| Message ID | 1544805179-2248-1-git-send-email-prime.zeng@hisilicon.com |
|---|---|
| State | New |
| Headers | show |
| Series | usb: dwc3: gadget: fix miss isoc issue introduced by IRQ latency | expand |
Zeng Tao <prime.zeng@hisilicon.com> writes: > If it's a busy system, some times when we start an isoc transfer, the > framenumber get from the event buffer may be already elasped, in this > case, we will get all the packets dropped due to miss isoc. And we turn > into transfer nothing, to fix this issue, we need to fix the framenumber > to make sure that it's not out of date. > > Signed-off-by: Liang Shengjun <liangshengjun@hisilicon.com> > Signed-off-by: Zeng Tao <prime.zeng@hisilicon.com> There's a patch going upstream already doing this: https://git.kernel.org/pub/scm/linux/kernel/git/balbi/usb.git/commit/?h=next&id=d53701067f048b8b11635e964b6d3bd9a248c622 -- balbi
>-----Original Message----- >From: Felipe Balbi [mailto:balbi@kernel.org] >Sent: Friday, December 14, 2018 4:52 PM >To: Zengtao (B) <prime.zeng@hisilicon.com> >Cc: liangshengjun <liangshengjun@hisilicon.com>; Zengtao (B) ><prime.zeng@hisilicon.com>; Greg Kroah-Hartman ><gregkh@linuxfoundation.org>; linux-usb@vger.kernel.org; >linux-kernel@vger.kernel.org >Subject: Re: [PATCH] usb: dwc3: gadget: fix miss isoc issue introduced by >IRQ latency > >* PGP Signed by an unknown key > >Zeng Tao <prime.zeng@hisilicon.com> writes: > >> If it's a busy system, some times when we start an isoc transfer, the >> framenumber get from the event buffer may be already elasped, in this >> case, we will get all the packets dropped due to miss isoc. And we >> turn into transfer nothing, to fix this issue, we need to fix the >> framenumber to make sure that it's not out of date. >> >> Signed-off-by: Liang Shengjun <liangshengjun@hisilicon.com> >> Signed-off-by: Zeng Tao <prime.zeng@hisilicon.com> > >There's a patch going upstream already doing this: > >https://git.kernel.org/pub/scm/linux/kernel/git/balbi/usb.git/commit/?h >=next&id=d53701067f048b8b11635e964b6d3bd9a248c622 > Sorry, I think I missed to update to the latest version. But I think my patch is more efficient. Because I just sync the frame from the HW, it won't fail and there is no need to extra tries. What do you think about it? Regards Zengtao >-- >balbi > >* Unknown Key >* 0xE11A9906
Hi, "Zengtao (B)" <prime.zeng@hisilicon.com> writes: >>-----Original Message----- >>From: Felipe Balbi [mailto:balbi@kernel.org] >>Sent: Friday, December 14, 2018 4:52 PM >>To: Zengtao (B) <prime.zeng@hisilicon.com> >>Cc: liangshengjun <liangshengjun@hisilicon.com>; Zengtao (B) >><prime.zeng@hisilicon.com>; Greg Kroah-Hartman >><gregkh@linuxfoundation.org>; linux-usb@vger.kernel.org; >>linux-kernel@vger.kernel.org >>Subject: Re: [PATCH] usb: dwc3: gadget: fix miss isoc issue introduced by >>IRQ latency >> >>* PGP Signed by an unknown key >> >>Zeng Tao <prime.zeng@hisilicon.com> writes: >> >>> If it's a busy system, some times when we start an isoc transfer, the >>> framenumber get from the event buffer may be already elasped, in this >>> case, we will get all the packets dropped due to miss isoc. And we >>> turn into transfer nothing, to fix this issue, we need to fix the >>> framenumber to make sure that it's not out of date. >>> >>> Signed-off-by: Liang Shengjun <liangshengjun@hisilicon.com> >>> Signed-off-by: Zeng Tao <prime.zeng@hisilicon.com> >> >>There's a patch going upstream already doing this: >> >>https://git.kernel.org/pub/scm/linux/kernel/git/balbi/usb.git/commit/?h >>=next&id=d53701067f048b8b11635e964b6d3bd9a248c622 >> > > Sorry, I think I missed to update to the latest version. But I think my > patch is more efficient. Because I just sync the frame from the HW, it > won't fail and there is no need to extra tries. > > What do you think about it? the 14 bits you use for your check is not representative of the actual micro-frame you're currently in. Thinh explained that in the discussion we had until we came to the patch I pointed you to above. Please look at the mailing list archives for details. -- balbi
