]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/commitdiff
iommu/amd: Make use of iova queue flushing
authorJoerg Roedel <jroedel@suse.de>
Thu, 10 Aug 2017 15:19:13 +0000 (17:19 +0200)
committerJoerg Roedel <jroedel@suse.de>
Tue, 15 Aug 2017 16:23:52 +0000 (18:23 +0200)
Rip out the implementation in the AMD IOMMU driver and use
the one in the common iova code instead.

Signed-off-by: Joerg Roedel <jroedel@suse.de>
drivers/iommu/amd_iommu.c

index 688e77576e5a50b3f2f137eec72cd721d6bf96fa..cabcaa506ed697d8b7166b774be930dd37497fbe 100644 (file)
@@ -137,20 +137,7 @@ struct kmem_cache *amd_iommu_irq_cache;
 static void update_domain(struct protection_domain *domain);
 static int protection_domain_init(struct protection_domain *domain);
 static void detach_device(struct device *dev);
-
-#define FLUSH_QUEUE_SIZE 256
-
-struct flush_queue_entry {
-       unsigned long iova_pfn;
-       unsigned long pages;
-       u64 counter; /* Flush counter when this entry was added to the queue */
-};
-
-struct flush_queue {
-       struct flush_queue_entry *entries;
-       unsigned head, tail;
-       spinlock_t lock;
-};
+static void iova_domain_flush_tlb(struct iova_domain *iovad);
 
 /*
  * Data container for a dma_ops specific protection domain
@@ -161,36 +148,6 @@ struct dma_ops_domain {
 
        /* IOVA RB-Tree */
        struct iova_domain iovad;
-
-       struct flush_queue __percpu *flush_queue;
-
-       /*
-        * We need two counter here to be race-free wrt. IOTLB flushing and
-        * adding entries to the flush queue.
-        *
-        * The flush_start_cnt is incremented _before_ the IOTLB flush starts.
-        * New entries added to the flush ring-buffer get their 'counter' value
-        * from here. This way we can make sure that entries added to the queue
-        * (or other per-cpu queues of the same domain) while the TLB is about
-        * to be flushed are not considered to be flushed already.
-        */
-       atomic64_t flush_start_cnt;
-
-       /*
-        * The flush_finish_cnt is incremented when an IOTLB flush is complete.
-        * This value is always smaller than flush_start_cnt. The queue_add
-        * function frees all IOVAs that have a counter value smaller than
-        * flush_finish_cnt. This makes sure that we only free IOVAs that are
-        * flushed out of the IOTLB of the domain.
-        */
-       atomic64_t flush_finish_cnt;
-
-       /*
-        * Timer to make sure we don't keep IOVAs around unflushed
-        * for too long
-        */
-       struct timer_list flush_timer;
-       atomic_t flush_timer_on;
 };
 
 static struct iova_domain reserved_iova_ranges;
@@ -1788,178 +1745,19 @@ static void free_gcr3_table(struct protection_domain *domain)
        free_page((unsigned long)domain->gcr3_tbl);
 }
 
-static void dma_ops_domain_free_flush_queue(struct dma_ops_domain *dom)
-{
-       int cpu;
-
-       for_each_possible_cpu(cpu) {
-               struct flush_queue *queue;
-
-               queue = per_cpu_ptr(dom->flush_queue, cpu);
-               kfree(queue->entries);
-       }
-
-       free_percpu(dom->flush_queue);
-
-       dom->flush_queue = NULL;
-}
-
-static int dma_ops_domain_alloc_flush_queue(struct dma_ops_domain *dom)
-{
-       int cpu;
-
-       atomic64_set(&dom->flush_start_cnt,  0);
-       atomic64_set(&dom->flush_finish_cnt, 0);
-
-       dom->flush_queue = alloc_percpu(struct flush_queue);
-       if (!dom->flush_queue)
-               return -ENOMEM;
-
-       /* First make sure everything is cleared */
-       for_each_possible_cpu(cpu) {
-               struct flush_queue *queue;
-
-               queue = per_cpu_ptr(dom->flush_queue, cpu);
-               queue->head    = 0;
-               queue->tail    = 0;
-               queue->entries = NULL;
-       }
-
-       /* Now start doing the allocation */
-       for_each_possible_cpu(cpu) {
-               struct flush_queue *queue;
-
-               queue = per_cpu_ptr(dom->flush_queue, cpu);
-               queue->entries = kzalloc(FLUSH_QUEUE_SIZE * sizeof(*queue->entries),
-                                        GFP_KERNEL);
-               if (!queue->entries) {
-                       dma_ops_domain_free_flush_queue(dom);
-                       return -ENOMEM;
-               }
-
-               spin_lock_init(&queue->lock);
-       }
-
-       return 0;
-}
-
 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
 {
-       atomic64_inc(&dom->flush_start_cnt);
        domain_flush_tlb(&dom->domain);
        domain_flush_complete(&dom->domain);
-       atomic64_inc(&dom->flush_finish_cnt);
 }
 
-static inline bool queue_ring_full(struct flush_queue *queue)
+static void iova_domain_flush_tlb(struct iova_domain *iovad)
 {
-       assert_spin_locked(&queue->lock);
-
-       return (((queue->tail + 1) % FLUSH_QUEUE_SIZE) == queue->head);
-}
+       struct dma_ops_domain *dom;
 
-#define queue_ring_for_each(i, q) \
-       for (i = (q)->head; i != (q)->tail; i = (i + 1) % FLUSH_QUEUE_SIZE)
-
-static inline unsigned queue_ring_add(struct flush_queue *queue)
-{
-       unsigned idx = queue->tail;
-
-       assert_spin_locked(&queue->lock);
-       queue->tail = (idx + 1) % FLUSH_QUEUE_SIZE;
-
-       return idx;
-}
-
-static inline void queue_ring_remove_head(struct flush_queue *queue)
-{
-       assert_spin_locked(&queue->lock);
-       queue->head = (queue->head + 1) % FLUSH_QUEUE_SIZE;
-}
-
-static void queue_ring_free_flushed(struct dma_ops_domain *dom,
-                                   struct flush_queue *queue)
-{
-       u64 counter = atomic64_read(&dom->flush_finish_cnt);
-       int idx;
-
-       queue_ring_for_each(idx, queue) {
-               /*
-                * This assumes that counter values in the ring-buffer are
-                * monotonously rising.
-                */
-               if (queue->entries[idx].counter >= counter)
-                       break;
-
-               free_iova_fast(&dom->iovad,
-                              queue->entries[idx].iova_pfn,
-                              queue->entries[idx].pages);
-
-               queue_ring_remove_head(queue);
-       }
-}
-
-static void queue_add(struct dma_ops_domain *dom,
-                     unsigned long address, unsigned long pages)
-{
-       struct flush_queue *queue;
-       unsigned long flags;
-       int idx;
-
-       pages     = __roundup_pow_of_two(pages);
-       address >>= PAGE_SHIFT;
-
-       queue = get_cpu_ptr(dom->flush_queue);
-       spin_lock_irqsave(&queue->lock, flags);
-
-       /*
-        * First remove the enries from the ring-buffer that are already
-        * flushed to make the below queue_ring_full() check less likely
-        */
-       queue_ring_free_flushed(dom, queue);
-
-       /*
-        * When ring-queue is full, flush the entries from the IOTLB so
-        * that we can free all entries with queue_ring_free_flushed()
-        * below.
-        */
-       if (queue_ring_full(queue)) {
-               dma_ops_domain_flush_tlb(dom);
-               queue_ring_free_flushed(dom, queue);
-       }
-
-       idx = queue_ring_add(queue);
-
-       queue->entries[idx].iova_pfn = address;
-       queue->entries[idx].pages    = pages;
-       queue->entries[idx].counter  = atomic64_read(&dom->flush_start_cnt);
-
-       spin_unlock_irqrestore(&queue->lock, flags);
-
-       if (atomic_cmpxchg(&dom->flush_timer_on, 0, 1) == 0)
-               mod_timer(&dom->flush_timer, jiffies + msecs_to_jiffies(10));
-
-       put_cpu_ptr(dom->flush_queue);
-}
-
-static void queue_flush_timeout(unsigned long data)
-{
-       struct dma_ops_domain *dom = (struct dma_ops_domain *)data;
-       int cpu;
-
-       atomic_set(&dom->flush_timer_on, 0);
+       dom = container_of(iovad, struct dma_ops_domain, iovad);
 
        dma_ops_domain_flush_tlb(dom);
-
-       for_each_possible_cpu(cpu) {
-               struct flush_queue *queue;
-               unsigned long flags;
-
-               queue = per_cpu_ptr(dom->flush_queue, cpu);
-               spin_lock_irqsave(&queue->lock, flags);
-               queue_ring_free_flushed(dom, queue);
-               spin_unlock_irqrestore(&queue->lock, flags);
-       }
 }
 
 /*
@@ -1973,11 +1771,6 @@ static void dma_ops_domain_free(struct dma_ops_domain *dom)
 
        del_domain_from_list(&dom->domain);
 
-       if (timer_pending(&dom->flush_timer))
-               del_timer(&dom->flush_timer);
-
-       dma_ops_domain_free_flush_queue(dom);
-
        put_iova_domain(&dom->iovad);
 
        free_pagetable(&dom->domain);
@@ -2013,16 +1806,11 @@ static struct dma_ops_domain *dma_ops_domain_alloc(void)
        init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
                         IOVA_START_PFN, DMA_32BIT_PFN);
 
-       /* Initialize reserved ranges */
-       copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
-
-       if (dma_ops_domain_alloc_flush_queue(dma_dom))
+       if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
                goto free_dma_dom;
 
-       setup_timer(&dma_dom->flush_timer, queue_flush_timeout,
-                   (unsigned long)dma_dom);
-
-       atomic_set(&dma_dom->flush_timer_on, 0);
+       /* Initialize reserved ranges */
+       copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
 
        add_domain_to_list(&dma_dom->domain);
 
@@ -2619,7 +2407,8 @@ static void __unmap_single(struct dma_ops_domain *dma_dom,
                domain_flush_tlb(&dma_dom->domain);
                domain_flush_complete(&dma_dom->domain);
        } else {
-               queue_add(dma_dom, dma_addr, pages);
+               pages = __roundup_pow_of_two(pages);
+               queue_iova(&dma_dom->iovad, dma_addr >> PAGE_SHIFT, pages, 0);
        }
 }