Add logic to consolidate TRBs for Synopsys xHC

This commit adds the documentation for consolidate-sgl, and snps,consolidate-sgl property. These when set enables the quirk for XHCI driver for consolidation of sg list into a temporary buffer when small buffer sizes are scattered over the sg list not making up to MPS or total transfer size within TRB cache size with Synopsys xHC.

Signed-off-by Tejas Joglekar <[email protected]>

---
 Documentation/devicetree/bindings/usb/dwc3.txt     | 3 +++
 Documentation/devicetree/bindings/usb/usb-xhci.txt | 3 +++
 2 files changed, 6 insertions(+)

Documentation/devicetree/bindings/usb/dwc3.txt +3 -0

index 9946ff9ba735..292d1f7969e4 100644
@@ -104,6 +104,9 @@ Optional properties:
 			this and tx-thr-num-pkt-prd to a valid, non-zero value
 			1-16 (DWC_usb31 programming guide section 1.2.3) to
 			enable periodic ESS TX threshold.
+ - snps,consolidate-sgl: enable sg list consolidation - host mode only. Set to use
+			SG buffers of at least MPS size by consolidating smaller SG
+			buffers list into a single buffer.
 
  - <DEPRECATED> tx-fifo-resize: determines if the FIFO *has* to be reallocated.
  - snps,incr-burst-type-adjustment: Value for INCR burst type of GSBUSCFG0

Documentation/devicetree/bindings/usb/usb-xhci.txt +0 -0

index 3f378951d624..a90d853557ee 100644

Quoted 24 lines · Tejas Joglekar

This commit adds the documentation for consolidate-sgl, and snps,consolidate-sgl property. These when set enables the quirk for XHCI driver for consolidation of sg list into a temporary buffer when small buffer sizes are scattered over the sg list not making up to MPS or total transfer size within TRB cache size with Synopsys xHC.

Signed-off-by Tejas Joglekar <[email protected]>

---
 Documentation/devicetree/bindings/usb/dwc3.txt     | 3 +++
 Documentation/devicetree/bindings/usb/usb-xhci.txt | 3 +++
 2 files changed, 6 insertions(+)

Documentation/devicetree/bindings/usb/dwc3.txt +3 -0

index 9946ff9ba735..292d1f7969e4 100644
@@ -104,6 +104,9 @@ Optional properties:
 			this and tx-thr-num-pkt-prd to a valid, non-zero value
 			1-16 (DWC_usb31 programming guide section 1.2.3) to
 			enable periodic ESS TX threshold.
+ - snps,consolidate-sgl: enable sg list consolidation - host mode only. Set to use
+			SG buffers of at least MPS size by consolidating smaller SG
+			buffers list into a single buffer.
 

Documentation/devicetree/bindings/usb/usb-xhci.txt Optional properties: · line 48

Quoted 13 lines

Documentation/devicetree/bindings/usb/dwc3.txt +0 -0

  - <DEPRECATED> tx-fifo-resize: determines if the FIFO *has* to be reallocated.
  - snps,incr-burst-type-adjustment: Value for INCR burst type of GSBUSCFG0

Documentation/devicetree/bindings/usb/usb-xhci.txt +3 -0

index 3f378951d624..a90d853557ee 100644
@@ -43,6 +43,9 @@ Optional properties:
   - quirk-broken-port-ped: set if the controller has broken port disable mechanism
   - imod-interval-ns: default interrupt moderation interval is 5000ns
   - phys : see usb-hcd.yaml in the current directory
+  - consolidate-sgl: indicate if you need to consolidate sg list into a
+    temporary buffer when small SG buffer sizes does not make upto MPS
+    size or total transfer size across the TRB cache size.

Shouldn't this refer to the fact that the hardware is broken? Otherwise why would anyone know if they should, or should not, enable this option?

thanks,

greg k-h

Hi,

Quoted 25 lines · Greg KH

Quoted 24 lines · Tejas Joglekar

This commit adds the documentation for consolidate-sgl, and snps,consolidate-sgl property. These when set enables the quirk for XHCI driver for consolidation of sg list into a temporary buffer when small buffer sizes are scattered over the sg list not making up to MPS or total transfer size within TRB cache size with Synopsys xHC.

Signed-off-by Tejas Joglekar <[email protected]>

---
 Documentation/devicetree/bindings/usb/dwc3.txt     | 3 +++
 Documentation/devicetree/bindings/usb/usb-xhci.txt | 3 +++
 2 files changed, 6 insertions(+)

Documentation/devicetree/bindings/usb/dwc3.txt +3 -0

index 9946ff9ba735..292d1f7969e4 100644
@@ -104,6 +104,9 @@ Optional properties:
 			this and tx-thr-num-pkt-prd to a valid, non-zero value
 			1-16 (DWC_usb31 programming guide section 1.2.3) to
 			enable periodic ESS TX threshold.
+ - snps,consolidate-sgl: enable sg list consolidation - host mode only. Set to use
+			SG buffers of at least MPS size by consolidating smaller SG
+			buffers list into a single buffer.
 

