When devices with different DMA masks are using the same domain, or for PCI devices where we usually try a speculative 32-bit allocation first, there is a fair possibility that the top PFN of the rcache stack at any given time may be unsuitable for the lower limit, prompting a fallback to allocating anew from the rbtree. Consequently, we may end up artifically increasing pressure on the 32-bit IOVA space as unused IOVAs accumulate lower down in the rcache stacks, while callers with 32-bit masks also impose unnecessary rbtree overhead.
In such cases, let's try a bit harder to satisfy the allocation locally first - scanning the whole stack should still be relatively inexpensive, and even rotating an entry up from the very bottom probably has less overall impact than going to the rbtree. Signed-off-by: Robin Murphy <[email protected]> --- drivers/iommu/iova.c | 19 ++++++++++++++++--- 1 file changed, 16 insertions(+), 3 deletions(-) diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c index 8f8b436afd81..a7af8273fa98 100644 --- a/drivers/iommu/iova.c +++ b/drivers/iommu/iova.c @@ -826,12 +826,25 @@ static bool iova_magazine_empty(struct iova_magazine *mag) static unsigned long iova_magazine_pop(struct iova_magazine *mag, unsigned long limit_pfn) { + int i; + unsigned long pfn; + BUG_ON(iova_magazine_empty(mag)); - if (mag->pfns[mag->size - 1] > limit_pfn) - return 0; + /* + * If we can pull a suitable pfn from anywhere in the stack, that's + * still probably preferable to falling back to the rbtree. + */ + for (i = mag->size - 1; mag->pfns[i] > limit_pfn; i--) + if (i == 0) + return 0; - return mag->pfns[--mag->size]; + pfn = mag->pfns[i]; + mag->size--; + for (; i < mag->size; i++) + mag->pfns[i] = mag->pfns[i + 1]; + + return pfn; } static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn) -- 2.13.4.dirty
