[PATCH -mm 0/4] fix iommu segment boundary problems

[PATCH -mm 0/4] fix iommu segment boundary problems

[FUJITA Tomonori]

This is the latter half of my iommu work to make the IOMMUs respect
LLDs restrictions.

IOMMUs allocate memory areas without considering a low level driver's
segment boundary limits. So we have some workarounds: splitting sg
segments again in LLDs; reserving all I/O space spanning 4GB boundary
in IOMMUs (with assumption that all the LLDs have 4GB boundary
restrictions). The goal is killing all the workarounds.

This patchset adds new accessors for segment_boundary_mask in
device_dma_parameters structure in the same way as the first half of
my work did for max_segment_size.

Currently, I fixed only swiotlb. Next, I'll generalize swiotlb's free
area management and convert all the IOMMUs to use it. Or I'll
generalize a free area management to use bitmap that most of the
IOMMUs use and convert them to use it.

This is against 2.6.24-rc3-mm1.

The first half of my iommu work is:

http://thread.gmane.org/gmane.linux.scsi/35602

[PATCH -mm 1/4] add accessors for segment_boundary_mask in device_dma_parameters

[FUJITA Tomonori]

This adds new accessors for segment_boundary_mask in
device_dma_parameters structure in the same way I did for
max_segment_size. So we can easily change where to place struct
device_dma_parameters in the future.

dma_get_segment boundary returns 0xffffffff if dma_parms in struct
device isn't set up properly. 0xffffffff is the default value used in
the block layer and the scsi mid layer.

Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
---
 include/linux/dma-mapping.h |   15 +++++++++++++++
 1 files changed, 15 insertions(+), 0 deletions(-)

diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
index 71972ca..7d157ed 100644
--- a/include/linux/dma-mapping.h
+++ b/include/linux/dma-mapping.h
@@ -75,6 +75,21 @@ static inline unsigned int dma_set_max_seg_size(struct device *dev,
 		return -EIO;
 }
 
+static inline unsigned long dma_get_seg_boundary(struct device *dev)
+{
+	return dev->dma_parms ?
+		dev->dma_parms->segment_boundary_mask : 0xffffffff;
+}
+
+static inline int dma_set_seg_boundary(struct device *dev, unsigned long mask)
+{
+	if (dev->dma_parms) {
+		dev->dma_parms->segment_boundary_mask = mask;
+		return 0;
+	} else
+		return -EIO;
+}
+
 /* flags for the coherent memory api */
 #define	DMA_MEMORY_MAP			0x01
 #define DMA_MEMORY_IO			0x02
-- 
1.5.3.4

[PATCH -mm 2/4] PCI: add dma segment boundary support

[FUJITA Tomonori]

This adds PCI's accessor for segment_boundary_mask in
device_dma_parameters.

The default segment_boundary is set to 0xffffffff, same to the block
layer's default value (and the scsi mid layer uses the same value).

Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
---
 drivers/pci/pci.c   |    8 ++++++++
 drivers/pci/probe.c |    1 +
 include/linux/pci.h |    2 ++
 3 files changed, 11 insertions(+), 0 deletions(-)

diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index de623cf..3b7e0e0 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -1435,6 +1435,14 @@ int pci_set_dma_max_seg_size(struct pci_dev *dev, unsigned int size)
 EXPORT_SYMBOL(pci_set_dma_max_seg_size);
 #endif
 
+#ifndef HAVE_ARCH_PCI_SET_DMA_SEGMENT_BOUNDARY
+int pci_set_dma_seg_boundary(struct pci_dev *dev, unsigned long mask)
+{
+	return dma_set_seg_boundary(&dev->dev, mask);
+}
+EXPORT_SYMBOL(pci_set_dma_seg_boundary);
+#endif
+
 /**
  * pcix_get_max_mmrbc - get PCI-X maximum designed memory read byte count
  * @dev: PCI device to query
diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
index aa343e1..2e8b539 100644
--- a/drivers/pci/probe.c
+++ b/drivers/pci/probe.c
@@ -987,6 +987,7 @@ void pci_device_add(struct pci_dev *dev, struct pci_bus *bus)
 	dev->dev.coherent_dma_mask = 0xffffffffull;
 
 	pci_set_dma_max_seg_size(dev, 65536);
+	pci_set_dma_seg_boundary(dev, 0xffffffff);
 
 	/* Fix up broken headers */
 	pci_fixup_device(pci_fixup_header, dev);
diff --git a/include/linux/pci.h b/include/linux/pci.h
index d56d0b6..a05a843 100644
--- a/include/linux/pci.h
+++ b/include/linux/pci.h
@@ -567,6 +567,7 @@ void pci_msi_off(struct pci_dev *dev);
 int pci_set_dma_mask(struct pci_dev *dev, u64 mask);
 int pci_set_consistent_dma_mask(struct pci_dev *dev, u64 mask);
 int pci_set_dma_max_seg_size(struct pci_dev *dev, unsigned int size);
+int pci_set_dma_seg_boundary(struct pci_dev *dev, unsigned long mask);
 int pcix_get_max_mmrbc(struct pci_dev *dev);
 int pcix_get_mmrbc(struct pci_dev *dev);
 int pcix_set_mmrbc(struct pci_dev *dev, int mmrbc);
@@ -753,6 +754,7 @@ static inline int pci_enable_device(struct pci_dev *dev) { return -EIO; }
 static inline void pci_disable_device(struct pci_dev *dev) { }
 static inline int pci_set_dma_mask(struct pci_dev *dev, u64 mask) { return -EIO; }
 static inline int pci_set_dma_max_seg_size(struct pci_dev *dev, unsigned int size) { return -EIO; }
+static inline int pci_set_dma_seg_boundary(struct pci_dev *dev, unsigned long mask) { return -EIO; }
 static inline int pci_assign_resource(struct pci_dev *dev, int i) { return -EBUSY;}
 static inline int __pci_register_driver(struct pci_driver *drv, struct module *owner) { return 0;}
 static inline int pci_register_driver(struct pci_driver *drv) { return 0;}
-- 
1.5.3.4