Quoted 17 lines

Documentation/devicetree/bindings/usb/usb-xhci.txt Optional properties: · line 48

Quoted 13 lines

Documentation/devicetree/bindings/usb/dwc3.txt +0 -0

  - <DEPRECATED> tx-fifo-resize: determines if the FIFO *has* to be reallocated.
  - snps,incr-burst-type-adjustment: Value for INCR burst type of GSBUSCFG0

Documentation/devicetree/bindings/usb/usb-xhci.txt +3 -0

index 3f378951d624..a90d853557ee 100644
@@ -43,6 +43,9 @@ Optional properties:
   - quirk-broken-port-ped: set if the controller has broken port disable mechanism
   - imod-interval-ns: default interrupt moderation interval is 5000ns
   - phys : see usb-hcd.yaml in the current directory
+  - consolidate-sgl: indicate if you need to consolidate sg list into a
+    temporary buffer when small SG buffer sizes does not make upto MPS
+    size or total transfer size across the TRB cache size.

Shouldn't this refer to the fact that the hardware is broken? Otherwise why would anyone know if they should, or should not, enable this option?

We have not seen issue with Linux environment for now. But with other OS with SG list with very small buffer sizes the xHC controller hang was seen. So currently introducing the binding as optional one. One could enable this option when xHC halt happens due to small SG list sizes.

Quoted 44 lines · Tejas Joglekar

Hi,

Quoted 25 lines · Greg KH

Quoted 24 lines · Tejas Joglekar

This commit adds the documentation for consolidate-sgl, and snps,consolidate-sgl property. These when set enables the quirk for XHCI driver for consolidation of sg list into a temporary buffer when small buffer sizes are scattered over the sg list not making up to MPS or total transfer size within TRB cache size with Synopsys xHC.

Signed-off-by Tejas Joglekar <[email protected]>

---
 Documentation/devicetree/bindings/usb/dwc3.txt     | 3 +++
 Documentation/devicetree/bindings/usb/usb-xhci.txt | 3 +++
 2 files changed, 6 insertions(+)

Documentation/devicetree/bindings/usb/dwc3.txt +3 -0

index 9946ff9ba735..292d1f7969e4 100644
@@ -104,6 +104,9 @@ Optional properties:
 			this and tx-thr-num-pkt-prd to a valid, non-zero value
 			1-16 (DWC_usb31 programming guide section 1.2.3) to
 			enable periodic ESS TX threshold.
+ - snps,consolidate-sgl: enable sg list consolidation - host mode only. Set to use
+			SG buffers of at least MPS size by consolidating smaller SG
+			buffers list into a single buffer.
 

Quoted 17 lines

Documentation/devicetree/bindings/usb/usb-xhci.txt Optional properties: · line 48

Quoted 13 lines

Documentation/devicetree/bindings/usb/dwc3.txt +0 -0

  - <DEPRECATED> tx-fifo-resize: determines if the FIFO *has* to be reallocated.
  - snps,incr-burst-type-adjustment: Value for INCR burst type of GSBUSCFG0

Documentation/devicetree/bindings/usb/usb-xhci.txt +3 -0

index 3f378951d624..a90d853557ee 100644
@@ -43,6 +43,9 @@ Optional properties:
   - quirk-broken-port-ped: set if the controller has broken port disable mechanism
   - imod-interval-ns: default interrupt moderation interval is 5000ns
   - phys : see usb-hcd.yaml in the current directory
+  - consolidate-sgl: indicate if you need to consolidate sg list into a
+    temporary buffer when small SG buffer sizes does not make upto MPS
+    size or total transfer size across the TRB cache size.

Shouldn't this refer to the fact that the hardware is broken? Otherwise why would anyone know if they should, or should not, enable this option?

Quoted 4 lines

We have not seen issue with Linux environment for now. But with other OS with SG list with very small buffer sizes the xHC controller hang was seen. So currently introducing the binding as optional one. One could enable this option when xHC halt happens due to small SG list sizes.

What I mean is this should be something like, "quirk-broken-sg-list-handler" or something like that. Otherwise how does anyone know if this really is needed or not. Reading this would seem like everyone would like to do this, as consolidating links sounds like a good optimization, when instead this really cause more memory allocations, making this possibly worse performance.

thanks,

greg k-h

This commit adds the documentation for consolidate-sgl, and snps,consolidate-sgl property. These when set enables the quirk for XHCI driver for consolidation of sg list into a temporary buffer when small buffer sizes are scattered over the sg list not making up to MPS or total transfer size within TRB cache size with Synopsys xHC.

