Hi, Benjamin Herrenschmidt <b...@kernel.crashing.org> writes: >> Benjamin Herrenschmidt <b...@kernel.crashing.org> writes: >> > So another question :-) >> > >> > Is it legal to break down a bulk (or ISO) IN request into packets that >> > are *smaller* than the Max EP size ? >> >> it depends if they are *supposed* to be two smaller packets or should be >> two parts of a single one. >> >> > For example, I have a max EP size of 512 bytes and the gadget gives me >> > an SG list where that's broken up into two segments (for example 2x256 >> > bytes). >> > >> > Now, I'm still trying to confirm with the vendor what happens here, but >> > I *suspect* that if I create 2 descriptors, the HW will break it down >> > into 2 packets smaller than the max EP size on the wire. >> >> If that's the case, then your HW can't handle SG lists. I suspect there >> will be some "chain" bit somewhere in the DMA descriptor which you can >> use to tell DMA that these two segments are part of a single transfer. > > No, there isn't, sadly. I've confirmed with the HW vendor. I might > still enable SG with a fallback to bounce buffering to a pre-allocated > packet buffer if I hit such a situation... It shouldn't be common for > mass storage (though do we use sglists for the mass storage gadget ?),
UASP does use it ;-) > it might be more with networking if our network function does > fragmented sends, I haven't checked. Well, even networking gadget is using a linear buffer. Nobody implemented support for sg lists yet. >> > I do have vague memories of seeing "client" drivers in the past with >> > make assumptions that some "replies" are fully contained in a single >> > packet, so I'm tempted to say that this won't work, and thus should >> > bounce such packets through a linear buffer but I'd like confirmation. >> >> yeah, USB makes that assumption. If host side requests 512 bytes and you >> end up breaking that into two packets, host will receive only the first >> 256 bytes, as short packets terminate transfers. > > Yup, that was indeed my assumption. Thanks for confirming. np -- balbi
signature.asc
Description: PGP signature
