Date:	Thu, 19 Nov 2015 19:14:46 +0800
From:	Baolin Wang <baolin.wang@...aro.org>
To:	Peter Hurley <peter@...leysoftware.com>
Cc:	Felipe Balbi <balbi@...com>, Greg KH <gregkh@...uxfoundation.org>,
	r.baldyga@...sung.com, fabio.estevam@...escale.com,
	Philip Oberstaller <Philip.Oberstaller@...tentrio.com>,
	scottwood@...escale.com, Mark Brown <broonie@...nel.org>,
	USB <linux-usb@...r.kernel.org>,
	LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] usb: gadget: Add the console support for usb-to-serial port

On 19 November 2015 at 18:28, Peter Hurley <peter@...leysoftware.com> wrote:
> On 11/18/2015 09:35 PM, Baolin Wang wrote:
>> On 18 November 2015 at 23:32, Peter Hurley <peter@...leysoftware.com> wrote:
>>> Hi Baolin,
>>>
>>> On 11/16/2015 02:05 AM, Baolin Wang wrote:
>>>> It dose not work when we want to use the usb-to-serial port based
>>>> on one usb gadget as a console. Thus this patch adds the console
>>>> initialization to support this request.
>>>
>>> I have some high level concerns.
>>>
>>> 1. I would defer registering the console until the port has at least been
>>>    allocated in gserial_alloc_line(), and unregister when the line is freed.
>>>    That also reduces many of the conditions that you shouldn't need to
>>>    check, like port number range and so on.
>>
>> The 'setup' callback of 'struct console' is just do some memory
>> allocation and member initialization, that no need to defer
>> registering the console in gserial_alloc_line(). But the
>> 'gs_console_connect()' which will use the port need to be called in
>> gserial_connect().
>
> My point here was why are you registering the console before the port
> table is even allocated and initialized?  The console can't possibly
> perform i/o that early because the port doesn't even exist.
> Which is why I suggested waiting until gserial_alloc_line() to
> register the console.
>
> A typical console setup() performs the cross-reference linking between
> the console data structure and the port table. The reason for that
> is the console needs to be cleaned up if the port is being torn down.
>
> For example, in gserial_disconnect() the port->port_usb is reset to NULL,
> and later in gserial_console_exit():
>
>         if (port && port->port_usb) {
>                 ....
>                 gs_request_free(req, ep);
>         }
>
> which means your leaking the request allocation when the port has been
> disconnected.
>

Yeah, that's right. I'll defer the console registration until
gserial_connect() and unregistration in disconnect.

>
>>>    Further, consider deferring the console registration until gserial_connect();
>>>    that would further simplify things. In this case, unregistration would
>>>    happen at disconnect.
>>
>> That sounds reasonable. I would try.
>>
>>>
>>> 2. Why are you using a kworker for actual device i/o when all of the i/o
>>>    is done with interrupts disabled anyway?
>>>    Getting rid of the work would eliminate using the 8K intermediate buffer
>>>    as well.
>>
>> If remove the kworker, there are some deadlocks to call the printk()
>> when in the process of transferring data with usb endpoint. So we need
>> a async kworker to complete the io or it can not work.
>
> The commit log should detail the major design choices, including why you
> used the kworker (because of re-entrancy issues with usb endpoint).
>

OK.