Signed-off-by Tejas Joglekar <[email protected]>

---
 Documentation/devicetree/bindings/usb/dwc3.txt     | 3 +++
 Documentation/devicetree/bindings/usb/usb-xhci.txt | 3 +++
 2 files changed, 6 insertions(+)

Documentation/devicetree/bindings/usb/dwc3.txt +0 -0

index 9946ff9ba735..292d1f7969e4 100644

Quoted 17 lines · Tejas Joglekar

This commit adds the documentation for consolidate-sgl, and snps,consolidate-sgl property. These when set enables the quirk for XHCI driver for consolidation of sg list into a temporary buffer when small buffer sizes are scattered over the sg list not making up to MPS or total transfer size within TRB cache size with Synopsys xHC.

Signed-off-by Tejas Joglekar <[email protected]>

---
 Documentation/devicetree/bindings/usb/dwc3.txt     | 3 +++
 Documentation/devicetree/bindings/usb/usb-xhci.txt | 3 +++
 2 files changed, 6 insertions(+)

Documentation/devicetree/bindings/usb/dwc3.txt +0 -0

index 9946ff9ba735..292d1f7969e4 100644

Documentation/devicetree/bindings/usb/dwc3.txt Optional properties: · line 109

Quoted 6 lines

Documentation/devicetree/bindings/usb/dwc3.txt +3 -0

@@ -104,6 +104,9 @@ Optional properties:
 			this and tx-thr-num-pkt-prd to a valid, non-zero value
 			1-16 (DWC_usb31 programming guide section 1.2.3) to
 			enable periodic ESS TX threshold.
+ - snps,consolidate-sgl: enable sg list consolidation - host mode only. Set to use
+			SG buffers of at least MPS size by consolidating smaller SG
+			buffers list into a single buffer.

The preference is not to keep adding properties for every single quirk or feature. These should be implied by specific compatibles. As Synopsys knows what quirks/errata/features are in each version of IP, the compatible strings should reflect those versions. (And yes, I'm sure there's customer ECO fixes that aren't reflected in the version, but that's why we have SoC specific compatibles too.) This is the only way we can fix quirks in the OS without doing DT updates. For comparison, do you want to have to update your PC BIOS so an OS can work-around issues?

Rob

Hi,

Quoted 25 lines · Rob Herring

Quoted 17 lines · Tejas Joglekar

This commit adds the documentation for consolidate-sgl, and snps,consolidate-sgl property. These when set enables the quirk for XHCI driver for consolidation of sg list into a temporary buffer when small buffer sizes are scattered over the sg list not making up to MPS or total transfer size within TRB cache size with Synopsys xHC.

Signed-off-by Tejas Joglekar <[email protected]>

---
 Documentation/devicetree/bindings/usb/dwc3.txt     | 3 +++
 Documentation/devicetree/bindings/usb/usb-xhci.txt | 3 +++
 2 files changed, 6 insertions(+)

Documentation/devicetree/bindings/usb/dwc3.txt +0 -0

index 9946ff9ba735..292d1f7969e4 100644

Documentation/devicetree/bindings/usb/dwc3.txt Optional properties: · line 109

Quoted 6 lines

Documentation/devicetree/bindings/usb/dwc3.txt +3 -0

@@ -104,6 +104,9 @@ Optional properties:
 			this and tx-thr-num-pkt-prd to a valid, non-zero value
 			1-16 (DWC_usb31 programming guide section 1.2.3) to
 			enable periodic ESS TX threshold.
+ - snps,consolidate-sgl: enable sg list consolidation - host mode only. Set to use
+			SG buffers of at least MPS size by consolidating smaller SG
+			buffers list into a single buffer.

Quoted 9 lines

The preference is not to keep adding properties for every single quirk or feature. These should be implied by specific compatibles. As Synopsys knows what quirks/errata/features are in each version of IP, the compatible strings should reflect those versions. (And yes, I'm sure there's customer ECO fixes that aren't reflected in the version, but that's why we have SoC specific compatibles too.) This is the only way we can fix quirks in the OS without doing DT updates. For comparison, do you want to have to update your PC BIOS so an OS can work-around issues?

Yes, I understand what you want to say here. But I think this compatible string does not work with the platform drivers with PCI based systems. So based on the vendor id and device id then I need to set the quirk required.

@Felipe: What do you suggest for setting up quirk without the DT update for dwc3?

Hi,

Quoted 40 lines · Tejas Joglekar

Hi,

Quoted 25 lines · Rob Herring

Quoted 17 lines · Tejas Joglekar

This commit adds the documentation for consolidate-sgl, and snps,consolidate-sgl property. These when set enables the quirk for XHCI driver for consolidation of sg list into a temporary buffer when small buffer sizes are scattered over the sg list not making up to MPS or total transfer size within TRB cache size with Synopsys xHC.

Signed-off-by Tejas Joglekar <[email protected]>

---
 Documentation/devicetree/bindings/usb/dwc3.txt     | 3 +++
 Documentation/devicetree/bindings/usb/usb-xhci.txt | 3 +++
 2 files changed, 6 insertions(+)

Documentation/devicetree/bindings/usb/dwc3.txt +0 -0

index 9946ff9ba735..292d1f7969e4 100644

Documentation/devicetree/bindings/usb/dwc3.txt Optional properties: · line 109

Quoted 6 lines

Documentation/devicetree/bindings/usb/dwc3.txt +3 -0

@@ -104,6 +104,9 @@ Optional properties:
 			this and tx-thr-num-pkt-prd to a valid, non-zero value
 			1-16 (DWC_usb31 programming guide section 1.2.3) to
 			enable periodic ESS TX threshold.
+ - snps,consolidate-sgl: enable sg list consolidation - host mode only. Set to use
+			SG buffers of at least MPS size by consolidating smaller SG
+			buffers list into a single buffer.

