md/raid6: move the spare page to a percpu allocation
authorDan Williams <dan.j.williams@intel.com>
Tue, 14 Jul 2009 18:48:22 +0000 (11:48 -0700)
committerDan Williams <dan.j.williams@intel.com>
Sun, 30 Aug 2009 02:09:26 +0000 (19:09 -0700)
In preparation for asynchronous handling of raid6 operations move the
spare page to a percpu allocation to allow multiple simultaneous
synchronous raid6 recovery operations.

Make this allocation cpu hotplug aware to maximize allocation
efficiency.

Signed-off-by: Dan Williams <dan.j.williams@intel.com>
drivers/md/raid5.c
drivers/md/raid5.h

index 9411466f71defadb48193487893d999e23cb9ce8..5359236a1ec76d94910db17ee1c14d48398bc0ec 100644 (file)
@@ -48,6 +48,7 @@
 #include <linux/raid/pq.h>
 #include <linux/async_tx.h>
 #include <linux/seq_file.h>
+#include <linux/cpu.h>
 #include "md.h"
 #include "raid5.h"
 #include "bitmap.h"
@@ -2565,14 +2566,15 @@ static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
 
 
 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
-                               struct stripe_head_state *s,
-                               struct r6_state *r6s, struct page *tmp_page,
-                               int disks)
+                                 struct stripe_head_state *s,
+                                 struct r6_state *r6s, int disks)
 {
        int update_p = 0, update_q = 0;
        struct r5dev *dev;
        int pd_idx = sh->pd_idx;
        int qd_idx = sh->qd_idx;
+       unsigned long cpu;
+       struct page *tmp_page;
 
        set_bit(STRIPE_HANDLE, &sh->state);
 
@@ -2583,78 +2585,75 @@ static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
         * case we can only check one of them, possibly using the
         * other to generate missing data
         */
-
-       /* If !tmp_page, we cannot do the calculations,
-        * but as we have set STRIPE_HANDLE, we will soon be called
-        * by stripe_handle with a tmp_page - just wait until then.
-        */
-       if (tmp_page) {
-               if (s->failed == r6s->q_failed) {
-                       /* The only possible failed device holds 'Q', so it
-                        * makes sense to check P (If anything else were failed,
-                        * we would have used P to recreate it).
-                        */
-                       compute_block_1(sh, pd_idx, 1);
-                       if (!page_is_zero(sh->dev[pd_idx].page)) {
-                               compute_block_1(sh, pd_idx, 0);
-                               update_p = 1;
-                       }
-               }
-               if (!r6s->q_failed && s->failed < 2) {
-                       /* q is not failed, and we didn't use it to generate
-                        * anything, so it makes sense to check it
-                        */
-                       memcpy(page_address(tmp_page),
-                              page_address(sh->dev[qd_idx].page),
-                              STRIPE_SIZE);
-                       compute_parity6(sh, UPDATE_PARITY);
-                       if (memcmp(page_address(tmp_page),
-                                  page_address(sh->dev[qd_idx].page),
-                                  STRIPE_SIZE) != 0) {
-                               clear_bit(STRIPE_INSYNC, &sh->state);
-                               update_q = 1;
-                       }
+       cpu = get_cpu();
+       tmp_page = per_cpu_ptr(conf->percpu, cpu)->spare_page;
+       if (s->failed == r6s->q_failed) {
+               /* The only possible failed device holds 'Q', so it
+                * makes sense to check P (If anything else were failed,
+                * we would have used P to recreate it).
+                */
+               compute_block_1(sh, pd_idx, 1);
+               if (!page_is_zero(sh->dev[pd_idx].page)) {
+                       compute_block_1(sh, pd_idx, 0);
+                       update_p = 1;
                }
-               if (update_p || update_q) {
-                       conf->mddev->resync_mismatches += STRIPE_SECTORS;
-                       if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
-                               /* don't try to repair!! */
-                               update_p = update_q = 0;
+       }
+       if (!r6s->q_failed && s->failed < 2) {
+               /* q is not failed, and we didn't use it to generate
+                * anything, so it makes sense to check it
+                */
+               memcpy(page_address(tmp_page),
+                      page_address(sh->dev[qd_idx].page),
+                      STRIPE_SIZE);
+               compute_parity6(sh, UPDATE_PARITY);
+               if (memcmp(page_address(tmp_page),
+                          page_address(sh->dev[qd_idx].page),
+                          STRIPE_SIZE) != 0) {
+                       clear_bit(STRIPE_INSYNC, &sh->state);
+                       update_q = 1;
                }
+       }
+       put_cpu();
 
