esp-idf/components/usb/maintainers.md
Darian Leung 0c758c8557 usb: Fix how the HCD handles sudden disconnection
This commit fixes how the USB Host HCD handles sudden disconnections.

Bugs:
- HW channels remain active when the port suddenly disconnects, and
previously the channel would be disabled by setting the disabled bit,
then waiting for a disabled interrupt. However, ISOC channels do not
generate the disabled interrupt when the port is invalid, thus leading
to tasks getting indefinitely blocked in hcd_pipe_command().

Fix:
On a sudden disconnection, forcibly treat all channels as halted even
if their HCCHAR.ChEna bit is still set. We do a soft reset after a port
error anyways, so the channels will eventually be reset.

Closes https://github.com/espressif/esp-idf/issues/7505
2021-11-02 14:30:58 +08:00

143 lines
12 KiB
Markdown

# USB Host Stack Maintainers Notes
This document is intended for future maintainers of the ESP-IDF USB Host stack.
Note: The underlying DWC_OTG controller will sometimes be referred to as the Host Controller in this document.
Note: Some static functions within the stack need to be called within critical sections. Therefore, the USB Host stack will prefix all such function names with an underscore (e.g., `_some_static_func()`) will need to be called in critical sections)
The host driver is currently split into the following layers, ordered from the lowest (furthest away from the user) to the highest layer (closest to the user).
* USB Host lower layer in `usbh_ll.h`
* USB HAL in `usbh_hal.h` and `usbh_hal.c`
* Host Controller Driver in `hcd.c` and `hcd.h`
* USB Host Driver in `usb_host.h` and `usb_host.c`
# DWC_OTG
Details regarding the DWC_OTG peripheral are covered in the data book and programming guide. However, this section adds some extra notes regarding the hardware's behavior that are not specified / cannot be easily found in the data book or programming guide.
## Implicit interrupt on short packet on INTERRUPT transfer.
- An interrupt channel will trigger an interrupt one of its QTDs is a short packet (event if the QTD did not set the interrupt on complete bit). This implicit interrupt is used to notify the software that it should halt the channel to
stop the remainder of the interrupt transfer.
## Channel interrupt on port errors
- If there are one or more channels active, and a port error interrupt occurs (such as disconnection, over-current), the active channels will not have an interrupt.
- The channels will remain active (i.e., `HCCHAR.ChEna` will still be set)
- Normal methods of disabling the channel (via setting `HCCHAR.ChDis` and waiting for an interrupt) will not work for ISOC channels (the interrupt will never be generated).
- Therefore, on port errors, just treat all the channels as halted and treat their in-flight transfers as failed. The soft reset that occurs after will reset all the channel's registers.
## Reset
- When resetting, if a disconnect occurs while a reset condition is being held. The disconnect is not detected until the reset condition is released. Once released, a `USBH_HAL_PORT_EVENT_DISCONN` is generated.
- During a second reset (i.e., there is already an enabled device).
- Once the reset condition is asserted, a `USBH_HAL_PORT_EVENT_DISABLED` is generated
- Once released, a `USBH_HAL_PORT_EVENT_ENABLED` is generated again.
# USB Host Lower Layer
* `usbh_ll.h` abstracts away the basic register operation of the DWC_OTG controller
* The LL provides register operations of the DWC OTG controller operating under Host Mode using scatter/gather internal DMA.
* For non-host mode specific register operations (i.e. global registers), the functions are prefixed with `usb_ll_...()`
* For host mode specific register operations, the functions are prefixed with `usbh_ll_...()`
# USB Host HAL
The HAL layer abstracts the DWC_OTG operating in Host Mode using Internal Scatter/Gather DMA. The HAL presents an abstraction of a single Host Port and multiple channels.
## HAL Host Port
- Models a single USB port where a single device can be attached
- Actions: Port can be powered ON/OFF, reset, suspended
- Events: When an interrupt occurs, call `usbh_hal_decode_intr()` to decoded the port interrupt.
- Port can detect various events such as connection, disconnection, enabled (i.e., connected device successfully reset), overcurrent etc.
## HAL Channels
- Channels are essentially the controllers abstraction of USB pipes. At any one point in time, a channel can be configured to map to a particular endpoint on a particular connected device (i.e., a particular device address).
- Channels have to be allocated and freed. It's possible to change a channel's endpoint characteristics (i.e., EP number, device address, direction, transfer type etc) so long as the channel is halted (i.e., not actively executing transfer descriptors).
- Use `usbh_hal_chan_alloc()` to allocate a channel
- Once allocated, use `usbh_hal_chan_set_ep_char()` to set the Endpoint characteristics of the channel (i.e., the information of the endpoint that the channel is communicating with). There are also some `usbh_hal_chan_set...()` functions to change a particular characteristic.
- Once the channel is no longer needed, call `usbh_hal_chan_free()` to free the channel
- Channels use a list of Queue Transfer Descriptors (QTDs) to executed USB transfers.
- A transfer descriptor list must be filled using `usbh_hal_xfer_desc_fill()`
- Once filled, a channel can be activated using `usbh_hal_chan_activate()`
- Once the channel is done (i.e., a descriptor with the `USBH_HAL_XFER_DESC_FLAG_HOC` is executed), a `USBH_HAL_CHAN_EVENT_CPLT` event is generated. The channel is now halted
- Call `usbh_hal_xfer_desc_parse()` to parse the results of the descriptor list
- If you need to halt the channel early (such as aborting a transfer), call `usbh_hal_chan_request_halt()`
- In case of a channel error event:
- Call `usbh_hal_chan_get_error()` to get the specific channel error that occurred
# Host Controller Driver (HCD)
The HCD (Host Controller Driver) abstracts the DWC_OTG as N number of ports and an arbitrary number of pipes that can be routed through one of the ports to a device. However, note that **the underlying hardware controller only has a single port, so technically only one port can ever be enabled**.
- In other words, the HCD essentially implements a root hub (not fully behavioral compliant) that contains a single port.