>
>
>>> 3. If for some reason i/o deferral is really necessary, consider using a kthread
>>>    kworker instead of the pooled kworker. The scheduler response will be _way_
>>>    better.
>>>
>>
>> OK, make sense.
>>
>>> 4. If for some reason i/o deferral is really necessary, use a circular buffer
>>>    for the intermediate buffer, preferably lockless since there is only
>>>    one producer and one consumer.
>>>
>>
>> Yeah, the circular buffer is better but more tricky. I would try.
>>
>>> Some other review comments below; please ignore anything other reviewers
>>> have already noted.
>>>
>>> Regards,
>>> Peter Hurley
>>>
>>>> Signed-off-by: Baolin Wang <baolin.wang@...aro.org>
>>>> ---
>>>>  drivers/usb/gadget/Kconfig             |    6 +
>>>>  drivers/usb/gadget/function/u_serial.c |  238 ++++++++++++++++++++++++++++++++
>>>>  2 files changed, 244 insertions(+)
>>>>
>>>> diff --git a/drivers/usb/gadget/Kconfig b/drivers/usb/gadget/Kconfig
>>>> index 33834aa..be5aab9 100644
>>>> --- a/drivers/usb/gadget/Kconfig
>>>> +++ b/drivers/usb/gadget/Kconfig
>>>> @@ -127,6 +127,12 @@ config USB_GADGET_STORAGE_NUM_BUFFERS
>>>>          a module parameter as well.
>>>>          If unsure, say 2.
>>>>
>>>> +config U_SERIAL_CONSOLE
>>>> +     bool "Serial gadget console support"
>>>> +     depends on USB_G_SERIAL
>>>> +     help
>>>> +        It supports the serial gadget can be used as a console.
>>>> +
>>>>  source "drivers/usb/gadget/udc/Kconfig"
>>>>
>>>>  #
>>>> diff --git a/drivers/usb/gadget/function/u_serial.c b/drivers/usb/gadget/function/u_serial.c
>>>> index f7771d8..4ade527 100644
>>>> --- a/drivers/usb/gadget/function/u_serial.c
>>>> +++ b/drivers/usb/gadget/function/u_serial.c
>>>> @@ -27,6 +27,7 @@
>>>>  #include <linux/slab.h>
>>>>  #include <linux/export.h>
>>>>  #include <linux/module.h>
>>>> +#include <linux/console.h>
>>>>
>>>>  #include "u_serial.h"
>>>>
>>>> @@ -79,6 +80,16 @@
>>>>   */
>>>>  #define QUEUE_SIZE           16
>>>>  #define WRITE_BUF_SIZE               8192            /* TX only */
>>>> +#define GS_BUFFER_SIZE               (4096)
>>>> +#define GS_CONSOLE_BUF_SIZE  (2 * GS_BUFFER_SIZE)
>>>> +
>>>> +struct gscons_info {
>>>> +     struct gs_port          *port;
>>>> +     struct tty_driver       *tty_driver;
>>>> +     struct work_struct      work;
>>>> +     int                     buf_tail;
>>>> +     char                    buf[GS_CONSOLE_BUF_SIZE];
>>>> +};
>>>
>>> Make struct gscons_info a static declaration instead of
>>> allocating it.
>>
>> But I think the dynamic allocation is more reasonable with reducing
>> some global variables.
>
> But introduces a failure mode that doesn't exist with the
> static definition.
>

OK. I'll consider that.

