Thread (28 messages) flat view 28 messages, 2 authors, 2016-09-15

Re: [PATCHv3 03/11] crypto: omap-sham: implement context export/import APIs

From: Tero Kristo <hidden>
Date: 2016-09-01 06:13:45
Also in: linux-arm-kernel, linux-omap

On 01/09/16 06:33, Herbert Xu wrote:
On Mon, Aug 29, 2016 at 05:11:35PM +0300, Tero Kristo wrote:
quoted
quoted
quoted
+static int omap_sham_export(struct ahash_request *req, void *out)
+{
+	struct omap_sham_reqctx *rctx = ahash_request_ctx(req);
+
+	while (omap_sham_flush(req) == -EINPROGRESS)
+		msleep(10);
Do we really need this? You must not call export until the previous
operation has completed.

Cheers,
Sorry about a late reply, I was out on vacation.

For OMAP SHAM, this is actually needed, because the driver still has
a very large internal buffer for performance reasons, and the whole
buffer can't be exported. The flush functionality pushes out
sufficient amount of data to the hardware, so that the rest of the
buffer can be exported to the available space.
It doesn't matter whether you have a buffer or not.  The point
is that the completion function should not be called until the
operation is actually complete.  This is the whole point of the
async interface.

As the user must not call export until the completion function
has been called, there should be no need to wait in the export
function.
Well, but the driver doesn't flush its buffers automatically, it caches 
data until it has sufficient amount available. So, assuming you want to 
do this:

sham_init
   sham_update 256 bytes
   sham_update 256 bytes
   wait until two above updates are complete
   sham_export

... the execution hangs at the wait phase as the driver is still waiting 
for more data to cache, and will never complete the two update requests. 
Currently, the driver is written in such way that it waits until it has 
enough data cached before starting to push it out to hardware, or waits 
until sham_final to be called. Pushing out small pieces of data causes 
severe performance degradation on the driver, as setting up the DMA 
operation itself is rather costly.

Either way, flush for the buffers is needed, I wonder if automatic flush 
should be added also based on some timer.

-Tero
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help