]> git.openfabrics.org - ~emulex/infiniband.git/commitdiff
powerpc/iommu: Reduce spinlock coverage in iommu_free
authorAnton Blanchard <anton@samba.org>
Sun, 3 Jun 2012 19:43:44 +0000 (19:43 +0000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Tue, 3 Jul 2012 04:14:47 +0000 (14:14 +1000)
This patch moves tce_free outside of the lock in iommu_free.

Some performance numbers were obtained with a Chelsio T3 adapter on
two POWER7 boxes, running a 100 session TCP round robin test.

Performance improved 25% with this patch applied.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/kernel/iommu.c

index 9c8967fa1e63474cf6d0aa905a36f945d5c389ab..d855cfc0732d5746cc5860dd1b80bdb0382c1816 100644 (file)
@@ -190,10 +190,7 @@ static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
         * not altered.
         */
        if (unlikely(build_fail)) {
-               spin_lock_irqsave(&(tbl->it_lock), flags);
                __iommu_free(tbl, ret, npages);
-               spin_unlock_irqrestore(&(tbl->it_lock), flags);
-
                return DMA_ERROR_CODE;
        }
 
@@ -207,8 +204,8 @@ static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
        return ret;
 }
 
-static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr, 
-                        unsigned int npages)
+static bool iommu_free_check(struct iommu_table *tbl, dma_addr_t dma_addr,
+                            unsigned int npages)
 {
        unsigned long entry, free_entry;
 
@@ -228,21 +225,53 @@ static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
                        printk(KERN_INFO "\tindex     = 0x%llx\n", (u64)tbl->it_index);
                        WARN_ON(1);
                }
-               return;
+
+               return false;
        }
 
+       return true;
+}
+
+static void __iommu_free_locked(struct iommu_table *tbl, dma_addr_t dma_addr,
+                        unsigned int npages)
+{
+       unsigned long entry, free_entry;
+
+       BUG_ON(!spin_is_locked(&tbl->it_lock));
+
+       entry = dma_addr >> IOMMU_PAGE_SHIFT;
+       free_entry = entry - tbl->it_offset;
+
+       if (!iommu_free_check(tbl, dma_addr, npages))
+               return;
+
        ppc_md.tce_free(tbl, entry, npages);
        bitmap_clear(tbl->it_map, free_entry, npages);
 }
 
-static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
-               unsigned int npages)
+static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
+                        unsigned int npages)
 {
+       unsigned long entry, free_entry;
        unsigned long flags;
 
+       entry = dma_addr >> IOMMU_PAGE_SHIFT;
+       free_entry = entry - tbl->it_offset;
+
+       if (!iommu_free_check(tbl, dma_addr, npages))
+               return;
+
+       ppc_md.tce_free(tbl, entry, npages);
+
        spin_lock_irqsave(&(tbl->it_lock), flags);
-       __iommu_free(tbl, dma_addr, npages);
+       bitmap_clear(tbl->it_map, free_entry, npages);
        spin_unlock_irqrestore(&(tbl->it_lock), flags);
+}
+
+static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
+               unsigned int npages)
+{
+       __iommu_free(tbl, dma_addr, npages);
 
        /* Make sure TLB cache is flushed if the HW needs it. We do
         * not do an mb() here on purpose, it is not needed on any of
@@ -390,7 +419,7 @@ int iommu_map_sg(struct device *dev, struct iommu_table *tbl,
                        vaddr = s->dma_address & IOMMU_PAGE_MASK;
                        npages = iommu_num_pages(s->dma_address, s->dma_length,
                                                 IOMMU_PAGE_SIZE);
-                       __iommu_free(tbl, vaddr, npages);
+                       __iommu_free_locked(tbl, vaddr, npages);
                        s->dma_address = DMA_ERROR_CODE;
                        s->dma_length = 0;
                }
@@ -425,7 +454,7 @@ void iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
                        break;
                npages = iommu_num_pages(dma_handle, sg->dma_length,
                                         IOMMU_PAGE_SIZE);
-               __iommu_free(tbl, dma_handle, npages);
+               __iommu_free_locked(tbl, dma_handle, npages);
                sg = sg_next(sg);
        }