-               /* now write out any block on a failed drive,
-                * or P or Q if they need it
-                */
+       if (update_p || update_q) {
+               conf->mddev->resync_mismatches += STRIPE_SECTORS;
+               if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
+                       /* don't try to repair!! */
+                       update_p = update_q = 0;
+       }
 
-               if (s->failed == 2) {
-                       dev = &sh->dev[r6s->failed_num[1]];
-                       s->locked++;
-                       set_bit(R5_LOCKED, &dev->flags);
-                       set_bit(R5_Wantwrite, &dev->flags);
-               }
-               if (s->failed >= 1) {
-                       dev = &sh->dev[r6s->failed_num[0]];
-                       s->locked++;
-                       set_bit(R5_LOCKED, &dev->flags);
-                       set_bit(R5_Wantwrite, &dev->flags);
-               }
+       /* now write out any block on a failed drive,
+        * or P or Q if they need it
+        */
 
-               if (update_p) {
-                       dev = &sh->dev[pd_idx];
-                       s->locked++;
-                       set_bit(R5_LOCKED, &dev->flags);
-                       set_bit(R5_Wantwrite, &dev->flags);
-               }
-               if (update_q) {
-                       dev = &sh->dev[qd_idx];
-                       s->locked++;
-                       set_bit(R5_LOCKED, &dev->flags);
-                       set_bit(R5_Wantwrite, &dev->flags);
-               }
-               clear_bit(STRIPE_DEGRADED, &sh->state);
+       if (s->failed == 2) {
+               dev = &sh->dev[r6s->failed_num[1]];
+               s->locked++;
+               set_bit(R5_LOCKED, &dev->flags);
+               set_bit(R5_Wantwrite, &dev->flags);
+       }
+       if (s->failed >= 1) {
+               dev = &sh->dev[r6s->failed_num[0]];
+               s->locked++;
+               set_bit(R5_LOCKED, &dev->flags);
+               set_bit(R5_Wantwrite, &dev->flags);
+       }
 
-               set_bit(STRIPE_INSYNC, &sh->state);
+       if (update_p) {
+               dev = &sh->dev[pd_idx];
+               s->locked++;
+               set_bit(R5_LOCKED, &dev->flags);
+               set_bit(R5_Wantwrite, &dev->flags);
+       }
+       if (update_q) {
+               dev = &sh->dev[qd_idx];
+               s->locked++;
+               set_bit(R5_LOCKED, &dev->flags);
+               set_bit(R5_Wantwrite, &dev->flags);
        }
+       clear_bit(STRIPE_DEGRADED, &sh->state);
+
+       set_bit(STRIPE_INSYNC, &sh->state);
 }
 
 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
@@ -3009,7 +3008,7 @@ static bool handle_stripe5(struct stripe_head *sh)
        return blocked_rdev == NULL;
 }
 
-static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
+static bool handle_stripe6(struct stripe_head *sh)
 {
        raid5_conf_t *conf = sh->raid_conf;
        int disks = sh->disks;
@@ -3164,7 +3163,7 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
         * data is available
         */
        if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
-               handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
+               handle_parity_checks6(conf, sh, &s, &r6s, disks);
 
        if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
                md_done_sync(conf->mddev, STRIPE_SECTORS,1);
@@ -3247,16 +3246,14 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
 }
 
 /* returns true if the stripe was handled */
-static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
+static bool handle_stripe(struct stripe_head *sh)
 {
        if (sh->raid_conf->level == 6)
-               return handle_stripe6(sh, tmp_page);
+               return handle_stripe6(sh);
        else
                return handle_stripe5(sh);
 }
 
-
-
 static void raid5_activate_delayed(raid5_conf_t *conf)
 {
        if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
@@ -4047,7 +4044,7 @@ static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *ski
        spin_unlock(&sh->lock);
 
        /* wait for any blocked device to be handled */
-       while(unlikely(!handle_stripe(sh, NULL)))
+       while (unlikely(!handle_stripe(sh)))
                ;
        release_stripe(sh);
 
@@ -4104,7 +4101,7 @@ static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
                        return handled;
                }
 
-               handle_stripe(sh, NULL);
+               handle_stripe(sh);
                release_stripe(sh);
                handled++;
        }
@@ -4168,7 +4165,7 @@ static void raid5d(mddev_t *mddev)
                spin_unlock_irq(&conf->device_lock);
                
                handled++;
-               handle_stripe(sh, conf->spare_page);
+               handle_stripe(sh);
                release_stripe(sh);
 
                spin_lock_irq(&conf->device_lock);
@@ -4309,15 +4306,104 @@ raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
        return sectors * (raid_disks - conf->max_degraded);
 }
 