- Pipes are "an association between an endpoint on a device and software on the host". URBs (USB Request Block) represent a USB transfer that can be enqueued into a pipe for transmission, and dequeued from a pipe when completed.
The HCD currently has the following limitations:
- HCD **does not "present the root hub and its behavior according to the hub class definition"**. We currently don't have a hub driver yet, so the port commands in the driver do not fully represent an interface of a USB hub as described in 10.4 of the USB2.0 spec.
- No more than 8 pipes can be allocated at any one time due to underlying Host Controllers 8 channel limit. In the future, we could make particular pipes share a single Host Controller channel.
## HCD Port
- An HCD port can be thought of as a simplified version of a port on the Root Hub of the host controller. However, the complexity of parsing Hub Requests is discarded in favor of port commands (`hcd_port_cmd_t`) as the current USB Host Stack does not support hubs yet.
- A port must first initialized before it can be used. A port is identified by its handled of type `hcd_port_handle_t`
- The port can be manipulated using commands such as:
- Powering the port ON/OFF
- Issuing reset/resume signals
- The various host port events are represented in the `hcd_port_event_t` enumeration
- When the following port events occur, the port will be put into the HCD_PORT_STATE_RECOVERY state. The port can be deinitialized from there, or recovered using `hcd_port_recover()`. All the pipes routed through the port must be freed before the port can be recovered.
- `HCD_PORT_EVENT_DISCONNECTION`
- `HCD_PORT_EVENT_ERROR`
- `HCD_PORT_EVENT_OVERCURRENT`
- The port's internal FIFOs (RX, Periodic TX, Non-periodic TX) must be after each port reset (a port reset seems to clear the FIFO sizing registers). The sizing of these FIFOs will affect the largest supported MPS of endpoints using that FIFO. For convenience, the HCD provides the `hcd_port_fifo_bias_t` enum that will set the FIFO sizes for you but biased towards a particular use case. For example, if the connected device has an IN endpoint with large MPS (e.g., 512 bytes), the FIFO should be biased as `HCD_PORT_FIFO_BIAS_RX`.
- The FIFO sizes will be set on port initialization (supplied in `hcd_port_config_t`)
- FIFOs can be resized after port reset using `hcd_port_set_fifo_bias()` but some restrictions apply (see API description).
## HCD Pipes
- Pipes can be opened to a particular endpoint based on a descriptor provided on allocation. If opening a default pipe, a `NULL` descriptor can be provided.
- URBs can be enqueued into a pipe. Pipes use a linked list internally, so there is (in theory) no limit to the number of URBs that can be enqueued.
- URBs need to be dequeued once they are completed.
- URBs need to have the transfer information (such as data buffer, transfer length in bytes) filled before they should be enqueued.
- URBs will be owned by the HCD until they are dequeued. Thus, users should not attempt to modify a URB object (and the URB's data buffer) until the URB is dequeued.
- The URB is defined in `usb_private.h` instead of `hcd.h` so that the same structure can shared throughout the entire Host stack. Each layer simply needs to pass the pointer of the URB to the next layer thus minimizing the amount of copying required.
## HCD SW Arch
The software arch for the HCD has the following properties and intentions:
- Some static functions may be blocking. Those functions will have a note in their descriptions.
- The HCD communicates events entirely through callbacks and polling/handling functions. The client can choose what type of data structure they want to use in the callbacks to keep track of port and pipe events. Callbacks don't even need to be used, and the HCD should be able to operate entirely on a polling basis.
- The port and each pipe have to be treated as completely separate entities (with their own handles and events). This allows clients to group these entities however it sees fit. For example, the client can:
- Let a hub driver manage the port, and manipulate the port via its port handle
- Group pipes into interfaces and expose the interface as a single entity
- The HCD will not internally allocate any tasks. It is up to the client to decide how to distribute workload (e.g., a single while loop polling a port and all its pipes vs or each pipe having its own task).
- The HCD uses an interrupt to process things that require low latency such as processing pipe transfer requests. Where possible, processing is deferred to the `hcd_port_handle_event()` to reduce ISR workload.
### Events
In order to communicate events to the client of the HCD, the HCD does not attempt to allocate any queues to store events that have occurred. Callbacks are used in order to achieve maximum flexibility. Within these callbacks, the client of the HCD is free too use whatever OS primitives they want to store/forward these events.
There are two types callbacks that the HCD offers:
- Port callback will run whenever an event a port occurs. `hcd_port_handle_event()` should be called after a port event occurs.
- A pipe callback on each pipe, that will run when a event occurs on a pipe.
The client of the HCD can also forego callbacks entirely and simply poll for port and pipe events using the `hcd_port_handle_event()` and `hcd_pipe_get_event()` respectively.
- Some of the port and pipe commands may need to block for a certain condition before the command is executed. For example when suspending the port, the port command must block until all pipes to finish their current transfers. The blocking and unblocking is handled by an internal event mechanism.
### Thread Safety
The HCD API is thread safe however the following limitations should be noted:
- It is the client's responsibility to ensure that `hcd_install()` is called before any other HCD function is called
- Likewise, it is the client's responsibility to ensure that events and pipes are cleared before calling `hcd_port_deinit()`.
- `hcd_port_command()` is thread safe, but only one port command can be executed at any one time. Therefore HCD internally used a mutex to protect against concurrent commands.
- If multiple threads attempt to execute a command on the sample one, all but one of those threads will return with an invalid state error.
- Blocking HCD functions should not be called from critical sections and interrupts (e.g., `hcd_port_command()` and `hcd_pipe_command()`).