>
>>>>  /* circular buffer */
>>>>  struct gs_buf {
>>>> @@ -118,6 +129,7 @@ struct gs_port {
>>>>
>>>>       /* REVISIT this state ... */
>>>>       struct usb_cdc_line_coding port_line_coding;    /* 8-N-1 etc */
>>>> +     struct usb_request      *console_req;
>>>>  };
>>>>
>>>>  static struct portmaster {
>>>> @@ -1054,6 +1066,7 @@ gs_port_alloc(unsigned port_num, struct usb_cdc_line_coding *coding)
>>>>
>>>>       port->port_num = port_num;
>>>>       port->port_line_coding = *coding;
>>>> +     port->console_req = NULL;
>>>>
>>>>       ports[port_num].port = port;
>>>>  out:
>>>> @@ -1143,6 +1156,227 @@ err:
>>>>  }
>>>>  EXPORT_SYMBOL_GPL(gserial_alloc_line);
>>>>
>>>> +#ifdef CONFIG_U_SERIAL_CONSOLE
>>>> +
>>>> +static struct usb_request *gs_request_new(struct usb_ep *ep, int buffer_size)
>>>                                                                 ^^^^^^^^^^^^^^^
>>> With only a single caller that uses a symbolic constant, is the
>>> 'buffer_size' parameter necessary?
>>>
>>
>> Yeah, I'll remove the 'buffer_size' parameter.
>>
>>>
>>>> +{
>>>> +     struct usb_request *req = usb_ep_alloc_request(ep, GFP_ATOMIC);
>>>> +
>>>
>>> Remove this newline.
>>
>> OK.
>>
>>>
>>>> +     if (!req)
>>>> +             return NULL;
>>>> +
>>>> +     /* now allocate buffers for the requests */
>>>
>>> Unnecessary comment.
>>
>> OK.
>>
>>>
>>>> +     req->buf = kmalloc(buffer_size, GFP_ATOMIC);
>>>> +     if (!req->buf) {
>>>> +             usb_ep_free_request(ep, req);
>>>> +             return NULL;
>>>> +     }
>>>> +
>>>> +     return req;
>>>> +}
>>>> +
>>>> +static void gs_request_free(struct usb_request *req, struct usb_ep *ep)
>>>> +{
>>>> +     if (req) {
>>>> +             kfree(req->buf);
>>>> +             usb_ep_free_request(ep, req);
>>>> +     }
>>>> +}
>>>> +
>>>> +static void gs_complete_out(struct usb_ep *ep, struct usb_request *req)
>>>> +{
>>>> +     if (req->status != 0 && req->status != -ECONNRESET)
>>>> +             return;
>>>> +}
>>>> +
>>>> +static struct console gserial_cons;
>>>> +static int gs_console_connect(void)
>>>
>>> Pass the console * as parameter, instead of forward declaring the console.
>>> Or initialize info directly from the static struct gscons_info address.
>>
>> Make sense.
>>
>>>
>>>> +{
>>>> +     struct gscons_info *info = gserial_cons.data;
>>>> +     int port_num = gserial_cons.index;
>>>> +     struct usb_request *req;
>>>> +     struct gs_port *port;
>>>> +     struct usb_ep *ep;
>>>> +
>>>> +     if (port_num >= MAX_U_SERIAL_PORTS || port_num < 0) {
>>>> +             pr_err("%s: port num [%d] exceeds the range.\n",
>>>> +                    __func__, port_num);
>>>> +             return -ENXIO;
>>>> +     }
>>>> +
>>>> +     port = ports[port_num].port;
>>>> +     if (!port) {
>>>> +             pr_err("%s: serial line [%d] not allocated.\n",
>>>> +                    __func__, port_num);
>>>> +             return -ENODEV;
>>>> +     }
>>>> +
>>>> +     if (!port->port_usb) {
>>>> +             pr_err("%s: no port usb.\n", __func__);
>>>> +             return -ENODEV;
>>>> +     }
>>>> +
>>>> +     ep = port->port_usb->in;
>>>> +     if (!ep) {
>>>> +             pr_err("%s: no usb endpoint.\n", __func__);
>>>> +             return -ENXIO;
>>>> +     }
>>>
>>> Looking at the caller, gserial_connect(), none of the error
>>> conditions above look possible.
>>
>> That's right. I'll remove these checks.
>>
>>>
>>>
>>>> +
>>>> +     req = port->console_req;
>>>> +     if (!req) {
>>>> +             req = gs_request_new(ep, GS_BUFFER_SIZE);
>>>
>>> Where is port->console_req assigned to?
>>
>> The connect may be called several times, if the req is allocated at
>> one time, there is no need to assign it.
>
> You've missed my point: where is
>
>                 port->console_req = ?????

Sorry, I missed here and will fix it.