[PATCH -mm 3/4] swiotlb: respect the segment boundary limits

[FUJITA Tomonori]

This patch makes swiotlb not allocate a memory area spanning LLD's
segment boundary.

is_span_boundary() judges whether a memory area spans LLD's segment
boundary. If map_single finds such a area, map_single tries to find
the next available memory area.

Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
---
 lib/swiotlb.c |   41 +++++++++++++++++++++++++++++++++++------
 1 files changed, 35 insertions(+), 6 deletions(-)

diff --git a/lib/swiotlb.c b/lib/swiotlb.c
index 1a8050a..4bb5a11 100644
--- a/lib/swiotlb.c
+++ b/lib/swiotlb.c
@@ -282,6 +282,15 @@ address_needs_mapping(struct device *hwdev, dma_addr_t addr)
 	return (addr & ~mask) != 0;
 }
 
+static inline unsigned int is_span_boundary(unsigned int index,
+					    unsigned int nslots,
+					    unsigned long offset_slots,
+					    unsigned long max_slots)
+{
+	unsigned long offset = (offset_slots + index) & (max_slots - 1);
+	return offset + nslots > max_slots;
+}
+
 /*
  * Allocates bounce buffer and returns its kernel virtual address.
  */
@@ -292,6 +301,16 @@ map_single(struct device *hwdev, char *buffer, size_t size, int dir)
 	char *dma_addr;
 	unsigned int nslots, stride, index, wrap;
 	int i;
+	unsigned long start_dma_addr;
+	unsigned long mask;
+	unsigned long offset_slots;
+	unsigned long max_slots;
+
+	mask = dma_get_seg_boundary(hwdev);
+	start_dma_addr = virt_to_bus(io_tlb_start) & mask;
+
+	offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
+	max_slots = ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
 
 	/*
 	 * For mappings greater than a page, we limit the stride (and
@@ -311,10 +330,17 @@ map_single(struct device *hwdev, char *buffer, size_t size, int dir)
 	 */
 	spin_lock_irqsave(&io_tlb_lock, flags);
 	{
-		wrap = index = ALIGN(io_tlb_index, stride);
-
+		index = ALIGN(io_tlb_index, stride);
 		if (index >= io_tlb_nslabs)
-			wrap = index = 0;
+			index = 0;
+
+		while (is_span_boundary(index, nslots, offset_slots,
+					max_slots)) {
+			index += stride;
+			if (index >= io_tlb_nslabs)
+				index = 0;
+		}
+		wrap = index;
 
 		do {
 			/*
@@ -341,9 +367,12 @@ map_single(struct device *hwdev, char *buffer, size_t size, int dir)
 
 				goto found;
 			}
-			index += stride;
-			if (index >= io_tlb_nslabs)
-				index = 0;
+			do {
+				index += stride;
+				if (index >= io_tlb_nslabs)
+					index = 0;
+			} while (is_span_boundary(index, nslots, offset_slots,
+						  max_slots));
 		} while (index != wrap);
 
 		spin_unlock_irqrestore(&io_tlb_lock, flags);
-- 
1.5.3.4

[PATCH -mm 4/4] call dma_set_seg_boundary in __scsi_alloc_queue

[FUJITA Tomonori]

This is a one-line patch to add the following to __scsi_alloc_queue():

dma_set_seg_boundary(dev, shost->dma_boundary);

This is the simplest approach but the result looks odd,
__scsi_alloc_queue() does:

blk_queue_segment_boundary(q, shost->dma_boundary);
dma_set_seg_boundary(dev, shost->dma_boundary);
blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));

I think that it would be better to set up segment boundary in the same
way as we did for the maximum segment size. That is, removing
shost->dma_boundary and LLDs call pci_set_dma_seg_boundary (or its
friends).

Then __scsi_alloc_queue() can set up both limits in the same way:

blk_queue_segment_boundary(q, dma_get_seg_boundary(dev));
blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));

killing dma_boundary in scsi_host_template needs a large patch for
libata (dma_boundary is used by only libata and sym53c8xx). I'll send
a patch to do that if it is acceptable. James and Jeff?

Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
---
 drivers/scsi/scsi_lib.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index 733176d..2a15a3b 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -1767,6 +1767,7 @@ struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
 	blk_queue_max_sectors(q, shost->max_sectors);
 	blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
 	blk_queue_segment_boundary(q, shost->dma_boundary);
+	dma_set_seg_boundary(dev, shost->dma_boundary);
 
 	blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
 
-- 
1.5.3.4