+static void raid5_free_percpu(raid5_conf_t *conf)
+{
+       struct raid5_percpu *percpu;
+       unsigned long cpu;
+
+       if (!conf->percpu)
+               return;
+
+       get_online_cpus();
+       for_each_possible_cpu(cpu) {
+               percpu = per_cpu_ptr(conf->percpu, cpu);
+               safe_put_page(percpu->spare_page);
+       }
+#ifdef CONFIG_HOTPLUG_CPU
+       unregister_cpu_notifier(&conf->cpu_notify);
+#endif
+       put_online_cpus();
+
+       free_percpu(conf->percpu);
+}
+
 static void free_conf(raid5_conf_t *conf)
 {
        shrink_stripes(conf);
-       safe_put_page(conf->spare_page);
+       raid5_free_percpu(conf);
        kfree(conf->disks);
        kfree(conf->stripe_hashtbl);
        kfree(conf);
 }
 
+#ifdef CONFIG_HOTPLUG_CPU
+static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
+                             void *hcpu)
+{
+       raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
+       long cpu = (long)hcpu;
+       struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
+
+       switch (action) {
+       case CPU_UP_PREPARE:
+       case CPU_UP_PREPARE_FROZEN:
+               if (!percpu->spare_page)
+                       percpu->spare_page = alloc_page(GFP_KERNEL);
+               if (!percpu->spare_page) {
+                       pr_err("%s: failed memory allocation for cpu%ld\n",
+                              __func__, cpu);
+                       return NOTIFY_BAD;
+               }
+               break;
+       case CPU_DEAD:
+       case CPU_DEAD_FROZEN:
+               safe_put_page(percpu->spare_page);
+               percpu->spare_page = NULL;
+               break;
+       default:
+               break;
+       }
+       return NOTIFY_OK;
+}
+#endif
+
+static int raid5_alloc_percpu(raid5_conf_t *conf)
+{
+       unsigned long cpu;
+       struct page *spare_page;
+       struct raid5_percpu *allcpus;
+       int err;
+
+       /* the only percpu data is the raid6 spare page */
+       if (conf->level != 6)
+               return 0;
+
+       allcpus = alloc_percpu(struct raid5_percpu);
+       if (!allcpus)
+               return -ENOMEM;
+       conf->percpu = allcpus;
+
+       get_online_cpus();
+       err = 0;
+       for_each_present_cpu(cpu) {
+               spare_page = alloc_page(GFP_KERNEL);
+               if (!spare_page) {
+                       err = -ENOMEM;
+                       break;
+               }
+               per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
+       }
+#ifdef CONFIG_HOTPLUG_CPU
+       conf->cpu_notify.notifier_call = raid456_cpu_notify;
+       conf->cpu_notify.priority = 0;
+       if (err == 0)
+               err = register_cpu_notifier(&conf->cpu_notify);
+#endif
+       put_online_cpus();
+
+       return err;
+}
+
 static raid5_conf_t *setup_conf(mddev_t *mddev)
 {
        raid5_conf_t *conf;
@@ -4372,11 +4458,10 @@ static raid5_conf_t *setup_conf(mddev_t *mddev)
        if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
                goto abort;
 
-       if (mddev->new_level == 6) {
-               conf->spare_page = alloc_page(GFP_KERNEL);
-               if (!conf->spare_page)
-                       goto abort;
-       }
+       conf->level = mddev->new_level;
+       if (raid5_alloc_percpu(conf) != 0)
+               goto abort;
+
        spin_lock_init(&conf->device_lock);
        init_waitqueue_head(&conf->wait_for_stripe);
        init_waitqueue_head(&conf->wait_for_overlap);
@@ -4412,7 +4497,6 @@ static raid5_conf_t *setup_conf(mddev_t *mddev)
        }
 
        conf->chunk_size = mddev->new_chunk;
-       conf->level = mddev->new_level;
        if (conf->level == 6)
                conf->max_degraded = 2;
        else
index 52ba99954decf0ceee11bc30da2213a7c6619983..07a7a4102f0545b39a2be49718882a76014434d1 100644 (file)
@@ -383,8 +383,13 @@ struct raid5_private_data {
                                            * (fresh device added).
                                            * Cleared when a sync completes.
                                            */
-
-       struct page             *spare_page; /* Used when checking P/Q in raid6 */
+       /* per cpu variables */
+       struct raid5_percpu {
+               struct page     *spare_page; /* Used when checking P/Q in raid6 */
+       } *percpu;
+#ifdef CONFIG_HOTPLUG_CPU
+       struct notifier_block   cpu_notify;
+#endif
 
        /*
         * Free stripes pool