Re-introduce TX FIFO resize for larger EP bursting
v9 cover Re-introduce TX FIFO resize for larger EP bursting
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.
Changes in V8:
- Rebased to usb-testing
- Using devm_kzalloc for adding txfifo property in dwc3-qcom
- Removed DWC3 QCOM ACPI property for enabling the txfifo resize
Changes in V7:
- Added a new property tx-fifo-max-num for limiting how much fifo space the resizing logic can allocate for endpoints with large burst values. This can differ across platforms, and tie in closely with overall system latency.
- Added recommended checks for DWC32.
- Added changes to set the tx-fifo-resize property from dwc3-qcom by default instead of modifying the current DTSI files.
- Added comments on all APIs/variables introduced.
- Updated the DWC3 YAML to include a better description of the tx-fifo-resize property and added an entry for tx-fifo-max-num.
Changes in V6:
- Rebased patches to usb-testing.
- Renamed to PATCH series instead of RFC.
- Checking for fs_descriptors instead of ss_descriptors for determining the endpoint count for a particular configuration.
- Re-ordered patch series to fix patch dependencies.
Changes in V5:
- Added check_config() logic, which is used to communicate the number of EPs used in a particular configuration. Based on this, the DWC3 gadget driver has the ability to know the maximum number of eps utilized in all configs. This helps reduce unnecessary allocation to unused eps, and will catch fifo allocation issues at bind() time.
- Fixed variable declaration to single line per variable, and reverse xmas.
- Created a helper for fifo clearing, which is used by ep0.c
Changes in V4:
- Removed struct dwc3* as an argument for dwc3_gadget_resize_tx_fifos()
- Removed WARN_ON(1) in case we run out of fifo space
Changes in V3:
- Removed "Reviewed-by" tags
- Renamed series back to RFC
- Modified logic to ensure that fifo_size is reset if we pass the minimum threshold. Tested with binding multiple FDs requesting 6 FIFOs.
Changes in V2:
- Modified TXFIFO resizing logic to ensure that each EP is reserved a FIFO.
- Removed dev_dbg() prints and fixed typos from patches
- Added some more description on the dt-bindings commit message
Currently, there is no functionality to allow for resizing the TXFIFOs, and relying on the HW default setting for the TXFIFO depth. In most cases, the HW default is probably sufficient, but for USB compositions that contain multiple functions that require EP bursting, the default settings might not be enough. Also to note, the current SW will assign an EP to a function driver w/o checking to see if the TXFIFO size for that particular EP is large enough. (this is a problem if there are multiple HW defined values for the TXFIFO size)
It is mentioned in the SNPS databook that a minimum of TX FIFO depth = 3 is required for an EP that supports bursting. Otherwise, there may be frequent occurences of bursts ending. For high bandwidth functions, such as data tethering (protocols that support data aggregation), mass storage, and media transfer protocol (over FFS), the bMaxBurst value can be large, and a bigger TXFIFO depth may prove to be beneficial in terms of USB throughput. (which can be associated to system access latency, etc...) It allows for a more consistent burst of traffic, w/o any interruptions, as data is readily available in the FIFO.
With testing done using the mass storage function driver, the results show that with a larger TXFIFO depth, the bandwidth increased significantly.
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMark
Results:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 193.60
| | 195.86
| | 184.77
| | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 287.35
| | 304.94
| | 289.64
| | 293.61-------------------------------------------
Wesley Cheng (5):
usb: gadget: udc: core: Introduce check_config to verify USB
configuration
usb: gadget: configfs: Check USB configuration before adding
usb: dwc3: Resize TX FIFOs to meet EP bursting requirements
usb: dwc3: dwc3-qcom: Enable tx-fifo-resize property by default
dt-bindings: usb: dwc3: Update dwc3 TX fifo properties
.../devicetree/bindings/usb/snps,dwc3.yaml | 15 +- drivers/usb/dwc3/core.c | 9 + drivers/usb/dwc3/core.h | 15 ++ drivers/usb/dwc3/dwc3-qcom.c | 9 + drivers/usb/dwc3/ep0.c | 2 + drivers/usb/dwc3/gadget.c | 212 +++++++++++++++++++++ drivers/usb/gadget/configfs.c | 22 +++ drivers/usb/gadget/udc/core.c | 25 +++ include/linux/usb/gadget.h | 5 + 9 files changed, 312 insertions(+), 2 deletions(-)
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
PATCH v9 1/5 usb: gadget: udc: core: Introduce check_config to verify USB configuration
Some UDCs may have constraints on how many high bandwidth endpoints it can support in a certain configuration. This API allows for the composite driver to pass down the total number of endpoints to the UDC so it can verify it has the required resources to support the configuration.
--- drivers/usb/gadget/udc/core.c | 25 +++++++++++++++++++++++++ include/linux/usb/gadget.h | 5 +++++ 2 files changed, 30 insertions(+)
drivers/usb/gadget/udc/core.c +25 -0
@@ -1003,6 +1003,31 @@ int usb_gadget_ep_match_desc(struct usb_gadget *gadget, }EXPORT_SYMBOL_GPL (usb_gadget_ep_match_desc); +/** +* usb_gadget_check_config - checks if the UDC can support the number of eps +* @gadget: controller to check the USB configuration +* @ep_map: bitmap of endpoints being requested by a USB configuration +* +* Ensure that a UDC is able to support the number of endpoints within a USB +* configuration, and that there are no resource limitations to support all +* requested eps. +* +* Returns zero on success, else a negative errno. +*/ +int usb_gadget_check_config (struct usb_gadget *gadget,unsigned long ep_map) +{ +int ret =0 ; + +if (!gadget->ops->check_config) + goto out; + + ret = gadget->ops->check_config (gadget, ep_map); + +out: +return ret; +} +EXPORT_SYMBOL_GPL (usb_gadget_check_config); +/* ------------------------------------------------------------------------- */ static void usb_gadget_state_work (struct work_struct *work)
include/linux/usb/gadget.h +5 -0
@@ -328,6 +328,7 @@ struct usb_gadget_ops {struct usb_ep *(*match_ep)(struct usb_gadget *,struct usb_endpoint_descriptor *,struct usb_ss_ep_comp_descriptor *); +int (*check_config)(struct usb_gadget *gadget,unsigned long ep_map); };/**
@@ -607,6 +608,7 @@ int usb_gadget_connect(struct usb_gadget *gadget);int usb_gadget_disconnect (struct usb_gadget *gadget);int usb_gadget_deactivate (struct usb_gadget *gadget);int usb_gadget_activate (struct usb_gadget *gadget); +int usb_gadget_check_config (struct usb_gadget *gadget,unsigned long ep_map);#else static inline int usb_gadget_frame_number (struct usb_gadget *gadget) {return 0 ; }
@@ -630,6 +632,9 @@ static inline int usb_gadget_deactivate(struct usb_gadget *gadget) {return 0 ; }static inline int usb_gadget_activate (struct usb_gadget *gadget) {return 0 ; } +static inline int usb_gadget_check_config (struct usb_gadget *gadget, +unsigned long ep_map) +{return 0 ; }#endif /* CONFIG_USB_GADGET */ /*-------------------------------------------------------------------------*/
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
PATCH v9 2/5 usb: gadget: configfs: Check USB configuration before adding
In reply to Wesley ChengEnsure that the USB gadget is able to support the configuration being added based on the number of endpoints required from all interfaces. This is for accounting for any bandwidth or space limitations.
--- drivers/usb/gadget/configfs.c | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
drivers/usb/gadget/configfs.c +22 -0
@@ -1374,6 +1374,7 @@ static int configfs_composite_bind(struct usb_gadget *gadget,struct usb_function *f;struct usb_function *tmp;struct gadget_config_name *cn; +unsigned long ep_map =0 ;if (gadget_is_otg (gadget)) c->descriptors = otg_desc;
@@ -1403,7 +1404,28 @@ static int configfs_composite_bind(struct usb_gadget *gadget,list_add (&f->list, &cfg->func_list); goto err_purge_funcs; } +if (f->fs_descriptors) { +struct usb_descriptor_header **d; + + d = f->fs_descriptors; +for (; *d; ++d) { +struct usb_endpoint_descriptor *ep; +int addr; + +if ((*d)->bDescriptorType !=USB_DT_ENDPOINT ) +continue ; + + ep = (struct usb_endpoint_descriptor *)*d; + addr = ((ep->bEndpointAddress &0x80 ) >>3 ) | + (ep->bEndpointAddress &0x0f ); +set_bit (addr, &ep_map); + } + } } + ret =usb_gadget_check_config (cdev->gadget, ep_map); +if (ret) + goto err_purge_funcs; +usb_ep_autoconfig_reset (cdev->gadget); }if (cdev->use_os_string) {
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
PATCH v9 3/5 usb: dwc3: Resize TX FIFOs to meet EP bursting requirements
In reply to Wesley ChengSome devices have USB compositions which may require multiple endpoints that support EP bursting. HW defined TX FIFO sizes may not always be sufficient for these compositions. By utilizing flexible TX FIFO allocation, this allows for endpoints to request the required FIFO depth to achieve higher bandwidth. With some higher bMaxBurst configurations, using a larger TX FIFO size results in better TX throughput.
By introducing the check_config() callback, the resizing logic can fetch the maximum number of endpoints used in the USB composition (can contain multiple configurations), which helps ensure that the resizing logic can fulfill the configuration(s), or return an error to the gadget layer otherwise during bind time.
--- drivers/usb/dwc3/core.c | 9 ++ drivers/usb/dwc3/core.h | 15 ++++ drivers/usb/dwc3/ep0.c | 2 + drivers/usb/dwc3/gadget.c | 212 ++++++++++++++++++++++++++++++++++++++++++++++ 4 files changed, 238 insertions(+)
drivers/usb/dwc3/core.c +9 -0
@@ -1267,6 +1267,7 @@ static void dwc3_get_properties(struct dwc3 *dwc)u8 rx_max_burst_prd;u8 tx_thr_num_pkt_prd;u8 tx_max_burst_prd; +u8 tx_fifo_resize_max_num;const char *usb_psy_name;int ret;
@@ -1282,6 +1283,8 @@ static void dwc3_get_properties(struct dwc3 *dwc)*/ hird_threshold =12 ; + tx_fifo_resize_max_num =6 ; + dwc->maximum_speed =usb_get_maximum_speed (dev); dwc->max_ssp_rate =usb_get_maximum_ssp_rate (dev); dwc->dr_mode =usb_get_dr_mode (dev);
@@ -1325,6 +1328,10 @@ static void dwc3_get_properties(struct dwc3 *dwc) &tx_thr_num_pkt_prd);device_property_read_u8 (dev,"snps,tx-max-burst-prd" , &tx_max_burst_prd); + dwc->do_fifo_resize =device_property_read_bool (dev, +"tx-fifo-resize" ); +device_property_read_u8 (dev,"tx-fifo-max-num" , + &tx_fifo_resize_max_num); dwc->disable_scramble_quirk =device_property_read_bool (dev,"snps,disable_scramble_quirk" );
@@ -1390,6 +1397,8 @@ static void dwc3_get_properties(struct dwc3 *dwc) dwc->tx_max_burst_prd = tx_max_burst_prd; dwc->imod_interval =0 ; + + dwc->tx_fifo_resize_max_num = tx_fifo_resize_max_num; }/* check whether the core supports IMOD */
drivers/usb/dwc3/core.h +15 -0
@@ -1023,6 +1023,7 @@ struct dwc3_scratchpad_array {* @rx_max_burst_prd: max periodic ESS receive burst size * @tx_thr_num_pkt_prd: periodic ESS transmit packet count * @tx_max_burst_prd: max periodic ESS transmit burst size +* @tx_fifo_resize_max_num: max number of fifos allocated during txfifo resize * @hsphy_interface: "utmi" or "ulpi" * @connected: true when we're connected to a host, false otherwise * @delayed_status: true when gadget driver asks for delayed status
@@ -1079,6 +1080,11 @@ struct dwc3_scratchpad_array {* @dis_split_quirk: set to disable split boundary. * @imod_interval: set the interrupt moderation interval in 250ns * increments or 0 to disable. +* @max_cfg_eps: current max number of IN eps used across all USB configs. +* @last_fifo_depth: last fifo depth used to determine next fifo ram start +* address. +* @num_ep_resized: carries the current number endpoints which have had its tx +* fifo resized. */ struct dwc3 {struct work_struct drd_work;
@@ -1234,6 +1240,7 @@ struct dwc3 {u8 rx_max_burst_prd;u8 tx_thr_num_pkt_prd;u8 tx_max_burst_prd; +u8 tx_fifo_resize_max_num;const char *hsphy_interface;
@@ -1247,6 +1254,7 @@ struct dwc3 { unsigned is_utmi_l1_suspend:1; unsigned is_fpga:1; unsigned pending_events:1; + unsigned do_fifo_resize:1; unsigned pullups_connected:1; unsigned setup_packet_pending:1; unsigned three_stage_setup:1;
@@ -1282,6 +1290,10 @@ struct dwc3 { unsigned dis_split_quirk:1;u16 imod_interval; + +int max_cfg_eps; +int last_fifo_depth; +int num_ep_resized; };#define INCRX_BURST_MODE 0
@@ -1513,6 +1525,7 @@ int dwc3_send_gadget_ep_cmd(struct dwc3_ep *dep, unsigned int cmd, struct dwc3_gadget_ep_cmd_params *params);int dwc3_send_gadget_generic_command (struct dwc3 *dwc,unsigned int cmd,u32 param); +void dwc3_gadget_clear_tx_fifos (struct dwc3 *dwc);#else static inline int dwc3_gadget_init (struct dwc3 *dwc) {return 0 ; }
@@ -1532,6 +1545,8 @@ static inline int dwc3_send_gadget_ep_cmd(struct dwc3_ep *dep, unsigned int cmd,static inline int dwc3_send_gadget_generic_command (struct dwc3 *dwc,int cmd,u32 param) {return 0 ; } +static inline void dwc3_gadget_clear_tx_fifos (struct dwc3 *dwc) +{ }#endif #if IS_ENABLED (CONFIG_USB_DWC3_DUAL_ROLE )
drivers/usb/dwc3/ep0.c +2 -0
@@ -616,6 +616,8 @@ static int dwc3_ep0_set_config(struct dwc3 *dwc, struct usb_ctrlrequest *ctrl)return -EINVAL ;case USB_STATE_ADDRESS : +dwc3_gadget_clear_tx_fifos (dwc); + ret =dwc3_ep0_delegate_req (dwc, ctrl);/* if the cfg matches and the cfg is non zero */ if (cfg && (!ret || (ret ==USB_GADGET_DELAYED_STATUS ))) {
drivers/usb/dwc3/gadget.c +212 -0
242 lines not on this page · raw message
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
PATCH v9 4/5 usb: dwc3: dwc3-qcom: Enable tx-fifo-resize property by default
In reply to Wesley ChengIn order to take advantage of the TX fifo resizing logic, manually add these properties to the DWC3 child node by default. This will allow the DWC3 gadget to resize the TX fifos for the IN endpoints, which help with performance.
--- drivers/usb/dwc3/dwc3-qcom.c | 9 +++++++++ 1 file changed, 9 insertions(+)
drivers/usb/dwc3/dwc3-qcom.c +9 -0
@@ -645,6 +645,7 @@ static int dwc3_qcom_of_register_core(struct platform_device *pdev)struct dwc3_qcom *qcom =platform_get_drvdata (pdev);struct device_node *np = pdev->dev.of_node, *dwc3_np;struct device *dev = &pdev->dev; +struct property *prop;int ret; dwc3_np =of_get_compatible_child (np,"snps,dwc3" );
@@ -653,6 +654,14 @@ static int dwc3_qcom_of_register_core(struct platform_device *pdev)return -ENODEV ; } + prop =devm_kzalloc (dev,sizeof (*prop),GFP_KERNEL ); +if (prop) { + prop->name ="tx-fifo-resize" ; + ret =of_add_property (dwc3_np, prop); +if (ret <0 ) +dev_info (dev,"unable to add tx-fifo-resize prop\n" ); + } + ret =of_platform_populate (np,NULL ,NULL , dev);if (ret) {dev_err (dev,"failed to register dwc3 core - %d\n" , ret);
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
PATCH v9 5/5 dt-bindings: usb: dwc3: Update dwc3 TX fifo properties
In reply to Wesley ChengUpdate the tx-fifo-resize property with a better description, while adding the tx-fifo-max-num, which is a new parameter allowing adjustments for the maximum number of packets the txfifo resizing logic can account for while resizing the endpoints.
--- Documentation/devicetree/bindings/usb/snps,dwc3.yaml | 15 +++++++++++++-- 1 file changed, 13 insertions(+), 2 deletions(-)
Documentation/devicetree/bindings/usb/snps,dwc3.yaml +13 -2
@@ -289,10 +289,21 @@ properties: maximum: 16 tx-fifo-resize: - description: Determines if the FIFO *has* to be reallocated - deprecated: true + description: Determines if the TX fifos can be dynamically resized depending + on the number of IN endpoints used and if bursting is supported. This + may help improve bandwidth on platforms with higher system latencies, as + increased fifo space allows for the controller to prefetch data into its + internal memory. type: boolean + tx-fifo-max-num: + description: Specifies the max number of packets the txfifo resizing logic + can account for when higher endpoint bursting is used. (bMaxBurst > 6) The + higher the number, the more fifo space the txfifo resizing logic will + allocate for that endpoint. + $ref: /schemas/types.yaml#/definitions/uint8 + minimum: 3 + snps,incr-burst-type-adjustment: description: Value for INCR burst type of GSBUSCFG0 register, undefined length INCR
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
Greg KH
maintainerQuoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.
This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]
653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0);
| ~~~~~~~~~~~~~^~~~~~~~~~
| |
| u32 {aka unsigned int}
In file included from drivers/usb/dwc3/debug.h:14,
from drivers/usb/dwc3/gadget.c:25:
drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’}
1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc)
| ~~~~~~~~~~~~~^~~drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]
660 | ret = of_add_property(dwc3_np, prop);
| ^~~~~~~~~~~~~~~
| of_get_propertyHow did you test these?
thanks,
greg k-h
Hi,
Quoted 20 lines · Greg KH
Quoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0); | ~~~~~~~~~~~~~^~~~~~~~~~ | | | u32 {aka unsigned int} In file included from drivers/usb/dwc3/debug.h:14, from drivers/usb/dwc3/gadget.c:25: drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’} 1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc) | ~~~~~~~~~~~~~^~~
Quoted 8 lines
drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]660 | ret = of_add_property(dwc3_np, prop); | ^~~~~~~~~~~~~~~ | of_get_propertyHow did you test these?
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
I still think this may cause other places to break down. The promise the databook makes is that increasing the FIFO size over 2x wMaxPacketSize should bring little to no benefit, if we're not maintaining that, I wonder if the problem is with some of the BUSCFG registers instead, where we configure interconnect bursting and the like.
Signature · 1 line
balbi
Greg KH
maintainerQuoted 46 lines · Felipe Balbi
Hi,
Quoted 20 lines · Greg KH
Quoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0); | ~~~~~~~~~~~~~^~~~~~~~~~ | | | u32 {aka unsigned int} In file included from drivers/usb/dwc3/debug.h:14, from drivers/usb/dwc3/gadget.c:25: drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’} 1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc) | ~~~~~~~~~~~~~^~~Quoted 8 lines
drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]660 | ret = of_add_property(dwc3_np, prop); | ^~~~~~~~~~~~~~~ | of_get_propertyHow did you test these?
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
Quoted 5 lines
I still think this may cause other places to break down. The promise the databook makes is that increasing the FIFO size over 2x wMaxPacketSize should bring little to no benefit, if we're not maintaining that, I wonder if the problem is with some of the BUSCFG registers instead, where we configure interconnect bursting and the like.
Good points.
Wesley, what kind of testing have you done on this on different devices?
thanks,
greg k-h
Hi Greg/Felipe,
Quoted 25 lines · Greg KH
Quoted 24 lines · Felipe Balbi
Hi,
Quoted 20 lines · Greg KH
Quoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0); | ~~~~~~~~~~~~~^~~~~~~~~~ | | | u32 {aka unsigned int} In file included from drivers/usb/dwc3/debug.h:14, from drivers/usb/dwc3/gadget.c:25: drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’} 1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc) | ~~~~~~~~~~~~~^~~
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]660 | ret = of_add_property(dwc3_np, prop); | ^~~~~~~~~~~~~~~ | of_get_propertyHow did you test these?
I ran these changes on our internal branches, which were probably missing some of the recent changes done to the DWC3 drivers. Will fix the above compile errors and re-submit.
In regards to how much these changes have been tested, we've been maintaining the TX FIFO resize logic downstream for a few years already, so its being used in end products. We also verify this with our internal testing, which has certain benchmarks we need to meet.
Quoted 8 lines
Quoted 8 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
Quoted 6 lines
Quoted 5 lines
I still think this may cause other places to break down. The promise the databook makes is that increasing the FIFO size over 2x wMaxPacketSize should bring little to no benefit, if we're not maintaining that, I wonder if the problem is with some of the BUSCFG registers instead, where we configure interconnect bursting and the like.
I've been referring mainly to the DWC3 programming guide for recommendations on how to improve USB performance in:
Section 3.3.5 System Bus Features to Improve USB Performance
At least when I ran the initial profiling, adjusting the RX/TX thresholds brought little to no benefits. Even in some of the examples, they have diagrams showing a TXFIFO size of 6 max packets (Figure 3-5).
I think its difficult to say that the TX fifo resizing won't help in
systems with limited, or shared resources where the bus latencies would be somewhat larger. By adjusting the TX FIFO size, the controller would be able to fetch more data from system memory into the memory within the controller, leading to less frequent end of bursts, etc... as data is readily available.
In terms of adjusting the AXI/AHB bursting, I would think the bandwidth increase would eventually be constrained based on your system's design.
We don't touch the GSBUSCFG registers, and leave them as is based off
the recommendations from the HW designers.
