We currently hold the IOMMU spinlock around tce_build and tce_flush.
This causes our spinlock hold times to be much higher than required
and can impact multiqueue adapters.
This patch moves tce_build and tce_flush outside of the lock in
iommu_alloc, and tce_flush 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 32% with this patch applied.
Signed-off-by: Anton Blanchard <an...@samba.org>
---
Index: linux-build/arch/powerpc/kernel/iommu.c
===================================================================
--- linux-build.orig/arch/powerpc/kernel/iommu.c 2012-06-04
10:38:38.045211977 +1000
+++ linux-build/arch/powerpc/kernel/iommu.c 2012-06-04 10:38:41.461284266
+1000
@@ -170,13 +170,11 @@ static dma_addr_t iommu_alloc(struct dev
int build_fail;
spin_lock_irqsave(&(tbl->it_lock), flags);
-
entry = iommu_range_alloc(dev, tbl, npages, NULL, mask, align_order);
+ spin_unlock_irqrestore(&(tbl->it_lock), flags);
- if (unlikely(entry == DMA_ERROR_CODE)) {
- spin_unlock_irqrestore(&(tbl->it_lock), flags);
+ if (unlikely(entry == DMA_ERROR_CODE))
return DMA_ERROR_CODE;
- }
entry += tbl->it_offset; /* Offset into real TCE table */
ret = entry << IOMMU_PAGE_SHIFT; /* Set the return dma address */
@@ -192,9 +190,10 @@ static dma_addr_t iommu_alloc(struct dev
* 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;
}
@@ -202,8 +201,6 @@ static dma_addr_t iommu_alloc(struct dev
if (ppc_md.tce_flush)
ppc_md.tce_flush(tbl);
- spin_unlock_irqrestore(&(tbl->it_lock), flags);
-
/* Make sure updates are seen by hardware */
mb();
@@ -244,8 +241,8 @@ static void iommu_free(struct iommu_tabl
unsigned long flags;
spin_lock_irqsave(&(tbl->it_lock), flags);
-
__iommu_free(tbl, dma_addr, npages);
+ spin_unlock_irqrestore(&(tbl->it_lock), flags);
/* 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
@@ -253,8 +250,6 @@ static void iommu_free(struct iommu_tabl
*/
if (ppc_md.tce_flush)
ppc_md.tce_flush(tbl);
-
- spin_unlock_irqrestore(&(tbl->it_lock), flags);
}
int iommu_map_sg(struct device *dev, struct iommu_table *tbl,
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev
