rcu/kvfree: Switch to a generic linked list API
authorUladzislau Rezki (Sony) <urezki@gmail.com>
Tue, 29 Nov 2022 15:58:19 +0000 (16:58 +0100)
committerPaul E. McKenney <paulmck@kernel.org>
Wed, 4 Jan 2023 01:48:41 +0000 (17:48 -0800)
This commit improves the readability and maintainability of the
kvfree_rcu() code by switching from an open-coded linked list to
the standard Linux-kernel circular doubly linked list.  This patch
does not introduce any functional change.

Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
kernel/rcu/tree.c

index 7d222acd85bfd5f56da9e3b6c2b52af522dd5015..4088b34ce9610bdd34b62a5e645a90f8ae7fc410 100644 (file)
@@ -2876,13 +2876,13 @@ EXPORT_SYMBOL_GPL(call_rcu);
 
 /**
  * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
+ * @list: List node. All blocks are linked between each other
  * @nr_records: Number of active pointers in the array
- * @next: Next bulk object in the block chain
  * @records: Array of the kvfree_rcu() pointers
  */
 struct kvfree_rcu_bulk_data {
+       struct list_head list;
        unsigned long nr_records;
-       struct kvfree_rcu_bulk_data *next;
        void *records[];
 };
 
@@ -2898,21 +2898,21 @@ struct kvfree_rcu_bulk_data {
  * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
  * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
  * @head_free: List of kfree_rcu() objects waiting for a grace period
- * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
+ * @bulk_head_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
  * @krcp: Pointer to @kfree_rcu_cpu structure
  */
 
 struct kfree_rcu_cpu_work {
        struct rcu_work rcu_work;
        struct rcu_head *head_free;
-       struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
+       struct list_head bulk_head_free[FREE_N_CHANNELS];
        struct kfree_rcu_cpu *krcp;
 };
 
 /**
  * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
  * @head: List of kfree_rcu() objects not yet waiting for a grace period
- * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
+ * @bulk_head: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
  * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
  * @lock: Synchronize access to this structure
  * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
@@ -2936,7 +2936,7 @@ struct kfree_rcu_cpu_work {
  */
 struct kfree_rcu_cpu {
        struct rcu_head *head;
-       struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS];
+       struct list_head bulk_head[FREE_N_CHANNELS];
        struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
        raw_spinlock_t lock;
        struct delayed_work monitor_work;
@@ -3031,12 +3031,13 @@ drain_page_cache(struct kfree_rcu_cpu *krcp)
 
 /*
  * This function is invoked in workqueue context after a grace period.
- * It frees all the objects queued on ->bkvhead_free or ->head_free.
+ * It frees all the objects queued on ->bulk_head_free or ->head_free.
  */
 static void kfree_rcu_work(struct work_struct *work)
 {
        unsigned long flags;
-       struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext;
+       struct kvfree_rcu_bulk_data *bnode, *n;
+       struct list_head bulk_head[FREE_N_CHANNELS];
        struct rcu_head *head, *next;
        struct kfree_rcu_cpu *krcp;
        struct kfree_rcu_cpu_work *krwp;
@@ -3048,10 +3049,8 @@ static void kfree_rcu_work(struct work_struct *work)
 
        raw_spin_lock_irqsave(&krcp->lock, flags);
        // Channels 1 and 2.
-       for (i = 0; i < FREE_N_CHANNELS; i++) {
-               bkvhead[i] = krwp->bkvhead_free[i];
-               krwp->bkvhead_free[i] = NULL;
-       }
+       for (i = 0; i < FREE_N_CHANNELS; i++)
+               list_replace_init(&krwp->bulk_head_free[i], &bulk_head[i]);
 
        // Channel 3.
        head = krwp->head_free;
@@ -3060,36 +3059,33 @@ static void kfree_rcu_work(struct work_struct *work)
 
        // Handle the first two channels.
        for (i = 0; i < FREE_N_CHANNELS; i++) {
-               for (; bkvhead[i]; bkvhead[i] = bnext) {
-                       bnext = bkvhead[i]->next;
-                       debug_rcu_bhead_unqueue(bkvhead[i]);
+               list_for_each_entry_safe(bnode, n, &bulk_head[i], list) {
+                       debug_rcu_bhead_unqueue(bnode);
 
                        rcu_lock_acquire(&rcu_callback_map);
                        if (i == 0) { // kmalloc() / kfree().
                                trace_rcu_invoke_kfree_bulk_callback(
-                                       rcu_state.name, bkvhead[i]->nr_records,
-                                       bkvhead[i]->records);
+                                       rcu_state.name, bnode->nr_records,
+                                       bnode->records);
 
-                               kfree_bulk(bkvhead[i]->nr_records,
-                                       bkvhead[i]->records);
+                               kfree_bulk(bnode->nr_records, bnode->records);
                        } else { // vmalloc() / vfree().
-                               for (j = 0; j < bkvhead[i]->nr_records; j++) {
+                               for (j = 0; j < bnode->nr_records; j++) {
                                        trace_rcu_invoke_kvfree_callback(
-                                               rcu_state.name,
-                                               bkvhead[i]->records[j], 0);
+                                               rcu_state.name, bnode->records[j], 0);
 