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
Thanks
Wesley Cheng
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
Hi,
Quoted 37 lines · Wesley Cheng
Hi Greg/Felipe,
Quoted 25 lines · Greg KH
Quoted 24 lines · Felipe Balbi
Hi,
Quoted 20 lines · Greg KH
Quoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0); | ~~~~~~~~~~~~~^~~~~~~~~~ | | | u32 {aka unsigned int} In file included from drivers/usb/dwc3/debug.h:14, from drivers/usb/dwc3/gadget.c:25: drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’} 1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc) | ~~~~~~~~~~~~~^~~Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]660 | ret = of_add_property(dwc3_np, prop); | ^~~~~~~~~~~~~~~ | of_get_propertyHow did you test these?
Quoted 8 lines
I ran these changes on our internal branches, which were probably missing some of the recent changes done to the DWC3 drivers. Will fix the above compile errors and re-submit.
In regards to how much these changes have been tested, we've been maintaining the TX FIFO resize logic downstream for a few years already, so its being used in end products. We also verify this with our internal testing, which has certain benchmarks we need to meet.
the problem with that is that you *know* which gadget is running there. You know everyone of those is going to run the android gadget. In a sense, all those multiple products are testing the same exact use case :-)
Quoted 14 lines
Quoted 7 lines
Quoted 7 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
Quoted 7 lines
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
The feature ends up being all-or-nothing, then :-) It sounds like we can be a little nicer in this regard.
Quoted 10 lines
Quoted 6 lines
Quoted 5 lines
I still think this may cause other places to break down. The promise the databook makes is that increasing the FIFO size over 2x wMaxPacketSize should bring little to no benefit, if we're not maintaining that, I wonder if the problem is with some of the BUSCFG registers instead, where we configure interconnect bursting and the like.
I've been referring mainly to the DWC3 programming guide for recommendations on how to improve USB performance in:
Section 3.3.5 System Bus Features to Improve USB Performance
dwc3 or dwc3.1? Either way, since I left Intel I don't have access to the databook anymore. I have to trust what you guys are telling me and, based on the description so far, I don't think we're doing the right thing (yet).
It would be nice if other users would test this patchset with different gadget drivers and different platforms to have some confidence that we're limiting possible regressions.
I would like for Thinh to comment from Synopsys side here.
Quoted 2 lines
At least when I ran the initial profiling, adjusting the RX/TX thresholds brought little to no benefits. Even in some of the examples,
right, the FIFO sizes shouldn't help much. At least that's what Paul told me several years ago. Thinh, has the recommendation changed?
Quoted 8 lines
they have diagrams showing a TXFIFO size of 6 max packets (Figure 3-5).
I think its difficult to say that the TX fifo resizing won't help in
systems with limited, or shared resources where the bus latencies would be somewhat larger. By adjusting the TX FIFO size, the controller would be able to fetch more data from system memory into the memory within the controller, leading to less frequent end of bursts, etc... as data is readily available.
Quoted 4 lines
In terms of adjusting the AXI/AHB bursting, I would think the bandwidth increase would eventually be constrained based on your system's design.
We don't touch the GSBUSCFG registers, and leave them as is based off
the recommendations from the HW designers.
Right, I want to touch those as little as possible too :-) However, to illustrate, the only reason I implemented FIFO resizing was because OMAP5 ES1 had TX FIFOs that were smaller than a full USB3 packet. HW Designer's recommendation can be bogus too ;-)
Quoted 7 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
Signature · 1 line
balbi
Greg KH
maintainerQuoted 6 lines · Felipe Balbi
Quoted 2 lines
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I used to think that too, until I saw some of the new crazy designs where lots of SoC connections internally to the device run on USB. Also look at the USB offload stuff as one example of how the voice sound path goes through the USB controller on the SoC on the latest Samsung Galaxy phones that are shipping now :(
There's also devices with the modem/network connection going over USB, along with other device types as well. Android Auto is crazy with almost everything hooked up directly with a USB connection to the host system running Linux.
Quoted 2 lines
I guess only developers are using USB during development to flash dev images heh.
I wish, we are reaching the point where the stability of the overall Android system depends on how well the USB controller works. We are a product of our success...
thanks,
greg k-h
Hi,
Quoted 8 lines · Greg KH
Quoted 6 lines · Felipe Balbi
Quoted 2 lines
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
Quoted 5 lines
I used to think that too, until I saw some of the new crazy designs where lots of SoC connections internally to the device run on USB. Also look at the USB offload stuff as one example of how the voice sound path goes through the USB controller on the SoC on the latest Samsung Galaxy phones that are shipping now :(
yeah, that's one reason NOT to touch the FIFO sizes :-) OMAP5 has, as mentioned before, processor trace offload via USB too. If we modify the FIFO configuration set by the HW designer we risk loosing those features.
Quoted 2 lines
There's also devices with the modem/network connection going over USB, along with other device types as well. Android Auto is crazy with
yeah, and there will be more coming. USB Debug class is already integrated in some SoCs, that gives us jtag-like access over USB.
Quoted 2 lines
almost everything hooked up directly with a USB connection to the host system running Linux.
that's running against USB host, though, right? Android is the host, not the gadget :-)
The FIFO sizes here are for the gadget side.
Quoted 10 lines
Quoted 2 lines
I guess only developers are using USB during development to flash dev images heh.
I wish, we are reaching the point where the stability of the overall Android system depends on how well the USB controller works. We are a product of our success...
thanks,
greg k-h
Signature · 1 line
balbi
Hi Felipe,
Quoted 51 lines · Felipe Balbi
Hi,
Quoted 37 lines · Wesley Cheng
Hi Greg/Felipe,
Quoted 25 lines · Greg KH
Quoted 24 lines · Felipe Balbi
Hi,
Quoted 20 lines · Greg KH
Quoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0); | ~~~~~~~~~~~~~^~~~~~~~~~ | | | u32 {aka unsigned int} In file included from drivers/usb/dwc3/debug.h:14, from drivers/usb/dwc3/gadget.c:25: drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’} 1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc) | ~~~~~~~~~~~~~^~~Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]660 | ret = of_add_property(dwc3_np, prop); | ^~~~~~~~~~~~~~~ | of_get_propertyHow did you test these?
Quoted 8 lines
I ran these changes on our internal branches, which were probably missing some of the recent changes done to the DWC3 drivers. Will fix the above compile errors and re-submit.
In regards to how much these changes have been tested, we've been maintaining the TX FIFO resize logic downstream for a few years already, so its being used in end products. We also verify this with our internal testing, which has certain benchmarks we need to meet.
Quoted 5 lines
the problem with that is that you *know* which gadget is running there. You know everyone of those is going to run the android gadget. In a sense, all those multiple products are testing the same exact use case :-)
Mmmm, the USB gadget has changed from since we've implemented it, such as going from Android gadget to Configfs. Don't forget, we do have other business segments that use this feature in other configurations as well :).
Quoted 22 lines
Quoted 14 lines
Quoted 7 lines
Quoted 7 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
Quoted 7 lines
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
Quoted 3 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can be a little nicer in this regard.
Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking at Andy Shevchenko's Github, the Intel tracer downstream changes were just to remove physical EP1 and 2 from the DWC3 endpoint list. If that was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs. However, I'm not sure how the changes would look like in the end, so I would like to wait later down the line to include that :).
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
Quoted 5 lines
I still think this may cause other places to break down. The promise the databook makes is that increasing the FIFO size over 2x wMaxPacketSize should bring little to no benefit, if we're not maintaining that, I wonder if the problem is with some of the BUSCFG registers instead, where we configure interconnect bursting and the like.
Quoted 9 lines
Quoted 3 lines
I've been referring mainly to the DWC3 programming guide for recommendations on how to improve USB performance in:
Section 3.3.5 System Bus Features to Improve USB Performancedwc3 or dwc3.1? Either way, since I left Intel I don't have access to the databook anymore. I have to trust what you guys are telling me and, based on the description so far, I don't think we're doing the right thing (yet).
Ah, I see. DWC3.1 and DWC3 both have that USB performance section. I can explain some of the points I made with a bit more detail. I thought you still had access to it.
Quoted 4 lines
It would be nice if other users would test this patchset with different gadget drivers and different platforms to have some confidence that we're limiting possible regressions.
Quoted 8 lines
I would like for Thinh to comment from Synopsys side here.
Quoted 2 lines
At least when I ran the initial profiling, adjusting the RX/TX thresholds brought little to no benefits. Even in some of the examples,
right, the FIFO sizes shouldn't help much. At least that's what Paul told me several years ago. Thinh, has the recommendation changed?
So when I mention the RX/TX thresholds, this is different than the FIFO resize. The RX/TX threshold is used by the controller to determine when to send or receive data based on the number of available FIFOs. So for the TX case, if we set the TX threshold, the controller will not start transmitting data over the link after X amount of packets are copied to the TXFIFO. So for example, a TXFIFO size of 6 w/ a TX threshold of 3, means that the controller will wait for 3 FIFO slots to be filled before it sends the data. So as you can see, with our configuration of TX FIFO size of 2 and TX threshold of 1, this would really be not beneficial to us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.
Quoted 13 lines
Quoted 8 lines
they have diagrams showing a TXFIFO size of 6 max packets (Figure 3-5).
I think its difficult to say that the TX fifo resizing won't help in
systems with limited, or shared resources where the bus latencies would be somewhat larger. By adjusting the TX FIFO size, the controller would be able to fetch more data from system memory into the memory within the controller, leading to less frequent end of bursts, etc... as data is readily available.Quoted 4 lines
In terms of adjusting the AXI/AHB bursting, I would think the bandwidth increase would eventually be constrained based on your system's design.
We don't touch the GSBUSCFG registers, and leave them as is based off
the recommendations from the HW designers.
Quoted 5 lines
Right, I want to touch those as little as possible too :-) However, to illustrate, the only reason I implemented FIFO resizing was because OMAP5 ES1 had TX FIFOs that were smaller than a full USB3 packet. HW Designer's recommendation can be bogus too ;-)
Haha...true, we question their designs only when there's something clearly wrong, but the AXI/AHB settings look good. :)
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
Quoted 7 lines
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a few years as well. The TXFIFO sizing really has helped with multimedia use cases, which use isoc endpoints, since esp. in those lower end CPU chips where latencies across the system are much larger, and a missed ISOC interval leads to a pop in your ear.
Thanks
Wesley Cheng
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
Hi,
Quoted 45 lines · Wesley Cheng
Quoted 39 lines
Quoted 30 lines
Quoted 21 lines
Quoted 21 lines
Quoted 20 lines · Greg KH
Quoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0); | ~~~~~~~~~~~~~^~~~~~~~~~ | | | u32 {aka unsigned int} In file included from drivers/usb/dwc3/debug.h:14, from drivers/usb/dwc3/gadget.c:25: drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’} 1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc) | ~~~~~~~~~~~~~^~~Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]660 | ret = of_add_property(dwc3_np, prop); | ^~~~~~~~~~~~~~~ | of_get_propertyHow did you test these?
Quoted 8 lines
I ran these changes on our internal branches, which were probably missing some of the recent changes done to the DWC3 drivers. Will fix the above compile errors and re-submit.
In regards to how much these changes have been tested, we've been maintaining the TX FIFO resize logic downstream for a few years already, so its being used in end products. We also verify this with our internal testing, which has certain benchmarks we need to meet.
Quoted 5 lines
the problem with that is that you *know* which gadget is running there. You know everyone of those is going to run the android gadget. In a sense, all those multiple products are testing the same exact use case :-)
Quoted 4 lines
Mmmm, the USB gadget has changed from since we've implemented it, such as going from Android gadget to Configfs. Don't forget, we do have other business segments that use this feature in other configurations as well :).
:)
Quoted 22 lines
Quoted 22 lines
Quoted 21 lines
Quoted 7 lines
Quoted 7 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
Quoted 6 lines
Quoted 3 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can be a little nicer in this regard.
Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
sure, I support that. But I want to make sure the first cut isn't likely to break things left and right :)
Hence, let's at least get more testing.
Quoted 2 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were just to remove physical EP1 and 2 from the DWC3 endpoint list. If that
right, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to skip that endpoint on that particular platform without interefering with everything else.
Of course, we still need to figure out a way to abstract the different dwc3 instantiations.
Quoted 2 lines
was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph, will be coming up with a solution that's elegant and works for all different instantiations of dwc3 (or musb, cdns3, etc).
Quoted 2 lines
However, I'm not sure how the changes would look like in the end, so I would like to wait later down the line to include that :).
Fair enough, I agree. Can we get some more testing of $subject, though? Did you test $subject with upstream too? Which gadget drivers did you use? How did you test
Quoted 17 lines
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
Quoted 5 lines
I still think this may cause other places to break down. The promise the databook makes is that increasing the FIFO size over 2x wMaxPacketSize should bring little to no benefit, if we're not maintaining that, I wonder if the problem is with some of the BUSCFG registers instead, where we configure interconnect bursting and the like.
Quoted 9 lines
Quoted 3 lines
I've been referring mainly to the DWC3 programming guide for recommendations on how to improve USB performance in:
Section 3.3.5 System Bus Features to Improve USB Performancedwc3 or dwc3.1? Either way, since I left Intel I don't have access to the databook anymore. I have to trust what you guys are telling me and, based on the description so far, I don't think we're doing the right thing (yet).
Quoted 3 lines
Ah, I see. DWC3.1 and DWC3 both have that USB performance section. I can explain some of the points I made with a bit more detail. I thought you still had access to it.
I wish :)
If Synopsys wants to give me access for the databook, I would not mind :-)
Quoted 9 lines
Quoted 9 lines
It would be nice if other users would test this patchset with different gadget drivers and different platforms to have some confidence that we're limiting possible regressions.
I would like for Thinh to comment from Synopsys side here.
Quoted 2 lines
At least when I ran the initial profiling, adjusting the RX/TX thresholds brought little to no benefits. Even in some of the examples,
Quoted 7 lines
Quoted 3 lines
right, the FIFO sizes shouldn't help much. At least that's what Paul told me several years ago. Thinh, has the recommendation changed?
So when I mention the RX/TX thresholds, this is different than the FIFO resize. The RX/TX threshold is used by the controller to determine when to send or receive data based on the number of available FIFOs. So for
oh right, I remember now :-
Quoted 5 lines
the TX case, if we set the TX threshold, the controller will not start transmitting data over the link after X amount of packets are copied to the TXFIFO. So for example, a TXFIFO size of 6 w/ a TX threshold of 3, means that the controller will wait for 3 FIFO slots to be filled before it sends the data. So as you can see, with our configuration of TX FIFO
yeah, makes sense.
Quoted 5 lines
size of 2 and TX threshold of 1, this would really be not beneficial to us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.
What I noticed with g_mass_storage is that we can amortize the cost of fetching data from memory, with a deeper request queue. Whenever I test(ed) g_mass_storage, I was doing so with 250 requests. And that was enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?
I feel that allocating more requests is a far simpler and more generic method that changing FIFO sizes :)
Quoted 13 lines
Quoted 13 lines
Quoted 12 lines
they have diagrams showing a TXFIFO size of 6 max packets (Figure 3-5).
I think its difficult to say that the TX fifo resizing won't help in
systems with limited, or shared resources where the bus latencies would be somewhat larger. By adjusting the TX FIFO size, the controller would be able to fetch more data from system memory into the memory within the controller, leading to less frequent end of bursts, etc... as data is readily available.In terms of adjusting the AXI/AHB bursting, I would think the bandwidth increase would eventually be constrained based on your system's design.
We don't touch the GSBUSCFG registers, and leave them as is based off
the recommendations from the HW designers.
Quoted 8 lines
Quoted 5 lines
Right, I want to touch those as little as possible too :-) However, to illustrate, the only reason I implemented FIFO resizing was because OMAP5 ES1 had TX FIFOs that were smaller than a full USB3 packet. HW Designer's recommendation can be bogus too ;-)
Haha...true, we question their designs only when there's something clearly wrong, but the AXI/AHB settings look good. :)
:)
Quoted 16 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
Quoted 7 lines
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
Quoted 7 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a few years as well. The TXFIFO sizing really has helped with multimedia use cases, which use isoc endpoints, since esp. in those lower end CPU chips where latencies across the system are much larger, and a missed ISOC interval leads to a pop in your ear.
This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Remember dwc3 can accomodate 255 requests + link for each endpoint. If our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.
Signature · 1 line
balbi
Hi Felipe,
Quoted 82 lines · Felipe Balbi
Hi,
Quoted 45 lines · Wesley Cheng
Quoted 39 lines
Quoted 30 lines
Quoted 21 lines
Quoted 21 lines
Quoted 20 lines · Greg KH
Quoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0); | ~~~~~~~~~~~~~^~~~~~~~~~ | | | u32 {aka unsigned int} In file included from drivers/usb/dwc3/debug.h:14, from drivers/usb/dwc3/gadget.c:25: drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’} 1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc) | ~~~~~~~~~~~~~^~~Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]660 | ret = of_add_property(dwc3_np, prop); | ^~~~~~~~~~~~~~~ | of_get_propertyHow did you test these?
Quoted 8 lines
I ran these changes on our internal branches, which were probably missing some of the recent changes done to the DWC3 drivers. Will fix the above compile errors and re-submit.
In regards to how much these changes have been tested, we've been maintaining the TX FIFO resize logic downstream for a few years already, so its being used in end products. We also verify this with our internal testing, which has certain benchmarks we need to meet.
Quoted 5 lines
the problem with that is that you *know* which gadget is running there. You know everyone of those is going to run the android gadget. In a sense, all those multiple products are testing the same exact use case :-)
Quoted 4 lines
Mmmm, the USB gadget has changed from since we've implemented it, such as going from Android gadget to Configfs. Don't forget, we do have other business segments that use this feature in other configurations as well :).
:)
Quoted 26 lines
Quoted 25 lines
Quoted 14 lines
Quoted 7 lines
Quoted 7 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
Quoted 7 lines
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
The feature ends up being all-or-nothing, then :-) It sounds like we can be a little nicer in this regard.
Quoted 8 lines
Quoted 2 lines
Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
sure, I support that. But I want to make sure the first cut isn't likely to break things left and right :)
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty big improvements on the TX path with this.
Quoted 13 lines
Quoted 2 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were just to remove physical EP1 and 2 from the DWC3 endpoint list. If that
right, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to skip that endpoint on that particular platform without interefering with everything else.
Of course, we still need to figure out a way to abstract the different dwc3 instantiations.
Quoted 7 lines
Quoted 2 lines
was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph, will be coming up with a solution that's elegant and works for all different instantiations of dwc3 (or musb, cdns3, etc).
Well, at least for the TX FIFO resizing logic, we'd only be needing to focus on the DWC3 implementation.
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We have a few special HW eps as well, which we try to maintain separately in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).
Quoted 7 lines
Quoted 2 lines
However, I'm not sure how the changes would look like in the end, so I would like to wait later down the line to include that :).
Fair enough, I agree. Can we get some more testing of $subject, though? Did you test $subject with upstream too? Which gadget drivers did you use? How did you test
The results that I included in the cover page was tested with the pure upstream kernel on our device. Below was using the ConfigFS gadget w/ a mass storage only composition.
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMark
Results:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 193.60
| | 195.86
| | 184.77
| | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 287.35
| | 304.94
| | 289.64
| | 293.61-------------------------------------------
We also have internal numbers which have shown similar improvements as well. Those are over networking/tethering interfaces, so testing IPERF loopback over TCP/UDP.
Quoted 58 lines
Quoted 17 lines
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
Quoted 5 lines
I still think this may cause other places to break down. The promise the databook makes is that increasing the FIFO size over 2x wMaxPacketSize should bring little to no benefit, if we're not maintaining that, I wonder if the problem is with some of the BUSCFG registers instead, where we configure interconnect bursting and the like.
Quoted 9 lines
Quoted 3 lines
I've been referring mainly to the DWC3 programming guide for recommendations on how to improve USB performance in:
Section 3.3.5 System Bus Features to Improve USB Performancedwc3 or dwc3.1? Either way, since I left Intel I don't have access to the databook anymore. I have to trust what you guys are telling me and, based on the description so far, I don't think we're doing the right thing (yet).
Quoted 3 lines
Ah, I see. DWC3.1 and DWC3 both have that USB performance section. I can explain some of the points I made with a bit more detail. I thought you still had access to it.
I wish :)
If Synopsys wants to give me access for the databook, I would not mind :-)
Quoted 9 lines
Quoted 9 lines
It would be nice if other users would test this patchset with different gadget drivers and different platforms to have some confidence that we're limiting possible regressions.
I would like for Thinh to comment from Synopsys side here.
Quoted 2 lines
At least when I ran the initial profiling, adjusting the RX/TX thresholds brought little to no benefits. Even in some of the examples,
Quoted 7 lines
Quoted 3 lines
right, the FIFO sizes shouldn't help much. At least that's what Paul told me several years ago. Thinh, has the recommendation changed?
So when I mention the RX/TX thresholds, this is different than the FIFO resize. The RX/TX threshold is used by the controller to determine when to send or receive data based on the number of available FIFOs. So for
oh right, I remember now :-
Quoted 5 lines
the TX case, if we set the TX threshold, the controller will not start transmitting data over the link after X amount of packets are copied to the TXFIFO. So for example, a TXFIFO size of 6 w/ a TX threshold of 3, means that the controller will wait for 3 FIFO slots to be filled before it sends the data. So as you can see, with our configuration of TX FIFO
yeah, makes sense.
Quoted 5 lines
size of 2 and TX threshold of 1, this would really be not beneficial to us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.
Quoted 9 lines
What I noticed with
g_mass_storageis that we can amortize the cost of fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?I feel that allocating more requests is a far simpler and more generic method that changing FIFO sizes :)
I wish I had a USB bus trace handy to show you, which would make it very clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So by increasing the number of USB requests, that will help if there was a bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.
So yes, this method of increasing the # of USB reqs will definitely help with situations such as HSUSB or in SSUSB when EP bursting isn't used. The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet, it should have a NumP value equal to the bMaxBurst reported in the EP desc. If we have a TXFIFO size of 2, then normally what I have seen is that after 2 data packets, the device issues a NRDY. So then we'd need to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With this constant NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant bursts for a request, until the request is done, or if the host runs out of RXFIFO. (ie no interruption [on the USB protocol level] during USB request data transfer)
Quoted 48 lines
Quoted 13 lines
Quoted 13 lines
Quoted 12 lines
they have diagrams showing a TXFIFO size of 6 max packets (Figure 3-5).
