dm kcopyd: per device
authorMikulas Patocka <mpatocka@redhat.com>
Thu, 24 Apr 2008 20:43:44 +0000 (21:43 +0100)
committerAlasdair G Kergon <agk@redhat.com>
Fri, 25 Apr 2008 12:26:49 +0000 (13:26 +0100)
Make one kcopyd thread per device.

The original shared kcopyd could deadlock.

Configuration:

drivers/md/kcopyd.c

index 4f2c61acf7c64323225b0cf98948e2f04eb0ea2d..3fb6c8334a82c64c6f488bde55f7c78e6a653a96 100644 (file)
 #include "kcopyd.h"
 #include "dm.h"
 
-static struct workqueue_struct *_kcopyd_wq;
-static struct work_struct _kcopyd_work;
-
-static void wake(void)
-{
-       queue_work(_kcopyd_wq, &_kcopyd_work);
-}
-
 /*-----------------------------------------------------------------
  * Each kcopyd client has its own little pool of preallocated
  * pages for kcopyd io.
@@ -50,8 +42,30 @@ struct dm_kcopyd_client {
 
        wait_queue_head_t destroyq;
        atomic_t nr_jobs;
+
+       struct workqueue_struct *kcopyd_wq;
+       struct work_struct kcopyd_work;
+
+/*
+ * We maintain three lists of jobs:
+ *
+ * i)   jobs waiting for pages
+ * ii)  jobs that have pages, and are waiting for the io to be issued.
+ * iii) jobs that have completed.
+ *
+ * All three of these are protected by job_lock.
+ */
+       spinlock_t job_lock;
+       struct list_head complete_jobs;
+       struct list_head io_jobs;
+       struct list_head pages_jobs;
 };
 
+static void wake(struct dm_kcopyd_client *kc)
+{
+       queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
+}
+
 static struct page_list *alloc_pl(void)
 {
        struct page_list *pl;
@@ -209,21 +223,6 @@ struct kcopyd_job {
 static struct kmem_cache *_job_cache;
 static mempool_t *_job_pool;
 
-/*
- * We maintain three lists of jobs:
- *
- * i)   jobs waiting for pages
- * ii)  jobs that have pages, and are waiting for the io to be issued.
- * iii) jobs that have completed.
- *
- * All three of these are protected by job_lock.
- */
-static DEFINE_SPINLOCK(_job_lock);
-
-static LIST_HEAD(_complete_jobs);
-static LIST_HEAD(_io_jobs);
-static LIST_HEAD(_pages_jobs);
-
 static int jobs_init(void)
 {
        _job_cache = KMEM_CACHE(kcopyd_job, 0);
@@ -241,10 +240,6 @@ static int jobs_init(void)
 
 static void jobs_exit(void)
 {
-       BUG_ON(!list_empty(&_complete_jobs));
-       BUG_ON(!list_empty(&_io_jobs));
-       BUG_ON(!list_empty(&_pages_jobs));
-
        mempool_destroy(_job_pool);
        kmem_cache_destroy(_job_cache);
        _job_pool = NULL;
@@ -255,18 +250,19 @@ static void jobs_exit(void)
  * Functions to push and pop a job onto the head of a given job
  * list.
  */
-static struct kcopyd_job *pop(struct list_head *jobs)
+static struct kcopyd_job *pop(struct list_head *jobs,
+                             struct dm_kcopyd_client *kc)
 {
        struct kcopyd_job *job = NULL;
        unsigned long flags;
 
-       spin_lock_irqsave(&_job_lock, flags);
+       spin_lock_irqsave(&kc->job_lock, flags);
 
        if (!list_empty(jobs)) {
                job = list_entry(jobs->next, struct kcopyd_job, list);
                list_del(&job->list);
        }
-       spin_unlock_irqrestore(&_job_lock, flags);
+       spin_unlock_irqrestore(&kc->job_lock, flags);
 
        return job;
 }
@@ -274,10 +270,11 @@ static struct kcopyd_job *pop(struct list_head *jobs)
 static void push(struct list_head *jobs, struct kcopyd_job *job)
 {
        unsigned long flags;
+       struct dm_kcopyd_client *kc = job->kc;
 
-       spin_lock_irqsave(&_job_lock, flags);
+       spin_lock_irqsave(&kc->job_lock, flags);
        list_add_tail(&job->list, jobs);
-       spin_unlock_irqrestore(&_job_lock, flags);
+       spin_unlock_irqrestore(&kc->job_lock, flags);
 }
 
 /*
@@ -310,6 +307,7 @@ static int run_complete_job(struct kcopyd_job *job)
 static void complete_io(unsigned long error, void *context)
 {
        struct kcopyd_job *job = (struct kcopyd_job *) context;
+       struct dm_kcopyd_client *kc = job->kc;
 
        if (error) {
                if (job->rw == WRITE)
@@ -318,21 +316,21 @@ static void complete_io(unsigned long error, void *context)
                        job->read_err = 1;
 
                if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
-                       push(&_complete_jobs, job);
-                       wake();
+                       push(&kc->complete_jobs, job);
+                       wake(kc);
                        return;
                }
        }
 
        if (job->rw == WRITE)
-               push(&_complete_jobs, job);
+               push(&kc->complete_jobs, job);
 
        else {
                job->rw = WRITE;
-               push(&_io_jobs, job);
+               push(&kc->io_jobs, job);
        }
 
-       wake();
+       wake(kc);
 }
 
 /*
@@ -369,7 +367,7 @@ static int run_pages_job(struct kcopyd_job *job)
        r = kcopyd_get_pages(job->kc, job->nr_pages, &job->pages);
        if (!r) {
                /* this job is ready for io */
-               push(&_io_jobs, job);
+               push(&job->kc->io_jobs, job);
                return 0;
        }
 
