bus: mvebu-mbus: important fixes and improvements

From: Nicolas Schichan <[email protected]>

Commit a0b5cd4ac2d6 ("bus: mvebu-mbus: use automatic I/O synchronization barriers") enabled the usage of automatic I/O synchronization barriers by enabling bit WIN_CTRL_SYNCBARRIER in the control registers of MBus windows, but on non io-coherent platforms (orion5x, kirkwood and dove) the WIN_CTRL_SYNCBARRIER bit in the window control register is either reserved (all windows except 6 and 7) or enables read-only protection (windows 6 and 7).

Signed-off-by Nicolas Schichan <[email protected]>

Reviewed-by Thomas Petazzoni <[email protected]>

Cc <[email protected]> # v4.0+

Fixes a0b5cd4ac2d6 ("bus: mvebu-mbus: use automatic I/O synchronization barriers")

Signed-off-by Thomas Petazzoni <[email protected]>

---
 drivers/bus/mvebu-mbus.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

drivers/bus/mvebu-mbus.c +3 -1

index fb9ec62..7fa4510 100644
@@ -70,6 +70,7 @@
  */
 #define WIN_CTRL_OFF		0x0000
 #define   WIN_CTRL_ENABLE       BIT(0)
+/* Only on HW I/O coherency capable platforms */
 #define   WIN_CTRL_SYNCBARRIER  BIT(1)
 #define   WIN_CTRL_TGT_MASK     0xf0
 #define   WIN_CTRL_TGT_SHIFT    4

@@ -323,8 +324,9 @@ static int mvebu_mbus_setup_window(struct mvebu_mbus_state *mbus,
 	ctrl = ((size - 1) & WIN_CTRL_SIZE_MASK) |
 		(attr << WIN_CTRL_ATTR_SHIFT)    |
 		(target << WIN_CTRL_TGT_SHIFT)   |
-		WIN_CTRL_SYNCBARRIER             |
 		WIN_CTRL_ENABLE;
+	if (mbus->hw_io_coherency)
+		ctrl |= WIN_CTRL_SYNCBARRIER;
 
 	writel(base & WIN_BASE_LOW, addr + WIN_BASE_OFF);
 	writel(ctrl, addr + WIN_CTRL_OFF);

From: Nicolas Schichan <[email protected]>

Commit a0b5cd4ac2d6 ("bus: mvebu-mbus: use automatic I/O synchronization barriers") enabled the usage of automatic I/O synchronization barriers by enabling bit WIN_CTRL_SYNCBARRIER in the control registers of MBus windows, but on non io-coherent platforms (orion5x, kirkwood and dove) the WIN_CTRL_SYNCBARRIER bit in the window control register is either reserved (all windows except 6 and 7) or enables read-only protection (windows 6 and 7).

Signed-off-by Nicolas Schichan <[email protected]>

Reviewed-by Thomas Petazzoni <[email protected]>

Cc <[email protected]> # v4.0+

Fixes a0b5cd4ac2d6 ("bus: mvebu-mbus: use automatic I/O synchronization barriers")

Signed-off-by Thomas Petazzoni <[email protected]>

Fixes 1737cac69369 ("bus: mvebu-mbus: make sure SDRAM CS for DMA don't overlap the MBus bridge window")

Cc <[email protected]> # v4.0+

---
 drivers/bus/mvebu-mbus.c | 105 ++++++++---------------------------------------
 1 file changed, 16 insertions(+), 89 deletions(-)

Quoted 5 lines · On Thursday 28 May 2015 10:40:13 Thomas Petazzoni wrote:

Fixes 1737cac69369 ("bus: mvebu-mbus: make sure SDRAM CS for DMA don't overlap the MBus bridge window")

Cc <[email protected]> # v4.0+

---
 drivers/bus/mvebu-mbus.c | 105 ++++++++---------------------------------------
 1 file changed, 16 insertions(+), 89 deletions(-)

Hmm, the stable kernel rules say that a patch cannot exceed 100 lines with context, so this one is technically too large.

Maybe Greg has a suggestion about what to do here. Is it possible to make an exception for a revert? In theory you could make a smaller version of the patch that adds an #if 0 instead of removing some of the code that was added, in order to get below the limit, but that seems counterproductive for minimizing the possible risk.

Quoted 5 lines · On Thursday 28 May 2015 10:40:13 Thomas Petazzoni wrote:

Fixes 1737cac69369 ("bus: mvebu-mbus: make sure SDRAM CS for DMA don't overlap the MBus bridge window")

Cc <[email protected]> # v4.0+

---
 drivers/bus/mvebu-mbus.c | 105 ++++++++---------------------------------------
 1 file changed, 16 insertions(+), 89 deletions(-)

Hmm, the stable kernel rules say that a patch cannot exceed 100 lines with context, so this one is technically too large.