I think its difficult to say that the TX fifo resizing won't help in
systems with limited, or shared resources where the bus latencies would be somewhat larger. By adjusting the TX FIFO size, the controller would be able to fetch more data from system memory into the memory within the controller, leading to less frequent end of bursts, etc... as data is readily available.In terms of adjusting the AXI/AHB bursting, I would think the bandwidth increase would eventually be constrained based on your system's design.
We don't touch the GSBUSCFG registers, and leave them as is based off
the recommendations from the HW designers.Quoted 8 lines
Quoted 5 lines
Right, I want to touch those as little as possible too :-) However, to illustrate, the only reason I implemented FIFO resizing was because OMAP5 ES1 had TX FIFOs that were smaller than a full USB3 packet. HW Designer's recommendation can be bogus too ;-)
Haha...true, we question their designs only when there's something clearly wrong, but the AXI/AHB settings look good. :)
:)
Quoted 16 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
Quoted 7 lines
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
Quoted 7 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a few years as well. The TXFIFO sizing really has helped with multimedia use cases, which use isoc endpoints, since esp. in those lower end CPU chips where latencies across the system are much larger, and a missed ISOC interval leads to a pop in your ear.
Quoted 8 lines
This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Remember dwc3 can accomodate 255 requests + link for each endpoint. If our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.
We're actually using both a deeper USB request queue + TX fifo resizing. :).
Thanks
Wesley Cheng
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
Hi,
Quoted 29 lines · Wesley Cheng
Quoted 29 lines
Quoted 28 lines
Quoted 22 lines
Quoted 14 lines
Quoted 7 lines
Quoted 7 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
Quoted 7 lines
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
Quoted 3 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can be a little nicer in this regard.
Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
Quoted 8 lines
Quoted 5 lines
sure, I support that. But I want to make sure the first cut isn't likely to break things left and right :)
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty big improvements on the TX path with this.
fingers crossed
Quoted 24 lines
Quoted 13 lines
Quoted 2 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were just to remove physical EP1 and 2 from the DWC3 endpoint list. If that
right, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to skip that endpoint on that particular platform without interefering with everything else.
Of course, we still need to figure out a way to abstract the different dwc3 instantiations.
Quoted 7 lines
Quoted 2 lines
was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph, will be coming up with a solution that's elegant and works for all different instantiations of dwc3 (or musb, cdns3, etc).
Well, at least for the TX FIFO resizing logic, we'd only be needing to focus on the DWC3 implementation.
Quoted 6 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We have a few special HW eps as well, which we try to maintain separately in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).
Awesome, as mentioned, the endpoint feature flags were added exactly to allow for these vendor-specific features :-)
I'm more than happy to help testing now that I finally got our SM8150 Surface Duo device tree accepted by Bjorn ;-)
Quoted 12 lines
Quoted 7 lines
Quoted 2 lines
However, I'm not sure how the changes would look like in the end, so I would like to wait later down the line to include that :).
Fair enough, I agree. Can we get some more testing of $subject, though? Did you test $subject with upstream too? Which gadget drivers did you use? How did you test
The results that I included in the cover page was tested with the pure upstream kernel on our device. Below was using the ConfigFS gadget w/ a mass storage only composition.
Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16
do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 193.60 | | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 287.35 | | 304.94 | | 289.64 | | 293.61
I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my memory could be failing.
Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 3 lines
We also have internal numbers which have shown similar improvements as well. Those are over networking/tethering interfaces, so testing IPERF loopback over TCP/UDP.
loopback iperf? That would skip the wire, no?
Quoted 23 lines
Quoted 6 lines
Quoted 5 lines
size of 2 and TX threshold of 1, this would really be not beneficial to us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.
Quoted 9 lines
What I noticed with
g_mass_storageis that we can amortize the cost of fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?I feel that allocating more requests is a far simpler and more generic method that changing FIFO sizes :)
I wish I had a USB bus trace handy to show you, which would make it very clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So by increasing the number of USB requests, that will help if there was a bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.
Quoted 4 lines
So yes, this method of increasing the # of USB reqs will definitely help with situations such as HSUSB or in SSUSB when EP bursting isn't used. The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.
Hmm, that's not what I remember. Perhaps the TRB cache size plays a role here too. I have clear memories of testing this very scenario of bursting (using g_mass_storage at the time) because I was curious about it. Back then, my tests showed no difference in behavior.
It could be nice if Heikki could test Intel parts with and without your changes on g_mass_storage with 250 requests.
Quoted 3 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet, it should have a NumP value equal to the bMaxBurst reported in the EP
Yes and no. Looking back at the history, we used to configure NUMP based on bMaxBurst, but it was changed later in commit 4e99472bc10bda9906526d725ff6d5f27b4ddca1 by yours truly because of a problem reported by John Youn.
And now we've come full circle. Because even if I believe more requests are enough for bursting, NUMP is limited by the RxFIFO size. This ends up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.
However, note that this is about RxFIFO size, not TxFIFO size. In fact, looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP (emphasis is mine):
"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to the maximum burst size supported by the endpoint as determined by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."
So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of this field.
For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this field may be treated by the host as informative only."
However, if I remember correctly (please verify dwc3 databook), NUMP in DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or TxFIFO size?
Quoted 9 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is that after 2 data packets, the device issues a NRDY. So then we'd need to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With this constant NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant bursts for a request, until the request is done, or if the host runs out of RXFIFO. (ie no interruption [on the USB protocol level] during USB request data transfer)
Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 28 lines
Quoted 28 lines
Quoted 16 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
Quoted 7 lines
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
Quoted 7 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a few years as well. The TXFIFO sizing really has helped with multimedia use cases, which use isoc endpoints, since esp. in those lower end CPU chips where latencies across the system are much larger, and a missed ISOC interval leads to a pop in your ear.
This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Quoted 6 lines
Quoted 4 lines
Remember dwc3 can accomodate 255 requests + link for each endpoint. If our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.
We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst behavior.
Signature · 1 line
balbi
Quoted 170 lines · Felipe Balbi
Hi,
Quoted 29 lines · Wesley Cheng
Quoted 29 lines
Quoted 28 lines
Quoted 22 lines
Quoted 14 lines
Quoted 7 lines
Quoted 7 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
Quoted 7 lines
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
Quoted 3 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can be a little nicer in this regard.
Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
Quoted 8 lines
Quoted 5 lines
sure, I support that. But I want to make sure the first cut isn't likely to break things left and right :)
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty big improvements on the TX path with this.
fingers crossed
Quoted 24 lines
Quoted 13 lines
Quoted 2 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were just to remove physical EP1 and 2 from the DWC3 endpoint list. If that
right, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to skip that endpoint on that particular platform without interefering with everything else.
Of course, we still need to figure out a way to abstract the different dwc3 instantiations.
Quoted 7 lines
Quoted 2 lines
was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph, will be coming up with a solution that's elegant and works for all different instantiations of dwc3 (or musb, cdns3, etc).
Well, at least for the TX FIFO resizing logic, we'd only be needing to focus on the DWC3 implementation.
Quoted 6 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We have a few special HW eps as well, which we try to maintain separately in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).
Awesome, as mentioned, the endpoint feature flags were added exactly to allow for these vendor-specific features :-)
I'm more than happy to help testing now that I finally got our SM8150 Surface Duo device tree accepted by Bjorn ;-)
Quoted 12 lines
Quoted 7 lines
Quoted 2 lines
However, I'm not sure how the changes would look like in the end, so I would like to wait later down the line to include that :).
Fair enough, I agree. Can we get some more testing of $subject, though? Did you test $subject with upstream too? Which gadget drivers did you use? How did you test
The results that I included in the cover page was tested with the pure upstream kernel on our device. Below was using the ConfigFS gadget w/ a mass storage only composition.
Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 193.60 | | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 287.35 | | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my memory could be failing.
Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 3 lines
We also have internal numbers which have shown similar improvements as well. Those are over networking/tethering interfaces, so testing IPERF loopback over TCP/UDP.
loopback iperf? That would skip the wire, no?
Quoted 23 lines
Quoted 6 lines
Quoted 5 lines
size of 2 and TX threshold of 1, this would really be not beneficial to us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.
Quoted 9 lines
What I noticed with
g_mass_storageis that we can amortize the cost of fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?I feel that allocating more requests is a far simpler and more generic method that changing FIFO sizes :)
I wish I had a USB bus trace handy to show you, which would make it very clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So by increasing the number of USB requests, that will help if there was a bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.
Quoted 4 lines
So yes, this method of increasing the # of USB reqs will definitely help with situations such as HSUSB or in SSUSB when EP bursting isn't used. The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.
Quoted 7 lines
Hmm, that's not what I remember. Perhaps the TRB cache size plays a role here too. I have clear memories of testing this very scenario of bursting (using
g_mass_storageat the time) because I was curious about it. Back then, my tests showed no difference in behavior.It could be nice if Heikki could test Intel parts with and without your changes on
g_mass_storagewith 250 requests.
Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
thanks,
Quoted 99 lines
Quoted 3 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet, it should have a NumP value equal to the bMaxBurst reported in the EP
Yes and no. Looking back at the history, we used to configure NUMP based on bMaxBurst, but it was changed later in commit
4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a problem reported by John Youn.And now we've come full circle. Because even if I believe more requests are enough for bursting, NUMP is limited by the RxFIFO size. This ends up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.
However, note that this is about RxFIFO size, not TxFIFO size. In fact, looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP (emphasis is mine):
"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to the maximum burst size supported by the endpoint as determined by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."
So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of this field.
For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this field may be treated by the host as informative only."
However, if I remember correctly (please verify dwc3 databook), NUMP in DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or TxFIFO size?
Quoted 9 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is that after 2 data packets, the device issues a NRDY. So then we'd need to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With this constant NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant bursts for a request, until the request is done, or if the host runs out of RXFIFO. (ie no interruption [on the USB protocol level] during USB request data transfer)
Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 28 lines
Quoted 28 lines
Quoted 16 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
Quoted 7 lines
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
Quoted 7 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a few years as well. The TXFIFO sizing really has helped with multimedia use cases, which use isoc endpoints, since esp. in those lower end CPU chips where latencies across the system are much larger, and a missed ISOC interval leads to a pop in your ear.
This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Quoted 6 lines
Quoted 4 lines
Remember dwc3 can accomodate 255 requests + link for each endpoint. If our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.
We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst behavior.
Signature · 1 line
heikki
Quoted 177 lines · Heikki Krogerus
Quoted 175 lines · Felipe Balbi
Hi,
Quoted 29 lines · Wesley Cheng
Quoted 29 lines
Quoted 28 lines
Quoted 22 lines
Quoted 14 lines
Quoted 7 lines
Quoted 7 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
Quoted 7 lines
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
Quoted 3 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can be a little nicer in this regard.
Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
Quoted 8 lines
Quoted 5 lines
sure, I support that. But I want to make sure the first cut isn't likely to break things left and right :)
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty big improvements on the TX path with this.
fingers crossed
Quoted 24 lines
Quoted 13 lines
Quoted 2 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were just to remove physical EP1 and 2 from the DWC3 endpoint list. If that
right, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to skip that endpoint on that particular platform without interefering with everything else.
Of course, we still need to figure out a way to abstract the different dwc3 instantiations.
Quoted 7 lines
Quoted 2 lines
was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph, will be coming up with a solution that's elegant and works for all different instantiations of dwc3 (or musb, cdns3, etc).
Well, at least for the TX FIFO resizing logic, we'd only be needing to focus on the DWC3 implementation.
Quoted 6 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We have a few special HW eps as well, which we try to maintain separately in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).
Awesome, as mentioned, the endpoint feature flags were added exactly to allow for these vendor-specific features :-)
I'm more than happy to help testing now that I finally got our SM8150 Surface Duo device tree accepted by Bjorn ;-)
Quoted 12 lines
Quoted 7 lines
Quoted 2 lines
However, I'm not sure how the changes would look like in the end, so I would like to wait later down the line to include that :).
Fair enough, I agree. Can we get some more testing of $subject, though? Did you test $subject with upstream too? Which gadget drivers did you use? How did you test
The results that I included in the cover page was tested with the pure upstream kernel on our device. Below was using the ConfigFS gadget w/ a mass storage only composition.
Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 193.60 | | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 287.35 | | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my memory could be failing.
Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 3 lines
We also have internal numbers which have shown similar improvements as well. Those are over networking/tethering interfaces, so testing IPERF loopback over TCP/UDP.
loopback iperf? That would skip the wire, no?
Quoted 23 lines
Quoted 6 lines
Quoted 5 lines
size of 2 and TX threshold of 1, this would really be not beneficial to us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.
Quoted 9 lines
What I noticed with
g_mass_storageis that we can amortize the cost of fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?I feel that allocating more requests is a far simpler and more generic method that changing FIFO sizes :)
I wish I had a USB bus trace handy to show you, which would make it very clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So by increasing the number of USB requests, that will help if there was a bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.
Quoted 4 lines
So yes, this method of increasing the # of USB reqs will definitely help with situations such as HSUSB or in SSUSB when EP bursting isn't used. The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.
Hmm, that's not what I remember. Perhaps the TRB cache size plays a role here too. I have clear memories of testing this very scenario of bursting (using
g_mass_storageat the time) because I was curious about it. Back then, my tests showed no difference in behavior.
Quoted 5 lines
Quoted 2 lines
It could be nice if Heikki could test Intel parts with and without your changes on
g_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
Quoted 102 lines
Quoted 95 lines
Quoted 3 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet, it should have a NumP value equal to the bMaxBurst reported in the EP
Yes and no. Looking back at the history, we used to configure NUMP based on bMaxBurst, but it was changed later in commit
4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a problem reported by John Youn.And now we've come full circle. Because even if I believe more requests are enough for bursting, NUMP is limited by the RxFIFO size. This ends up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.
However, note that this is about RxFIFO size, not TxFIFO size. In fact, looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP (emphasis is mine):
"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to the maximum burst size supported by the endpoint as determined by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."
So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of this field.
For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this field may be treated by the host as informative only."
However, if I remember correctly (please verify dwc3 databook), NUMP in DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or TxFIFO size?
Quoted 9 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is that after 2 data packets, the device issues a NRDY. So then we'd need to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With this constant NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant bursts for a request, until the request is done, or if the host runs out of RXFIFO. (ie no interruption [on the USB protocol level] during USB request data transfer)
Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 33 lines
Quoted 32 lines
Quoted 16 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
Quoted 7 lines
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
Quoted 7 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a few years as well. The TXFIFO sizing really has helped with multimedia use cases, which use isoc endpoints, since esp. in those lower end CPU chips where latencies across the system are much larger, and a missed ISOC interval leads to a pop in your ear.
This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Remember dwc3 can accomodate 255 requests + link for each endpoint. If our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.
Quoted 4 lines
Quoted 1 line
We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst behavior.
--
heikki
Signature · 2 lines
With Best Regards, Andy Shevchenko
Hi
Op 11-06-2021 om 15:21 schreef Andy Shevchenko:
Quoted 159 lines
Quoted 157 lines · Heikki Krogerus
Quoted 156 lines · Felipe Balbi
Hi,
Quoted 25 lines · Wesley Cheng
Quoted 25 lines
Quoted 25 lines
Quoted 20 lines
Quoted 13 lines
Quoted 7 lines
Quoted 7 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
Quoted 7 lines
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
Quoted 3 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can be a little nicer in this regard.
Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
Quoted 7 lines
Quoted 5 lines
sure, I support that. But I want to make sure the first cut isn't likely to break things left and right :)
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty big improvements on the TX path with this.
fingers crossed
Quoted 21 lines
Quoted 9 lines
Quoted 2 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were just to remove physical EP1 and 2 from the DWC3 endpoint list. If that
right, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to skip that endpoint on that particular platform without interefering with everything else.
Quoted 9 lines
Of course, we still need to figure out a way to abstract the different dwc3 instantiations.
Quoted 2 lines
was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph, will be coming up with a solution that's elegant and works for all different instantiations of dwc3 (or musb, cdns3, etc).
Well, at least for the TX FIFO resizing logic, we'd only be needing to focus on the DWC3 implementation.
Quoted 6 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We have a few special HW eps as well, which we try to maintain separately in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).
Awesome, as mentioned, the endpoint feature flags were added exactly to allow for these vendor-specific features :-)
I'm more than happy to help testing now that I finally got our SM8150 Surface Duo device tree accepted by Bjorn ;-)
Quoted 10 lines
Quoted 6 lines
Quoted 2 lines
However, I'm not sure how the changes would look like in the end, so I would like to wait later down the line to include that :).
Fair enough, I agree. Can we get some more testing of $subject, though? Did you test $subject with upstream too? Which gadget drivers did you use? How did you test
The results that I included in the cover page was tested with the pure upstream kernel on our device. Below was using the ConfigFS gadget w/ a mass storage only composition.
Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 193.60 | | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 287.35 | | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my memory could be failing.
Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 3 lines
We also have internal numbers which have shown similar improvements as well. Those are over networking/tethering interfaces, so testing IPERF loopback over TCP/UDP.
loopback iperf? That would skip the wire, no?
Quoted 21 lines
Quoted 5 lines
Quoted 5 lines
size of 2 and TX threshold of 1, this would really be not beneficial to us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.
Quoted 9 lines
What I noticed with
g_mass_storageis that we can amortize the cost of fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?I feel that allocating more requests is a far simpler and more generic method that changing FIFO sizes :)
I wish I had a USB bus trace handy to show you, which would make it very clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So by increasing the number of USB requests, that will help if there was a bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.
Quoted 4 lines
So yes, this method of increasing the # of USB reqs will definitely help with situations such as HSUSB or in SSUSB when EP bursting isn't used. The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.
Hmm, that's not what I remember. Perhaps the TRB cache size plays a role here too. I have clear memories of testing this very scenario of bursting (using
g_mass_storageat the time) because I was curious about it. Back then, my tests showed no difference in behavior.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
It could be nice if Heikki could test Intel parts with and without your changes on
g_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
I just tested on 5.13.0-rc4 on Intel Edison (x86_64). All 5 patches apply. Switching between host/gadget works, no connections dropping, no errors in dmesg.
In host mode I connect a smsc9504 eth+4p hub. In gadget mode I have composite device created from configfs with gser / eem / mass_storage / uac2.
Tested with iperf3 performance in host (93.6Mbits/sec) and gadget (207Mbits/sec) mode. Compared to v5.10.41 without patches host (93.4Mbits/sec) and gadget (198Mbits/sec).
Gadget seems to be a little faster with the patches, but that might also be caused by something else, on v5.10.41 I see the bitrate bouncing between 207 and 199.
I saw a mention to test iperf3 to self (loopback). 3.09 Gbits/sec. With v5.10.41 3.07Gbits/sec. Not bad for a 500MHz device.
With gnome-disks I did a read access benchmark 35.4MB/s, with v5.10.41 34.7MB/s. This might be limited by Edison's internal eMMC speed (when booting U-Boot reads the kernel with 21.4 MiB/s).
Quoted 95 lines
Quoted 93 lines
Quoted 84 lines
Quoted 3 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet, it should have a NumP value equal to the bMaxBurst reported in the EP
Yes and no. Looking back at the history, we used to configure NUMP based on bMaxBurst, but it was changed later in commit
4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a problem reported by John Youn.And now we've come full circle. Because even if I believe more requests are enough for bursting, NUMP is limited by the RxFIFO size. This ends up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.
However, note that this is about RxFIFO size, not TxFIFO size. In fact, looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP (emphasis is mine):
"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to the maximum burst size supported by the endpoint as determined by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."
So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of this field.
For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this field may be treated by the host as informative only."
However, if I remember correctly (please verify dwc3 databook), NUMP in DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or TxFIFO size?
Quoted 9 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is that after 2 data packets, the device issues a NRDY. So then we'd need to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With this constant NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant bursts for a request, until the request is done, or if the host runs out of RXFIFO. (ie no interruption [on the USB protocol level] during USB request data transfer)
Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 24 lines
Quoted 24 lines
Quoted 13 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
Quoted 9 lines
Quoted 2 lines
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
Quoted 7 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a few years as well. The TXFIFO sizing really has helped with multimedia use cases, which use isoc endpoints, since esp. in those lower end CPU chips where latencies across the system are much larger, and a missed ISOC interval leads to a pop in your ear.
This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Quoted 7 lines
Quoted 5 lines
Quoted 4 lines
Remember dwc3 can accomodate 255 requests + link for each endpoint. If our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.
We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst behavior.
--
heikki
Quoted 5 lines · Ferry Toth
Op 11-06-2021 om 15:21 schreef Andy Shevchenko:
Quoted 4 lines
Quoted 2 lines · Heikki Krogerus
Quoted 1 line · Felipe Balbi
Wesley Cheng <[email protected]> writes:
...
Quoted 92 lines
Quoted 68 lines
Quoted 68 lines
Quoted 68 lines
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 193.60 | | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 287.35 | | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my memory could be failing.
Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 3 lines
We also have internal numbers which have shown similar improvements as well. Those are over networking/tethering interfaces, so testing IPERF loopback over TCP/UDP.
loopback iperf? That would skip the wire, no?
Quoted 21 lines
Quoted 5 lines
Quoted 5 lines
size of 2 and TX threshold of 1, this would really be not beneficial to us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.
Quoted 9 lines
What I noticed with
g_mass_storageis that we can amortize the cost of fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?I feel that allocating more requests is a far simpler and more generic method that changing FIFO sizes :)
I wish I had a USB bus trace handy to show you, which would make it very clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So by increasing the number of USB requests, that will help if there was a bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.
Quoted 4 lines
So yes, this method of increasing the # of USB reqs will definitely help with situations such as HSUSB or in SSUSB when EP bursting isn't used. The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.
Hmm, that's not what I remember. Perhaps the TRB cache size plays a role here too. I have clear memories of testing this very scenario of bursting (using
g_mass_storageat the time) because I was curious about it. Back then, my tests showed no difference in behavior.Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
It could be nice if Heikki could test Intel parts with and without your changes on
g_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
I just tested on 5.13.0-rc4 on Intel Edison (
x86_64). All 5 patches apply. Switching between host/gadget works, no connections dropping, no errors in dmesg.In host mode I connect a smsc9504 eth+4p hub. In gadget mode I have composite device created from configfs with gser / eem /
mass_storage/ uac2.Tested with iperf3 performance in host (93.6Mbits/sec) and gadget (207Mbits/sec) mode. Compared to v5.10.41 without patches host (93.4Mbits/sec) and gadget (198Mbits/sec).
