[PATCH] iommu/iova: Remove size-alignment for large allocations

[Chris Wilson]

fi-cml-drallion is configured with highly restricted PCI windows:

<6>[    1.773230] PCI host bridge to bus 0000:00
<6>[    1.773276] pci_bus 0000:00: root bus resource [io  0x0000-0x0cf7 window]
<6>[    1.773335] pci_bus 0000:00: root bus resource [io  0x0d00-0xffff window]
<6>[    1.773394] pci_bus 0000:00: root bus resource [mem 0x000a0000-0x000bffff window]
<6>[    1.773457] pci_bus 0000:00: root bus resource [mem 0x000c0000-0x000c3fff window]
<6>[    1.773520] pci_bus 0000:00: root bus resource [mem 0x000c4000-0x000c7fff window]
<6>[    1.773583] pci_bus 0000:00: root bus resource [mem 0x000c8000-0x000cbfff window]
<6>[    1.773645] pci_bus 0000:00: root bus resource [mem 0x000cc000-0x000cffff window]
<6>[    1.773707] pci_bus 0000:00: root bus resource [mem 0x000d0000-0x000d3fff window]
<6>[    1.773770] pci_bus 0000:00: root bus resource [mem 0x000d4000-0x000d7fff window]
<6>[    1.773832] pci_bus 0000:00: root bus resource [mem 0x000d8000-0x000dbfff window]
<6>[    1.773894] pci_bus 0000:00: root bus resource [mem 0x000dc000-0x000dffff window]
<6>[    1.773956] pci_bus 0000:00: root bus resource [mem 0x000e0000-0x000e3fff window]
<6>[    1.774014] pci_bus 0000:00: root bus resource [mem 0x000e4000-0x000e7fff window]
<6>[    1.774076] pci_bus 0000:00: root bus resource [mem 0x000e8000-0x000ebfff window]
<6>[    1.774138] pci_bus 0000:00: root bus resource [mem 0x000ec000-0x000effff window]
<6>[    1.774200] pci_bus 0000:00: root bus resource [mem 0x000f0000-0x000fffff window]
<6>[    1.774267] pci_bus 0000:00: root bus resource [mem 0x9f800000-0xdfffffff window]
<6>[    1.774330] pci_bus 0000:00: root bus resource [mem 0x25e800000-0x7fffffffff window]
<6>[    1.774395] pci_bus 0000:00: root bus resource [mem 0xfc800000-0xfe7fffff window]
<6>[    1.774458] pci_bus 0000:00: root bus resource [mem 0xfed40000-0xfed47fff window]
<6>[    1.774521] pci_bus 0000:00: root bus resource [bus 00-ff]

that limits the usable DMA address space to <10G. Combined with the
fragmentation from top-down allocations starting at 4G, an attempt to
allocate a size-aligned (SZ_4G+SZ_4K) request fails, despite having a 6G
hole. The size-alignment is only required for the rcache, so remove the
restriction for large allocations.

It would also be useful for the limited DMA32 [SAC] allocation requests
to be allocated bottom-up with the extra large allocations going top
down.

Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/2929
---
 drivers/iommu/dma-iommu.c |  8 --------
 drivers/iommu/iova.c      | 31 ++++++++++++++++++++++++++-----
 2 files changed, 26 insertions(+), 13 deletions(-)

diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index 4078358ed66e..34353d8a785d 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -436,14 +436,6 @@ static dma_addr_t iommu_dma_alloc_iova(struct iommu_domain *domain,
 
 	shift = iova_shift(iovad);
 	iova_len = size >> shift;
-	/*
-	 * Freeing non-power-of-two-sized allocations back into the IOVA caches
-	 * will come back to bite us badly, so we have to waste a bit of space
-	 * rounding up anything cacheable to make sure that can't happen. The
-	 * order of the unadjusted size will still match upon freeing.
-	 */
-	if (iova_len < (1 << (IOVA_RANGE_CACHE_MAX_SIZE - 1)))
-		iova_len = roundup_pow_of_two(iova_len);
 
 	dma_limit = min_not_zero(dma_limit, dev->bus_dma_limit);
 
diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
index d20b8b333d30..e2b2d9825ba1 100644
--- a/drivers/iommu/iova.c
+++ b/drivers/iommu/iova.c
@@ -426,15 +426,27 @@ unsigned long
 alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
 		unsigned long limit_pfn, bool flush_rcache)
 {
-	unsigned long iova_pfn;
+	unsigned long iova_pfn, iova_len;
 	struct iova *new_iova;
+	bool align = false;
 
-	iova_pfn = iova_rcache_get(iovad, size, limit_pfn + 1);
-	if (iova_pfn)
-		return iova_pfn;
+	/*
+	 * Freeing non-power-of-two-sized allocations back into the IOVA caches
+	 * will come back to bite us badly, so we have to waste a bit of space
+	 * rounding up anything cacheable to make sure that can't happen. The
+	 * order of the unadjusted size will still match upon freeing.
+	 */
+	if (size < (1 << (IOVA_RANGE_CACHE_MAX_SIZE - 1))) {
+		iova_len = roundup_pow_of_two(size);
+		iova_pfn = iova_rcache_get(iovad, iova_len, limit_pfn + 1);
+		if (iova_pfn)
+			return iova_pfn;
+
+		align = true;
+	}
 
 retry:
-	new_iova = alloc_iova(iovad, size, limit_pfn, true);
+	new_iova = alloc_iova(iovad, iova_len, limit_pfn, align);
 	if (!new_iova) {
 		unsigned int cpu;
 
@@ -446,6 +458,9 @@ alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
 		for_each_online_cpu(cpu)
 			free_cpu_cached_iovas(cpu, iovad);
 		free_global_cached_iovas(iovad);
+
+		align = false;
+		iova_len = size;
 		goto retry;
 	}
 
@@ -918,6 +933,12 @@ static bool iova_rcache_insert(struct iova_domain *iovad, unsigned long pfn,
 	if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
 		return false;
 
+	if (!is_power_of_2(size))
+		return false;
+
+	if (!IS_ALIGNED(pfn, size))
+		return false;
+
 	return __iova_rcache_insert(iovad, &iovad->rcaches[log_size], pfn);
 }
 
-- 
2.20.1