This reverts commit 1737cac69369 ("bus: mvebu-mbus: make sure SDRAM CS for DMA don't overlap the MBus bridge window"), because it breaks DMA on platforms having more than 2 GB of RAM.

This commit changed the information reported to DMA masters device drivers through the mv_mbus_dram_info() function so that the returned DRAM ranges do not overlap with I/O windows.

This was necessary as a preparation to support the new CESA Crypto Engine driver, which will use DMA for cryptographic operations. But since it does DMA with the SRAM which is mapped as an I/O window, having DRAM ranges overlapping with I/O windows was problematic.

To solve this, the above mentioned commit changed the mvebu-mbus to adjust the DRAM ranges so that they don't overlap with the I/O windows. However, by doing this, we re-adjust the DRAM ranges in a way that makes them have a size that is no longer a power of two. While this is perfectly fine for the Crypto Engine, which supports DRAM ranges with a granularity of 64 KB, it breaks basically all other DMA masters, which expect power of two sizes for the DRAM ranges.

Due to this, if the installed system memory is 4 GB, in two chip-selects of 2 GB, the second DRAM range will be reduced from 2 GB to a little bit less than 2 GB to not overlap with the I/O windows, in a way that results in a DRAM range that doesn't have a power of two size. This means that whenever you do a DMA transfer with an address located in the [ 2 GB ; 4 GB ] area, it will freeze the system. Any serious DMA activity like simply running:

  for i in $(seq 1 64) ; do dd if=/dev/urandom of=file$i bs=1M count=16 ; done

in an ext3 partition mounted over a SATA drive will freeze the system.

Since the new CESA crypto driver that uses DMA has not been merged yet, the easiest fix is to simply revert this commit. A follow-up commit will introduce a different solution for the CESA crypto driver.

Signed-off-by Thomas Petazzoni <[email protected]>

Fixes 1737cac69369 ("bus: mvebu-mbus: make sure SDRAM CS for DMA don't overlap the MBus bridge window")

Cc <[email protected]> # v4.0+

---
 drivers/bus/mvebu-mbus.c | 105 ++++++++---------------------------------------
 1 file changed, 16 insertions(+), 89 deletions(-)

drivers/bus/mvebu-mbus.c +16 -89

index 7fa4510..6f047dc 100644
@@ -58,7 +58,6 @@
 #include <linux/debugfs.h>
 #include <linux/log2.h>
 #include <linux/syscore_ops.h>
-#include <linux/memblock.h>
 
 /*
  * DDR target is the same on all platforms.

@@ -103,9 +102,7 @@
 
 /* Relative to mbusbridge_base */
 #define MBUS_BRIDGE_CTRL_OFF	0x0
-#define  MBUS_BRIDGE_SIZE_MASK  0xffff0000
 #define MBUS_BRIDGE_BASE_OFF	0x4
-#define  MBUS_BRIDGE_BASE_MASK  0xffff0000
 
 /* Maximum number of windows, for all known platforms */
 #define MBUS_WINS_MAX           20

@@ -579,106 +576,36 @@ static unsigned int armada_xp_mbus_win_remap_offset(int win)
 		return MVEBU_MBUS_NO_REMAP;
 }
 
-/*
- * Use the memblock information to find the MBus bridge hole in the
- * physical address space.
- */
-static void __init
-mvebu_mbus_find_bridge_hole(uint64_t *start, uint64_t *end)
-{
-	struct memblock_region *r;
-	uint64_t s = 0;
-
-	for_each_memblock(memory, r) {
-		/*
-		 * This part of the memory is above 4 GB, so we don't
-		 * care for the MBus bridge hole.
-		 */
-		if (r->base >= 0x100000000)
-			continue;
-
-		/*
-		 * The MBus bridge hole is at the end of the RAM under
-		 * the 4 GB limit.
-		 */
-		if (r->base + r->size > s)
-			s = r->base + r->size;
-	}
-
-	*start = s;
-	*end = 0x100000000;
-}
-
 static void __init
 mvebu_mbus_default_setup_cpu_target(struct mvebu_mbus_state *mbus)
 {
 	int i;
 	int cs;
-	uint64_t mbus_bridge_base, mbus_bridge_end;
 
 	mvebu_mbus_dram_info.mbus_dram_target_id = TARGET_DDR;
 