Gadget seems to be a little faster with the patches, but that might also be caused by something else, on v5.10.41 I see the bitrate bouncing between 207 and 199.
Quoted 6 lines
I saw a mention to test iperf3 to self (loopback). 3.09 Gbits/sec. With v5.10.41 3.07Gbits/sec. Not bad for a 500MHz device.
With gnome-disks I did a read access benchmark 35.4MB/s, with v5.10.41 34.7MB/s. This might be limited by Edison's internal eMMC speed (when booting U-Boot reads the kernel with 21.4 MiB/s).
Ferry, thank you very much for this information and testing efforts!
Quoted 91 lines
Quoted 91 lines
Quoted 91 lines
Quoted 91 lines
Quoted 3 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet, it should have a NumP value equal to the bMaxBurst reported in the EP
Yes and no. Looking back at the history, we used to configure NUMP based on bMaxBurst, but it was changed later in commit
4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a problem reported by John Youn.And now we've come full circle. Because even if I believe more requests are enough for bursting, NUMP is limited by the RxFIFO size. This ends up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.
However, note that this is about RxFIFO size, not TxFIFO size. In fact, looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP (emphasis is mine):
"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to the maximum burst size supported by the endpoint as determined by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."
So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of this field.
For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this field may be treated by the host as informative only."
However, if I remember correctly (please verify dwc3 databook), NUMP in DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or TxFIFO size?
Quoted 9 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is that after 2 data packets, the device issues a NRDY. So then we'd need to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With this constant NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant bursts for a request, until the request is done, or if the host runs out of RXFIFO. (ie no interruption [on the USB protocol level] during USB request data transfer)
Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 24 lines
Quoted 24 lines
Quoted 13 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
Quoted 9 lines
Quoted 2 lines
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
Quoted 7 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a few years as well. The TXFIFO sizing really has helped with multimedia use cases, which use isoc endpoints, since esp. in those lower end CPU chips where latencies across the system are much larger, and a missed ISOC interval leads to a pop in your ear.
This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Quoted 5 lines
Quoted 4 lines
Remember dwc3 can accomodate 255 requests + link for each endpoint. If our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.
We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst behavior.
Signature · 2 lines
With Best Regards, Andy Shevchenko
Quoted 223 lines · Ferry Toth
Hi
Op 11-06-2021 om 15:21 schreef Andy Shevchenko:
Quoted 196 lines
Quoted 194 lines · Heikki Krogerus
Quoted 193 lines · Felipe Balbi
Hi,
Quoted 43 lines · Wesley Cheng
Quoted 43 lines
Quoted 43 lines
Quoted 36 lines
Quoted 23 lines
Quoted 13 lines
Quoted 13 lines
to be honest, I don't think these should go in (apart from the build
failure) because it's likely to break instantiations of the core with
differing FIFO sizes. Some instantiations even have some endpoints with
dedicated functionality that requires the default FIFO size configured
during coreConsultant instantiation. I know of at OMAP5 and some Intel
implementations which have dedicated endpoints for processor tracing.Quoted 10 lines
Quoted 10 lines
With OMAP5, these endpoints are configured at the top of the available
endpoints, which means that if a gadget driver gets loaded and takes
over most of the FIFO space because of this resizing, processor tracing
will have a hard time running. That being said, processor tracing isn't
supported in upstream at this moment.Quoted 13 lines
I agree that the application of this logic may differ between vendors,
hence why I wanted to keep this controllable by the DT property, so that
for those which do not support this use case can leave it disabled. The
logic is there to ensure that for a given USB configuration, for each EP
it would have at least 1 TX FIFO. For USB configurations which don't
utilize all available IN EPs, it would allow re-allocation of internal
memory to EPs which will actually be in use.Quoted 4 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can
be a little nicer in this regard.Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
Quoted 8 lines
Quoted 6 lines
sure, I support that. But I want to make sure the first cut isn't likely
to break things left and right :)Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty big improvements on the TX path with this.
fingers crossed
Quoted 26 lines
Quoted 16 lines
Quoted 4 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were
just to remove physical EP1 and 2 from the DWC3 endpoint list. If thatright, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to skip that endpoint on that particular platform without interefering with
everything else.Of course, we still need to figure out a way to abstract the different
dwc3 instantiations.Quoted 7 lines
Quoted 2 lines
was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph, will be
coming up with a solution that's elegant and works for all different instantiations of dwc3 (or musb, cdns3, etc).Well, at least for the TX FIFO resizing logic, we'd only be needing to focus on the DWC3 implementation.
Quoted 6 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We have a few special HW eps as well, which we try to maintain separately in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).
Awesome, as mentioned, the endpoint feature flags were added exactly to allow for these vendor-specific features :-)
I'm more than happy to help testing now that I finally got our SM8150 Surface Duo device tree accepted by Bjorn ;-)
Quoted 13 lines
Quoted 8 lines
Quoted 3 lines
However, I'm not sure how the changes would look like in the end, so I
would like to wait later down the line to include that :).Fair enough, I agree. Can we get some more testing of $subject, though?
Did you test $subject with upstream too? Which gadget drivers did you use? How did you testThe results that I included in the cover page was tested with the pure upstream kernel on our device. Below was using the ConfigFS gadget w/ a
mass storage only composition.Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 193.60| | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 287.35| | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my memory could be failing.
Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 4 lines
We also have internal numbers which have shown similar improvements as well. Those are over networking/tethering interfaces, so testing IPERF
loopback over TCP/UDP.loopback iperf? That would skip the wire, no?
Quoted 28 lines
Quoted 14 lines
Quoted 6 lines
size of 2 and TX threshold of 1, this would really be not beneficial to
us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.What I noticed with
g_mass_storageis that we can amortize the cost of
fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was
enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?Quoted 4 lines
I feel that allocating more requests is a far simpler and more generic
method that changing FIFO sizes :)I wish I had a USB bus trace handy to show you, which would make it very
clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So
by increasing the number of USB requests, that will help if there was a
bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.Quoted 5 lines
So yes, this method of increasing the # of USB reqs will definitely help
with situations such as HSUSB or in SSUSB when EP bursting isn't used. The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.Hmm, that's not what I remember. Perhaps the TRB cache size plays a role
here too. I have clear memories of testing this very scenario of bursting (usingg_mass_storageat the time) because I was curious about it. Back then, my tests showed no difference in behavior.Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
It could be nice if Heikki could test Intel parts with and without your changes on
g_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
I just tested on 5.13.0-rc4 on Intel Edison (
x86_64). All 5 patches apply. Switching between host/gadget works, no connections dropping, no errors in dmesg.In host mode I connect a smsc9504 eth+4p hub. In gadget mode I have composite device created from configfs with gser / eem /
mass_storage/ uac2.Tested with iperf3 performance in host (93.6Mbits/sec) and gadget (207Mbits/sec) mode. Compared to v5.10.41 without patches host (93.4Mbits/sec) and gadget (198Mbits/sec).
Gadget seems to be a little faster with the patches, but that might also be caused by something else, on v5.10.41 I see the bitrate bouncing between 207 and 199.
Quoted 7 lines
I saw a mention to test iperf3 to self (loopback). 3.09 Gbits/sec. With v5.10.41 3.07Gbits/sec. Not bad for a 500MHz device.
With gnome-disks I did a read access benchmark 35.4MB/s, with v5.10.41 34.7MB/s. This might be limited by Edison's internal eMMC speed (when booting U-Boot reads the kernel with 21.4 MiB/s).
Hi Ferry,
Thanks for the testing. Just to double check, did you also enable the property, which enabled the TXFIFO resize feature on the platform? For example, for the QCOM SM8150 platform, we're adding the following to our device tree node:
tx-fifo-resize
If not, then your results at least confirms that w/o the property present, the changes won't break anything :). Thanks again for the initial testing!
Thanks
Wesley Cheng
Quoted 115 lines
Quoted 115 lines
Quoted 113 lines
Quoted 111 lines
Quoted 4 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet,
it should have a NumP value equal to the bMaxBurst reported in the EPYes and no. Looking back at the history, we used to configure NUMP based
on bMaxBurst, but it was changed later in commit4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a problem reported by John Youn.And now we've come full circle. Because even if I believe more requests are enough for bursting, NUMP is limited by the RxFIFO size. This ends up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.
However, note that this is about RxFIFO size, not TxFIFO size. In fact, looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP (emphasis is mine):
"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to the maximum burst size supported by the endpoint as determined by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."
So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of this field.
For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this field may be treated by the host as informative only."
However, if I remember correctly (please verify dwc3 databook), NUMP in DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or
TxFIFO size?Quoted 6 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is
that after 2 data packets, the device issues a NRDY. So then we'd need
to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With thisQuoted 8 lines
constant
NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant
bursts for a request, until the request is done, or if the host runs out
of RXFIFO. (ie no interruption [on the USB protocol level] during USB request data transfer)Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 34 lines
Quoted 34 lines
Quoted 19 lines
Quoted 15 lines
Quoted 9 lines
Quoted 5 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user
devices for the past few years, so its been through a lot of testing :).all with the same gadget driver. Also, who uses USB on android devices
these days? Most of the data transfer goes via WiFi or Bluetooth, anyway
:-)Quoted 4 lines
I guess only developers are using USB during development to flash dev
images heh.Quoted 11 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a
few years as well. The TXFIFO sizing really has helped with multimedia
use cases, which use isoc endpoints, since esp. in those lower end CPU
chips where latencies across the system are much larger, and a missed
ISOC interval leads to a pop in your ear.This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Quoted 7 lines
Quoted 5 lines
Remember dwc3 can accomodate 255 requests + link for each endpoint. If
our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst behavior.
Signature · 1 line
heikki
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
Op 14-06-2021 om 20:58 schreef Wesley Cheng:
Quoted 238 lines · Ferry Toth
Quoted 222 lines
Hi
Op 11-06-2021 om 15:21 schreef Andy Shevchenko:
Quoted 196 lines
Quoted 194 lines · Heikki Krogerus
Quoted 193 lines · Felipe Balbi
Hi,
Quoted 43 lines · Wesley Cheng
Quoted 43 lines
Quoted 43 lines
Quoted 36 lines
Quoted 23 lines
Quoted 13 lines
Quoted 13 lines
to be honest, I don't think these should go in (apart from the build
failure) because it's likely to break instantiations of the core with
differing FIFO sizes. Some instantiations even have some endpoints with
dedicated functionality that requires the default FIFO size configured
during coreConsultant instantiation. I know of at OMAP5 and some Intel
implementations which have dedicated endpoints for processor tracing.Quoted 10 lines
Quoted 10 lines
With OMAP5, these endpoints are configured at the top of the available
endpoints, which means that if a gadget driver gets loaded and takes
over most of the FIFO space because of this resizing, processor tracing
will have a hard time running. That being said, processor tracing isn't
supported in upstream at this moment.Quoted 13 lines
I agree that the application of this logic may differ between vendors,
hence why I wanted to keep this controllable by the DT property, so that
for those which do not support this use case can leave it disabled. The
logic is there to ensure that for a given USB configuration, for each EP
it would have at least 1 TX FIFO. For USB configurations which don't
utilize all available IN EPs, it would allow re-allocation of internal
memory to EPs which will actually be in use.Quoted 4 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can
be a little nicer in this regard.Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
Quoted 8 lines
Quoted 6 lines
sure, I support that. But I want to make sure the first cut isn't likely
to break things left and right :)Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty big improvements on the TX path with this.
fingers crossed
Quoted 26 lines
Quoted 16 lines
Quoted 4 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were
just to remove physical EP1 and 2 from the DWC3 endpoint list. If thatright, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to skip that endpoint on that particular platform without interefering with
everything else.Of course, we still need to figure out a way to abstract the different
dwc3 instantiations.Quoted 7 lines
Quoted 2 lines
was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph, will be
coming up with a solution that's elegant and works for all different instantiations of dwc3 (or musb, cdns3, etc).Well, at least for the TX FIFO resizing logic, we'd only be needing to focus on the DWC3 implementation.
Quoted 6 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We have a few special HW eps as well, which we try to maintain separately in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).
Awesome, as mentioned, the endpoint feature flags were added exactly to allow for these vendor-specific features :-)
I'm more than happy to help testing now that I finally got our SM8150 Surface Duo device tree accepted by Bjorn ;-)
Quoted 13 lines
Quoted 8 lines
Quoted 3 lines
However, I'm not sure how the changes would look like in the end, so I
would like to wait later down the line to include that :).Fair enough, I agree. Can we get some more testing of $subject, though?
Did you test $subject with upstream too? Which gadget drivers did you use? How did you testThe results that I included in the cover page was tested with the pure upstream kernel on our device. Below was using the ConfigFS gadget w/ a
mass storage only composition.Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 193.60| | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 287.35| | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my memory could be failing.
Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 4 lines
We also have internal numbers which have shown similar improvements as well. Those are over networking/tethering interfaces, so testing IPERF
loopback over TCP/UDP.loopback iperf? That would skip the wire, no?
Quoted 28 lines
Quoted 14 lines
Quoted 6 lines
size of 2 and TX threshold of 1, this would really be not beneficial to
us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.What I noticed with
g_mass_storageis that we can amortize the cost of
fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was
enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?Quoted 4 lines
I feel that allocating more requests is a far simpler and more generic
method that changing FIFO sizes :)I wish I had a USB bus trace handy to show you, which would make it very
clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So
by increasing the number of USB requests, that will help if there was a
bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.Quoted 5 lines
So yes, this method of increasing the # of USB reqs will definitely help
with situations such as HSUSB or in SSUSB when EP bursting isn't used. The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.Hmm, that's not what I remember. Perhaps the TRB cache size plays a role
here too. I have clear memories of testing this very scenario of bursting (usingg_mass_storageat the time) because I was curious about it. Back then, my tests showed no difference in behavior.Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
It could be nice if Heikki could test Intel parts with and without your changes on
g_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
I just tested on 5.13.0-rc4 on Intel Edison (
x86_64). All 5 patches apply. Switching between host/gadget works, no connections dropping, no errors in dmesg.In host mode I connect a smsc9504 eth+4p hub. In gadget mode I have composite device created from configfs with gser / eem /
mass_storage/ uac2.Tested with iperf3 performance in host (93.6Mbits/sec) and gadget (207Mbits/sec) mode. Compared to v5.10.41 without patches host (93.4Mbits/sec) and gadget (198Mbits/sec).
Gadget seems to be a little faster with the patches, but that might also be caused by something else, on v5.10.41 I see the bitrate bouncing between 207 and 199.
Quoted 7 lines
I saw a mention to test iperf3 to self (loopback). 3.09 Gbits/sec. With v5.10.41 3.07Gbits/sec. Not bad for a 500MHz device.
With gnome-disks I did a read access benchmark 35.4MB/s, with v5.10.41 34.7MB/s. This might be limited by Edison's internal eMMC speed (when booting U-Boot reads the kernel with 21.4 MiB/s).
Hi Ferry,
Thanks for the testing. Just to double check, did you also enable the property, which enabled the TXFIFO resize feature on the platform? For example, for the QCOM SM8150 platform, we're adding the following to our device tree node:
Quoted 5 lines
tx-fifo-resize
If not, then your results at least confirms that w/o the property present, the changes won't break anything :). Thanks again for the initial testing!
No I didn't. Afaik we don't have a devicetree property to set.
But I'd be happy to test that as well. But where to set the property?
dwc3_pci_mrfld_properties[] in dwc3-pci?
Quoted 117 lines
Thanks
Wesley ChengQuoted 114 lines
Quoted 114 lines
Quoted 113 lines
Quoted 111 lines
Quoted 4 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet,
it should have a NumP value equal to the bMaxBurst reported in the EPYes and no. Looking back at the history, we used to configure NUMP based
on bMaxBurst, but it was changed later in commit4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a problem reported by John Youn.And now we've come full circle. Because even if I believe more requests are enough for bursting, NUMP is limited by the RxFIFO size. This ends up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.
However, note that this is about RxFIFO size, not TxFIFO size. In fact, looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP (emphasis is mine):
"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to the maximum burst size supported by the endpoint as determined by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."
So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of this field.
For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this field may be treated by the host as informative only."
However, if I remember correctly (please verify dwc3 databook), NUMP in DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or
TxFIFO size?Quoted 6 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is
that after 2 data packets, the device issues a NRDY. So then we'd need
to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With thisQuoted 8 lines
constant
NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant
bursts for a request, until the request is done, or if the host runs out
of RXFIFO. (ie no interruption [on the USB protocol level] during USB request data transfer)Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 34 lines
Quoted 34 lines
Quoted 19 lines
Quoted 15 lines
Quoted 9 lines
Quoted 5 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user
devices for the past few years, so its been through a lot of testing :).all with the same gadget driver. Also, who uses USB on android devices
these days? Most of the data transfer goes via WiFi or Bluetooth, anyway
:-)Quoted 4 lines
I guess only developers are using USB during development to flash dev
images heh.Quoted 11 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a
few years as well. The TXFIFO sizing really has helped with multimedia
use cases, which use isoc endpoints, since esp. in those lower end CPU
chips where latencies across the system are much larger, and a missed
ISOC interval leads to a pop in your ear.This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Quoted 7 lines
Quoted 5 lines
Remember dwc3 can accomodate 255 requests + link for each endpoint. If
our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst behavior.
Signature · 1 line
heikki
Quoted 264 lines · Ferry Toth
Op 14-06-2021 om 20:58 schreef Wesley Cheng:
Quoted 256 lines · Ferry Toth
Quoted 240 lines
Hi
Op 11-06-2021 om 15:21 schreef Andy Shevchenko:
Quoted 213 lines
Quoted 211 lines · Heikki Krogerus
Quoted 210 lines · Felipe Balbi
Hi,
Quoted 47 lines · Wesley Cheng
Quoted 47 lines
Quoted 36 lines
Quoted 36 lines
Quoted 36 lines
Quoted 13 lines
Quoted 13 lines
to be honest, I don't think these should go in (apart from
the build
failure) because it's likely to break instantiations of the
core with
differing FIFO sizes. Some instantiations even have some
endpoints with
dedicated functionality that requires the default FIFO size
configured
during coreConsultant instantiation. I know of at OMAP5 and
some Intel
implementations which have dedicated endpoints for processor
tracing.Quoted 10 lines
Quoted 10 lines
With OMAP5, these endpoints are configured at the top of the
available
endpoints, which means that if a gadget driver gets loaded
and takes
over most of the FIFO space because of this resizing,
processor tracing
will have a hard time running. That being said, processor
tracing isn't
supported in upstream at this moment.I agree that the application of this logic may differ between vendors,
hence why I wanted to keep this controllable by the DT property, so that
for those which do not support this use case can leave it disabled. The
logic is there to ensure that for a given USB configuration, for each EP
it would have at least 1 TX FIFO. For USB configurations which don't
utilize all available IN EPs, it would allow re-allocation of internal
memory to EPs which will actually be in use.Quoted 7 lines
Quoted 4 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can
be a little nicer in this regard.Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
sure, I support that. But I want to make sure the first cut isn't likely
to break things left and right :)Quoted 5 lines
Quoted 2 lines
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty
big improvements on the TX path with this.fingers crossed
Quoted 30 lines
Quoted 17 lines
Quoted 4 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were
just to remove physical EP1 and 2 from the DWC3 endpoint list. If thatright, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to
skip that endpoint on that particular platform without interefering with
everything else.Of course, we still need to figure out a way to abstract the different
dwc3 instantiations.Quoted 9 lines
Quoted 3 lines
was the change which ended up upstream for the Intel tracer then we
could improve the logic to avoid re-sizing those particular EPs.The problem then, just as I mentioned in the previous paragraph, will be
coming up with a solution that's elegant and works for all different
instantiations of dwc3 (or musb, cdns3, etc).Well, at least for the TX FIFO resizing logic, we'd only be needing to
focus on the DWC3 implementation.Quoted 8 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We
have a few special HW eps as well, which we try to maintain separately
in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).Awesome, as mentioned, the endpoint feature flags were added exactly to
allow for these vendor-specific features :-)I'm more than happy to help testing now that I finally got our SM8150 Surface Duo device tree accepted by Bjorn ;-)
Quoted 15 lines
Quoted 9 lines
Quoted 3 lines
However, I'm not sure how the changes would look like in the end, so I
would like to wait later down the line to include that :).Fair enough, I agree. Can we get some more testing of $subject, though?
Did you test $subject with upstream too? Which gadget drivers did you
use? How did you testThe results that I included in the cover page was tested with the pure
upstream kernel on our device. Below was using the ConfigFS gadget w/ a
mass storage only composition.Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 193.60| | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 287.35| | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my
memory could be failing.Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 5 lines
We also have internal numbers which have shown similar improvements as
well. Those are over networking/tethering interfaces, so testing IPERF
loopback over TCP/UDP.loopback iperf? That would skip the wire, no?
Quoted 31 lines
Quoted 17 lines
Quoted 8 lines
size of 2 and TX threshold of 1, this would really be not beneficial to
us, because we can only change the TX threshold to 2 at max, and at
least in my observations, once we have to go out to system memory to
fetch the next data packet, that latency takes enough time for the controller to end the current burst.What I noticed with
g_mass_storageis that we can amortize the cost of
fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was
enough to give me very good performance. Never had to poke at TX FIFO
resizing. Did you try something like this too?Quoted 4 lines
I feel that allocating more requests is a far simpler and more generic
method that changing FIFO sizes :)I wish I had a USB bus trace handy to show you, which would make it very
clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So
by increasing the number of USB requests, that will help if there was a
bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.Quoted 6 lines
So yes, this method of increasing the # of USB reqs will definitely help
with situations such as HSUSB or in SSUSB when EP bursting isn't used.
The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.Hmm, that's not what I remember. Perhaps the TRB cache size plays a role
here too. I have clear memories of testing this very scenario of bursting (usingg_mass_storageat the time) because I was curious about
it. Back then, my tests showed no difference in behavior.Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
It could be nice if Heikki could test Intel parts with and without your
changes ong_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
I just tested on 5.13.0-rc4 on Intel Edison (
x86_64). All 5 patches apply. Switching between host/gadget works, no connections dropping, no errors in dmesg.In host mode I connect a smsc9504 eth+4p hub. In gadget mode I have composite device created from configfs with gser / eem /
mass_storage/ uac2.Tested with iperf3 performance in host (93.6Mbits/sec) and gadget (207Mbits/sec) mode. Compared to v5.10.41 without patches host (93.4Mbits/sec) and gadget (198Mbits/sec).
