid for the
+	 * duration of the life of the context
+	 */
+	ocxl_afu_get(afu);
+	*context = ctx;
+	return 0;
+}
+EXPORT_SYMBOL_GPL(ocxl_context_alloc);
+
+/*
+ * Callback for when a translation fault triggers an error
+ * data:	a pointer to the context which triggered the fault
+ * addr:	the address that triggered the error
+ * dsisr:	the value of the PPC64 dsisr register
+ */
+static void xsl_fault_error(void *data, u64 addr, u64 dsisr)
+{
+	struct ocxl_context *ctx = data;
+
+	mutex_lock(&ctx->xsl_error_lock);
+	ctx->xsl_error.addr = addr;
+	ctx->xsl_error.dsisr = dsisr;
+	ctx->xsl_error.count++;
+	mutex_unlock(&ctx->xsl_error_lock);
+
+	wake_up_all(&ctx->events_wq);
+}
+
+int ocxl_context_attach(struct ocxl_context *ctx, u64 amr, struct mm_struct *mm)
+{
+	int rc;
+	unsigned long pidr = 0;
+	struct pci_dev *dev;
+
+	// Locks both status & tidr
+	mutex_lock(&ctx->status_mutex);
+	if (ctx->status != OPENED) {
+		rc = -EIO;
+		goto out;
+	}
+
+	if (mm)
+		pidr = mm->context.id;
+
+	dev = to_pci_dev(ctx->afu->fn->dev.parent);
+	rc = ocxl_link_add_pe(ctx->afu->fn->link, ctx->pasid, pidr, ctx->tidr,
+			      amr, pci_dev_id(dev), mm, xsl_fault_error, ctx);
+	if (rc)
+		goto out;
+
+	ctx->status = ATTACHED;
+out:
+	mutex_unlock(&ctx->status_mutex);
+	return rc;
+}
+EXPORT_SYMBOL_GPL(ocxl_context_attach);
+
+static vm_fault_t map_afu_irq(struct vm_area_struct *vma, unsigned long address,
+		u64 offset, struct ocxl_context *ctx)
+{
+	u64 trigger_addr;
+	int irq_id = ocxl_irq_offset_to_id(ctx, offset);
+
+	trigger_addr = ocxl_afu_irq_get_addr(ctx, irq_id);
+	if (!trigger_addr)
+		return VM_FAULT_SIGBUS;
+
+	return vmf_insert_pfn(vma, address, trigger_addr >> PAGE_SHIFT);
+}
+
+static vm_fault_t map_pp_mmio(struct vm_area_struct *vma, unsigned long address,
+		u64 offset, struct ocxl_context *ctx)
+{
+	u64 pp_mmio_addr;
+	int pasid_off;
+	vm_fault_t ret;
+
+	if (offset >= ctx->afu->config.pp_mmio_stride)
+		return VM_FAULT_SIGBUS;
+
+	mutex_lock(&ctx->status_mutex);
+	if (ctx->status != ATTACHED) {
+		mutex_unlock(&ctx->status_mutex);
+		pr_debug("%s: Context not attached, failing mmio mmap\n",
+			__func__);
+		return VM_FAULT_SIGBUS;
+	}
+
+	pasid_off = ctx->pasid - ctx->afu->pasid_base;
+	pp_mmio_addr = ctx->afu->pp_mmio_start +
+		pasid_off * ctx->afu->config.pp_mmio_stride +
+		offset;
+
+	ret = vmf_insert_pfn(vma, address, pp_mmio_addr >> PAGE_SHIFT);
+	mutex_unlock(&ctx->status_mutex);
+	return ret;
+}
+
+static vm_fault_t ocxl_mmap_fault(struct vm_fault *vmf)
+{
+	struct vm_area_struct *vma = vmf->vma;
+	struct ocxl_context *ctx = vma->vm_file->private_data;
+	u64 offset;
+	vm_fault_t ret;
+
+	offset = vmf->pgoff << PAGE_SHIFT;
+	pr_debug("%s: pasid %d address 0x%lx offset 0x%llx\n", __func__,
+		ctx->pasid, vmf->address, offset);
+
+	if (offset < ctx->afu->irq_base_offset)
+		ret = map_pp_mmio(vma, vmf->address, offset, ctx);
+	else
+		ret = map_afu_irq(vma, vmf->address, offset, ctx);
+	return ret;
+}
+
+static const struct vm_operations_struct ocxl_vmops = {
+	.fault = ocxl_mmap_fault,
+};
+
+static int check_mmap_afu_irq(struct ocxl_context *ctx,
+			struct vm_area_struct *vma)
+{
+	int irq_id = ocxl_irq_offset_to_id(ctx, vma->vm_pgoff << PAGE_SHIFT);
+
+	/* only one page */
+	if (vma_pages(vma) != 1)
+		return -EINVAL;
+
+	/* check offset validty */
+	if (!ocxl_afu_irq_get_addr(ctx, irq_id))
+		return -EINVAL;
+
+	/*
+	 * trigger page should only be accessible in write mode.