@@ -384,12 +382,13 @@ static int run_pages_job(struct kcopyd_job *job)
  * Run through a list for as long as possible.  Returns the count
  * of successful jobs.
  */
-static int process_jobs(struct list_head *jobs, int (*fn) (struct kcopyd_job *))
+static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
+                       int (*fn) (struct kcopyd_job *))
 {
        struct kcopyd_job *job;
        int r, count = 0;
 
-       while ((job = pop(jobs))) {
+       while ((job = pop(jobs, kc))) {
 
                r = fn(job);
 
@@ -399,7 +398,7 @@ static int process_jobs(struct list_head *jobs, int (*fn) (struct kcopyd_job *))
                                job->write_err = (unsigned long) -1L;
                        else
                                job->read_err = 1;
-                       push(&_complete_jobs, job);
+                       push(&kc->complete_jobs, job);
                        break;
                }
 
@@ -421,8 +420,11 @@ static int process_jobs(struct list_head *jobs, int (*fn) (struct kcopyd_job *))
 /*
  * kcopyd does this every time it's woken up.
  */
-static void do_work(struct work_struct *ignored)
+static void do_work(struct work_struct *work)
 {
+       struct dm_kcopyd_client *kc = container_of(work,
+                                       struct dm_kcopyd_client, kcopyd_work);
+
        /*
         * The order that these are called is *very* important.
         * complete jobs can free some pages for pages jobs.
@@ -430,9 +432,9 @@ static void do_work(struct work_struct *ignored)
         * list.  io jobs call wake when they complete and it all
         * starts again.
         */
-       process_jobs(&_complete_jobs, run_complete_job);
-       process_jobs(&_pages_jobs, run_pages_job);
-       process_jobs(&_io_jobs, run_io_job);
+       process_jobs(&kc->complete_jobs, kc, run_complete_job);
+       process_jobs(&kc->pages_jobs, kc, run_pages_job);
+       process_jobs(&kc->io_jobs, kc, run_io_job);
 }
 
 /*
@@ -442,9 +444,10 @@ static void do_work(struct work_struct *ignored)
  */
 static void dispatch_job(struct kcopyd_job *job)
 {
-       atomic_inc(&job->kc->nr_jobs);
-       push(&_pages_jobs, job);
-       wake();
+       struct dm_kcopyd_client *kc = job->kc;
+       atomic_inc(&kc->nr_jobs);
+       push(&kc->pages_jobs, job);
+       wake(kc);
 }
 
 #define SUB_JOB_SIZE 128
@@ -625,15 +628,7 @@ static int kcopyd_init(void)
                return r;
        }
 
-       _kcopyd_wq = create_singlethread_workqueue("kcopyd");
-       if (!_kcopyd_wq) {
-               jobs_exit();
-               mutex_unlock(&kcopyd_init_lock);
-               return -ENOMEM;
-       }
-
        kcopyd_clients++;
-       INIT_WORK(&_kcopyd_work, do_work);
        mutex_unlock(&kcopyd_init_lock);
        return 0;
 }
@@ -644,8 +639,6 @@ static void kcopyd_exit(void)
        kcopyd_clients--;
        if (!kcopyd_clients) {
                jobs_exit();
-               destroy_workqueue(_kcopyd_wq);
-               _kcopyd_wq = NULL;
        }
        mutex_unlock(&kcopyd_init_lock);
 }
@@ -662,15 +655,31 @@ int dm_kcopyd_client_create(unsigned int nr_pages,
 
        kc = kmalloc(sizeof(*kc), GFP_KERNEL);
        if (!kc) {
+               r = -ENOMEM;
                kcopyd_exit();
-               return -ENOMEM;
+               return r;
        }
 
        spin_lock_init(&kc->lock);
+       spin_lock_init(&kc->job_lock);
+       INIT_LIST_HEAD(&kc->complete_jobs);
+       INIT_LIST_HEAD(&kc->io_jobs);
+       INIT_LIST_HEAD(&kc->pages_jobs);
+
+       INIT_WORK(&kc->kcopyd_work, do_work);
+       kc->kcopyd_wq = create_singlethread_workqueue("kcopyd");
+       if (!kc->kcopyd_wq) {
+               r = -ENOMEM;
+               kfree(kc);
+               kcopyd_exit();
+               return r;
+       }
+
        kc->pages = NULL;
        kc->nr_pages = kc->nr_free_pages = 0;
        r = client_alloc_pages(kc, nr_pages);
        if (r) {
+               destroy_workqueue(kc->kcopyd_wq);
                kfree(kc);
                kcopyd_exit();
                return r;
@@ -680,6 +689,7 @@ int dm_kcopyd_client_create(unsigned int nr_pages,
        if (IS_ERR(kc->io_client)) {
                r = PTR_ERR(kc->io_client);
                client_free_pages(kc);
+               destroy_workqueue(kc->kcopyd_wq);
                kfree(kc);
                kcopyd_exit();
                return r;
@@ -699,6 +709,10 @@ void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
        /* Wait for completion of all jobs submitted by this client. */
        wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
 
+       BUG_ON(!list_empty(&kc->complete_jobs));
+       BUG_ON(!list_empty(&kc->io_jobs));
+       BUG_ON(!list_empty(&kc->pages_jobs));
+       destroy_workqueue(kc->kcopyd_wq);
        dm_io_client_destroy(kc->io_client);
        client_free_pages(kc);
        client_del(kc);