-                                       vfree(bkvhead[i]->records[j]);
+                                       vfree(bnode->records[j]);
                                }
                        }
                        rcu_lock_release(&rcu_callback_map);
 
                        raw_spin_lock_irqsave(&krcp->lock, flags);
-                       if (put_cached_bnode(krcp, bkvhead[i]))
-                               bkvhead[i] = NULL;
+                       if (put_cached_bnode(krcp, bnode))
+                               bnode = NULL;
                        raw_spin_unlock_irqrestore(&krcp->lock, flags);
 
-                       if (bkvhead[i])
-                               free_page((unsigned long) bkvhead[i]);
+                       if (bnode)
+                               free_page((unsigned long) bnode);
 
                        cond_resched_tasks_rcu_qs();
                }
@@ -3125,7 +3121,7 @@ need_offload_krc(struct kfree_rcu_cpu *krcp)
        int i;
 
        for (i = 0; i < FREE_N_CHANNELS; i++)
-               if (krcp->bkvhead[i])
+               if (!list_empty(&krcp->bulk_head[i]))
                        return true;
 
        return !!krcp->head;
@@ -3162,21 +3158,20 @@ static void kfree_rcu_monitor(struct work_struct *work)
        for (i = 0; i < KFREE_N_BATCHES; i++) {
                struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
 
-               // Try to detach bkvhead or head and attach it over any
+               // Try to detach bulk_head or head and attach it over any
                // available corresponding free channel. It can be that
                // a previous RCU batch is in progress, it means that
                // immediately to queue another one is not possible so
                // in that case the monitor work is rearmed.
-               if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) ||
-                       (krcp->bkvhead[1] && !krwp->bkvhead_free[1]) ||
+               if ((!list_empty(&krcp->bulk_head[0]) && list_empty(&krwp->bulk_head_free[0])) ||
+                       (!list_empty(&krcp->bulk_head[1]) && list_empty(&krwp->bulk_head_free[1])) ||
                                (krcp->head && !krwp->head_free)) {
+
                        // Channel 1 corresponds to the SLAB-pointer bulk path.
                        // Channel 2 corresponds to vmalloc-pointer bulk path.
                        for (j = 0; j < FREE_N_CHANNELS; j++) {
-                               if (!krwp->bkvhead_free[j]) {
-                                       krwp->bkvhead_free[j] = krcp->bkvhead[j];
-                                       krcp->bkvhead[j] = NULL;
-                               }
+                               if (list_empty(&krwp->bulk_head_free[j]))
+                                       list_replace_init(&krcp->bulk_head[j], &krwp->bulk_head_free[j]);
                        }
 
                        // Channel 3 corresponds to both SLAB and vmalloc
@@ -3288,10 +3283,11 @@ add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
                return false;
 
        idx = !!is_vmalloc_addr(ptr);
+       bnode = list_first_entry_or_null(&(*krcp)->bulk_head[idx],
+               struct kvfree_rcu_bulk_data, list);
 
        /* Check if a new block is required. */
-       if (!(*krcp)->bkvhead[idx] ||
-                       (*krcp)->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
+       if (!bnode || bnode->nr_records == KVFREE_BULK_MAX_ENTR) {
                bnode = get_cached_bnode(*krcp);
                if (!bnode && can_alloc) {
                        krc_this_cpu_unlock(*krcp, *flags);
@@ -3315,18 +3311,13 @@ add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
                if (!bnode)
                        return false;
 
-               /* Initialize the new block. */
+               // Initialize the new block and attach it.
                bnode->nr_records = 0;
-               bnode->next = (*krcp)->bkvhead[idx];
-
-               /* Attach it to the head. */
-               (*krcp)->bkvhead[idx] = bnode;
+               list_add(&bnode->list, &(*krcp)->bulk_head[idx]);
        }
 
        /* Finally insert. */
-       (*krcp)->bkvhead[idx]->records
-               [(*krcp)->bkvhead[idx]->nr_records++] = ptr;
-
+       bnode->records[bnode->nr_records++] = ptr;
        return true;
 }
 
@@ -4761,7 +4752,7 @@ struct workqueue_struct *rcu_gp_wq;
 static void __init kfree_rcu_batch_init(void)
 {
        int cpu;
-       int i;
+       int i, j;
 
        /* Clamp it to [0:100] seconds interval. */
        if (rcu_delay_page_cache_fill_msec < 0 ||
@@ -4781,8 +4772,14 @@ static void __init kfree_rcu_batch_init(void)
                for (i = 0; i < KFREE_N_BATCHES; i++) {
                        INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
                        krcp->krw_arr[i].krcp = krcp;
+
+                       for (j = 0; j < FREE_N_CHANNELS; j++)
+                               INIT_LIST_HEAD(&krcp->krw_arr[i].bulk_head_free[j]);
                }
 
+               for (i = 0; i < FREE_N_CHANNELS; i++)
+                       INIT_LIST_HEAD(&krcp->bulk_head[i]);
+
                INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
                INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
                krcp->initialized = true;