incrase usbdevfs bulk buffer size
PATCH 1/1 incrase usbdevfs bulk buffer size
Increase the usbdevfs buffer size limit.
I hit this limit with running ehci in the VM. The a single ehci qTD transfer buffer can be 5 pages long, that is 20K. The current devio bulk transfer limit is 16K. I am not sure why the bulk transfer limit is 16K. I can complicate the user space part to work around that, but it seems much simpler just allow usbdevfs to accept bigger buffers.
I did some limited usage test with ehci controller. It seems works for me, it did use a bigger buffer and nothing blow up.
drivers/usb/core/devio.c +1 -1
@@ -72,7 +72,7 @@ MODULE_PARM_DESC (usbfs_snoop, "true to } while (0) -#define MAX_USBFS_BUFFER_SIZE 16384 +#define MAX_USBFS_BUFFER_SIZE (32*1024)static inline int connected (struct usb_device *dev) {
Re: PATCH] incrase usbdevfs bulk buffer size
Quoted 2 lines · Christopher Li
I hit this limit with running ehci in the VM. The a single ehci qTD transfer buffer can be 5 pages long, that is 20K. [...]
Even 16K is too much, IMHO.
Quoted 2 lines
I can complicate the user space part to work around that, but it seems much simpler just allow usbdevfs to accept bigger buffers.
It seems, yes. However, I assure you that this is not going to work for anyone who has anything reasonable swapped out, because of the kmalloc().
16K is an order 2 allocation on systems with 4KB pages, such as Opteron. It kinda sorta works, but not really.
This looks like a requirement to think about a better API. Also, the things that Harald was trying to fix, with pids and signals, just tell me that something was really wrong from the start.
-- Pete
Re: PATCH] incrase usbdevfs bulk buffer size
Quoted 10 lines · Pete Zaitcev
Quoted 2 lines · Christopher Li
I hit this limit with running ehci in the VM. The a single ehci qTD transfer buffer can be 5 pages long, that is 20K. [...]
Even 16K is too much, IMHO.
Quoted 2 lines
I can complicate the user space part to work around that, but it seems much simpler just allow usbdevfs to accept bigger buffers.
Quoted 8 lines
It seems, yes. However, I assure you that this is not going to work for anyone who has anything reasonable swapped out, because of the
kmalloc().16K is an order 2 allocation on systems with 4KB pages, such as Opteron. It kinda sorta works, but not really.
This looks like a requirement to think about a better API. Also,
I think the API is kind of fine in this aspect. The usbdevfs should be able to take bigger than 16K, but the internal copy of the urb does not have to use kmalloc on data buffers.
Quoted 2 lines
the things that Harald was trying to fix, with pids and signals, just tell me that something was really wrong from the start.
More detail please?
Chris
Re: PATCH] incrase usbdevfs bulk buffer size
Quoted 8 lines · Christopher Li
Quoted 4 lines
16K is an order 2 allocation on systems with 4KB pages, such as Opteron. It kinda sorta works, but not really.
This looks like a requirement to think about a better API. []
I think the API is kind of fine in this aspect. The usbdevfs should be able to take bigger than 16K, but the internal copy of the urb does not have to use kmalloc on data buffers.
You miss an important detail here, namely that single URBs do not have a capability to transfer to a discotiguous buffer. As long as you try to map one transfer insive VMware to one URB, one and only one kmalloc has to be done. But if splitting the transfer is acceptable, there is no reason to up the maximum buffer size in usbfs. But the above is IMHO only, and perhaps I miss something as well.
-- Pete
Re: PATCH] incrase usbdevfs bulk buffer size
Quoted 8 lines · Pete Zaitcev
Quoted 4 lines · Christopher Li
I think the API is kind of fine in this aspect. The usbdevfs should be able to take bigger than 16K, but the internal copy of the urb does not have to use kmalloc on data buffers.
You miss an important detail here, namely that single URBs do not have a capability to transfer to a discotiguous buffer. As long as you try
That is exactly my point that the kernel should not limit itself on only using contiguous buffers. Every USB controller can handle discrete DMA buffer why shouldn't the kernel? Obviously it should be nice to address the contiguous buffer restriction before bump up the bulk transfer limit.
Quoted 1 line
to map one transfer insive VMware to one URB, one and only one kmalloc
The current usbdevfs does a extra copy between from the user space urb to the kernel space urb. So it does not matter if the user space urb is contiguous or not. If the kernel can handle discrete dma internally, the usbdevfs could use it and maintain it's current API.
BTW, That is not VMware choice how the data buffer was arranged. It is the guest.
Quoted 1 line
has to be done. But if splitting the transfer is acceptable, there is
I still think fixing the kernel to allow address scatter-getter buffer and allow bigger buffer size is the right thing to do.
Chris