>
>
>>>> +             if (!req) {
>>>> +                     pr_err("%s: request fail.\n", __func__);
>>>
>>> Remove redundant error message; the allocator has already done so.
>>
>> OK.
>>
>>>
>>>> +                     return -ENOMEM;
>>>> +             }
>>>> +             req->complete = gs_complete_out;
>>>> +     }
>>>> +
>>>> +     info->port = port;
>>>> +
>>>> +     pr_debug("%s: port[%d] console connect!\n", __func__, port_num);
>>>> +     return 0;
>>>> +}
>>>> +
>>>> +static void gs_console_work(struct work_struct *work)
>>>> +{
>>>> +     struct gscons_info *info = container_of(work, struct gscons_info, work);
>>>> +     struct gs_port *port = info->port;
>>>> +     struct usb_request *req;
>>>> +     struct usb_ep *ep;
>>>> +     int xfer, ret, count;
>>>> +     char *p;
>>>> +
>>>> +     if (!port || !port->port_usb)
>>>> +             return;
>>>> +
>>>> +     req = port->console_req;
>>>> +     ep = port->port_usb->in;
>>>> +     if (!req || !ep)
>>>> +             return;
>>>> +
>>>> +     spin_lock_irq(&port->port_lock);
>>>> +     count = info->buf_tail;
>>>> +     p = info->buf;
>>>> +
>>>> +     while (count > 0 && !port->write_busy) {
>>>> +             if (count > GS_BUFFER_SIZE)
>>>> +                     xfer = GS_BUFFER_SIZE;
>>>> +             else
>>>> +                     xfer = count;
>>>> +
>>>> +             memcpy(req->buf, p, xfer);
>>>> +             req->length = xfer;
>>>> +
>>>> +             port->write_busy = true;
>>>> +             spin_unlock(&port->port_lock);
>>>> +             ret = usb_ep_queue(ep, req, GFP_ATOMIC);
>>>> +             spin_lock(&port->port_lock);
>>>> +             port->write_busy = false;
>>>> +             if (ret < 0)
>>>> +                     break;
>>>> +
>>>> +             p += xfer;
>>>> +             count -= xfer;
>>>> +     }
>>>> +
>>>> +     info->buf_tail -= count;
>
> I'm not seeing how info->buf_tail is ever reset to 0.
>
>         count = info->buf_tail
>
>         while (count > 0) {
>                 ....
>                 count -= xfer;
>         }
>
> At loop exit count == 0, so
>
>         info->buf_tail -= count;
>
> never decreases buf_tail?
>

Sorry, that's a mistake and I'll fix that.

>
>>>> +     spin_unlock_irq(&port->port_lock);
>>>> +}
>>>> +
>>>> +static int gs_console_setup(struct console *co, char *options)
>>>> +{
>>>> +     struct gscons_info *gscons_info;
>>>> +
>>>> +     gscons_info = kzalloc(sizeof(struct gscons_info), GFP_KERNEL);
>>>> +     if (!gscons_info)
>>>> +             return -ENOMEM;
>>>> +
>>>> +     gscons_info->port = NULL;
>>>
>>> Redundant.
>>
>> Will remove it.
>>
>>>
>>>> +     gscons_info->tty_driver = gs_tty_driver;
>>>> +     INIT_WORK(&gscons_info->work, gs_console_work);
>>>> +     gscons_info->buf_tail = 0;
>>>
>>> Redundant.
>>
>> Will remove it.
>>
>>>
>>>> +     co->data = gscons_info;
>>>> +
>>>> +     return 0;
>>>> +}
>>>> +
>>>> +static void gs_console_write(struct console *co,
>>>> +                          const char *buf, unsigned count)
>>>> +{
>>>> +     struct gscons_info *info = co->data;
>>>> +     int avail, xfer;
>>>> +     char *p;
>>>> +
>>>> +     avail = GS_CONSOLE_BUF_SIZE - info->buf_tail;
>>>> +     if (count > avail)
>>>> +             xfer = avail;
>>>> +     else
>>>> +             xfer = count;
>>>
>>>         xfer = min(count, GS_CONSOLE_BUF_SIZE - info->buf_tail);
>>
>> Yeah, that's right.
>>
>>>
>>>> +
>>>> +     p = &info->buf[info->buf_tail];
>>>> +     memcpy(p, buf, xfer);
>>>> +     info->buf_tail += xfer;
>>>
>>> What is preventing concurrently running work and this from
>>> using/modifying the info->buf and info->buf_tail simultaneously?
>>
>> Like I said above, it will meet deadlocks if running the work
>> directly, then introduce the kworker.
>
> You've missed my point here:
>
> CPU 0                                 CPU 1
> --------------------------------      -------------------------------
> gs_console_write()                    gs_console_work()
>
>    info->buf_tail += xfer                info->buf_tail -= count;
>
>
> Neither of these operations are atomic so what will the value of
> info->buf_tail be? For example:
>
>    A <= LOAD(info->buf_tail)
>                                          B <= LOAD(info->buf_tail)
>    A <= A + xfer                         B <= B - count
>    STORE(A, info->buf_tail)
>                                          STORE(B, info->buf_tail)
>
> The result is as if info->buf_tail += xfer never happened.
>

