RE: [PATCH v7 3/5] usb: host: ohci-sm501: init genalloc for local memory

[Laurentiu Tudor <laurentiu.tudor@nxp.com>]

Hello,

> -----Original Message-----
> From: Guenter Roeck <groeck7@gmail.com> On Behalf Of Guenter Roeck
> Sent: Thursday, June 13, 2019 9:06 PM
> 
> On 6/13/19 8:34 AM, Fredrik Noring wrote:
> > Hi Guenter,
> >
> >> Thanks for the confirmation. Do you see the problem only with the
> >> ohci-sm501 driver or also with others ?
> >
> > All are likely affected, but it depends, because I believe the problem
> is
> > that the USB subsystem runs out of memory. Please try the attached patch!
> >
> > The pool assumed 4096 byte page alignment for every allocation, which is
> > excessive given that many requests are for 16 and 32 bytes. In the patch
> > below, I have turned down the order to 5, which is good enough for the
> ED
> > and TD structures of the OHCI, but not enough for the HCCA that needs
> 256
> > byte alignment. With some luck, the WARN_ON_ONCE will not trigger in
> your
> > test, though. If it does, you may try to increase the order from 5 to 8.
> >
> 
> You are right, the patch below fixes the problem. I did not get the
> warning
> with order==5. Nevertheless, I also tested with order==8; that works as
> well.

Sorry for the late reply, I was OOO for past week+ and many thanks for taking a look at this.
So if my understanding is correct, an order of PAGE_SHIFT is too large and leads to waste of memory and in the end to an out of memory condition. This leaves me wondering what a safe order value would be.
I'll try to look into the legacy dma coherent allocation code maybe I can gather some info on the subject.

---
Best Regards, Laurentiu

> 
> > I have observed strange things happen when the USB subsystem runs out of
> > memory. The mass storage drivers often seem to busy-wait on -ENOMEM,
> > consuming a lot of processor resources. It would be much more efficient
> > to sleep waiting for memory to become available.
> >
> > In your case I suspect that allocation failures are not correctly
> > attributed. Certain kinds of temporary freezes may also occur, as the
> > various devices are reset due to host memory allocation errors.
> >  > Fredrik
> >
> > diff --git a/drivers/usb/core/hcd.c b/drivers/usb/core/hcd.c
> > --- a/drivers/usb/core/hcd.c
> > +++ b/drivers/usb/core/hcd.c
> > @@ -3011,7 +3011,7 @@ int usb_hcd_setup_local_mem(struct usb_hcd *hcd,
> phys_addr_t phys_addr,
> >   	int err;
> >   	void __iomem *local_mem;
> >
> > -	hcd->localmem_pool = devm_gen_pool_create(hcd->self.sysdev,
> PAGE_SHIFT,
> > +	hcd->localmem_pool = devm_gen_pool_create(hcd->self.sysdev, 5,
> >   						  dev_to_node(hcd->self.sysdev),
> >   						  dev_name(hcd->self.sysdev));
> >   	if (IS_ERR(hcd->localmem_pool))
> > diff --git a/drivers/usb/host/ohci-hcd.c b/drivers/usb/host/ohci-hcd.c
> > --- a/drivers/usb/host/ohci-hcd.c
> > +++ b/drivers/usb/host/ohci-hcd.c
> > @@ -517,6 +517,7 @@ static int ohci_init (struct ohci_hcd *ohci)
> >   						GFP_KERNEL);
> >   	if (!ohci->hcca)
> >   		return -ENOMEM;
> > +	WARN_ON_ONCE(ohci->hcca_dma & 0xff);
> >
> >   	if ((ret = ohci_mem_init (ohci)) < 0)
> >   		ohci_stop (hcd);
> >