-	mvebu_mbus_find_bridge_hole(&mbus_bridge_base, &mbus_bridge_end);
-
 	for (i = 0, cs = 0; i < 4; i++) {
-		u64 base = readl(mbus->sdramwins_base + DDR_BASE_CS_OFF(i));
-		u64 size = readl(mbus->sdramwins_base + DDR_SIZE_CS_OFF(i));
-		u64 end;
-		struct mbus_dram_window *w;
-
-		/* Ignore entries that are not enabled */
-		if (!(size & DDR_SIZE_ENABLED))
-			continue;
-
-		/*
-		 * Ignore entries whose base address is above 2^32,
-		 * since devices cannot DMA to such high addresses
-		 */
-		if (base & DDR_BASE_CS_HIGH_MASK)
-			continue;
-
-		base = base & DDR_BASE_CS_LOW_MASK;
-		size = (size | ~DDR_SIZE_MASK) + 1;
-		end = base + size;
-
-		/*
-		 * Adjust base/size of the current CS to make sure it
-		 * doesn't overlap with the MBus bridge hole. This is
-		 * particularly important for devices that do DMA from
-		 * DRAM to a SRAM mapped in a MBus window, such as the
-		 * CESA cryptographic engine.
-		 */
+		u32 base = readl(mbus->sdramwins_base + DDR_BASE_CS_OFF(i));
+		u32 size = readl(mbus->sdramwins_base + DDR_SIZE_CS_OFF(i));
 
 		/*
-		 * The CS is fully enclosed inside the MBus bridge
-		 * area, so ignore it.
+		 * We only take care of entries for which the chip
+		 * select is enabled, and that don't have high base
+		 * address bits set (devices can only access the first
+		 * 32 bits of the memory).
 		 */
-		if (base >= mbus_bridge_base && end <= mbus_bridge_end)
-			continue;
+		if ((size & DDR_SIZE_ENABLED) &&
+		    !(base & DDR_BASE_CS_HIGH_MASK)) {
+			struct mbus_dram_window *w;
 
-		/*
-		 * Beginning of CS overlaps with end of MBus, raise CS
-		 * base address, and shrink its size.
-		 */
-		if (base >= mbus_bridge_base && end > mbus_bridge_end) {
-			size -= mbus_bridge_end - base;
-			base = mbus_bridge_end;
+			w = &mvebu_mbus_dram_info.cs[cs++];
+			w->cs_index = i;
+			w->mbus_attr = 0xf & ~(1 << i);
+			if (mbus->hw_io_coherency)
+				w->mbus_attr |= ATTR_HW_COHERENCY;
+			w->base = base & DDR_BASE_CS_LOW_MASK;
+			w->size = (size | ~DDR_SIZE_MASK) + 1;
 		}
-
-		/*
-		 * End of CS overlaps with beginning of MBus, shrink
-		 * CS size.
-		 */
-		if (base < mbus_bridge_base && end > mbus_bridge_base)
-			size -= end - mbus_bridge_base;
-
-		w = &mvebu_mbus_dram_info.cs[cs++];
-		w->cs_index = i;
-		w->mbus_attr = 0xf & ~(1 << i);
-		if (mbus->hw_io_coherency)
-			w->mbus_attr |= ATTR_HW_COHERENCY;
-		w->base = base;
-		w->size = size;
 	}
 	mvebu_mbus_dram_info.num_cs = cs;
 }

Hi Thomas,

Quoted 37 lines · Thomas Petazzoni

This reverts commit 1737cac69369 ("bus: mvebu-mbus: make sure SDRAM CS for DMA don't overlap the MBus bridge window"), because it breaks DMA on platforms having more than 2 GB of RAM.

This commit changed the information reported to DMA masters device drivers through the mv_mbus_dram_info() function so that the returned DRAM ranges do not overlap with I/O windows.

This was necessary as a preparation to support the new CESA Crypto Engine driver, which will use DMA for cryptographic operations. But since it does DMA with the SRAM which is mapped as an I/O window, having DRAM ranges overlapping with I/O windows was problematic.

To solve this, the above mentioned commit changed the mvebu-mbus to adjust the DRAM ranges so that they don't overlap with the I/O windows. However, by doing this, we re-adjust the DRAM ranges in a way that makes them have a size that is no longer a power of two. While this is perfectly fine for the Crypto Engine, which supports DRAM ranges with a granularity of 64 KB, it breaks basically all other DMA masters, which expect power of two sizes for the DRAM ranges.

Due to this, if the installed system memory is 4 GB, in two chip-selects of 2 GB, the second DRAM range will be reduced from 2 GB to a little bit less than 2 GB to not overlap with the I/O windows, in a way that results in a DRAM range that doesn't have a power of two size. This means that whenever you do a DMA transfer with an address located in the [ 2 GB ; 4 GB ] area, it will freeze the system. Any serious DMA activity like simply running:

  for i in $(seq 1 64) ; do dd if=/dev/urandom of=file$i bs=1M count=16 ; done

in an ext3 partition mounted over a SATA drive will freeze the system.

Since the new CESA crypto driver that uses DMA has not been merged yet, the easiest fix is to simply revert this commit. A follow-up commit will introduce a different solution for the CESA crypto driver.

Quoted 3 lines

Signed-off-by Thomas Petazzoni <[email protected]>

Fixes 1737cac69369 ("bus: mvebu-mbus: make sure SDRAM CS for DMA don't overlap the MBus bridge window")

Cc <[email protected]> # v4.0+

applied on mvebu/fixes

I hope that the 100 lines rules will be not a problem for this case