Make sense. I would remove the buffer counter and replace with a
circular buffer. Very thanks for your good suggestions.

>
>>>> +
>>>> +     schedule_work(&info->work);
>>>> +}
>>>> +
>>>> +static struct tty_driver *gs_console_device(struct console *co, int *index)
>>>> +{
>>>> +     struct gscons_info *info = co->data;
>>>> +
>>>> +     *index = co->index;
>>>> +     return info->tty_driver;
>>>> +}
>>>> +
>>>> +static struct console gserial_cons = {
>>>> +     .name =         "ttyGS",
>>>> +     .write =        gs_console_write,
>>>> +     .device =       gs_console_device,
>>>> +     .setup =        gs_console_setup,
>>>> +     .flags =        CON_PRINTBUFFER,
>>>> +     .index =        -1,
>>>> +};
>>>> +
>>>> +static void gserial_console_init(void)
>>>> +{
>>>> +     register_console(&gserial_cons);
>>>> +}
>>>> +
>>>> +static void gserial_console_exit(void)
>>>> +{
>>>> +     struct gscons_info *info = gserial_cons.data;
>>>> +     struct gs_port *port = info->port;
>>>> +     struct usb_request *req;
>>>> +     struct usb_ep *ep;
>>>> +
>>>> +     if (port && port->port_usb) {
>>>> +             req = port->console_req;
>>>> +             ep = port->port_usb->in;
>>>> +             gs_request_free(req, ep);
>>>> +     }
>>>> +
>>>> +     kfree(info);
>>>> +     unregister_console(&gserial_cons);
>>>> +}
>>>> +
>>>> +#else
>>>> +
>>>> +static int gs_console_connect(void)
>>>> +{
>>>> +     return 0;
>>>> +}
>>>> +
>>>> +static void gserial_console_init(void)
>>>> +{
>>>> +}
>>>> +
>>>> +static void gserial_console_exit(void)
>>>> +{
>>>> +}
>>>> +
>>>> +#endif
>>>> +
>>>>  /**
>>>>   * gserial_connect - notify TTY I/O glue that USB link is active
>>>>   * @gser: the function, set up with endpoints and descriptors
>>>> @@ -1219,6 +1453,7 @@ int gserial_connect(struct gserial *gser, u8 port_num)
>>>>                       gser->disconnect(gser);
>>>>       }
>>>>
>>>> +     status = gs_console_connect();
>>>>       spin_unlock_irqrestore(&port->port_lock, flags);
>>>>
>>>>       return status;
>>>> @@ -1320,6 +1555,8 @@ static int userial_init(void)
>>>>               goto fail;
>>>>       }
>>>>
>>>> +     gserial_console_init();
>>>> +
>>>>       pr_debug("%s: registered %d ttyGS* device%s\n", __func__,
>>>>                       MAX_U_SERIAL_PORTS,
>>>>                       (MAX_U_SERIAL_PORTS == 1) ? "" : "s");
>>>> @@ -1334,6 +1571,7 @@ module_init(userial_init);
>>>>
>>>>  static void userial_cleanup(void)
>>>>  {
>>>> +     gserial_console_exit();
>>>>       tty_unregister_driver(gs_tty_driver);
>>>>       put_tty_driver(gs_tty_driver);
>>>>       gs_tty_driver = NULL;
>>>>
>>>
>>
>>
>>
>



-- 
Baolin.wang
Best Regards
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists