Hi, Bin Liu <b-liu@xxxxxx> writes: >> >> Felipe Balbi <felipe.balbi@xxxxxxxxxxxxxxx> writes: >> >> > Bin Liu <b-liu@xxxxxx> writes: >> >> >> On Fri, Mar 18, 2016 at 08:59:48AM +0200, Felipe Balbi wrote: >> >> >>> >> >> >>> Hi, >> >> >>> >> >> >>> Bin Liu <binmlist@xxxxxxxxx> writes: >> >> >>> > [ text/plain ] >> >> >>> > Hi, >> >> >>> > >> >> >>> > On Fri, Mar 11, 2016 at 6:54 AM, Felipe Balbi >> >> >>> > <felipe.balbi@xxxxxxxxxxxxxxx> wrote: >> >> >>> >> previously we were using a maximum of 32 TRBs per >> >> >>> >> endpoint. With each TRB being 16 bytes long, we were >> >> >>> >> using 512 bytes of memory for each endpoint. >> >> >>> >> >> >> >>> >> However, SLAB/SLUB will always allocate PAGE_SIZE >> >> >>> >> chunks. In order to better utilize the memory we >> >> >>> >> allocate and to allow deeper queues for gadgets >> >> >>> >> which would benefit from it (g_ether comes to mind), >> >> >>> >> let's increase the maximum to 256 TRBs which rounds >> >> >>> >> up to 4096 bytes for each endpoint. >> >> >>> > >> >> >>> > Do we want to increase the same for event ring buffers as >> >> >>> > while, which is allocated by dma_alloc_coherent(), which >> >> >>> > is also at PAGE_SIZE chunks, right? >> >> >>> >> >> >>> I could, but that's much less important. Currently we have up to 2 >> >> >> >> >> >> I agree it is less important, the only issue I see is wasting of memory. >> >> >> The device I am working on right now has two dwc3 controllers, each >> >> >> allocates 16 event buffers, so for the total of 128KB allocated pages, >> >> >> only 8KB is really used, 120KB is wasted. >> >> >> >> >> >> Seems dma pool makes more sense in here? >> >> > >> >> > I don't know. I think the real thing is that I really need to revisit >> >> > that part of the code/databook. The whole "multiple interrupters" seems >> >> > like it's only really necessary for host side. Which means that we could >> >> > drop all the loops for multiple event buffers and always use a single >> >> > one. >> >> > >> >> > Do you wanna test the following ? >> >> > >> >> > diff --git a/drivers/usb/dwc3/core.c b/drivers/usb/dwc3/core.c >> >> > index 17fd81447c9f..ebb3ee9c06f1 100644 >> >> > --- a/drivers/usb/dwc3/core.c >> >> > +++ b/drivers/usb/dwc3/core.c >> >> > @@ -237,8 +237,7 @@ static int dwc3_alloc_event_buffers(struct dwc3 *dwc, unsigned length) >> >> > int num; >> >> > int i; >> >> > >> >> > - num = DWC3_NUM_INT(dwc->hwparams.hwparams1); >> >> > - dwc->num_event_buffers = num; >> >> > + dwc->num_event_buffers = 1; >> >> > >> >> > dwc->ev_buffs = devm_kzalloc(dwc->dev, sizeof(*dwc->ev_buffs) * num, >> >> > GFP_KERNEL); >> >> > >> >> > I'll re-read what these bits actually mean. I have a strong feeling we >> >> > could (should?) be ignoring them for the peripheral side. >> >> >> >> Okay, so when we're configuring the endpoints, we could route endpoint >> >> interrupts to different event buffers. I really think that's really >> >> unimportant for us, specially since we end up using a single IRQ line. >> >> >> >> I guess I'll just go ahead and remove that code. If it turns out we >> >> decide to use it, we shouldn't really be using a loop in the hardirq >> >> handler anyway. >> > >> > Sounds good to me, I only see one evt buffer is used in all my devices, >> > even thought multi buffers are allocated based on hwparams1. >> >> I sent some patches yesterday. You might wanna give it a review ;-) > > The 3 patches are all look good to me. I bet you already tested it, so I > didn't do so. ;) yeah, tested with 3 intel platforms. -- balbi
Attachment:
signature.asc
Description: PGP signature