+	 *
+	 * It's a bit theoretical, as a page mmaped with only
+	 * PROT_WRITE is currently readable, but it doesn't hurt.
+	 */
+	if ((vma->vm_flags & VM_READ) || (vma->vm_flags & VM_EXEC) ||
+		!(vma->vm_flags & VM_WRITE))
+		return -EINVAL;
+	vm_flags_clear(vma, VM_MAYREAD | VM_MAYEXEC);
+	return 0;
+}
+
+static int check_mmap_mmio(struct ocxl_context *ctx,
+			struct vm_area_struct *vma)
+{
+	if ((vma_pages(vma) + vma->vm_pgoff) >
+		(ctx->afu->config.pp_mmio_stride >> PAGE_SHIFT))
+		return -EINVAL;
+	return 0;
+}
+
+int ocxl_context_mmap(struct ocxl_context *ctx, struct vm_area_struct *vma)
+{
+	int rc;
+
+	if ((vma->vm_pgoff << PAGE_SHIFT) < ctx->afu->irq_base_offset)
+		rc = check_mmap_mmio(ctx, vma);
+	else
+		rc = check_mmap_afu_irq(ctx, vma);
+	if (rc)
+		return rc;
+
+	vm_flags_set(vma, VM_IO | VM_PFNMAP);
+	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+	vma->vm_ops = &ocxl_vmops;
+	return 0;
+}
+
+int ocxl_context_detach(struct ocxl_context *ctx)
+{
+	struct pci_dev *dev;
+	int afu_control_pos;
+	enum ocxl_context_status status;
+	int rc;
+
+	mutex_lock(&ctx->status_mutex);
+	status = ctx->status;
+	ctx->status = CLOSED;
+	mutex_unlock(&ctx->status_mutex);
+	if (status != ATTACHED)
+		return 0;
+
+	dev = to_pci_dev(ctx->afu->fn->dev.parent);
+	afu_control_pos = ctx->afu->config.dvsec_afu_control_pos;
+
+	mutex_lock(&ctx->afu->afu_control_lock);
+	rc = ocxl_config_terminate_pasid(dev, afu_control_pos, ctx->pasid);
+	mutex_unlock(&ctx->afu->afu_control_lock);
+	trace_ocxl_terminate_pasid(ctx->pasid, rc);
+	if (rc) {
+		/*
+		 * If we timeout waiting for the AFU to terminate the
+		 * pasid, then it's dangerous to clean up the Process
+		 * Element entry in the SPA, as it may be referenced
+		 * in the future by the AFU. In which case, we would
+		 * checkstop because of an invalid PE access (FIR
+		 * register 2, bit 42). So leave the PE
+		 * defined. Caller shouldn't free the context so that
+		 * PASID remains allocated.
+		 *
+		 * A link reset will be required to cleanup the AFU
+		 * and the SPA.
+		 */
+		if (rc == -EBUSY)
+			return rc;
+	}
+	rc = ocxl_link_remove_pe(ctx->afu->fn->link, ctx->pasid);
+	if (rc) {
+		dev_warn(&dev->dev,
+			"Couldn't remove PE entry cleanly: %d\n", rc);
+	}
+	return 0;
+}
+EXPORT_SYMBOL_GPL(ocxl_context_detach);
+
+void ocxl_context_detach_all(struct ocxl_afu *afu)
+{
+	struct ocxl_context *ctx;
+	int tmp;
+
+	mutex_lock(&afu->contexts_lock);
+	idr_for_each_entry(&afu->contexts_idr, ctx, tmp) {
+		ocxl_context_detach(ctx);
+		/*
+		 * We are force detaching - remove any active mmio
+		 * mappings so userspace cannot interfere with the
+		 * card if it comes back.  Easiest way to exercise
+		 * this is to unbind and rebind the driver via sysfs
+		 * while it is in use.
+		 */
+		mutex_lock(&ctx->mapping_lock);
+		if (ctx->mapping)
+			unmap_mapping_range(ctx->mapping, 0, 0, 1);
+		mutex_unlock(&ctx->mapping_lock);
+	}
+	mutex_unlock(&afu->contexts_lock);
+}
+
+void ocxl_context_free(struct ocxl_context *ctx)
+{
+	mutex_lock(&ctx->afu->contexts_lock);
+	ctx->afu->pasid_count--;
+	idr_remove(&ctx->afu->contexts_idr, ctx->pasid);
+	mutex_unlock(&ctx->afu->contexts_lock);
+
+	ocxl_afu_irq_free_all(ctx);
+	idr_destroy(&ctx->irq_idr);
+	/* reference to the AFU taken in ocxl_context_alloc() */
+	ocxl_afu_put(ctx->afu);
+	kfree(ctx);
+}
+EXPORT_SYMBOL_GPL(ocxl_context_free);