Gadget seems to be a little faster with the patches, but that might also be caused by something else, on v5.10.41 I see the bitrate bouncing between 207 and 199.
Quoted 7 lines
I saw a mention to test iperf3 to self (loopback). 3.09 Gbits/sec. With v5.10.41 3.07Gbits/sec. Not bad for a 500MHz device.
With gnome-disks I did a read access benchmark 35.4MB/s, with v5.10.41 34.7MB/s. This might be limited by Edison's internal eMMC speed (when booting U-Boot reads the kernel with 21.4 MiB/s).
Hi Ferry,
Thanks for the testing. Just to double check, did you also enable the property, which enabled the TXFIFO resize feature on the platform? For example, for the QCOM SM8150 platform, we're adding the following to our device tree node:
Quoted 5 lines
tx-fifo-resize
If not, then your results at least confirms that w/o the property present, the changes won't break anything :). Thanks again for the initial testing!
Quoted 6 lines
No I didn't. Afaik we don't have a devicetree property to set.
But I'd be happy to test that as well. But where to set the property?
dwc3_pci_mrfld_properties[] in dwc3-pci?
Hi Ferry,
Not too sure which DWC3 driver is used for the Intel platform, but I believe that should be the one. (if that's what is normally used) We'd just need to add an entry w/ the below:
PROPERTY_ENTRY_BOOL("tx-fifo-resize")
Thanks
Wesley Cheng
Quoted 132 lines
Quoted 132 lines
Thanks
Wesley ChengQuoted 129 lines
Quoted 129 lines
Quoted 128 lines
Quoted 121 lines
Quoted 5 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet,
it should have a NumP value equal to the bMaxBurst reported in the EPYes and no. Looking back at the history, we used to configure NUMP based
on bMaxBurst, but it was changed later in commit4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a problem reported by John Youn.And now we've come full circle. Because even if I believe more requests
are enough for bursting, NUMP is limited by the RxFIFO size. This ends
up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.However, note that this is about RxFIFO size, not TxFIFO size. In fact,
looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP
(emphasis is mine):"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to
the maximum burst size supported by the endpoint as determined
by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of
this field.For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this field
may be treated by the host as informative only."However, if I remember correctly (please verify dwc3 databook), NUMP in
DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or
TxFIFO size?Quoted 7 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is
that after 2 data packets, the device issues a NRDY. So then we'd need
to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With this constantQuoted 8 lines
NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant
bursts for a request, until the request is done, or if the host runs out
of RXFIFO. (ie no interruption [on the USB protocol level] during USB
request data transfer)Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 42 lines
Quoted 42 lines
Quoted 19 lines
Quoted 15 lines
Quoted 9 lines
Quoted 5 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user
devices for the past few years, so its been through a lot of testing :).all with the same gadget driver. Also, who uses USB on android devices
these days? Most of the data transfer goes via WiFi or Bluetooth, anyway
:-)Quoted 4 lines
I guess only developers are using USB during development to flash dev
images heh.Quoted 13 lines
I used to be a customer facing engineer, so honestly I did see some
really interesting and crazy designs. Again, we do have non-Android
products that use the same code, and it has been working in there for a
few years as well. The TXFIFO sizing really has helped with multimedia
use cases, which use isoc endpoints, since esp. in those lower end CPU
chips where latencies across the system are much larger, and a missed
ISOC interval leads to a pop in your ear.This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in
knowing if a deeper request queue would also help here.Remember dwc3 can accomodate 255 requests + link for each endpoint. If
our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.Quoted 5 lines
Quoted 2 lines
We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst
behavior.--
heikki
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
Ferry Toth
AppliedHi
Op 15-06-2021 om 06:22 schreef Wesley Cheng:
Quoted 258 lines · Ferry Toth
Quoted 256 lines
Op 14-06-2021 om 20:58 schreef Wesley Cheng:
Quoted 255 lines · Ferry Toth
Quoted 240 lines
Hi
Op 11-06-2021 om 15:21 schreef Andy Shevchenko:
Quoted 213 lines
Quoted 211 lines · Heikki Krogerus
Quoted 210 lines · Felipe Balbi
Hi,
Quoted 47 lines · Wesley Cheng
Quoted 47 lines
Quoted 36 lines
Quoted 36 lines
Quoted 36 lines
Quoted 23 lines
to be honest, I don't think these should go in (apart from
the build
failure) because it's likely to break instantiations of the
core with
differing FIFO sizes. Some instantiations even have some
endpoints with
dedicated functionality that requires the default FIFO size
configured
during coreConsultant instantiation. I know of at OMAP5 and
some Intel
implementations which have dedicated endpoints for processor
tracing.With OMAP5, these endpoints are configured at the top of the
available
endpoints, which means that if a gadget driver gets loaded
and takes
over most of the FIFO space because of this resizing,
processor tracing
will have a hard time running. That being said, processor
tracing isn't
supported in upstream at this moment.I agree that the application of this logic may differ between vendors,
hence why I wanted to keep this controllable by the DT property, so that
for those which do not support this use case can leave it disabled. The
logic is there to ensure that for a given USB configuration, for each EP
it would have at least 1 TX FIFO. For USB configurations which don't
utilize all available IN EPs, it would allow re-allocation of internal
memory to EPs which will actually be in use.Quoted 7 lines
Quoted 4 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can
be a little nicer in this regard.Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
sure, I support that. But I want to make sure the first cut isn't likely
to break things left and right :)Quoted 5 lines
Quoted 2 lines
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty
big improvements on the TX path with this.fingers crossed
Quoted 30 lines
Quoted 17 lines
Quoted 4 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were
just to remove physical EP1 and 2 from the DWC3 endpoint list. If thatright, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to
skip that endpoint on that particular platform without interefering with
everything else.Of course, we still need to figure out a way to abstract the different
dwc3 instantiations.Quoted 9 lines
Quoted 3 lines
was the change which ended up upstream for the Intel tracer then we
could improve the logic to avoid re-sizing those particular EPs.The problem then, just as I mentioned in the previous paragraph, will be
coming up with a solution that's elegant and works for all different
instantiations of dwc3 (or musb, cdns3, etc).Well, at least for the TX FIFO resizing logic, we'd only be needing to
focus on the DWC3 implementation.Quoted 8 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We
have a few special HW eps as well, which we try to maintain separately
in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).Awesome, as mentioned, the endpoint feature flags were added exactly to
allow for these vendor-specific features :-)I'm more than happy to help testing now that I finally got our SM8150 Surface Duo device tree accepted by Bjorn ;-)
Quoted 15 lines
Quoted 9 lines
Quoted 3 lines
However, I'm not sure how the changes would look like in the end, so I
would like to wait later down the line to include that :).Fair enough, I agree. Can we get some more testing of $subject, though?
Did you test $subject with upstream too? Which gadget drivers did you
use? How did you testThe results that I included in the cover page was tested with the pure
upstream kernel on our device. Below was using the ConfigFS gadget w/ a
mass storage only composition.Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 193.60| | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 287.35| | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my
memory could be failing.Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 5 lines
We also have internal numbers which have shown similar improvements as
well. Those are over networking/tethering interfaces, so testing IPERF
loopback over TCP/UDP.loopback iperf? That would skip the wire, no?
Quoted 31 lines
Quoted 17 lines
Quoted 8 lines
size of 2 and TX threshold of 1, this would really be not beneficial to
us, because we can only change the TX threshold to 2 at max, and at
least in my observations, once we have to go out to system memory to
fetch the next data packet, that latency takes enough time for the controller to end the current burst.What I noticed with
g_mass_storageis that we can amortize the cost of
fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was
enough to give me very good performance. Never had to poke at TX FIFO
resizing. Did you try something like this too?Quoted 4 lines
I feel that allocating more requests is a far simpler and more generic
method that changing FIFO sizes :)I wish I had a USB bus trace handy to show you, which would make it very
clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So
by increasing the number of USB requests, that will help if there was a
bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.Quoted 6 lines
So yes, this method of increasing the # of USB reqs will definitely help
with situations such as HSUSB or in SSUSB when EP bursting isn't used.
The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.Hmm, that's not what I remember. Perhaps the TRB cache size plays a role
here too. I have clear memories of testing this very scenario of bursting (usingg_mass_storageat the time) because I was curious about
it. Back then, my tests showed no difference in behavior.Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
It could be nice if Heikki could test Intel parts with and without your
changes ong_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
I just tested on 5.13.0-rc4 on Intel Edison (
x86_64). All 5 patches apply. Switching between host/gadget works, no connections dropping, no errors in dmesg.In host mode I connect a smsc9504 eth+4p hub. In gadget mode I have composite device created from configfs with gser / eem /
mass_storage/ uac2.Tested with iperf3 performance in host (93.6Mbits/sec) and gadget (207Mbits/sec) mode. Compared to v5.10.41 without patches host (93.4Mbits/sec) and gadget (198Mbits/sec).
Gadget seems to be a little faster with the patches, but that might also be caused by something else, on v5.10.41 I see the bitrate bouncing between 207 and 199.
Quoted 7 lines
I saw a mention to test iperf3 to self (loopback). 3.09 Gbits/sec. With v5.10.41 3.07Gbits/sec. Not bad for a 500MHz device.
With gnome-disks I did a read access benchmark 35.4MB/s, with v5.10.41 34.7MB/s. This might be limited by Edison's internal eMMC speed (when booting U-Boot reads the kernel with 21.4 MiB/s).
Hi Ferry,
Thanks for the testing. Just to double check, did you also enable the property, which enabled the TXFIFO resize feature on the platform? For example, for the QCOM SM8150 platform, we're adding the following to our device tree node:
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
tx-fifo-resize
If not, then your results at least confirms that w/o the property present, the changes won't break anything :). Thanks again for the initial testing!
I applied the patch now to 5.13.0-rc5 + the following:
drivers/usb/dwc3/dwc3-pci.c +0 -0
@@ -124,6 +124,7 @@ static const struct property_entry dwc3_pci_mrfld_properties[] = {
PROPERTY_ENTRY_BOOL("snps,dis_u3_susphy_quirk"), PROPERTY_ENTRY_BOOL("snps,dis_u2_susphy_quirk"), PROPERTY_ENTRY_BOOL("snps,usb2-gadget-lpm-disable"), + PROPERTY_ENTRY_BOOL("tx-fifo-resize"), PROPERTY_ENTRY_BOOL("linux,sysdev_is_parent"), {}
};
and when switching to gadget mode unfortunately received the following oops:
BUG: unable to handle page fault for address: 00000000202043f2 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI CPU: 0 PID: 617 Comm: conf-gadget.sh Not tainted 5.13.0-rc5-edison-acpi-standard #1 Hardware name: Intel Corporation Merrifield/BODEGA BAY, BIOS 542 2015.01.21:18.19.48 RIP: 0010:dwc3_gadget_check_config+0x33/0x80 Code: 59 04 00 00 04 74 61 48 c1 ee 10 48 89 f7 f3 48 0f b8 c7 48 89 c7 39 81 60 04 00 00 7d 4a 89 81 60 04 00 00 8b 81 08 04 00 00 <81> b8 e8 03 00 00 32 33 00 00 0f b6 b0 09 04 00 00 75 0d 8b 80 20 RSP: 0018:ffffb5550038fda0 EFLAGS: 00010297 RAX: 000000002020400a RBX: ffffa04502627348 RCX: ffffa04507354028 RDX: 0000000000000000 RSI: 000000000000003c RDI: 0000000000000004 RBP: ffffa04508ac0550 R08: ffffa04503a75b2c R09: 0000000000000000 R10: 0000000000000216 R11: 000000000002eba0 R12: ffffa04508ac0550 R13: dead000000000100 R14: ffffa04508ac0600 R15: ffffa04508ac0520 FS: 00007f7471e2f740(0000) GS:ffffa0453e200000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00000000202043f2 CR3: 0000000003f38000 CR4: 00000000001006f0 Call Trace: configfs_composite_bind+0x2f4/0x430 [libcomposite] udc_bind_to_driver+0x64/0x180 usb_gadget_probe_driver+0x114/0x150 gadget_dev_desc_UDC_store+0xbc/0x130 [libcomposite] configfs_write_file+0xcd/0x140 vfs_write+0xbb/0x250 ksys_write+0x5a/0xd0 do_syscall_64+0x40/0x80 entry_SYSCALL_64_after_hwframe+0x44/0xae RIP: 0033:0x7f7471f1ff53 Code: 8b 15 21 cf 0c 00 f7 d8 64 89 02 48 c7 c0 ff ff ff ff eb b7 0f 1f 00 64 8b 04 25 18 00 00 00 85 c0 75 14 b8 01 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 55 c3 0f 1f 40 00 48 83 ec 28 48 89 54 24 18 RSP: 002b:00007fffa3dcd328 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 RAX: ffffffffffffffda RBX: 000000000000000c RCX: 00007f7471f1ff53 RDX: 000000000000000c RSI: 00005614d615a770 RDI: 0000000000000001 RBP: 00005614d615a770 R08: 000000000000000a R09: 00007f7471fb20c0 R10: 0000000000000000 R11: 0000000000000246 R12: 000000000000000c R13: 00007f7471fee520 R14: 000000000000000c R15: 00007f7471fee720 Modules linked in: usb_f_uac2 u_audio usb_f_mass_storage usb_f_eem u_ether usb_f_serial u_serial libcomposite rfcomm iptable_nat bnep snd_sof_nocodec spi_pxa2xx_platform dw_dmac smsc snd_sof_pci_intel_tng snd_sof_pci snd_sof_acpi_intel_byt snd_sof_intel_ipc snd_sof_acpi smsc95xx snd_sof pwm_lpss_pci pwm_lpss snd_sof_xtensa_dsp snd_intel_dspcfg snd_soc_acpi_intel_match snd_soc_acpi dw_dmac_pci intel_mrfld_pwrbtn intel_mrfld_adc dw_dmac_core spi_pxa2xx_pci brcmfmac brcmutil leds_gpio hci_uart btbcm ti_ads7950 industrialio_triggered_buffer kfifo_buf ledtrig_timer ledtrig_heartbeat mmc_block extcon_intel_mrfld sdhci_pci cqhci sdhci led_class intel_soc_pmic_mrfld mmc_core btrfs libcrc32c xor zstd_compress zlib_deflate raid6_pq CR2: 00000000202043f2 ---[ end trace 5c11fe50dca92ad4 ]---
Quoted 148 lines
Quoted 6 lines
No I didn't. Afaik we don't have a devicetree property to set.
But I'd be happy to test that as well. But where to set the property?
dwc3_pci_mrfld_properties[] in dwc3-pci?
Hi Ferry,
Not too sure which DWC3 driver is used for the Intel platform, but I believe that should be the one. (if that's what is normally used) We'd just need to add an entry w/ the below:
PROPERTY_ENTRY_BOOL("tx-fifo-resize")Thanks
Wesley ChengQuoted 131 lines
Quoted 131 lines
Thanks
Wesley ChengQuoted 128 lines
Quoted 128 lines
Quoted 128 lines
Quoted 126 lines
Quoted 5 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet,
it should have a NumP value equal to the bMaxBurst reported in the EPYes and no. Looking back at the history, we used to configure NUMP based
on bMaxBurst, but it was changed later in commit4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a problem reported by John Youn.And now we've come full circle. Because even if I believe more requests
are enough for bursting, NUMP is limited by the RxFIFO size. This ends
up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.However, note that this is about RxFIFO size, not TxFIFO size. In fact,
looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP
(emphasis is mine):"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or
equal to
the maximum burst size supported by the endpoint as
determined
by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the
description of
this field.For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this
field
may be treated by the host as informative only."However, if I remember correctly (please verify dwc3 databook), NUMP in
DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or
TxFIFO size?Quoted 7 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is
that after 2 data packets, the device issues a NRDY. So then we'd need
to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With this constantQuoted 8 lines
NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant
bursts for a request, until the request is done, or if the host runs out
of RXFIFO. (ie no interruption [on the USB protocol level] during USB
request data transfer)Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 37 lines
Quoted 37 lines
Quoted 19 lines
Quoted 15 lines
Quoted 9 lines
Quoted 5 lines
Good points.
Wesley, what kind of testing have you done on this on
different devices?As mentioned above, these changes are currently present on end user
devices for the past few years, so its been through a lot of testing :).all with the same gadget driver. Also, who uses USB on android devices
these days? Most of the data transfer goes via WiFi or Bluetooth, anyway
:-)Quoted 4 lines
I guess only developers are using USB during development to flash dev
images heh.Quoted 13 lines
I used to be a customer facing engineer, so honestly I did see some
really interesting and crazy designs. Again, we do have non-Android
products that use the same code, and it has been working in there for a
few years as well. The TXFIFO sizing really has helped with multimedia
use cases, which use isoc endpoints, since esp. in those lower end CPU
chips where latencies across the system are much larger, and a missed
ISOC interval leads to a pop in your ear.This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in
knowing if a deeper request queue would also help here.Quoted 7 lines
Quoted 5 lines
Remember dwc3 can accomodate 255 requests + link for each endpoint. If
our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst
behavior.Signature · 1 line
heikki
Quoted 279 lines · Ferry Toth
Hi
Op 15-06-2021 om 06:22 schreef Wesley Cheng:
Quoted 276 lines · Ferry Toth
Quoted 274 lines
Op 14-06-2021 om 20:58 schreef Wesley Cheng:
Quoted 273 lines · Ferry Toth
Quoted 256 lines
Hi
Op 11-06-2021 om 15:21 schreef Andy Shevchenko:
Quoted 226 lines
Quoted 224 lines · Heikki Krogerus
Quoted 223 lines · Felipe Balbi
Hi,
Quoted 48 lines · Wesley Cheng
Quoted 48 lines
Quoted 37 lines
Quoted 37 lines
Quoted 37 lines
to be honest, I don't think these should go in (apart from
the build
failure) because it's likely to break instantiations of the
core with
differing FIFO sizes. Some instantiations even have some
endpoints with
dedicated functionality that requires the default FIFO size
configured
during coreConsultant instantiation. I know of at OMAP5 and
some Intel
implementations which have dedicated endpoints for processor
tracing.With OMAP5, these endpoints are configured at the top of the
available
endpoints, which means that if a gadget driver gets loaded
and takes
over most of the FIFO space because of this resizing,
processor tracing
will have a hard time running. That being said, processor
tracing isn't
supported in upstream at this moment.I agree that the application of this logic may differ between
vendors,
hence why I wanted to keep this controllable by the DT
property, so that
for those which do not support this use case can leave it
disabled. The
logic is there to ensure that for a given USB configuration,
for each EP
it would have at least 1 TX FIFO. For USB configurations
which
don't
utilize all available IN EPs, it would allow re-allocation of
internal
memory to EPs which will actually be in use.Quoted 7 lines
Quoted 4 lines
The feature ends up being all-or-nothing, then :-) It sounds like we can
be a little nicer in this regard.Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
sure, I support that. But I want to make sure the first cut isn't likely
to break things left and right :)Quoted 5 lines
Quoted 2 lines
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty
big improvements on the TX path with this.fingers crossed
Quoted 31 lines
Quoted 18 lines
Quoted 4 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were
just to remove physical EP1 and 2 from the DWC3 endpoint list. If thatright, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before
allowing it to be enabled. Then the UDC driver could tell UDC core to
skip that endpoint on that particular platform without interefering with
everything else.Of course, we still need to figure out a way to abstract the different
dwc3 instantiations.Quoted 9 lines
Quoted 3 lines
was the change which ended up upstream for the Intel tracer then we
could improve the logic to avoid re-sizing those particular EPs.The problem then, just as I mentioned in the previous paragraph, will be
coming up with a solution that's elegant and works for all different
instantiations of dwc3 (or musb, cdns3, etc).Well, at least for the TX FIFO resizing logic, we'd only be needing to
focus on the DWC3 implementation.Quoted 10 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific
endpoints and how they can co-exist with other "normal" endpoints. We
have a few special HW eps as well, which we try to maintain separately
in our DWC3 vendor driver, but it isn't the most convenient, or most
pretty method :).Awesome, as mentioned, the endpoint feature flags were added exactly to
allow for these vendor-specific features :-)I'm more than happy to help testing now that I finally got our SM8150
Surface Duo device tree accepted by Bjorn ;-)Quoted 17 lines
Quoted 10 lines
Quoted 4 lines
However, I'm not sure how the changes would look like in the end,
so I
would like to wait later down the line to include that :).Fair enough, I agree. Can we get some more testing of $subject, though?
Did you test $subject with upstream too? Which gadget drivers did you
use? How did you testThe results that I included in the cover page was tested with the pure
upstream kernel on our device. Below was using the ConfigFS gadget
w/ a
mass storage only composition.Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 193.60| | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 287.35| | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my
memory could be failing.Then again, I never ran with CrystalDiskMark, I was using my own tool
(it's somewhere in github. If you care, I can look up the URL).Quoted 5 lines
We also have internal numbers which have shown similar improvements as
well. Those are over networking/tethering interfaces, so testing IPERF
loopback over TCP/UDP.loopback iperf? That would skip the wire, no?
Quoted 34 lines
Quoted 19 lines
Quoted 9 lines
size of 2 and TX threshold of 1, this would really be not beneficial to
us, because we can only change the TX threshold to 2 at max, and at
least in my observations, once we have to go out to system memory to
fetch the next data packet, that latency takes enough time for the
controller to end the current burst.What I noticed with
g_mass_storageis that we can amortize the cost of
fetching data from memory, with a deeper request queue. Whenever I
test(ed)g_mass_storage, I was doing so with 250 requests. And that was
enough to give me very good performance. Never had to poke at TX FIFO
resizing. Did you try something like this too?Quoted 4 lines
I feel that allocating more requests is a far simpler and more generic
method that changing FIFO sizes :)I wish I had a USB bus trace handy to show you, which would make it
very
clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So
by increasing the number of USB requests, that will help if there was a
bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.Quoted 8 lines
So yes, this method of increasing the # of USB reqs will definitely
help
with situations such as HSUSB or in SSUSB when EP bursting isn't used.
