Re: [U-Boot] [PATCH 2/3] usb: dfu: introduce dfuMANIFEST state

17 Mar 2014

Hi Heiko,
...
Hello Lukasz,
Am 14.03.2014 11:47, schrieb Lukasz Majewski:
...
Hi Heiko,
...
on nand flash using ubi, after the download of the new image into
the flash, the "rest" of the nand sectors get erased while flushing
the medium. With current u-boot version dfu-util may show:
Starting download:
[##################################################] finished!
state(7) = dfuMANIFEST, status(0) = No error condition is present
unable to read DFU status
as get_status is not answered while erasing sectors, if erasing
needs some time.
So do the following changes to prevent this:

introduce dfuManifest state
 According to dfu specification
 ( http://www.usb.org/developers/devclass_docs/usbdfu10.pdf )

section 7: "the device enters the dfuMANIFEST-SYNC state and awaits
the solicitation of the status report by the host. Upon receipt of
the anticipated DFU_GETSTATUS, the device enters the dfuMANIFEST
state, where it completes its reprogramming operations."

when stepping into dfuManifest state, sending a PollTimeout
 DFU_MANIFEST_POLL_TIMEOUT in ms, to the host, so the host
 (dfu-util) waits the PollTimeout before sending a get_status

again.
Signed-off-by: Heiko Schocherhs@denx.de
Cc: Lukasz Majewskil.majewski@samsung.com
Cc: Kyungmin Parkkyungmin.park@samsung.com
Cc: Marek Vasutmarex@denx.de

README                     |  5 ++++
  drivers/dfu/dfu.c          |  1 -
  drivers/usb/gadget/f_dfu.c | 60
+++++++++++++++++++++++++++++++++++++++-------
include/dfu.h              |  3 +++ 4 files changed, 59
insertions(+), 10 deletions(-)

diff --git a/README b/README
index 216f0c7..5076248 100644
--- a/README
+++ b/README
@@ -1525,6 +1525,11 @@ The following options need to be configured:
   	this to the maximum filesize (in bytes) for the
buffer. Default is 4 MiB if undefined.

DFU_MANIFEST_POLL_TIMEOUT


Poll timeout [ms], which the device sends to the



host when

entering dfuMANIFEST state. Host waits this



timeout, before

sending again an USB request to the device.




Could you also add information about the DFU_DEFAULT_POLL_TIMEOUT to
the README, please (I've forgotten to do that)?
Done.
[...]
...
...
diff --git a/drivers/usb/gadget/f_dfu.c
b/drivers/usb/gadget/f_dfu.c index a045864..f4bf918 100644
--- a/drivers/usb/gadget/f_dfu.c
+++ b/drivers/usb/gadget/f_dfu.c
@@ -164,15 +164,22 @@ static void dnload_request_complete(struct
usb_ep *ep, struct usb_request *req)
   dfu_write(dfu_get_entity(f_dfu->altsetting), req->buf,
   	  req->length, f_dfu->blk_seq_num);
+}

if (req->length == 0)
puts("DOWNLOAD ... OK\nCtrl+C to exit ...\n");



+static void dnload_request_flush(struct usb_ep *ep, struct
usb_request *req) +{

struct f_dfu *f_dfu = req->context;


req->length = 0;

dfu_flush(dfu_get_entity(f_dfu->altsetting), req->buf,

  req->length, f_dfu->blk_seq_num);

}
static void handle_getstatus(struct usb_request *req)
{
struct dfu_status *dstat = (struct dfu_status *)req->buf;
struct f_dfu *f_dfu = req->context;

int setpolltimeout = 1;


We can go away with this flag. Please see below.
...
switch (f_dfu->dfu_state) {
   case DFU_STATE_dfuDNLOAD_SYNC:
@@ -180,17 +187,24 @@ static void handle_getstatus(struct
usb_request *req) f_dfu->dfu_state = DFU_STATE_dfuDNLOAD_IDLE;
   	break;
   case DFU_STATE_dfuMANIFEST_SYNC:

f_dfu->dfu_state = DFU_STATE_dfuMANIFEST;

break;
case DFU_STATE_dfuMANIFEST:
dfu_set_poll_timeout(dstat,



DFU_MANIFEST_POLL_TIMEOUT);

setpolltimeout = 0;

default:
break;
}


dfu_set_poll_timeout(dstat, 0);


if (setpolltimeout) {
dfu_set_poll_timeout(dstat, 0);





if (f_dfu->poll_timeout)
if (!(f_dfu->blk_seq_num %


      (dfu_get_buf_size() / DFU_USB_BUFSIZ)))


	dfu_set_poll_timeout(dstat,



f_dfu->poll_timeout);

if (f_dfu->poll_timeout)


	if (!(f_dfu->blk_seq_num %


	      (dfu_get_buf_size() /



DFU_USB_BUFSIZ)))

		dfu_set_poll_timeout(dstat,




f_dfu->poll_timeout);