Quoted 9 lines

The preference is not to keep adding properties for every single quirk or feature. These should be implied by specific compatibles. As Synopsys knows what quirks/errata/features are in each version of IP, the compatible strings should reflect those versions. (And yes, I'm sure there's customer ECO fixes that aren't reflected in the version, but that's why we have SoC specific compatibles too.) This is the only way we can fix quirks in the OS without doing DT updates. For comparison, do you want to have to update your PC BIOS so an OS can work-around issues?

Yes, I understand what you want to say here. But I think this compatible string does not work with the platform drivers with PCI based systems. So based on the vendor id and device id then I need to set the quirk required.

Quoted 1 line

@Felipe: What do you suggest for setting up quirk without the DT update for dwc3?

We have been using Synopsys controller's revision register, but that's not visible to xhci driver and we don't have a separate compatible for each synopsys version on the xhci driver side. One option would be to add "snps,xhci-foo-bar" to xhci-plat and use that.

Hi,

Quoted 44 lines · Felipe Balbi

Hi,

Quoted 40 lines · Tejas Joglekar

Hi,

Quoted 25 lines · Rob Herring

Quoted 17 lines · Tejas Joglekar

This commit adds the documentation for consolidate-sgl, and snps,consolidate-sgl property. These when set enables the quirk for XHCI driver for consolidation of sg list into a temporary buffer when small buffer sizes are scattered over the sg list not making up to MPS or total transfer size within TRB cache size with Synopsys xHC.

Signed-off-by Tejas Joglekar <[email protected]>

---
 Documentation/devicetree/bindings/usb/dwc3.txt     | 3 +++
 Documentation/devicetree/bindings/usb/usb-xhci.txt | 3 +++
 2 files changed, 6 insertions(+)

Documentation/devicetree/bindings/usb/dwc3.txt +0 -0

index 9946ff9ba735..292d1f7969e4 100644

Documentation/devicetree/bindings/usb/dwc3.txt Optional properties: · line 109

Quoted 6 lines

Documentation/devicetree/bindings/usb/dwc3.txt +3 -0

@@ -104,6 +104,9 @@ Optional properties:
 			this and tx-thr-num-pkt-prd to a valid, non-zero value
 			1-16 (DWC_usb31 programming guide section 1.2.3) to
 			enable periodic ESS TX threshold.
+ - snps,consolidate-sgl: enable sg list consolidation - host mode only. Set to use
+			SG buffers of at least MPS size by consolidating smaller SG
+			buffers list into a single buffer.

Quoted 9 lines

The preference is not to keep adding properties for every single quirk or feature. These should be implied by specific compatibles. As Synopsys knows what quirks/errata/features are in each version of IP, the compatible strings should reflect those versions. (And yes, I'm sure there's customer ECO fixes that aren't reflected in the version, but that's why we have SoC specific compatibles too.) This is the only way we can fix quirks in the OS without doing DT updates. For comparison, do you want to have to update your PC BIOS so an OS can work-around issues?

Yes, I understand what you want to say here. But I think this compatible string does not work with the platform drivers with PCI based systems. So based on the vendor id and device id then I need to set the quirk required.

Quoted 7 lines

Quoted 1 line

@Felipe: What do you suggest for setting up quirk without the DT update for dwc3?

We have been using Synopsys controller's revision register, but that's not visible to xhci driver and we don't have a separate compatible for each synopsys version on the xhci driver side. One option would be to add "snps,xhci-foo-bar" to xhci-plat and use that.

@Felipe: Thanks for suggestion

@Mathias: Do you prefer parameter passed this way? I can send the v2 patch set, after your input.

The Synopsys xHC has an internal TRB cache of size TRB_CACHE_SIZE for each endpoint. The default value for TRB_CACHE_SIZE is 16 for SS and 8 for HS. The controller loads and updates the TRB cache from the transfer ring in system memory whenever the driver issues a start transfer or update transfer command.

For chained TRBs, the Synopsys xHC requires that the total amount of bytes for all TRBs loaded in the TRB cache be greater than or equal to 1 MPS. Or the chain ends within the TRB cache (with a last TRB).

If this requirement is not met, the controller will not be able to send or receive a packet and it will hang causing a driver timeout and error.

This can be a problem if a class driver queues SG requests with many small-buffer entries. The XHCI driver will create a chained TRB for each entry which may trigger this issue.

This patch adds logic to the XHCI driver to detect and prevent this from happening.

For every (TRB_CACHE_SIZE - 2), we check the total buffer size of the SG list and if the last window of (TRB_CACHE_SIZE - 2) SG list length and we don't make up at least 1 MPS, we create a temporary buffer to consolidate full SG list into the buffer.

We check at (TRB_CACHE_SIZE - 2) window because it is possible that there would be a link and/or event data TRB that take up to 2 of the cache entries.

We discovered this issue with devices on other platforms but have not yet come across any device that triggers this on Linux. But it could be a real problem now or in the future. All it takes is N number of small chained TRBs. And other instances of the Synopsys IP may have smaller values for the TRB_CACHE_SIZE which would exacerbate the problem.

Signed-off-by Tejas Joglekar <[email protected]>

---

Resending as 'umlaut' in email are not accepted by some servers.

 drivers/usb/core/hcd.c       |   8 +++
 drivers/usb/host/xhci-ring.c |   2 +-
 drivers/usb/host/xhci.c      | 128 +++++++++++++++++++++++++++++++++++++++++++
 drivers/usb/host/xhci.h      |   4 ++
 4 files changed, 141 insertions(+), 1 deletion(-)

drivers/usb/core/hcd.c +0 -0

index aa45840d8273..fdd257a2b8a6 100644

drivers/usb/core/hcd.c +1 -0

+
 				n = dma_map_sg(
 						hcd->self.sysdev,
 						urb->sg,

drivers/usb/host/xhci-ring.c +1 -1

index a78787bb5133..2fad9474912a 100644
@@ -3291,7 +3291,7 @@ int xhci_queue_bulk_tx(struct xhci_hcd *xhci, gfp_t mem_flags,
 
 	full_len = urb->transfer_buffer_length;
 	/* If we have scatter/gather list, we use it. */
-	if (urb->num_sgs) {
+	if (urb->num_sgs && !(urb->transfer_flags & URB_DMA_MAP_SINGLE)) {
 		num_sgs = urb->num_mapped_sgs;
 		sg = urb->sg;
 		addr = (u64) sg_dma_address(sg);

drivers/usb/host/xhci.c +7 -0

index fe38275363e0..94fddbd06179 100644
@@ -1256,6 +1256,109 @@ EXPORT_SYMBOL_GPL(xhci_resume);
 
 /*-------------------------------------------------------------------------*/
 
+static int xhci_map_temp_buffer(struct usb_hcd *hcd, struct urb *urb)
+{
+	void *temp;
+	int ret = 0;
+	unsigned int len;
+	unsigned int buf_len;
+	enum dma_data_direction dir;

Quoted 6 lines · Tejas Joglekar

The Synopsys xHC has an internal TRB cache of size TRB_CACHE_SIZE for each endpoint. The default value for TRB_CACHE_SIZE is 16 for SS and 8 for HS. The controller loads and updates the TRB cache from the transfer ring in system memory whenever the driver issues a start transfer or update transfer command.

Quoted 6 lines

For chained TRBs, the Synopsys xHC requires that the total amount of bytes for all TRBs loaded in the TRB cache be greater than or equal to 1 MPS. Or the chain ends within the TRB cache (with a last TRB).

If this requirement is not met, the controller will not be able to send or receive a packet and it will hang causing a driver timeout and error.

Sounds like broken hardware, or is this requirement in the xhci spec?

Quoted 42 lines

This can be a problem if a class driver queues SG requests with many small-buffer entries. The XHCI driver will create a chained TRB for each entry which may trigger this issue.

This patch adds logic to the XHCI driver to detect and prevent this from happening.

For every (TRB_CACHE_SIZE - 2), we check the total buffer size of the SG list and if the last window of (TRB_CACHE_SIZE - 2) SG list length and we don't make up at least 1 MPS, we create a temporary buffer to consolidate full SG list into the buffer.

We check at (TRB_CACHE_SIZE - 2) window because it is possible that there would be a link and/or event data TRB that take up to 2 of the cache entries.

We discovered this issue with devices on other platforms but have not yet come across any device that triggers this on Linux. But it could be a real problem now or in the future. All it takes is N number of small chained TRBs. And other instances of the Synopsys IP may have smaller values for the TRB_CACHE_SIZE which would exacerbate the problem.

Signed-off-by Tejas Joglekar <[email protected]>

---

Resending as 'umlaut' in email are not accepted by some servers.

 drivers/usb/core/hcd.c       |   8 +++
 drivers/usb/host/xhci-ring.c |   2 +-
 drivers/usb/host/xhci.c      | 128 +++++++++++++++++++++++++++++++++++++++++++
 drivers/usb/host/xhci.h      |   4 ++
 4 files changed, 141 insertions(+), 1 deletion(-)

drivers/usb/core/hcd.c +0 -0

index aa45840d8273..fdd257a2b8a6 100644
@@ -1459,6 +1459,14 @@ int usb_hcd_map_urb_for_dma(struct usb_hcd *hcd, struct urb *urb,
 					return -EINVAL;
 				}
 

drivers/usb/core/hcd.c int usb_hcd_map_urb_for_dma(struct usb_hcd *hcd, struct urb *urb, · line 1464

Quoted 3 lines

drivers/usb/core/hcd.c +3 -0

+				/*
+				 * If SG is consolidate into single buffer
+				 * return early

I do not understand this comment.

drivers/usb/core/hcd.c int usb_hcd_map_urb_for_dma(struct usb_hcd *hcd, struct urb *urb, · line 1468

Quoted 5 lines

drivers/usb/core/hcd.c +4 -0

+				 */
+				if ((urb->transfer_flags &
+				     URB_DMA_MAP_SINGLE))
+					return ret;

Why? Isn't this now going to affect other host controllers (like all of them?)