The TXFIFO resize comes into play for SSUSB, which utilizes endpoint
bursting.Hmm, that's not what I remember. Perhaps the TRB cache size plays a role
here too. I have clear memories of testing this very scenario of bursting (usingg_mass_storageat the time) because I was curious about
it. Back then, my tests showed no difference in behavior.Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
It could be nice if Heikki could test Intel parts with and without your
changes ong_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
I just tested on 5.13.0-rc4 on Intel Edison (
x86_64). All 5 patches apply. Switching between host/gadget works, no connections dropping, no
errors in dmesg.In host mode I connect a smsc9504 eth+4p hub. In gadget mode I have composite device created from configfs with gser / eem /
mass_storage/
uac2.Tested with iperf3 performance in host (93.6Mbits/sec) and gadget (207Mbits/sec) mode. Compared to v5.10.41 without patches host (93.4Mbits/sec) and gadget (198Mbits/sec).
Gadget seems to be a little faster with the patches, but that might also
be caused by something else, on v5.10.41 I see the bitrate bouncing between 207 and 199.Quoted 8 lines
I saw a mention to test iperf3 to self (loopback). 3.09 Gbits/sec. With
v5.10.41 3.07Gbits/sec. Not bad for a 500MHz device.With gnome-disks I did a read access benchmark 35.4MB/s, with v5.10.41 34.7MB/s. This might be limited by Edison's internal eMMC speed (when booting U-Boot reads the kernel with 21.4 MiB/s).
Hi Ferry,
Thanks for the testing. Just to double check, did you also enable the property, which enabled the TXFIFO resize feature on the platform? For example, for the QCOM SM8150 platform, we're adding the following to our
device tree node:
Quoted 8 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
tx-fifo-resize
If not, then your results at least confirms that w/o the property present, the changes won't break anything :). Thanks again for the initial testing!
I applied the patch now to 5.13.0-rc5 + the following:
Hi Ferry,
Quick question...there was a compile error with the V9 patch series, as it was using the dwc3_mwidth() incorrectly. I will update this with the proper use of the mdwidth, but which patch version did you use?
Thanks
Wesley Cheng
Quoted 233 lines
drivers/usb/dwc3/dwc3-pci.c +1
-0@@ -124,6 +124,7 @@ static const struct property_entry dwc3_pci_mrfld_properties[] = { PROPERTY_ENTRY_BOOL("snps,dis_u3_susphy_quirk"), PROPERTY_ENTRY_BOOL("snps,dis_u2_susphy_quirk"), PROPERTY_ENTRY_BOOL("snps,usb2-gadget-lpm-disable"), + PROPERTY_ENTRY_BOOL("tx-fifo-resize"), PROPERTY_ENTRY_BOOL("linux,sysdev_is_parent"), {}};
and when switching to gadget mode unfortunately received the following oops:
BUG: unable to handle page fault for address: 00000000202043f2 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI CPU: 0 PID: 617 Comm: conf-gadget.sh Not tainted 5.13.0-rc5-edison-acpi-standard #1 Hardware name: Intel Corporation Merrifield/BODEGA BAY, BIOS 542 2015.01.21:18.19.48 RIP: 0010:dwc3_gadget_check_config+0x33/0x80 Code: 59 04 00 00 04 74 61 48 c1 ee 10 48 89 f7 f3 48 0f b8 c7 48 89 c739 81 60 04 00 00 7d 4a 89 81 60 04 00 00 8b 81 08 04 00 00 <81> b8 e8 03 00 00 32 33 00 00 0f b6 b0 09 04 00 00 75 0d 8b 80 20
RSP: 0018:ffffb5550038fda0 EFLAGS: 00010297
RAX:000000002020400aRBX:ffffa04502627348RCX:ffffa04507354028
RDX: 0000000000000000 RSI:000000000000003cRDI: 0000000000000004
RBP:ffffa04508ac0550R08:ffffa04503a75b2cR09: 0000000000000000
R10: 0000000000000216 R11:000000000002eba0R12:ffffa04508ac0550
R13:dead000000000100R14:ffffa04508ac0600R15:ffffa04508ac0520FS: 00007f7471e2f740(0000) GS:ffffa0453e200000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2:00000000202043f2CR3:0000000003f38000CR4:00000000001006f0Call Trace: configfs_composite_bind+0x2f4/0x430 [libcomposite] udc_bind_to_driver+0x64/0x180 usb_gadget_probe_driver+0x114/0x150 gadget_dev_desc_UDC_store+0xbc/0x130 [libcomposite] configfs_write_file+0xcd/0x140 vfs_write+0xbb/0x250 ksys_write+0x5a/0xd0 do_syscall_64+0x40/0x80 entry_SYSCALL_64_after_hwframe+0x44/0xae RIP: 0033:0x7f7471f1ff53 Code: 8b 15 21 cf 0c 00 f7 d8 64 89 02 48 c7 c0 ff ff ff ff eb b7 0f 1f00 64 8b 04 25 18 00 00 00 85 c0 75 14 b8 01 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 55 c3 0f 1f 40 00 48 83 ec 28 48 89 54 24 18
RSP: 002b:00007fffa3dcd328 EFLAGS: 00000246ORIG_RAX: 0000000000000001
RAX: ffffffffffffffda RBX:000000000000000cRCX:00007f7471f1ff53
RDX:000000000000000cRSI:00005614d615a770RDI: 0000000000000001
RBP:00005614d615a770R08:000000000000000aR09:00007f7471fb20c0
R10: 0000000000000000 R11: 0000000000000246 R12:000000000000000c
R13:00007f7471fee520R14:000000000000000cR15:00007f7471fee720Modules linked in: usb_f_uac2 u_audio usb_f_mass_storage usb_f_eem
u_etherusb_f_serialu_seriallibcomposite rfcommiptable_natbnepsnd_sof_nocodecspi_pxa2xx_platformdw_dmacsmscsnd_sof_pci_intel_tngsnd_sof_pcisnd_sof_acpi_intel_bytsnd_sof_intel_ipcsnd_sof_acpismsc95xxsnd_sofpwm_lpss_pcipwm_lpsssnd_sof_xtensa_dspsnd_intel_dspcfgsnd_soc_acpi_intel_matchsnd_soc_acpidw_dmac_pciintel_mrfld_pwrbtnintel_mrfld_adcdw_dmac_corespi_pxa2xx_pcibrcmfmac brcmutilleds_gpiohci_uartbtbcmti_ads7950industrialio_triggered_bufferkfifo_bufledtrig_timerledtrig_heartbeatmmc_blockextcon_intel_mrfldsdhci_pcicqhci sdhciled_classintel_soc_pmic_mrfldmmc_corebtrfs libcrc32c xorzstd_compresszlib_deflateraid6_pq
CR2:00000000202043f2---[ end trace 5c11fe50dca92ad4 ]---Quoted 160 lines
Quoted 6 lines
No I didn't. Afaik we don't have a devicetree property to set.
But I'd be happy to test that as well. But where to set the property?
dwc3_pci_mrfld_properties[] in dwc3-pci?
Hi Ferry,
Not too sure which DWC3 driver is used for the Intel platform, but I believe that should be the one. (if that's what is normally used) We'd just need to add an entry w/ the below:
PROPERTY_ENTRY_BOOL("tx-fifo-resize")Thanks
Wesley ChengQuoted 143 lines
Quoted 143 lines
Thanks
Wesley ChengQuoted 140 lines
Quoted 140 lines
Quoted 140 lines
Quoted 133 lines
Quoted 5 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet,
it should have a NumP value equal to the bMaxBurst reported in the EPYes and no. Looking back at the history, we used to configure NUMP based
on bMaxBurst, but it was changed later in commit4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a
problem reported by John Youn.And now we've come full circle. Because even if I believe more requests
are enough for bursting, NUMP is limited by the RxFIFO size. This ends
up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.However, note that this is about RxFIFO size, not TxFIFO size. In fact,
looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP
(emphasis is mine):"Number of Packets (NumP). This field is used to indicate the
number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to
the maximum burst size supported by the endpoint as determined
by the value in the bMaxBurst field in the Endpoint Companion
Descriptor (refer to Section 9.6.7)."So, NumP is for the receiver, not the transmitter. Could you clarify
what you mean here?/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of
this field.For an IN endpoint this field is set by the endpoint to the
number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater
than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint
Companion Descriptor. Note that the value reported in this field
may be treated by the host as informative only."However, if I remember correctly (please verify dwc3 databook), NUMP in
DCFG was only for receive buffers. Thin, John, how does dwc3 compute
NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or
TxFIFO size?Quoted 9 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is
that after 2 data packets, the device issues a NRDY. So then we'd need
to send an ERDY once data is available within the FIFO, and the same
sequence happens until the USB request is complete. With this constant
NRDY/ERDY handshake going on, you actually see that the bus isQuoted 8 lines
under
utilized. When we increase an EP's FIFO size, then you'll see constant
bursts for a request, until the request is done, or if the host runs out
of RXFIFO. (ie no interruption [on the USB protocol level] during USB
request data transfer)Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 44 lines
Quoted 44 lines
Quoted 19 lines
Quoted 15 lines
Quoted 9 lines
Good points.
Wesley, what kind of testing have you done on this on
different devices?As mentioned above, these changes are currently present on end
user
devices for the past few years, so its been through a lot of
testing :).all with the same gadget driver. Also, who uses USB on android devices
these days? Most of the data transfer goes via WiFi or Bluetooth, anyway
:-)Quoted 4 lines
I guess only developers are using USB during development to flash dev
images heh.Quoted 14 lines
I used to be a customer facing engineer, so honestly I did see some
really interesting and crazy designs. Again, we do have non-Android
products that use the same code, and it has been working in there
for a
few years as well. The TXFIFO sizing really has helped with multimedia
use cases, which use isoc endpoints, since esp. in those lower end CPU
chips where latencies across the system are much larger, and a missed
ISOC interval leads to a pop in your ear.This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in
knowing if a deeper request queue would also help here.Remember dwc3 can accomodate 255 requests + link for each endpoint. If
our gadget driver uses a low number of requests, we're never really
using the TRB ring in our benefit.Quoted 5 lines
Quoted 2 lines
We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst
behavior.--
heikki
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 294 lines · Ferry Toth
Hi
Op 17-06-2021 om 06:25 schreef Wesley Cheng:
Quoted 291 lines · Ferry Toth
Quoted 278 lines
Hi
Op 15-06-2021 om 06:22 schreef Wesley Cheng:
Quoted 275 lines · Ferry Toth
Quoted 274 lines
Op 14-06-2021 om 20:58 schreef Wesley Cheng:
Quoted 273 lines · Ferry Toth
Quoted 256 lines
Hi
Op 11-06-2021 om 15:21 schreef Andy Shevchenko:
Quoted 226 lines
Quoted 224 lines · Heikki Krogerus
Quoted 223 lines · Felipe Balbi
Hi,
Quoted 48 lines · Wesley Cheng
Quoted 48 lines
Quoted 44 lines
to be honest, I don't think these should go in (apart from
the build
failure) because it's likely to break instantiations of the
core with
differing FIFO sizes. Some instantiations even have some
endpoints with
dedicated functionality that requires the default FIFO size
configured
during coreConsultant instantiation. I know of at OMAP5 and
some Intel
implementations which have dedicated endpoints for processor
tracing.With OMAP5, these endpoints are configured at the top of the
available
endpoints, which means that if a gadget driver gets loaded
and takes
over most of the FIFO space because of this resizing,
processor tracing
will have a hard time running. That being said, processor
tracing isn't
supported in upstream at this moment.I agree that the application of this logic may differ between
vendors,
hence why I wanted to keep this controllable by the DT
property, so that
for those which do not support this use case can leave it
disabled. The
logic is there to ensure that for a given USB configuration,
for each EP
it would have at least 1 TX FIFO. For USB configurations
which
don't
utilize all available IN EPs, it would allow re-allocation of
internal
memory to EPs which will actually be in use.
The feature ends up being all-or-nothing, then :-) It sounds
like we can
be a little nicer in this regard.Don't get me wrong, I think once those features become available
upstream, we can improve the logic. From what I remember when
lookingsure, I support that. But I want to make sure the first cut isn't likely
to break things left and right :)Quoted 5 lines
Quoted 2 lines
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty
big improvements on the TX path with this.fingers crossed
Quoted 31 lines
Quoted 18 lines
Quoted 4 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes
were
just to remove physical EP1 and 2 from the DWC3 endpoint list.
If thatright, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before
allowing it to be enabled. Then the UDC driver could tell UDC core to
skip that endpoint on that particular platform without interefering with
everything else.Of course, we still need to figure out a way to abstract the different
dwc3 instantiations.Quoted 9 lines
Quoted 3 lines
was the change which ended up upstream for the Intel tracer
then we
could improve the logic to avoid re-sizing those particular EPs.The problem then, just as I mentioned in the previous paragraph, will be
coming up with a solution that's elegant and works for all different
instantiations of dwc3 (or musb, cdns3, etc).Well, at least for the TX FIFO resizing logic, we'd only be needing to
focus on the DWC3 implementation.Quoted 10 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific
endpoints and how they can co-exist with other "normal" endpoints. We
have a few special HW eps as well, which we try to maintain separately
in our DWC3 vendor driver, but it isn't the most convenient, or most
pretty method :).Awesome, as mentioned, the endpoint feature flags were added exactly to
allow for these vendor-specific features :-)I'm more than happy to help testing now that I finally got our SM8150
Surface Duo device tree accepted by Bjorn ;-)Quoted 17 lines
Quoted 10 lines
Quoted 4 lines
However, I'm not sure how the changes would look like in the
end,
so I
would like to wait later down the line to include that :).Fair enough, I agree. Can we get some more testing of $subject, though?
Did you test $subject with upstream too? Which gadget drivers did you
use? How did you testThe results that I included in the cover page was tested with the pure
upstream kernel on our device. Below was using the ConfigFS gadget
w/ a
mass storage only composition.Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 193.60| | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 287.35| | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my
memory could be failing.Then again, I never ran with CrystalDiskMark, I was using my own tool
(it's somewhere in github. If you care, I can look up the URL).Quoted 5 lines
We also have internal numbers which have shown similar improvements as
well. Those are over networking/tethering interfaces, so testing IPERF
loopback over TCP/UDP.loopback iperf? That would skip the wire, no?
Quoted 34 lines
Quoted 19 lines
Quoted 9 lines
size of 2 and TX threshold of 1, this would really be not
beneficial to
us, because we can only change the TX threshold to 2 at max,
and at
least in my observations, once we have to go out to system
memory to
fetch the next data packet, that latency takes enough time
for the
controller to end the current burst.What I noticed with
g_mass_storageis that we can amortize the cost of
fetching data from memory, with a deeper request queue. Whenever I
test(ed)g_mass_storage, I was doing so with 250 requests. And that was
enough to give me very good performance. Never had to poke at TX FIFO
resizing. Did you try something like this too?Quoted 4 lines
I feel that allocating more requests is a far simpler and more generic
method that changing FIFO sizes :)I wish I had a USB bus trace handy to show you, which would make it
very
clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So
by increasing the number of USB requests, that will help if there was a
bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.Quoted 8 lines
So yes, this method of increasing the # of USB reqs will definitely
help
with situations such as HSUSB or in SSUSB when EP bursting isn't used.
The TXFIFO resize comes into play for SSUSB, which utilizes endpoint
bursting.Hmm, that's not what I remember. Perhaps the TRB cache size plays a role
here too. I have clear memories of testing this very scenario of bursting (usingg_mass_storageat the time) because I was curious about
it. Back then, my tests showed no difference in behavior.Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
It could be nice if Heikki could test Intel parts with and without your
changes ong_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
I just tested on 5.13.0-rc4 on Intel Edison (
x86_64). All 5 patches apply. Switching between host/gadget works, no connections dropping, no
errors in dmesg.In host mode I connect a smsc9504 eth+4p hub. In gadget mode I have composite device created from configfs with gser / eem /
mass_storage/
uac2.Tested with iperf3 performance in host (93.6Mbits/sec) and gadget (207Mbits/sec) mode. Compared to v5.10.41 without patches host (93.4Mbits/sec) and gadget (198Mbits/sec).
Gadget seems to be a little faster with the patches, but that might also
be caused by something else, on v5.10.41 I see the bitrate bouncing between 207 and 199.Quoted 8 lines
I saw a mention to test iperf3 to self (loopback). 3.09 Gbits/sec. With
v5.10.41 3.07Gbits/sec. Not bad for a 500MHz device.With gnome-disks I did a read access benchmark 35.4MB/s, with v5.10.41 34.7MB/s. This might be limited by Edison's internal eMMC speed (when booting U-Boot reads the kernel with 21.4 MiB/s).
Hi Ferry,
Thanks for the testing. Just to double check, did you also enable the property, which enabled the TXFIFO resize feature on the platform? For example, for the QCOM SM8150 platform, we're adding the following to our
device tree node:Quoted 7 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
tx-fifo-resize
If not, then your results at least confirms that w/o the property present, the changes won't break anything :). Thanks again for the initial testing!
I applied the patch now to 5.13.0-rc5 + the following:
Hi Ferry,
Quick question...there was a compile error with the V9 patch series, as it was using the
dwc3_mwidth()incorrectly. I will update this with the proper use of the mdwidth, but which patch version did you use?
Hi Ferry,
Quoted 3 lines
The V9 set gets applied to 5.13.0-rc5 by Yocto, if it doesn't apply it stops the build. I didn't notice any compile errors, they stop the whole build process too. But warnings are ignored. I'll check the logs to be sure.
Ah, ok. I think the incorrect usage of the API will result as a warning as seen in Greg's compile log:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]
653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0);
This is probably why the page fault occurs, as we're not passing in the DWC3 struct. I will send out a V10 shortly after testing it on my device.
Thanks
Wesley Cheng
Quoted 236 lines
Quoted 236 lines
Thanks
Wesley ChengQuoted 233 lines
drivers/usb/dwc3/dwc3-pci.c +1
-0@@ -124,6 +124,7 @@ static const struct property_entry dwc3_pci_mrfld_properties[] = { PROPERTY_ENTRY_BOOL("snps,dis_u3_susphy_quirk"), PROPERTY_ENTRY_BOOL("snps,dis_u2_susphy_quirk"), PROPERTY_ENTRY_BOOL("snps,usb2-gadget-lpm-disable"), + PROPERTY_ENTRY_BOOL("tx-fifo-resize"), PROPERTY_ENTRY_BOOL("linux,sysdev_is_parent"), {}};
and when switching to gadget mode unfortunately received the following oops:
BUG: unable to handle page fault for address: 00000000202043f2 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI CPU: 0 PID: 617 Comm: conf-gadget.sh Not tainted 5.13.0-rc5-edison-acpi-standard #1 Hardware name: Intel Corporation Merrifield/BODEGA BAY, BIOS 542 2015.01.21:18.19.48 RIP: 0010:dwc3_gadget_check_config+0x33/0x80 Code: 59 04 00 00 04 74 61 48 c1 ee 10 48 89 f7 f3 48 0f b8 c7 48 89 c739 81 60 04 00 00 7d 4a 89 81 60 04 00 00 8b 81 08 04 00 00 <81> b8 e8 03 00 00 32 33 00 00 0f b6 b0 09 04 00 00 75 0d 8b 80 20
RSP: 0018:ffffb5550038fda0 EFLAGS: 00010297
RAX:000000002020400aRBX:ffffa04502627348RCX:ffffa04507354028
RDX: 0000000000000000 RSI:000000000000003cRDI: 0000000000000004
RBP:ffffa04508ac0550R08:ffffa04503a75b2cR09: 0000000000000000
R10: 0000000000000216 R11:000000000002eba0R12:ffffa04508ac0550
R13:dead000000000100R14:ffffa04508ac0600R15:ffffa04508ac0520FS: 00007f7471e2f740(0000) GS:ffffa0453e200000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2:00000000202043f2CR3:0000000003f38000CR4:00000000001006f0Call Trace: configfs_composite_bind+0x2f4/0x430 [libcomposite] udc_bind_to_driver+0x64/0x180 usb_gadget_probe_driver+0x114/0x150 gadget_dev_desc_UDC_store+0xbc/0x130 [libcomposite] configfs_write_file+0xcd/0x140 vfs_write+0xbb/0x250 ksys_write+0x5a/0xd0 do_syscall_64+0x40/0x80 entry_SYSCALL_64_after_hwframe+0x44/0xae RIP: 0033:0x7f7471f1ff53 Code: 8b 15 21 cf 0c 00 f7 d8 64 89 02 48 c7 c0 ff ff ff ff eb b7 0f 1f00 64 8b 04 25 18 00 00 00 85 c0 75 14 b8 01 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 55 c3 0f 1f 40 00 48 83 ec 28 48 89 54 24 18
RSP: 002b:00007fffa3dcd328 EFLAGS: 00000246ORIG_RAX: 0000000000000001
RAX: ffffffffffffffda RBX:000000000000000cRCX:00007f7471f1ff53
RDX:000000000000000cRSI:00005614d615a770RDI: 0000000000000001
RBP:00005614d615a770R08:000000000000000aR09:00007f7471fb20c0
R10: 0000000000000000 R11: 0000000000000246 R12:000000000000000c
R13:00007f7471fee520R14:000000000000000cR15:00007f7471fee720Modules linked in: usb_f_uac2 u_audio usb_f_mass_storage usb_f_eem
u_etherusb_f_serialu_seriallibcomposite rfcommiptable_natbnepsnd_sof_nocodecspi_pxa2xx_platformdw_dmacsmscsnd_sof_pci_intel_tngsnd_sof_pcisnd_sof_acpi_intel_bytsnd_sof_intel_ipcsnd_sof_acpismsc95xxsnd_sofpwm_lpss_pcipwm_lpsssnd_sof_xtensa_dspsnd_intel_dspcfgsnd_soc_acpi_intel_matchsnd_soc_acpidw_dmac_pciintel_mrfld_pwrbtnintel_mrfld_adcdw_dmac_corespi_pxa2xx_pcibrcmfmac brcmutilleds_gpiohci_uartbtbcmti_ads7950industrialio_triggered_bufferkfifo_bufledtrig_timerledtrig_heartbeatmmc_blockextcon_intel_mrfldsdhci_pcicqhci sdhciled_classintel_soc_pmic_mrfldmmc_corebtrfs libcrc32c xorzstd_compresszlib_deflateraid6_pq
CR2:00000000202043f2---[ end trace 5c11fe50dca92ad4 ]---Quoted 160 lines
Quoted 6 lines
No I didn't. Afaik we don't have a devicetree property to set.