}

What about this solution:
dfu_set_poll_timeout(dstat, 0);
switch (f_dfu->dfu_state) {
   case DFU_STATE_dfuDNLOAD_SYNC:
   case DFU_STATE_dfuDNBUSY:
   	f_dfu->dfu_state = DFU_STATE_dfuDNLOAD_IDLE;
   	break;
   case DFU_STATE_dfuMANIFEST_SYNC:
   	f_dfu->dfu_state = DFU_STATE_dfuMANIFEST;
   	break;
   case DFU_STATE_dfuMANIFEST:
   	dfu_set_poll_timeout(dstat,
DFU_MANIFEST_POLL_TIMEOUT); default:
   	break;
   }
if (f_dfu->poll_timeout)
   	if (!(f_dfu->blk_seq_num %
   	      (dfu_get_buf_size() / DFU_USB_BUFSIZ)))
   		dfu_set_poll_timeout(dstat,f_dfu->poll_timeout);
Then we could avoid setpolltimeout flag.
Done.
...
...
/* send status response */
   dstat->bStatus = f_dfu->dfu_status;
@@ -446,10 +460,11 @@ static int state_dfu_manifest_sync(struct
f_dfu *f_dfu, switch (ctrl->bRequest) {
   case USB_REQ_DFU_GETSTATUS:
   	/* We're MainfestationTolerant */

f_dfu->dfu_state = DFU_STATE_dfuIDLE;




f_dfu->dfu_state = DFU_STATE_dfuMANIFEST;

handle_getstatus(req);
f_dfu->blk_seq_num = 0;
value = RET_STAT_LEN;
req->complete = dnload_request_flush;

break;
case USB_REQ_DFU_GETSTATE:
handle_getstate(req);

@@ -463,6 +478,33 @@ static int state_dfu_manifest_sync(struct
f_dfu *f_dfu, return value;
  }
+static int state_dfu_manifest(struct f_dfu *f_dfu,

	      const struct usb_ctrlrequest *ctrl,


	      struct usb_gadget *gadget,


	      struct usb_request *req)



+{

int value = 0;

switch (ctrl->bRequest) {
case USB_REQ_DFU_GETSTATUS:
/* We're MainfestationTolerant */



Please look into the comments globally - we do now support the
DFU_STATE_dfuMANIFEST_* states
Hmm.. Yes, but we are MainfestationTolerant ... so the comment is
Ok, or?
My understanding of "ManifestationTolerant" is as follow:
"We didn't used the dfuMANIFEST state and immediately went to dfuIDLE
state from dfuDOWNLOAD-IDLE, so we are 'ManifestationTolerant'".
...
...
...

f_dfu->dfu_state = DFU_STATE_dfuIDLE;


handle_getstatus(req);


f_dfu->blk_seq_num = 0;


value = RET_STAT_LEN;


puts("DOWNLOAD ... OK\nCtrl+C to exit ...\n");


break;


case USB_REQ_DFU_GETSTATE:
handle_getstate(req);


break;


default:
f_dfu->dfu_state = DFU_STATE_dfuERROR;


value = RET_STALL;


break;


}
return value;

+}

static int state_dfu_upload_idle(struct f_dfu *f_dfu,
  		 const struct usb_ctrlrequest

*ctrl, struct usb_gadget *gadget,
@@ -539,7 +581,7 @@ static dfu_state_fn dfu_state[] = {
   state_dfu_dnbusy,        /* DFU_STATE_dfuDNBUSY */
   state_dfu_dnload_idle,   /* DFU_STATE_dfuDNLOAD_IDLE */
   state_dfu_manifest_sync, /* DFU_STATE_dfuMANIFEST_SYNC */

NULL,                    /* DFU_STATE_dfuMANIFEST */


state_dfu_manifest,	 /* DFU_STATE_dfuMANIFEST */
NULL,                    /*

DFU_STATE_dfuMANIFEST_WAIT_RST */ state_dfu_upload_idle,   /*
DFU_STATE_dfuUPLOAD_IDLE */ state_dfu_error          /*
DFU_STATE_dfuERROR */ diff --git a/include/dfu.h b/include/dfu.h
index 272a245..6c71ecb 100644
--- a/include/dfu.h
+++ b/include/dfu.h
@@ -80,6 +80,9 @@ static inline unsigned int get_mmc_blk_size(int
dev) #ifndef DFU_DEFAULT_POLL_TIMEOUT
  #define DFU_DEFAULT_POLL_TIMEOUT 0
  #endif
+#ifndef DFU_MANIFEST_POLL_TIMEOUT
+#define DFU_MANIFEST_POLL_TIMEOUT	DFU_DEFAULT_POLL_TIMEOUT
+#endif
struct dfu_entity {
   char			name[DFU_NAME_SIZE];
I did some preliminary DFU testing and it seems to not introduce any
regressions.
Thanks!
...
I'm looking forward for v2.
I am ready for sending it, but need response from you as I think the
comment is correct ...
Please explain why it is correct in your opinion.
...
BTW:
With the DFU_MANIFEST_POLL_TIMEOUT solution, we have a static timeout,
which is suboptimal, as we have a big MTD partittion and we must have
a "big" timeout for erasing (in the worst case) the complete
partition. Also this big timeout is (on nand using ubi) only needed
for the nand ubi partiton, not for the raw partitions containing for
example u-boot ... where it could be 0 ...
So, the problem is with setting a worst case value for all kinds of
your dfu entities?
...
Better would be, to have the information how long does it take to
erase one sector (define?) and how many sectors we have to erase, and
sending this value to the host... Do you see a chance to get this
information within the dfu code?
We have f_dfu->poll_timeout field in the dfu function. Value from
it sets the poll timeout to be waited by the host.
Maybe it would be better to reuse this field to set it accordingly for
MANIFEST_POLL_TIMEOUT. Now it is used for setting timeout when we store
big chunks of data (32 MiB).
...
Thinking about it, possibly we need a callback from the mtd layer,
which gives back the info, how long the flush would take, as this
is a media/partition size dependent timeout ...
I think that we could add this callback - called e.g. (*poll_timeout)
to the struct [mmc|nand]_internal_data. Then some helper functions
would be defined at dfu_[mmc|nand].c and used at f_dfu.c
Other question is if we accept current patches and wait for above
improvement to follow of drop them and wait for solution addressing
those issues.
What do you think?
...
bye,
Heiko
-- 
Best regards,

Lukasz Majewski

Samsung R&D Institute Poland (SRPOL) | Linux Platform Group

    