Quoted 36 lines

drivers/usb/core/hcd.c +1 -0

+
 				n = dma_map_sg(
 						hcd->self.sysdev,
 						urb->sg,

drivers/usb/host/xhci-ring.c +1 -1

index a78787bb5133..2fad9474912a 100644
@@ -3291,7 +3291,7 @@ int xhci_queue_bulk_tx(struct xhci_hcd *xhci, gfp_t mem_flags,
 
 	full_len = urb->transfer_buffer_length;
 	/* If we have scatter/gather list, we use it. */
-	if (urb->num_sgs) {
+	if (urb->num_sgs && !(urb->transfer_flags & URB_DMA_MAP_SINGLE)) {
 		num_sgs = urb->num_mapped_sgs;
 		sg = urb->sg;
 		addr = (u64) sg_dma_address(sg);

drivers/usb/host/xhci.c +11 -0

index fe38275363e0..94fddbd06179 100644
@@ -1256,6 +1256,109 @@ EXPORT_SYMBOL_GPL(xhci_resume);
 
 /*-------------------------------------------------------------------------*/
 
+static int xhci_map_temp_buffer(struct usb_hcd *hcd, struct urb *urb)
+{
+	void *temp;
+	int ret = 0;
+	unsigned int len;
+	unsigned int buf_len;
+	enum dma_data_direction dir;
+	struct xhci_hcd *xhci;
+
+	xhci = hcd_to_xhci(hcd);
+	dir = usb_urb_dir_in(urb) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;

drivers/usb/host/xhci.c EXPORT_SYMBOL_GPL(xhci_resume); · line 1275

Quoted 6 lines

drivers/usb/host/xhci.c +6 -0

+	buf_len = urb->transfer_buffer_length;
+
+	temp = kzalloc_node(buf_len, GFP_ATOMIC,
+			    dev_to_node(hcd->self.sysdev));
+	if (!temp) {
+		xhci_warn(xhci, "Failed to create temp buffer, HC may fail\n");

Didn't kzalloc just warn before this?

And isn't this whole thing going to cause a lot more memory allocations per submission than before?

drivers/usb/host/xhci.c EXPORT_SYMBOL_GPL(xhci_resume); · line 1283

Quoted 8 lines

drivers/usb/host/xhci.c +8 -0

+		return -ENOMEM;
+	}
+
+	if (usb_urb_dir_out(urb)) {
+		len = sg_pcopy_to_buffer(urb->sg, urb->num_sgs,
+					 temp, buf_len, 0);
+		if (len != buf_len)
+			xhci_warn(xhci, "Wrong temp buffer write length\n");

How could this happen? And if it does, why spam the kernel log about it and yet not return an error?

Quoted 4 lines

drivers/usb/host/xhci.c +4 -0

+	}
+
+	urb->transfer_buffer = temp;
+	urb->transfer_dma = dma_map_single(hcd->self.sysdev,

drivers/usb/host/xhci.c EXPORT_SYMBOL_GPL(xhci_resume); · line 1293

Quoted 6 lines

drivers/usb/host/xhci.c +6 -0

+					   urb->transfer_buffer,
+					   urb->transfer_buffer_length,
+					   dir);
+	if (dma_mapping_error(hcd->self.sysdev,
+			      urb->transfer_dma)) {
+		xhci_err(xhci, "dma mapping error\n");

Again, didn't dma_mapping_error() spit out a message?

drivers/usb/host/xhci.c EXPORT_SYMBOL_GPL(xhci_resume); · line 1301

Quoted 8 lines

drivers/usb/host/xhci.c +8 -0

+		ret = -EAGAIN;
+		kfree(temp);
+	} else {
+		urb->transfer_flags |= URB_DMA_MAP_SINGLE;
+	}
+
+	return ret;
+}

thanks,

greg k-h

Hi,

Quoted 11 lines · Greg KH

Quoted 10 lines · Tejas Joglekar

The Synopsys xHC has an internal TRB cache of size TRB_CACHE_SIZE for each endpoint. The default value for TRB_CACHE_SIZE is 16 for SS and 8 for HS. The controller loads and updates the TRB cache from the transfer ring in system memory whenever the driver issues a start transfer or update transfer command.

For chained TRBs, the Synopsys xHC requires that the total amount of bytes for all TRBs loaded in the TRB cache be greater than or equal to 1 MPS. Or the chain ends within the TRB cache (with a last TRB).