But I'd be happy to test that as well. But where to set the property?
dwc3_pci_mrfld_properties[] in dwc3-pci?
Hi Ferry,
Not too sure which DWC3 driver is used for the Intel platform, but I believe that should be the one. (if that's what is normally used) We'd just need to add an entry w/ the below:
PROPERTY_ENTRY_BOOL("tx-fifo-resize")Thanks
Wesley ChengQuoted 143 lines
Quoted 143 lines
Thanks
Wesley ChengQuoted 140 lines
Quoted 140 lines
Quoted 140 lines
Quoted 133 lines
Quoted 5 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet,
it should have a NumP value equal to the bMaxBurst reported in the EPYes and no. Looking back at the history, we used to configure NUMP based
on bMaxBurst, but it was changed later in commit4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a
problem reported by John Youn.And now we've come full circle. Because even if I believe more requests
are enough for bursting, NUMP is limited by the RxFIFO size. This ends
up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.However, note that this is about RxFIFO size, not TxFIFO size. In fact,
looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP
(emphasis is mine):"Number of Packets (NumP). This field is used to indicate the
number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to
the maximum burst size supported by the endpoint as determined
by the value in the bMaxBurst field in the Endpoint Companion
Descriptor (refer to Section 9.6.7)."So, NumP is for the receiver, not the transmitter. Could you clarify
what you mean here?/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of
this field.For an IN endpoint this field is set by the endpoint to the
number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater
than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint
Companion Descriptor. Note that the value reported in this field
may be treated by the host as informative only."However, if I remember correctly (please verify dwc3 databook), NUMP in
DCFG was only for receive buffers. Thin, John, how does dwc3 compute
NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or
TxFIFO size?Quoted 9 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is
that after 2 data packets, the device issues a NRDY. So then we'd need
to send an ERDY once data is available within the FIFO, and the same
sequence happens until the USB request is complete. With this constant
NRDY/ERDY handshake going on, you actually see that the bus isQuoted 8 lines
under
utilized. When we increase an EP's FIFO size, then you'll see constant
bursts for a request, until the request is done, or if the host runs out
of RXFIFO. (ie no interruption [on the USB protocol level] during USB
request data transfer)Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 44 lines
Quoted 44 lines
Quoted 33 lines
Good points.
Wesley, what kind of testing have you done on this on
different devices?As mentioned above, these changes are currently present on end
user
devices for the past few years, so its been through a lot of
testing :).
all with the same gadget driver. Also, who uses USB on android
devices
these days? Most of the data transfer goes via WiFi or
Bluetooth, anyway
:-)I guess only developers are using USB during development to
flash dev
images heh.I used to be a customer facing engineer, so honestly I did see
some
really interesting and crazy designs. Again, we do have
non-Android
products that use the same code, and it has been working in
there
for a
few years as well. The TXFIFO sizing really has helped with
multimedia
use cases, which use isoc endpoints, since esp. in those lower
end CPU
chips where latencies across the system are much larger, and a
missed
ISOC interval leads to a pop in your ear.This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in
knowing if a deeper request queue would also help here.Remember dwc3 can accomodate 255 requests + link for each endpoint. If
our gadget driver uses a low number of requests, we're never really
using the TRB ring in our benefit.Quoted 5 lines
Quoted 2 lines
We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst
behavior.--
heikki
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
Hi
Op 17-06-2021 om 10:30 schreef Wesley Cheng:
Quoted 301 lines · Ferry Toth
Quoted 294 lines
Hi
Op 17-06-2021 om 06:25 schreef Wesley Cheng:
Quoted 291 lines · Ferry Toth
Quoted 278 lines
Hi
Op 15-06-2021 om 06:22 schreef Wesley Cheng:
Quoted 275 lines · Ferry Toth
Quoted 274 lines
Op 14-06-2021 om 20:58 schreef Wesley Cheng:
Quoted 273 lines · Ferry Toth
Quoted 256 lines
Hi
Op 11-06-2021 om 15:21 schreef Andy Shevchenko:
Quoted 226 lines
Quoted 224 lines · Heikki Krogerus
Quoted 223 lines · Felipe Balbi
Hi,
Quoted 48 lines · Wesley Cheng
Quoted 48 lines
to be honest, I don't think these should go in (apart from
the build
failure) because it's likely to break instantiations of the
core with
differing FIFO sizes. Some instantiations even have some
endpoints with
dedicated functionality that requires the default FIFO size
configured
during coreConsultant instantiation. I know of at OMAP5 and
some Intel
implementations which have dedicated endpoints for processor
tracing.With OMAP5, these endpoints are configured at the top of the
available
endpoints, which means that if a gadget driver gets loaded
and takes
over most of the FIFO space because of this resizing,
processor tracing
will have a hard time running. That being said, processor
tracing isn't
supported in upstream at this moment.I agree that the application of this logic may differ between
vendors,
hence why I wanted to keep this controllable by the DT
property, so that
for those which do not support this use case can leave it
disabled. The
logic is there to ensure that for a given USB configuration,
for each EP
it would have at least 1 TX FIFO. For USB configurations
which
don't
utilize all available IN EPs, it would allow re-allocation of
internal
memory to EPs which will actually be in use.
The feature ends up being all-or-nothing, then :-) It sounds
like we can
be a little nicer in this regard.Don't get me wrong, I think once those features become available
upstream, we can improve the logic. From what I remember when
looking
sure, I support that. But I want to make sure the first cut isn't
likely
to break things left and right :)Quoted 5 lines
Quoted 2 lines
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty
big improvements on the TX path with this.fingers crossed
Quoted 31 lines
Quoted 27 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes
were
just to remove physical EP1 and 2 from the DWC3 endpoint list.
If that
right, that's the reason why we introduced the endpoint feature
flags. The end goal was that the UDC would be able to have custom
feature flags paired with ->validate_endpoint() or whatever
before
allowing it to be enabled. Then the UDC driver could tell UDC
core to
skip that endpoint on that particular platform without
interefering with
everything else.Of course, we still need to figure out a way to abstract the
different
dwc3 instantiations.was the change which ended up upstream for the Intel tracer
then we
could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph,
will be
coming up with a solution that's elegant and works for all
different
instantiations of dwc3 (or musb, cdns3, etc).Well, at least for the TX FIFO resizing logic, we'd only be needing to
focus on the DWC3 implementation.Quoted 10 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific
endpoints and how they can co-exist with other "normal" endpoints. We
have a few special HW eps as well, which we try to maintain separately
in our DWC3 vendor driver, but it isn't the most convenient, or most
pretty method :).Awesome, as mentioned, the endpoint feature flags were added exactly to
allow for these vendor-specific features :-)I'm more than happy to help testing now that I finally got our SM8150
Surface Duo device tree accepted by Bjorn ;-)Quoted 17 lines
Quoted 10 lines
However, I'm not sure how the changes would look like in the
end,
so I
would like to wait later down the line to include that :).
Fair enough, I agree. Can we get some more testing of $subject,
though?
Did you test $subject with upstream too? Which gadget drivers
did you
use? How did you testThe results that I included in the cover page was tested with the pure
upstream kernel on our device. Below was using the ConfigFS gadget
w/ a
mass storage only composition.Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 193.60| | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------
Sequential|1 GB x |
Read |9 loops | 287.35| | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my
memory could be failing.Then again, I never ran with CrystalDiskMark, I was using my own tool
(it's somewhere in github. If you care, I can look up the URL).Quoted 5 lines
We also have internal numbers which have shown similar improvements as
well. Those are over networking/tethering interfaces, so testing IPERF
loopback over TCP/UDP.loopback iperf? That would skip the wire, no?
Quoted 34 lines
Quoted 23 lines
size of 2 and TX threshold of 1, this would really be not
beneficial to
us, because we can only change the TX threshold to 2 at max,
and at
least in my observations, once we have to go out to system
memory to
fetch the next data packet, that latency takes enough time
for the
controller to end the current burst.
What I noticed with g_mass_storage is that we can amortize the
cost of
fetching data from memory, with a deeper request queue.
Whenever I
test(ed) g_mass_storage, I was doing so with 250 requests. And
that was
enough to give me very good performance. Never had to poke at TX
FIFO
resizing. Did you try something like this too?I feel that allocating more requests is a far simpler and more
generic
method that changing FIFO sizes :)I wish I had a USB bus trace handy to show you, which would make it
very
clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So
by increasing the number of USB requests, that will help if there was a
bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.Quoted 8 lines
So yes, this method of increasing the # of USB reqs will definitely
help
with situations such as HSUSB or in SSUSB when EP bursting isn't used.
The TXFIFO resize comes into play for SSUSB, which utilizes endpoint
bursting.Hmm, that's not what I remember. Perhaps the TRB cache size plays a role
here too. I have clear memories of testing this very scenario of bursting (usingg_mass_storageat the time) because I was curious about
it. Back then, my tests showed no difference in behavior.Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
It could be nice if Heikki could test Intel parts with and without your
changes ong_mass_storagewith 250 requests.Andy, you have a system at hand that has the DWC3 block enabled, right? Can you help out here?
I'm not sure if i will have time soon, I Cc'ed to Ferry who has a few more test cases (I have only one or two) and maybe can help. But I'll keep this in mind.
I just tested on 5.13.0-rc4 on Intel Edison (
x86_64). All 5 patches apply. Switching between host/gadget works, no connections dropping, no
errors in dmesg.In host mode I connect a smsc9504 eth+4p hub. In gadget mode I have composite device created from configfs with gser / eem /
mass_storage/
uac2.Tested with iperf3 performance in host (93.6Mbits/sec) and gadget (207Mbits/sec) mode. Compared to v5.10.41 without patches host (93.4Mbits/sec) and gadget (198Mbits/sec).
Gadget seems to be a little faster with the patches, but that might also
be caused by something else, on v5.10.41 I see the bitrate bouncing between 207 and 199.Quoted 8 lines
I saw a mention to test iperf3 to self (loopback). 3.09 Gbits/sec. With
v5.10.41 3.07Gbits/sec. Not bad for a 500MHz device.With gnome-disks I did a read access benchmark 35.4MB/s, with v5.10.41 34.7MB/s. This might be limited by Edison's internal eMMC speed (when booting U-Boot reads the kernel with 21.4 MiB/s).
Hi Ferry,
Thanks for the testing. Just to double check, did you also enable the property, which enabled the TXFIFO resize feature on the platform? For example, for the QCOM SM8150 platform, we're adding the following to our
device tree node:Quoted 7 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
tx-fifo-resize
If not, then your results at least confirms that w/o the property present, the changes won't break anything :). Thanks again for the initial testing!
I applied the patch now to 5.13.0-rc5 + the following:
Hi Ferry,
Quick question...there was a compile error with the V9 patch series, as it was using the
dwc3_mwidth()incorrectly. I will update this with the proper use of the mdwidth, but which patch version did you use?Hi Ferry,
Quoted 3 lines
The V9 set gets applied to 5.13.0-rc5 by Yocto, if it doesn't apply it stops the build. I didn't notice any compile errors, they stop the whole build process too. But warnings are ignored. I'll check the logs to be sure.
Quoted 7 lines
Ah, ok. I think the incorrect usage of the API will result as a warning as seen in Greg's compile log:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]
653 | mdwidth =dwc3_mdwidth(dwc->hwparams.hwparams0);
I found exactly this warning in my logs. Sorry for this noise, I was watching for an error.
Quoted 242 lines
This is probably why the page fault occurs, as we're not passing in the DWC3 struct. I will send out a V10 shortly after testing it on my device.
Thanks
Wesley ChengQuoted 236 lines
Quoted 236 lines
Thanks
Wesley ChengQuoted 233 lines
drivers/usb/dwc3/dwc3-pci.c +1
-0@@ -124,6 +124,7 @@ static const struct property_entry dwc3_pci_mrfld_properties[] = { PROPERTY_ENTRY_BOOL("snps,dis_u3_susphy_quirk"), PROPERTY_ENTRY_BOOL("snps,dis_u2_susphy_quirk"), PROPERTY_ENTRY_BOOL("snps,usb2-gadget-lpm-disable"), + PROPERTY_ENTRY_BOOL("tx-fifo-resize"), PROPERTY_ENTRY_BOOL("linux,sysdev_is_parent"), {}};
and when switching to gadget mode unfortunately received the following
oops:BUG: unable to handle page fault for address: 00000000202043f2 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI CPU: 0 PID: 617 Comm: conf-gadget.sh Not tainted 5.13.0-rc5-edison-acpi-standard #1 Hardware name: Intel Corporation Merrifield/BODEGA BAY, BIOS 542 2015.01.21:18.19.48 RIP: 0010:dwc3_gadget_check_config+0x33/0x80 Code: 59 04 00 00 04 74 61 48 c1 ee 10 48 89 f7 f3 48 0f b8 c7 48 89 c739 81 60 04 00 00 7d 4a 89 81 60 04 00 00 8b 81 08 04 00 00 <81> b8 e8 03 00 00 32 33 00 00 0f b6 b0 09 04 00 00 75 0d 8b 80 20
RSP: 0018:ffffb5550038fda0 EFLAGS: 00010297
RAX:000000002020400aRBX:ffffa04502627348RCX:ffffa04507354028
RDX: 0000000000000000 RSI:000000000000003cRDI: 0000000000000004
RBP:ffffa04508ac0550R08:ffffa04503a75b2cR09: 0000000000000000
R10: 0000000000000216 R11:000000000002eba0R12:ffffa04508ac0550
R13:dead000000000100R14:ffffa04508ac0600R15:ffffa04508ac0520FS: 00007f7471e2f740(0000) GS:ffffa0453e200000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2:00000000202043f2CR3:0000000003f38000CR4:00000000001006f0Call Trace: configfs_composite_bind+0x2f4/0x430 [libcomposite] udc_bind_to_driver+0x64/0x180 usb_gadget_probe_driver+0x114/0x150 gadget_dev_desc_UDC_store+0xbc/0x130 [libcomposite] configfs_write_file+0xcd/0x140 vfs_write+0xbb/0x250 ksys_write+0x5a/0xd0 do_syscall_64+0x40/0x80 entry_SYSCALL_64_after_hwframe+0x44/0xae RIP: 0033:0x7f7471f1ff53 Code: 8b 15 21 cf 0c 00 f7 d8 64 89 02 48 c7 c0 ff ff ff ff eb b7 0f 1f00 64 8b 04 25 18 00 00 00 85 c0 75 14 b8 01 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 55 c3 0f 1f 40 00 48 83 ec 28 48 89 54 24 18
RSP: 002b:00007fffa3dcd328 EFLAGS: 00000246ORIG_RAX: 0000000000000001
RAX: ffffffffffffffda RBX:000000000000000cRCX:00007f7471f1ff53
RDX:000000000000000cRSI:00005614d615a770RDI: 0000000000000001
RBP:00005614d615a770R08:000000000000000aR09:00007f7471fb20c0
R10: 0000000000000000 R11: 0000000000000246 R12:000000000000000c
R13:00007f7471fee520R14:000000000000000cR15:00007f7471fee720Modules linked in: usb_f_uac2 u_audio usb_f_mass_storage usb_f_eem
u_etherusb_f_serialu_seriallibcomposite rfcommiptable_natbnepsnd_sof_nocodecspi_pxa2xx_platformdw_dmacsmscsnd_sof_pci_intel_tngsnd_sof_pcisnd_sof_acpi_intel_bytsnd_sof_intel_ipcsnd_sof_acpismsc95xxsnd_sofpwm_lpss_pcipwm_lpsssnd_sof_xtensa_dspsnd_intel_dspcfgsnd_soc_acpi_intel_matchsnd_soc_acpidw_dmac_pciintel_mrfld_pwrbtnintel_mrfld_adcdw_dmac_corespi_pxa2xx_pcibrcmfmac brcmutilleds_gpiohci_uartbtbcmti_ads7950industrialio_triggered_bufferkfifo_bufledtrig_timerledtrig_heartbeatmmc_blockextcon_intel_mrfldsdhci_pcicqhci sdhciled_classintel_soc_pmic_mrfldmmc_corebtrfs libcrc32c xorzstd_compresszlib_deflateraid6_pq
CR2:00000000202043f2---[ end trace 5c11fe50dca92ad4 ]---Quoted 160 lines
Quoted 6 lines
No I didn't. Afaik we don't have a devicetree property to set.
But I'd be happy to test that as well. But where to set the property?
dwc3_pci_mrfld_properties[] in dwc3-pci?
Hi Ferry,
Not too sure which DWC3 driver is used for the Intel platform, but I believe that should be the one. (if that's what is normally used) We'd just need to add an entry w/ the below:
PROPERTY_ENTRY_BOOL("tx-fifo-resize")Thanks
Wesley ChengQuoted 143 lines
Quoted 143 lines
Thanks
Wesley ChengQuoted 140 lines
Quoted 140 lines
Quoted 140 lines
Quoted 138 lines
Quoted 5 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet,
it should have a NumP value equal to the bMaxBurst reported in the EPYes and no. Looking back at the history, we used to configure NUMP based
on bMaxBurst, but it was changed later in commit4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a
problem reported by John Youn.And now we've come full circle. Because even if I believe more requests
are enough for bursting, NUMP is limited by the RxFIFO size. This ends
up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.However, note that this is about RxFIFO size, not TxFIFO size. In fact,
looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP
(emphasis is mine):"Number of Packets (NumP). This field is used to
indicate the
number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or
equal to
the maximum burst size supported by the endpoint as
determined
by the value in the bMaxBurst field in the Endpoint
Companion
Descriptor (refer to Section 9.6.7)."So, NumP is for the receiver, not the transmitter. Could you clarify
what you mean here?/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the
description of
this field.For an IN endpoint this field is set by the endpoint to
the
number of packets it can transmit when the host resumes transactions to it. This field shall not have a value
greater
than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the
Endpoint
Companion Descriptor. Note that the value reported in this
field
may be treated by the host as informative only."However, if I remember correctly (please verify dwc3 databook), NUMP in
DCFG was only for receive buffers. Thin, John, how does dwc3 compute
NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or
TxFIFO size?Quoted 9 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is
that after 2 data packets, the device issues a NRDY. So then we'd need
to send an ERDY once data is available within the FIFO, and the same
sequence happens until the USB request is complete. With this constant
NRDY/ERDY handshake going on, you actually see that the bus isQuoted 8 lines
under
utilized. When we increase an EP's FIFO size, then you'll see constant
bursts for a request, until the request is done, or if the host runs out
of RXFIFO. (ie no interruption [on the USB protocol level] during USB
request data transfer)Unfortunately I don't have access to a USB sniffer anymore :-(
Quoted 38 lines
Quoted 38 lines
Good points.
Wesley, what kind of testing have you done on this on
different devices?As mentioned above, these changes are currently present on end
user
devices for the past few years, so its been through a lot of
testing :).
all with the same gadget driver. Also, who uses USB on android
devices
these days? Most of the data transfer goes via WiFi or
Bluetooth, anyway
:-)I guess only developers are using USB during development to
flash dev
images heh.I used to be a customer facing engineer, so honestly I did see
some
really interesting and crazy designs. Again, we do have
non-Android
products that use the same code, and it has been working in
there
for a
few years as well. The TXFIFO sizing really has helped with
multimedia
use cases, which use isoc endpoints, since esp. in those lower
end CPU
chips where latencies across the system are much larger, and a
missed
ISOC interval leads to a pop in your ear.
This is good background information. Thanks for bringing this
up. Admitedly, we still have ISOC issues with dwc3. I'm
interested in
knowing if a deeper request queue would also help here.Quoted 8 lines
Quoted 6 lines
Remember dwc3 can accomodate 255 requests + link for each
endpoint. If
our gadget driver uses a low number of requests, we're never
really
using the TRB ring in our benefit.We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst
behavior.Signature · 1 line
heikki
Quoted 41 lines · Felipe Balbi
Hi,
Quoted 37 lines · Wesley Cheng
Quoted 26 lines
Quoted 26 lines
Quoted 25 lines
Quoted 14 lines
Quoted 7 lines
Quoted 7 lines
to be honest, I don't think these should go in (apart from the build failure) because it's likely to break instantiations of the core with differing FIFO sizes. Some instantiations even have some endpoints with dedicated functionality that requires the default FIFO size configured during coreConsultant instantiation. I know of at OMAP5 and some Intel implementations which have dedicated endpoints for processor tracing.
Quoted 6 lines
Quoted 6 lines
With OMAP5, these endpoints are configured at the top of the available endpoints, which means that if a gadget driver gets loaded and takes over most of the FIFO space because of this resizing, processor tracing will have a hard time running. That being said, processor tracing isn't supported in upstream at this moment.
Quoted 7 lines
I agree that the application of this logic may differ between vendors, hence why I wanted to keep this controllable by the DT property, so that for those which do not support this use case can leave it disabled. The logic is there to ensure that for a given USB configuration, for each EP it would have at least 1 TX FIFO. For USB configurations which don't utilize all available IN EPs, it would allow re-allocation of internal memory to EPs which will actually be in use.
The feature ends up being all-or-nothing, then :-) It sounds like we can be a little nicer in this regard.
Quoted 8 lines
Quoted 2 lines
Don't get me wrong, I think once those features become available upstream, we can improve the logic. From what I remember when looking
sure, I support that. But I want to make sure the first cut isn't likely to break things left and right :)
Hence, let's at least get more testing.
Sure, I'd hope that the other users of DWC3 will also see some pretty big improvements on the TX path with this.
Hi Felipe,
Sorry for the delayed response. I went into the office to capture a USB trace to better show you the difference with and without the TXFIFO resize changes. Let me address your comments below first before showing the traces.
Quoted 3 lines
fingers crossed
Unfortunately, based on Ferry's testing, it looks like the Intel HW platform itself doesn't have a SS capable port. Although, we did get some good information from it, as we found that math is different between controller revisions.
Quoted 49 lines
Quoted 24 lines
Quoted 13 lines
Quoted 2 lines
at Andy Shevchenko's Github, the Intel tracer downstream changes were just to remove physical EP1 and 2 from the DWC3 endpoint list. If that
right, that's the reason why we introduced the endpoint feature flags. The end goal was that the UDC would be able to have custom feature flags paired with ->validate_endpoint() or whatever before allowing it to be enabled. Then the UDC driver could tell UDC core to skip that endpoint on that particular platform without interefering with everything else.
