iommu/amd: Rip out old queue flushing code
authorJoerg Roedel <jroedel@suse.de>
Fri, 2 Jun 2017 12:28:01 +0000 (14:28 +0200)
committerJoerg Roedel <jroedel@suse.de>
Thu, 8 Jun 2017 12:38:57 +0000 (14:38 +0200)
The queue flushing is pretty inefficient when it flushes the
queues for all cpus at once. Further it flushes all domains
from all IOMMUs for all CPUs, which is overkill as well.

Rip it out to make room for something more efficient.

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

index 1efbef7f3b61a2fca4347a20f6ad6cdb15e28c66..ec9da25c06a14d471ab3b9c84f841b246d62dbca 100644 (file)
@@ -89,25 +89,6 @@ LIST_HEAD(ioapic_map);
 LIST_HEAD(hpet_map);
 LIST_HEAD(acpihid_map);
 
-#define FLUSH_QUEUE_SIZE 256
-
-struct flush_queue_entry {
-       unsigned long iova_pfn;
-       unsigned long pages;
-       struct dma_ops_domain *dma_dom;
-};
-
-struct flush_queue {
-       spinlock_t lock;
-       unsigned next;
-       struct flush_queue_entry *entries;
-};
-
-static DEFINE_PER_CPU(struct flush_queue, flush_queue);
-
-static atomic_t queue_timer_on;
-static struct timer_list queue_timer;
-
 /*
  * Domain for untranslated devices - only allocated
  * if iommu=pt passed on kernel cmd line.
@@ -2253,92 +2234,6 @@ static struct iommu_group *amd_iommu_device_group(struct device *dev)
  *
  *****************************************************************************/
 
-static void __queue_flush(struct flush_queue *queue)
-{
-       struct protection_domain *domain;
-       unsigned long flags;
-       int idx;
-
-       /* First flush TLB of all known domains */
-       spin_lock_irqsave(&amd_iommu_pd_lock, flags);
-       list_for_each_entry(domain, &amd_iommu_pd_list, list)
-               domain_flush_tlb(domain);
-       spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
-
-       /* Wait until flushes have completed */
-       domain_flush_complete(NULL);
-
-       for (idx = 0; idx < queue->next; ++idx) {
-               struct flush_queue_entry *entry;
-
-               entry = queue->entries + idx;
-
-               free_iova_fast(&entry->dma_dom->iovad,
-                               entry->iova_pfn,
-                               entry->pages);
-
-               /* Not really necessary, just to make sure we catch any bugs */
-               entry->dma_dom = NULL;
-       }
-
-       queue->next = 0;
-}
-
-static void queue_flush_all(void)
-{
-       int cpu;
-
-       for_each_possible_cpu(cpu) {
-               struct flush_queue *queue;
-               unsigned long flags;
-
-               queue = per_cpu_ptr(&flush_queue, cpu);
-               spin_lock_irqsave(&queue->lock, flags);
-               if (queue->next > 0)
-                       __queue_flush(queue);
-               spin_unlock_irqrestore(&queue->lock, flags);
-       }
-}
-
-static void queue_flush_timeout(unsigned long unsused)
-{
-       atomic_set(&queue_timer_on, 0);
-       queue_flush_all();
-}
-
-static void queue_add(struct dma_ops_domain *dma_dom,
-                     unsigned long address, unsigned long pages)
-{
-       struct flush_queue_entry *entry;
-       struct flush_queue *queue;
-       unsigned long flags;
-       int idx;
-
-       pages     = __roundup_pow_of_two(pages);
-       address >>= PAGE_SHIFT;
-
-       queue = get_cpu_ptr(&flush_queue);
-       spin_lock_irqsave(&queue->lock, flags);
-
-       if (queue->next == FLUSH_QUEUE_SIZE)
-               __queue_flush(queue);
-
-       idx   = queue->next++;
-       entry = queue->entries + idx;
-
-       entry->iova_pfn = address;
-       entry->pages    = pages;
-       entry->dma_dom  = dma_dom;
-
-       spin_unlock_irqrestore(&queue->lock, flags);
-
-       if (atomic_cmpxchg(&queue_timer_on, 0, 1) == 0)
-               mod_timer(&queue_timer, jiffies + msecs_to_jiffies(10));
-
-       put_cpu_ptr(&flush_queue);
-}
-
-
 /*
  * In the dma_ops path we only have the struct device. This function
  * finds the corresponding IOMMU, the protection domain and the
@@ -2490,7 +2385,10 @@ 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);
+               /* Keep the if() around, we need it later again */
+               dma_ops_free_iova(dma_dom, dma_addr, pages);
+               domain_flush_tlb(&dma_dom->domain);
+               domain_flush_complete(&dma_dom->domain);
        }
 }
 
@@ -2825,7 +2723,7 @@ static int init_reserved_iova_ranges(void)
 
 int __init amd_iommu_init_api(void)
 {
-       int ret, cpu, err = 0;
+       int ret, err = 0;
 
        ret = iova_cache_get();
        if (ret)
@@ -2835,18 +2733,6 @@ int __init amd_iommu_init_api(void)
        if (ret)
                return ret;
 
-       for_each_possible_cpu(cpu) {
-               struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
-
-               queue->entries = kzalloc(FLUSH_QUEUE_SIZE *
-                                        sizeof(*queue->entries),
-                                        GFP_KERNEL);
-               if (!queue->entries)
-                       goto out_put_iova;
-
-               spin_lock_init(&queue->lock);
-       }
-
        err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
        if (err)
                return err;
@@ -2858,23 +2744,12 @@ int __init amd_iommu_init_api(void)
        err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
        if (err)
                return err;
-       return 0;
-
-out_put_iova:
-       for_each_possible_cpu(cpu) {
-               struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
 
-               kfree(queue->entries);
-       }
-
-       return -ENOMEM;
+       return 0;
 }
 
 int __init amd_iommu_init_dma_ops(void)
 {
-       setup_timer(&queue_timer, queue_flush_timeout, 0);
-       atomic_set(&queue_timer_on, 0);
-
        swiotlb        = iommu_pass_through ? 1 : 0;
        iommu_detected = 1;
 
@@ -3030,12 +2905,6 @@ static void amd_iommu_domain_free(struct iommu_domain *dom)
 
        switch (dom->type) {
        case IOMMU_DOMAIN_DMA:
-               /*
-                * First make sure the domain is no longer referenced from the
-                * flush queue
-                */
-               queue_flush_all();
-
                /* Now release the domain */
                dma_dom = to_dma_ops_domain(domain);
                dma_ops_domain_free(dma_dom);