Quoted 5 lines

Quoted 2 lines

If this requirement is not met, the controller will not be able to send or receive a packet and it will hang causing a driver timeout and error.

Sounds like broken hardware, or is this requirement in the xhci spec?

Not a xhci spec requirement, but behavior of Synopsys xHC. We have not seen actual failure on Linux yet but it is possible in future if SG list with very small buffer size is given for transfer.

Quoted 41 lines

Quoted 41 lines

This can be a problem if a class driver queues SG requests with many small-buffer entries. The XHCI driver will create a chained TRB for each entry which may trigger this issue.

This patch adds logic to the XHCI driver to detect and prevent this from happening.

For every (TRB_CACHE_SIZE - 2), we check the total buffer size of the SG list and if the last window of (TRB_CACHE_SIZE - 2) SG list length and we don't make up at least 1 MPS, we create a temporary buffer to consolidate full SG list into the buffer.

We check at (TRB_CACHE_SIZE - 2) window because it is possible that there would be a link and/or event data TRB that take up to 2 of the cache entries.

We discovered this issue with devices on other platforms but have not yet come across any device that triggers this on Linux. But it could be a real problem now or in the future. All it takes is N number of small chained TRBs. And other instances of the Synopsys IP may have smaller values for the TRB_CACHE_SIZE which would exacerbate the problem.

Signed-off-by Tejas Joglekar <[email protected]>

---

Resending as 'umlaut' in email are not accepted by some servers.

 drivers/usb/core/hcd.c       |   8 +++
 drivers/usb/host/xhci-ring.c |   2 +-
 drivers/usb/host/xhci.c      | 128 +++++++++++++++++++++++++++++++++++++++++++
 drivers/usb/host/xhci.h      |   4 ++
 4 files changed, 141 insertions(+), 1 deletion(-)

drivers/usb/core/hcd.c +0 -0

index aa45840d8273..fdd257a2b8a6 100644
@@ -1459,6 +1459,14 @@ int usb_hcd_map_urb_for_dma(struct usb_hcd *hcd, struct urb *urb,
 					return -EINVAL;
 				}
 

Quoted 5 lines

drivers/usb/core/hcd.c int usb_hcd_map_urb_for_dma(struct usb_hcd *hcd, struct urb *urb, · line 1464

Quoted 3 lines

drivers/usb/core/hcd.c +3 -0

+				/*
+				 * If SG is consolidate into single buffer
+				 * return early

I do not understand this comment.

The SG list is copied to a temporary buffer, and buffer is DMA mapped so we should not map the SG list again, and return without any mapping here.

Quoted 9 lines

drivers/usb/core/hcd.c int usb_hcd_map_urb_for_dma(struct usb_hcd *hcd, struct urb *urb, · line 1468

Quoted 5 lines

drivers/usb/core/hcd.c +4 -0

+				 */
+				if ((urb->transfer_flags &
+				     URB_DMA_MAP_SINGLE))
+					return ret;

Why? Isn't this now going to affect other host controllers (like all of them?)

When urb->num_sgs is greater than 0, other than my quirk dma_map function no one will set the DMA transfer flag as URB_DMA_MAP_SINGLE. So it would not be called by all HC's. Even when the SG list does not have very small buffer sizes this quirk will not set the URB_DMA_MAP_SINGLE transfer flag.

Quoted 36 lines

Quoted 36 lines

drivers/usb/core/hcd.c +1 -0

+
 				n = dma_map_sg(
 						hcd->self.sysdev,
 						urb->sg,

drivers/usb/host/xhci-ring.c +1 -1

index a78787bb5133..2fad9474912a 100644
@@ -3291,7 +3291,7 @@ int xhci_queue_bulk_tx(struct xhci_hcd *xhci, gfp_t mem_flags,
 
 	full_len = urb->transfer_buffer_length;
 	/* If we have scatter/gather list, we use it. */
-	if (urb->num_sgs) {
+	if (urb->num_sgs && !(urb->transfer_flags & URB_DMA_MAP_SINGLE)) {
 		num_sgs = urb->num_mapped_sgs;
 		sg = urb->sg;
 		addr = (u64) sg_dma_address(sg);

drivers/usb/host/xhci.c +11 -0

index fe38275363e0..94fddbd06179 100644
@@ -1256,6 +1256,109 @@ EXPORT_SYMBOL_GPL(xhci_resume);
 
 /*-------------------------------------------------------------------------*/
 
+static int xhci_map_temp_buffer(struct usb_hcd *hcd, struct urb *urb)
+{
+	void *temp;
+	int ret = 0;
+	unsigned int len;
+	unsigned int buf_len;
+	enum dma_data_direction dir;
+	struct xhci_hcd *xhci;
+
+	xhci = hcd_to_xhci(hcd);
+	dir = usb_urb_dir_in(urb) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;

Quoted 9 lines

drivers/usb/host/xhci.c EXPORT_SYMBOL_GPL(xhci_resume); · line 1275

Quoted 6 lines

drivers/usb/host/xhci.c +6 -0

+	buf_len = urb->transfer_buffer_length;
+
+	temp = kzalloc_node(buf_len, GFP_ATOMIC,
+			    dev_to_node(hcd->self.sysdev));
+	if (!temp) {
+		xhci_warn(xhci, "Failed to create temp buffer, HC may fail\n");

Didn't kzalloc just warn before this?

Yes, It should.

Quoted 3 lines

And isn't this whole thing going to cause a lot more memory allocations per submission than before?

If buffer sizes for SG list are very small (less than MPS size per TRB_CACHE_SIZE) yes, it will have more memory allocations.

Quoted 21 lines

Quoted 12 lines

Quoted 12 lines

drivers/usb/host/xhci.c +11 -0

+static int xhci_map_temp_buffer(struct usb_hcd *hcd, struct urb *urb)
+{
+	void *temp;
+	int ret = 0;
+	unsigned int len;
+	unsigned int buf_len;
+	enum dma_data_direction dir;
+	struct xhci_hcd *xhci;
+
+	xhci = hcd_to_xhci(hcd);
+	dir = usb_urb_dir_in(urb) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;

Quoted 9 lines

drivers/usb/host/xhci.c EXPORT_SYMBOL_GPL(xhci_resume); · line 1275

Quoted 6 lines

drivers/usb/host/xhci.c +6 -0

+	buf_len = urb->transfer_buffer_length;
+
+	temp = kzalloc_node(buf_len, GFP_ATOMIC,
+			    dev_to_node(hcd->self.sysdev));
+	if (!temp) {
+		xhci_warn(xhci, "Failed to create temp buffer, HC may fail\n");

Didn't kzalloc just warn before this?

Quoted 1 line

Yes, It should.

Then do not spit out another message please.

Quoted 5 lines

Quoted 3 lines

And isn't this whole thing going to cause a lot more memory allocations per submission than before?

If buffer sizes for SG list are very small (less than MPS size per TRB_CACHE_SIZE) yes, it will have more memory allocations.

That's not good :(

The Synopsys xHC has an internal TRB cache of size TRB_CACHE_SIZE for each endpoint. The default value for TRB_CACHE_SIZE is 16 for SS and 8 for HS. The controller loads and updates the TRB cache from the transfer ring in system memory whenever the driver issues a start transfer or update transfer command.

For chained TRBs, the Synopsys xHC requires that the total amount of bytes for all TRBs loaded in the TRB cache be greater than or equal to 1 MPS. Or the chain ends within the TRB cache (with a last TRB).

If this requirement is not met, the controller will not be able to send or receive a packet and it will hang causing a driver timeout and error.

This patch set adds logic to the XHCI driver to detect and prevent this from happening along with the quirk to enable this logic for Synopsys HAPS platform.

Based on Mathias's feedback on previous implementation where consolidation was done in TRB cache, with this patch series the implementation is done during mapping of the URB by consolidating the SG list into a temporary buffer if the SG list buffer sizes within TRB_CACHE_SIZE is less than MPS.

Tejas Joglekar (4):
dt-bindings: usb: Add snps,consolidate-sgl & consolidate-sgl
usb: xhci: Set quirk for XHCI_CONSOLIDATE_SG_LIST
usb: dwc3: Add device property consolidate-sgl
usb: xhci: Use temporary buffer to consolidate SG

Resending as 'umlaut' is not accepted by some servers in email.

Quoted 30 lines · Tejas Joglekar

The Synopsys xHC has an internal TRB cache of size TRB_CACHE_SIZE for each endpoint. The default value for TRB_CACHE_SIZE is 16 for SS and 8 for HS. The controller loads and updates the TRB cache from the transfer ring in system memory whenever the driver issues a start transfer or update transfer command.

For chained TRBs, the Synopsys xHC requires that the total amount of bytes for all TRBs loaded in the TRB cache be greater than or equal to 1 MPS. Or the chain ends within the TRB cache (with a last TRB).

If this requirement is not met, the controller will not be able to send or receive a packet and it will hang causing a driver timeout and error.

This patch set adds logic to the XHCI driver to detect and prevent this from happening along with the quirk to enable this logic for Synopsys HAPS platform.

Based on Mathias's feedback on previous implementation where consolidation was done in TRB cache, with this patch series the implementation is done during mapping of the URB by consolidating the SG list into a temporary buffer if the SG list buffer sizes within TRB_CACHE_SIZE is less than MPS.

Tejas Joglekar (4):
dt-bindings: usb: Add snps,consolidate-sgl & consolidate-sgl
usb: xhci: Set quirk for XHCI_CONSOLIDATE_SG_LIST
usb: dwc3: Add device property consolidate-sgl
usb: xhci: Use temporary buffer to consolidate SG

Quoted 1 line

Resending as 'umlaut' is not accepted by some servers in email.

Only patch 4/4 seems to have gone through :(