s390/pci_dma: improve lazy flush for unmap
Lazy unmap (defer tlb flush after unmap until dma address reuse) can
greatly reduce the number of RPCIT instructions in the best case. In
reality we are often far away from the best case scenario because our
implementation suffers from the following problem:
To create dma addresses we maintain an iommu bitmap and a pointer into
that bitmap to mark the start of the next search. That pointer moves from
the start to the end of that bitmap and we issue a global tlb flush
once that pointer wraps around. To prevent address reuse before we issue
the tlb flush we even have to move the next pointer during unmaps - when
clearing a bit > next. This could lead to a situation where we only use
the rear part of that bitmap and issue more tlb flushes than expected.
To fix this we no longer clear bits during unmap but maintain a 2nd
bitmap which we use to mark addresses that can't be reused until we issue
the global tlb flush after wrap around.
Signed-off-by: Sebastian Ott <[email protected]>
Reviewed-by: Gerald Schaefer <[email protected]>
Signed-off-by: Martin Schwidefsky <[email protected]>
diff --git a/arch/s390/pci/pci_dma.c b/arch/s390/pci/pci_dma.c
index 9e5f2ec..7350c8b 100644
--- a/arch/s390/pci/pci_dma.c
+++ b/arch/s390/pci/pci_dma.c
@@ -257,20 +257,28 @@
spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
offset = __dma_alloc_iommu(dev, zdev->next_bit, size);
if (offset == -1) {
+ if (!zdev->tlb_refresh && !s390_iommu_strict) {
+ /* global flush before DMA addresses are reused */
+ if (zpci_refresh_global(zdev))
+ goto out_error;
+
+ bitmap_andnot(zdev->iommu_bitmap, zdev->iommu_bitmap,
+ zdev->lazy_bitmap, zdev->iommu_pages);
+ bitmap_zero(zdev->lazy_bitmap, zdev->iommu_pages);
+ }
/* wrap-around */
offset = __dma_alloc_iommu(dev, 0, size);
- if (offset == -1) {
- spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
- return DMA_ERROR_CODE;
- }
- if (!zdev->tlb_refresh && !s390_iommu_strict)
- /* global flush after wrap-around with lazy unmap */
- zpci_refresh_global(zdev);
+ if (offset == -1)
+ goto out_error;
}
zdev->next_bit = offset + size;
spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
return zdev->start_dma + offset * PAGE_SIZE;
+
+out_error:
+ spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
+ return DMA_ERROR_CODE;
}
static void dma_free_address(struct device *dev, dma_addr_t dma_addr, int size)
@@ -283,13 +291,12 @@
spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
if (!zdev->iommu_bitmap)
goto out;
- bitmap_clear(zdev->iommu_bitmap, offset, size);
- /*
- * Lazy flush for unmap: need to move next_bit to avoid address re-use
- * until wrap-around.
- */
- if (!s390_iommu_strict && offset >= zdev->next_bit)
- zdev->next_bit = offset + size;
+
+ if (zdev->tlb_refresh || s390_iommu_strict)
+ bitmap_clear(zdev->iommu_bitmap, offset, size);
+ else
+ bitmap_set(zdev->lazy_bitmap, offset, size);
+
out:
spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
}
@@ -557,7 +564,14 @@
rc = -ENOMEM;
goto free_dma_table;
}
+ if (!zdev->tlb_refresh && !s390_iommu_strict) {
+ zdev->lazy_bitmap = vzalloc(zdev->iommu_pages / 8);
+ if (!zdev->lazy_bitmap) {
+ rc = -ENOMEM;
+ goto free_bitmap;
+ }
+ }
rc = zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma,
(u64) zdev->dma_table);
if (rc)
@@ -567,6 +581,8 @@
free_bitmap:
vfree(zdev->iommu_bitmap);
zdev->iommu_bitmap = NULL;
+ vfree(zdev->lazy_bitmap);
+ zdev->lazy_bitmap = NULL;
free_dma_table:
dma_free_cpu_table(zdev->dma_table);
zdev->dma_table = NULL;
@@ -588,6 +604,9 @@
zdev->dma_table = NULL;
vfree(zdev->iommu_bitmap);
zdev->iommu_bitmap = NULL;
+ vfree(zdev->lazy_bitmap);
+ zdev->lazy_bitmap = NULL;
+
zdev->next_bit = 0;
}
