From: Raz Manor Raz.Manor@valens.com
[ Upstream commit ef5e2fa9f65befa12f1113c734602d2c1964d2a5 ]
In the function scan_dma_completions() there is a reusage of tmp variable. That coused a wrong value being used in some case when reading a short packet terminated transaction from an endpoint, in 2 concecutive reads.
This was my logic for the patch:
The req->td->dmadesc equals to 0 iff: -- There was a transaction ending with a short packet, and -- The read() to read it was shorter than the transaction length, and -- The read() to complete it is longer than the residue. I believe this is true from the printouts of various cases, but I can't be positive it is correct.
Entering this if, there should be no more data in the endpoint (a short packet terminated the transaction). If there is, the transaction wasn't really done and we should exit and wait for it to finish entirely. That is the inner if. That inner if should never happen, but it is there to be on the safe side. That is why it is marked with the comment /* paranoia */. The size of the data available in the endpoint is ep->dma->dmacount and it is read to tmp. This entire clause is based on my own educated guesses.
If we passed that inner if without breaking in the original code, than tmp & DMA_BYTE_MASK_COUNT== 0. That means we will always pass dma bytes count of 0 to dma_done(), meaning all the requested bytes were read.
dma_done() reports back to the upper layer that the request (read()) was done and how many bytes were read. In the original code that would always be the request size, regardless of the actual size of the data. That did not make sense to me at all.
However, the original value of tmp is req->td->dmacount, which is the dmacount value when the request's dma transaction was finished. And that is a much more reasonable value to report back to the caller.
To recreate the problem: Read from a bulk out endpoint in a loop, 1024 * n bytes in each iteration. Connect the PLX to a host you can control. Send to that endpoint 1024 * n + x bytes, such that 0 < x < 1024 * n and (x % 1024) != 0 You would expect the first read() to return 1024 * n and the second read() to return x. But you will get the first read to return 1024 * n and the second one to return 1024 * n. That is true for every positive integer n.
Cc: Felipe Balbi balbi@kernel.org Cc: Greg Kroah-Hartman gregkh@linuxfoundation.org Cc: linux-usb@vger.kernel.org Signed-off-by: Raz Manor Raz.Manor@valens.com Signed-off-by: Felipe Balbi felipe.balbi@linux.intel.com Signed-off-by: Sasha Levin alexander.levin@verizon.com --- drivers/usb/gadget/udc/net2280.c | 25 +++++++++++++------------ 1 file changed, 13 insertions(+), 12 deletions(-)
diff --git a/drivers/usb/gadget/udc/net2280.c b/drivers/usb/gadget/udc/net2280.c index 33f3987218f7..d133252ef2c3 100644 --- a/drivers/usb/gadget/udc/net2280.c +++ b/drivers/usb/gadget/udc/net2280.c @@ -1146,15 +1146,15 @@ static int scan_dma_completions(struct net2280_ep *ep) */ while (!list_empty(&ep->queue)) { struct net2280_request *req; - u32 tmp; + u32 req_dma_count;
req = list_entry(ep->queue.next, struct net2280_request, queue); if (!req->valid) break; rmb(); - tmp = le32_to_cpup(&req->td->dmacount); - if ((tmp & BIT(VALID_BIT)) != 0) + req_dma_count = le32_to_cpup(&req->td->dmacount); + if ((req_dma_count & BIT(VALID_BIT)) != 0) break;
/* SHORT_PACKET_TRANSFERRED_INTERRUPT handles "usb-short" @@ -1163,40 +1163,41 @@ static int scan_dma_completions(struct net2280_ep *ep) */ if (unlikely(req->td->dmadesc == 0)) { /* paranoia */ - tmp = readl(&ep->dma->dmacount); - if (tmp & DMA_BYTE_COUNT_MASK) + u32 const ep_dmacount = readl(&ep->dma->dmacount); + + if (ep_dmacount & DMA_BYTE_COUNT_MASK) break; /* single transfer mode */ - dma_done(ep, req, tmp, 0); + dma_done(ep, req, req_dma_count, 0); num_completed++; break; } else if (!ep->is_in && (req->req.length % ep->ep.maxpacket) && !(ep->dev->quirks & PLX_PCIE)) {
- tmp = readl(&ep->regs->ep_stat); + u32 const ep_stat = readl(&ep->regs->ep_stat); /* AVOID TROUBLE HERE by not issuing short reads from * your gadget driver. That helps avoids errata 0121, * 0122, and 0124; not all cases trigger the warning. */ - if ((tmp & BIT(NAK_OUT_PACKETS)) == 0) { + if ((ep_stat & BIT(NAK_OUT_PACKETS)) == 0) { ep_warn(ep->dev, "%s lost packet sync!\n", ep->ep.name); req->req.status = -EOVERFLOW; } else { - tmp = readl(&ep->regs->ep_avail); - if (tmp) { + u32 const ep_avail = readl(&ep->regs->ep_avail); + if (ep_avail) { /* fifo gets flushed later */ ep->out_overflow = 1; ep_dbg(ep->dev, "%s dma, discard %d len %d\n", - ep->ep.name, tmp, + ep->ep.name, ep_avail, req->req.length); req->req.status = -EOVERFLOW; } } } - dma_done(ep, req, tmp, 0); + dma_done(ep, req, req_dma_count, 0); num_completed++; }