Hi Zengtao, On 12/14/2018 3:24 AM, Felipe Balbi wrote: > Hi, > > "Zengtao (B)" <prime.zeng@hisilicon.com> writes: >>> -----Original Message----- >>> From: Felipe Balbi [mailto:balbi@kernel.org] >>> Sent: Friday, December 14, 2018 4:52 PM >>> To: Zengtao (B) <prime.zeng@hisilicon.com> >>> Cc: liangshengjun <liangshengjun@hisilicon.com>; Zengtao (B) >>> <prime.zeng@hisilicon.com>; Greg Kroah-Hartman >>> <gregkh@linuxfoundation.org>; linux-usb@vger.kernel.org; >>> linux-kernel@vger.kernel.org >>> Subject: Re: [PATCH] usb: dwc3: gadget: fix miss isoc issue introduced by >>> IRQ latency >>> >>> * PGP Signed by an unknown key >>> >>> Zeng Tao <prime.zeng@hisilicon.com> writes: >>> >>>> If it's a busy system, some times when we start an isoc transfer, the >>>> framenumber get from the event buffer may be already elasped, in this >>>> case, we will get all the packets dropped due to miss isoc. And we >>>> turn into transfer nothing, to fix this issue, we need to fix the >>>> framenumber to make sure that it's not out of date. >>>> >>>> Signed-off-by: Liang Shengjun <liangshengjun@hisilicon.com> >>>> Signed-off-by: Zeng Tao <prime.zeng@hisilicon.com> >>> There's a patch going upstream already doing this: >>> >>> https://git.kernel.org/pub/scm/linux/kernel/git/balbi/usb.git/commit/?h >>> =next&id=d53701067f048b8b11635e964b6d3bd9a248c622 >>> >> Sorry, I think I missed to update to the latest version. But I think my >> patch is more efficient. Because I just sync the frame from the HW, it >> won't fail and there is no need to extra tries. >> >> What do you think about it? > the 14 bits you use for your check is not representative of the actual > micro-frame you're currently in. Thinh explained that in the discussion > we had until we came to the patch I pointed you to above. Please look at > the mailing list archives for details. > There are several issues with this patch. 1) Your frame elapsed time check is not based on interval but statically DWC3_EVENT_PRAM_SOFFN / 2. That's about 1 second. So it doesn't account for isoc transfers with large service interval of 1 sec or more. 2) This function __dwc3_gadget_target_frame_elapsed() should have comments for what it does. The name implies that this function checks for eframe > cframe, and not eframe > cframe + 1s. 3) If this check fails, then it will do DWC3_ALIGN_FRAME() at least twice. The isoc transfer will start 1 more interval into the future than it needs to. Also, I think the role of this check should be from the controller as it has more information and its own logic to decide if the selected future uframe has elapsed. BR, Thinh
Hi Thinh: >-----Original Message----- >From: Thinh Nguyen [mailto:thinh.nguyen@synopsys.com] >Sent: Saturday, December 15, 2018 5:43 AM >To: Felipe Balbi <balbi@kernel.org>; Zengtao (B) ><prime.zeng@hisilicon.com> >Cc: liangshengjun <liangshengjun@hisilicon.com>; Greg Kroah-Hartman ><gregkh@linuxfoundation.org>; linux-usb@vger.kernel.org; >linux-kernel@vger.kernel.org; Thinh Nguyen ><thinh.nguyen@synopsys.com> >Subject: Re: [PATCH] usb: dwc3: gadget: fix miss isoc issue introduced by >IRQ latency > >Hi Zengtao, > >On 12/14/2018 3:24 AM, Felipe Balbi wrote: >> Hi, >> >> "Zengtao (B)" <prime.zeng@hisilicon.com> writes: >>>> -----Original Message----- >>>> From: Felipe Balbi [mailto:balbi@kernel.org] >>>> Sent: Friday, December 14, 2018 4:52 PM >>>> To: Zengtao (B) <prime.zeng@hisilicon.com> >>>> Cc: liangshengjun <liangshengjun@hisilicon.com>; Zengtao (B) >>>> <prime.zeng@hisilicon.com>; Greg Kroah-Hartman >>>> <gregkh@linuxfoundation.org>; linux-usb@vger.kernel.org; >>>> linux-kernel@vger.kernel.org >>>> Subject: Re: [PATCH] usb: dwc3: gadget: fix miss isoc issue >>>> introduced by IRQ latency >>>> >>>> * PGP Signed by an unknown key >>>> >>>> Zeng Tao <prime.zeng@hisilicon.com> writes: >>>> >>>>> If it's a busy system, some times when we start an isoc transfer, >>>>> the framenumber get from the event buffer may be already elasped, >>>>> in this case, we will get all the packets dropped due to miss isoc. >>>>> And we turn into transfer nothing, to fix this issue, we need to >>>>> fix the framenumber to make sure that it's not out of date. >>>>> >>>>> Signed-off-by: Liang Shengjun <liangshengjun@hisilicon.com> >>>>> Signed-off-by: Zeng Tao <prime.zeng@hisilicon.com> >>>> There's a patch going upstream already doing this: >>>> >>>> >https://git.kernel.org/pub/scm/linux/kernel/git/balbi/usb.git/commit >>>> /?h >>>> =next&id=d53701067f048b8b11635e964b6d3bd9a248c622 >>>> >>> Sorry, I think I missed to update to the latest version. But I think >>> my patch is more efficient. Because I just sync the frame from the >>> HW, it won't fail and there is no need to extra tries. >>> >>> What do you think about it? >> the 14 bits you use for your check is not representative of the actual >> micro-frame you're currently in. Thinh explained that in the >> discussion we had until we came to the patch I pointed you to above. >> Please look at the mailing list archives for details. >> > >There are several issues with this patch. >1) Your frame elapsed time check is not based on interval but statically >DWC3_EVENT_PRAM_SOFFN / 2. That's about 1 second. So it doesn't >account for isoc transfers with large service interval of 1 sec or more. This is just for checking whether the Frame number is overflow or not. So 1 second should a reason value. >2) This function __dwc3_gadget_target_frame_elapsed() should have >comments for what it does. The name implies that this function checks >for eframe > cframe, and not eframe > cframe + 1s. eframe > cframe + 1s is used to deal with the Frame number overflow. >3) If this check fails, then it will do DWC3_ALIGN_FRAME() at least twice. >The isoc transfer will start 1 more interval into the future than it needs to. > If the interval is one, 1 more interval should be more reasonable because the core always fetch the trb one frame ahead. >Also, I think the role of this check should be from the controller as it has >more information and its own logic to decide if the selected future uframe >has elapsed. Yes, agree, but I think if the sw can be used to do the same thing and more effective, Why not? BR Zengtao
diff --git a/drivers/usb/dwc3/core.h b/drivers/usb/dwc3/core.h index 5bfb625..8742d96 100644 --- a/drivers/usb/dwc3/core.h +++ b/drivers/usb/dwc3/core.h @@ -452,6 +452,9 @@ #define DWC3_DSTS_RXFIFOEMPTY BIT(17) +#define DWC3_EVENT_PRAM_MAX_SOFFN 0x3fff +#define DWC3_EVENT_PRAM_SOFFN_MASK 0x3fff + #define DWC3_DSTS_SOFFN_MASK (0x3fff << 3) #define DWC3_DSTS_SOFFN(n) (((n) & DWC3_DSTS_SOFFN_MASK) >> 3) diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c index 9f92ee0..b63bd72 100644 --- a/drivers/usb/dwc3/gadget.c +++ b/drivers/usb/dwc3/gadget.c @@ -1263,6 +1263,15 @@ static int __dwc3_gadget_get_frame(struct dwc3 *dwc) return DWC3_DSTS_SOFFN(reg); } +static bool __dwc3_gadget_target_frame_elapsed(struct dwc3_ep *dep) +{ + u16 cframe = __dwc3_gadget_get_frame(dep->dwc); + u16 eframe = dep->frame_number & DWC3_EVENT_PRAM_SOFFN_MASK; + + return (((eframe - cframe) & DWC3_EVENT_PRAM_SOFFN_MASK) + > DWC3_EVENT_PRAM_MAX_SOFFN / 2); +} + static void __dwc3_gadget_start_isoc(struct dwc3_ep *dep) { if (list_empty(&dep->pending_list)) { @@ -1272,6 +1281,9 @@ static void __dwc3_gadget_start_isoc(struct dwc3_ep *dep) return; } + while (__dwc3_gadget_target_frame_elapsed(dep)) + dep->frame_number = DWC3_ALIGN_FRAME(dep); + dep->frame_number = DWC3_ALIGN_FRAME(dep); __dwc3_gadget_kick_transfer(dep); }