On Sun, 12 Dec 2010 17:15:21 -0500 (EST)
[email protected] wrote:
> Since your patch can cause alot of extra data to be sent, I suggest looking
> into this patch before your usbmon become publicized.
Usbmon was publicised for years now, but let's see.
> Corrects isoc monitor data payload to represent the "actual_length"s
> of urb buffer data instead of "length" of buffer data.
> Since isoc records are a series of fragments, uninitialized buffer
> data could be sent as monitor data.
As an aside, there is no security or privacy issue with fetching
the "unitialized" data (it is the same ring buffer, so unrelated
kernel memory does not leak).
> - if (urb->num_sgs == 0) {
> - mon_copy_to_buff(rp, offset, urb->transfer_buffer, length);
> - length = 0;
> - } else {
> + if (!ndesc && urb->num_sgs > 0) {
> + struct scatterlist *sg;
> /* If IOMMU coalescing occurred, we cannot trust sg_page */
>[............]
> *flag = 'D';
> + } else {
> + if (ndesc) {
> + struct usb_iso_packet_descriptor *fp;
>[............]
> + }
> + else {
> + mon_copy_to_buff(rp, offset, buf, length);
> + length = 0;
> + }
> }
This looks obviously incorrect. If anyone ever submits an ISO with
the newfanged s/g URB, we're going to copy the scatterlist (if not
crash).
> + fp = urb->iso_frame_desc;
> + for (i=ndesc; length > 0 && --i >= 0; ++fp) {
> + this_ofs = fp->offset;
> + this_len = min_t(unsigned int, fp->actual_length, length);
> + offset = mon_copy_to_buff(rp, offset, buf+this_ofs, this_len);
> + length -= this_len;
> + }
This is no better. It is not going to save anything from outgoing
transfers, where actual_lengh is not set.
In any case, the whole excersie seems rather pointless to me.
Even for the numbers that Marton presented, I was not sure it was
worth to rescan the descriptors, only to save a few kilobytes per
URB. It was 19KB total for bz#22182. In the event, we saved almost
all of it: the existing code only transfers 4170 bytes of 19200.
Now all this new code to save 4KB? No way.
-- Pete
On Sun, 12 Dec 2010, Pete Zaitcev wrote:
> This looks obviously incorrect. If anyone ever submits an ISO with
> the newfanged s/g URB, we're going to copy the scatterlist (if not
> crash).
That won't happen. We support scatter-gather URBs really just for bulk
transfers. They can also be used for interrupt (though I don't know
why anyone would want to), and just possibly for control (not supported
by all HCDs), but not for isochronous transfers.
Alan Stern