Of course, we still need to figure out a way to abstract the different dwc3 instantiations.
Quoted 7 lines
Quoted 2 lines
was the change which ended up upstream for the Intel tracer then we could improve the logic to avoid re-sizing those particular EPs.
The problem then, just as I mentioned in the previous paragraph, will be coming up with a solution that's elegant and works for all different instantiations of dwc3 (or musb, cdns3, etc).
Well, at least for the TX FIFO resizing logic, we'd only be needing to focus on the DWC3 implementation.
Quoted 6 lines
You bring up another good topic that I'll eventually needing to be taking a look at, which is a nice way we can handle vendor specific endpoints and how they can co-exist with other "normal" endpoints. We have a few special HW eps as well, which we try to maintain separately in our DWC3 vendor driver, but it isn't the most convenient, or most pretty method :).
Awesome, as mentioned, the endpoint feature flags were added exactly to allow for these vendor-specific features :-)
I'm more than happy to help testing now that I finally got our SM8150 Surface Duo device tree accepted by Bjorn ;-)
Quoted 12 lines
Quoted 7 lines
Quoted 2 lines
However, I'm not sure how the changes would look like in the end, so I would like to wait later down the line to include that :).
Fair enough, I agree. Can we get some more testing of $subject, though? Did you test $subject with upstream too? Which gadget drivers did you use? How did you test
The results that I included in the cover page was tested with the pure upstream kernel on our device. Below was using the ConfigFS gadget w/ a mass storage only composition.
Quoted 9 lines
Quoted 5 lines
Test Parameters:
- Platform: Qualcomm SM8150
- bMaxBurst = 6
- USB req size = 256kB
- Num of USB reqs = 16do you mind testing with the regular request size (16KiB) and 250 requests? I think we can even do 15 bursts in that case.
Let's go over the trace. If you are still convinced this would help with the particular scenario we're looking at, then I can run this test with the above.
Quoted 31 lines
Quoted 19 lines
- USB Speed = Super-Speed
- Function Driver: Mass Storage (w/ ramdisk)
- Test Application: CrystalDiskMarkResults:
TXFIFO Depth = 3 max packets
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 193.60 | | 195.86 | | 184.77 | | 193.60-------------------------------------------
TXFIFO Depth = 6 max packets
Quoted 7 lines
Test Case | Data Size | AVG tput (in MB/s)
-------------------------------------------Sequential|1 GB x | Read |9 loops | 287.35 | | 304.94 | | 289.64 | | 293.61I remember getting close to 400MiB/sec with Intel platforms without resizing FIFOs and I'm sure the FIFO size was set to 2x1024, though my memory could be failing.
Quoted 9 lines
Then again, I never ran with CrystalDiskMark, I was using my own tool (it's somewhere in github. If you care, I can look up the URL).
Quoted 3 lines
We also have internal numbers which have shown similar improvements as well. Those are over networking/tethering interfaces, so testing IPERF loopback over TCP/UDP.
loopback iperf? That would skip the wire, no?
The iperf server (receiver) would be running on the PC, and the iperf client would be running on our device (transmitter).
Quoted 81 lines
Quoted 23 lines
Quoted 6 lines
Quoted 5 lines
size of 2 and TX threshold of 1, this would really be not beneficial to us, because we can only change the TX threshold to 2 at max, and at least in my observations, once we have to go out to system memory to fetch the next data packet, that latency takes enough time for the controller to end the current burst.
Quoted 9 lines
What I noticed with
g_mass_storageis that we can amortize the cost of fetching data from memory, with a deeper request queue. Whenever I test(ed)g_mass_storage, I was doing so with 250 requests. And that was enough to give me very good performance. Never had to poke at TX FIFO resizing. Did you try something like this too?I feel that allocating more requests is a far simpler and more generic method that changing FIFO sizes :)
I wish I had a USB bus trace handy to show you, which would make it very clear how the USB bus is currently utilized with TXFIFO size 2 vs 6. So by increasing the number of USB requests, that will help if there was a bottleneck at the SW level where the application/function driver utilizing the DWC3 was submitting data much faster than the HW was processing them.
Quoted 4 lines
So yes, this method of increasing the # of USB reqs will definitely help with situations such as HSUSB or in SSUSB when EP bursting isn't used. The TXFIFO resize comes into play for SSUSB, which utilizes endpoint bursting.
Hmm, that's not what I remember. Perhaps the TRB cache size plays a role here too. I have clear memories of testing this very scenario of bursting (using
g_mass_storageat the time) because I was curious about it. Back then, my tests showed no difference in behavior.It could be nice if Heikki could test Intel parts with and without your changes on
g_mass_storagewith 250 requests.Quoted 3 lines
Now with endpoint bursting, if the function notifies the host that bursting is supported, when the host sends the ACK for the Data Packet, it should have a NumP value equal to the bMaxBurst reported in the EP
Yes and no. Looking back at the history, we used to configure NUMP based on bMaxBurst, but it was changed later in commit
4e99472bc10bda9906526d725ff6d5f27b4ddca1by yours truly because of a problem reported by John Youn.And now we've come full circle. Because even if I believe more requests are enough for bursting, NUMP is limited by the RxFIFO size. This ends up supporting your claim that we need RxFIFO resizing if we want to squeeze more throughput out of the controller.
However, note that this is about RxFIFO size, not TxFIFO size. In fact, looking at Table 8-13 of USB 3.1 r1.0, we read the following about NumP (emphasis is mine):
"Number of Packets (NumP). This field is used to indicate the number of Data Packet buffers that the **receiver** can accept. The value in this field shall be less than or equal to the maximum burst size supported by the endpoint as determined by the value in the bMaxBurst field in the Endpoint Companion Descriptor (refer to Section 9.6.7)."
So, NumP is for the receiver, not the transmitter. Could you clarify what you mean here?
/me keeps reading
Hmm, table 8-15 tries to clarify:
"Number of Packets (NumP).
For an OUT endpoint, refer to Table 8-13 for the description of this field.
For an IN endpoint this field is set by the endpoint to the number of packets it can transmit when the host resumes transactions to it. This field shall not have a value greater than the maximum burst size supported by the endpoint as indicated by the value in the bMaxBurst field in the Endpoint Companion Descriptor. Note that the value reported in this field may be treated by the host as informative only."
Quoted 5 lines
However, if I remember correctly (please verify dwc3 databook), NUMP in DCFG was only for receive buffers. Thin, John, how does dwc3 compute NumP for TX/IN endpoints? Is that computed as a function of DCFG.NUMP or TxFIFO size?
Sorry for confusing you. So you are right about NumP being applicable to the receiver path, and the PC USB host controller also will have its own RXFIFO, which for the IN direction correlates to the NumP value being sent by the host within the ACK. The point I was trying to make is that, the bus utilization is not hampered by the host running out of RXFIFO (getting ACKs w/ NumP=0), but with the device always pre-maturely ending the burst.
Quoted 10 lines
Quoted 9 lines
desc. If we have a TXFIFO size of 2, then normally what I have seen is that after 2 data packets, the device issues a NRDY. So then we'd need to send an ERDY once data is available within the FIFO, and the same sequence happens until the USB request is complete. With this constant NRDY/ERDY handshake going on, you actually see that the bus is under utilized. When we increase an EP's FIFO size, then you'll see constant bursts for a request, until the request is done, or if the host runs out of RXFIFO. (ie no interruption [on the USB protocol level] during USB request data transfer)
Quoted 2 lines
Unfortunately I don't have access to a USB sniffer anymore :-(
So I went ahead and captured USB Lecroy log with the following conditions:
- bMaxBurst value = 6
- TXFIFOSZ = 6 max packets (with resize) / TXFIFOSZ = 1 max packet (w/o)
- Test case = USB tethering w/ IPERF loopback (between PC and device)
- USB request size = 32kB
- Speed = USB3.1 gen 1
Trace Hierarchy:
Transfer
--> Transaction
--> Packet
1 transfer has multiple transactions, and 1 transaction has multiple packets. This is just how Lecroy does the packet groupings to help make the traces more readable. For now, we can focus on the "Packet" entries.
With TXFIFO resize:
Transfer(1062) Left("Left") G1(x1) Bulk(IN) ADDR(15) ENDP(4) _______| Bytes Transferred(31584) Time Stamp(0 . 400 271 196) _______|_______________________________________________________________________ Transaction(1768) Left("Left") G1(x1) IN ADDR(15) ENDP(4) Data(31584 bytes) _______| Time Stamp(0 . 400 271 196) _______|_______________________________________________________________________L
Packet(268690) Left("Left") Dir G1(x1) TP ACK(1) ADDR(15) ENDP(4)
_______| Dir(In) SeqN(31) NumP(3) Stream ID(0x0000) PP(Pnd) LCW (Hseq:2) _______| Duration(40.100 ns) Time(358.000 ns) Time Stamp(0 . 400 271 196) _______|_______________________________________________________________________R
Packet(268692) Right("Right") Dir G1(x1) DP Data Len(1024) ADDR(15) _______| ENDP(4) Dir(In) SeqN(31) EoB(N) Stream ID(0x0000) PP(Not Pnd)
_______| LCW (Hseq:3) Data(1024 bytes) Duration( 2.117 us) Time(
2.140 us)
_______| Time Stamp(0 . 400 271 554) _______|_______________________________________________________________________R
Packet(268696) Right("Right") Dir G1(x1) DP Data Len(1024) ADDR(15) _______| ENDP(4) Dir(In) SeqN(0) EoB(N) Stream ID(0x0000) PP(Not Pnd)
_______| LCW (Hseq:4) Data(1024 bytes) Duration( 2.117 us)
Time(262.000 ns)
_______| Time Stamp(0 . 400 273 694) _______|_______________________________________________________________________L
Packet(268699) Left("Left") Dir G1(x1) TP ACK(1) ADDR(15) ENDP(4)
_______| Dir(In) SeqN(0) NumP(3) Stream ID(0x0000) PP(Pnd) LCW (Hseq:3) _______| Duration(40.100 ns) Time( 1.890 us) Time Stamp(0 . 400 273 956) _______|_______________________________________________________________________R
Packet(268702) Right("Right") Dir G1(x1) DP Data Len(1024) ADDR(15) _______| ENDP(4) Dir(In) SeqN(1) EoB(N) Stream ID(0x0000) PP(Not Pnd)
_______| LCW (Hseq:5) Data(1024 bytes) Duration( 2.117 us)
Time(258.000 ns)
_______| Time Stamp(0 . 400 275 846) _______|_______________________________________________________________________L
Packet(268705) Left("Left") Dir G1(x1) TP ACK(1) ADDR(15) ENDP(4)
_______| Dir(In) SeqN(1) NumP(3) Stream ID(0x0000) PP(Pnd) LCW (Hseq:4) _______| Duration(40.100 ns) Time( 1.918 us) Time Stamp(0 . 400 276 104) _______|_______________________________________________________________________R
Packet(268708) Right("Right") Dir G1(x1) DP Data Len(1024) ADDR(15) _______| ENDP(4) Dir(In) SeqN(2) EoB(N) Stream ID(0x0000) PP(Not Pnd)
_______| LCW (Hseq:6) Data(1024 bytes) Duration( 2.117 us)
Time(246.000 ns)
_______| Time Stamp(0 . 400 278 022) _______|_______________________________________________________________________L
Packet(268711) Left("Left") Dir G1(x1) TP ACK(1) ADDR(15) ENDP(4)
_______| Dir(In) SeqN(2) NumP(2) Stream ID(0x0000) PP(Pnd) LCW (Hseq:5) _______| Duration(40.100 ns) Time( 1.962 us) Time Stamp(0 . 400 278 268) _______|_______________________________________________________________________R
Observations:
- Within a transfer, there are no data packets w/ the EoB set to yes.
- Host never runs out of RXFIFO (NumP never reaches 0)
- Packets within a transaction is never interrupted with a NRDY. (followed by an ERDY from the device to continue the transaction)
==========================================================================
Without TXFIFO resize:
Transfer(1542) Left("Left") G1(x1) Bulk(IN) ADDR(19) ENDP(4) _______| Bytes Transferred(31584) Time Stamp(0 . 619 833 722) _______|_______________________________________________________________________ Transaction(7677) Left("Left") G1(x1) IN ADDR(19) ENDP(4) Condition(Flow Ctrl)
_______| Data(1024 bytes) Time Stamp(0 . 619 833 722) _______|_______________________________________________________________________L
Packet(415331) Left("Left") Dir G1(x1) TP ACK(1) ADDR(19) ENDP(4)
_______| Dir(In) SeqN(13) NumP(4) Stream ID(0x0000) PP(Pnd) LCW (Hseq:3) _______| Duration(40.100 ns) Time(396.000 ns) Time Stamp(0 . 619 833 722) _______|_______________________________________________________________________R
Packet(415338) Right("Right") Dir G1(x1) DP Data Len(1024) ADDR(19) _______| ENDP(4) Dir(In) SeqN(13) EoB(Y) Stream ID(0x0000) PP(Not Pnd)
_______| LCW (Hseq:5) Data(1024 bytes) Duration( 2.117 us) Time(
4.012 us)
_______| Time Stamp(0 . 619 834 118) _______|_______________________________________________________________________L
Packet(415349) Left("Left") Dir G1(x1) TP ACK(1) ADDR(19) ENDP(4)
_______| Dir(In) SeqN(14) NumP(0) Stream ID(0x0000) PP(Pnd) LCW (Hseq:6) _______| Duration(40.100 ns) Time(460.000 ns) Time Stamp(0 . 619 838 130) _______|_______________________________________________________________________
Transaction(7681) Right("Right") G1(x1) EP Ready ADDR(19) ENDP(4) Dir(IN) _______| Time Stamp(0 . 619 838 590) _______|_______________________________________________________________________R
Packet(415354) Right("Right") Dir G1(x1) TP ERDY(3) ADDR(19) ENDP(4)
_______| Dir(In) NumP(7) Stream ID(0x0000) LCW (Hseq:7)
Duration(40.100 ns)
_______| Time( 2.148 us) Time Stamp(0 . 619 838 590) _______|_______________________________________________________________________ Transaction(7682) Left("Left") G1(x1) IN ADDR(19) ENDP(4) Condition(Flow Ctrl)
_______| Data(1024 bytes) Time Stamp(0 . 619 840 738) _______|_______________________________________________________________________L
Packet(415358) Left("Left") Dir G1(x1) TP ACK(1) ADDR(19) ENDP(4)
_______| Dir(In) SeqN(14) NumP(3) Stream ID(0x0000) PP(Pnd) LCW (Hseq:0) _______| Duration(40.100 ns) Time(428.000 ns) Time Stamp(0 . 619 840 738) _______|_______________________________________________________________________R
Packet(415365) Right("Right") Dir G1(x1) DP Data Len(1024) ADDR(19) _______| ENDP(4) Dir(In) SeqN(14) EoB(Y) Stream ID(0x0000) PP(Not Pnd)
_______| LCW (Hseq:1) Data(1024 bytes) Duration( 2.117 us) Time(
3.980 us)
_______| Time Stamp(0 . 619 841 166) _______|_______________________________________________________________________L
Packet(415376) Left("Left") Dir G1(x1) TP ACK(1) ADDR(19) ENDP(4)
_______| Dir(In) SeqN(15) NumP(0) Stream ID(0x0000) PP(Pnd) LCW (Hseq:3) _______| Duration(40.100 ns) Time(444.000 ns) Time Stamp(0 . 619 845 146) _______|_______________________________________________________________________
Transaction(7683) Right("Right") G1(x1) EP Ready ADDR(19) ENDP(4) Dir(IN) _______| Time Stamp(0 . 619 845 590) _______|_______________________________________________________________________R
Packet(415383) Right("Right") Dir G1(x1) TP ERDY(3) ADDR(19) ENDP(4)
_______| Dir(In) NumP(7) Stream ID(0x0000) LCW (Hseq:4)
Duration(40.100 ns)
_______| Time( 3.964 us) Time Stamp(0 . 619 845 590) _______|_______________________________________________________________________ Transaction(7684) Left("Left") G1(x1) IN ADDR(19) ENDP(4) Condition(Flow Ctrl)
_______| Data(1024 bytes) Time Stamp(0 . 619 849 554) _______|_______________________________________________________________________L
Packet(415394) Left("Left") Dir G1(x1) TP ACK(1) ADDR(19) ENDP(4)
_______| Dir(In) SeqN(15) NumP(1) Stream ID(0x0000) PP(Pnd) LCW (Hseq:6)
_______| Duration(40.100 ns) Time(396.000 ns) Time Stamp(0 . 619 849 554)
Observations:
- Since the current setting has TXFIFO size to only have 1 max packet, the device sets the EoB to yes during the data packet.
- Within a transfer, there are multiple "transactions" as Lecroy groups transactions whenever there is a NRDY --> ERDY transition.
- Frequent NRDY --> ERDY transitions coming from the device
I hope this clears up what the TXFIFO resize is actually helping with. It keeps the burst continuously going w/o having to do a NRDY-->ERDY handshake. (which is unnecessary overhead) Even though we allocate more USB requests from the SW, the SW has no control over how the request is sent over the link.
Thanks
Wesley Cheng
Quoted 38 lines
Quoted 28 lines
Quoted 28 lines
Quoted 16 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Good points.
Wesley, what kind of testing have you done on this on different devices?
As mentioned above, these changes are currently present on end user devices for the past few years, so its been through a lot of testing :).
Quoted 7 lines
all with the same gadget driver. Also, who uses USB on android devices these days? Most of the data transfer goes via WiFi or Bluetooth, anyway :-)
I guess only developers are using USB during development to flash dev images heh.
Quoted 7 lines
I used to be a customer facing engineer, so honestly I did see some really interesting and crazy designs. Again, we do have non-Android products that use the same code, and it has been working in there for a few years as well. The TXFIFO sizing really has helped with multimedia use cases, which use isoc endpoints, since esp. in those lower end CPU chips where latencies across the system are much larger, and a missed ISOC interval leads to a pop in your ear.
This is good background information. Thanks for bringing this up. Admitedly, we still have ISOC issues with dwc3. I'm interested in knowing if a deeper request queue would also help here.
Quoted 6 lines
Quoted 4 lines
Remember dwc3 can accomodate 255 requests + link for each endpoint. If our gadget driver uses a low number of requests, we're never really using the TRB ring in our benefit.
We're actually using both a deeper USB request queue + TX fifo resizing. :).
okay, great. Let's see what John and/or Thinh respond WRT dwc3 TX Burst behavior.
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
Hey Wesley,
Quoted 9 lines · Greg KH
Quoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]
Quoted 9 lines
653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0);
| ~~~~~~~~~~~~~^~~~~~~~~~
| |
| u32 {aka unsigned int}
In file included from drivers/usb/dwc3/debug.h:14,
from drivers/usb/dwc3/gadget.c:25:
drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’}
1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc)
| ~~~~~~~~~~~~~^~~I'm guessing you were previously using the DWC3_MDWIDTH macro which operated on the 'hwparams0' reg value directly, but probably had to switch it to the dwc3_mdwidth() inline function that Thinh had replaced it with recently. Forgot to compile-test I bet? :)
Quoted 5 lines
drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]660 | ret = of_add_property(dwc3_np, prop); | ^~~~~~~~~~~~~~~ | of_get_property
Scratched my head on this one a bit, since 'of_add_property' is clearly declared in <linux/of.h> which dwc3-qcom.c directly includes. Then I looked closer and saw the declaration only in case of #ifdef CONFIG_OF and noticed it doesn't have a corresponding no-op static inline definition in the case of !CONFIG_OF. Again I'm guessing here that Greg must have built on a non-OF config. We should probably include a patch that adds the stub.
Thanks,
Jack
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
Hi Jack,
Quoted 22 lines · Jack Pham
Hey Wesley,
Quoted 9 lines · Greg KH
Quoted 3 lines · Wesley Cheng
Changes in V9:
- Fixed incorrect patch in series. Removed changes in DTSI, as dwc3-qcom will add the property by default from the kernel.This patch series has one build failure and one warning added:
drivers/usb/dwc3/gadget.c: In function ‘dwc3_gadget_calc_tx_fifo_size’:drivers/usb/dwc3/gadget.c:653:45: warning: passing argument 1 of ‘dwc3_mdwidth’ makes pointer from integer without a cast [-Wint-conversion]Quoted 9 lines
653 | mdwidth = dwc3_mdwidth(dwc->hwparams.hwparams0); | ~~~~~~~~~~~~~^~~~~~~~~~ | | | u32 {aka unsigned int} In file included from drivers/usb/dwc3/debug.h:14, from drivers/usb/dwc3/gadget.c:25: drivers/usb/dwc3/core.h:1493:45: note: expected ‘struct dwc3 *’ but argument is of type ‘u32’ {aka ‘unsigned int’} 1493 | static inline u32 dwc3_mdwidth(struct dwc3 *dwc) | ~~~~~~~~~~~~~^~~
Quoted 5 lines
I'm guessing you were previously using the
DWC3_MDWIDTHmacro which operated on the 'hwparams0' reg value directly, but probably had to switch it to thedwc3_mdwidth()inline function that Thinh had replaced it with recently. Forgot to compile-test I bet? :)
Ah, looks like that's the case. I tried this on our internal branches, which didn't have Thinh's change, which is probably why it worked in the first place. Will fix this.
Quoted 6 lines
Quoted 5 lines
drivers/usb/dwc3/dwc3-qcom.c: In function ‘dwc3_qcom_of_register_core’:drivers/usb/dwc3/dwc3-qcom.c:660:23: error: implicit declaration of function ‘of_add_property’; did you mean ‘of_get_property’? [-Werror=implicit-function-declaration]660 | ret = of_add_property(dwc3_np, prop); | ^~~~~~~~~~~~~~~ | of_get_property
Quoted 8 lines
Scratched my head on this one a bit, since '
of_add_property' is clearly declared in <linux/of.h> whichdwc3-qcom.cdirectly includes. Then I looked closer and saw the declaration only in case of #ifdefCONFIG_OFand noticed it doesn't have a corresponding no-op static inline definition in the case of !CONFIG_OF. Again I'm guessing here that Greg must have built on a non-OF config. We should probably include a patch that adds the stub.
Nice catch, will add the stub.
Thanks
Wesley Cheng
Quoted 3 lines
Thanks,
Jack
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project