David Brownell wrote:
On Saturday 17 December 2005 1:18 pm, Vitaly Wool wrote:
Greetings,
the patch attached changes the way transfers are chained in the SPI
core. Namely, they are turned into linked lists instead of array. The
reason behind is that we'd like to be able to use lightweight memory
allocation mechanism to use it in interrupt context.
Hmm, color me confused. Is there something preventing a driver from
having its own freelist (or whatever), in cases where kmalloc doesn't
suffice? If not, why should the core change? And what sort of driver
measurements are you doing, to conclude that kmalloc doesn't suffice?
Can't get what you're talking about. A freelist of what?
Basically the idea of the custom lightweight allocation is the
following: allocate a page and divide into regions of the same size,
initialize and use these regions as a stack.
Next, why the current model prevents us from doing that. The array may
be of any size thus we can't divide the page into regions of the same
size: we can't predict what the maximum size will be. And this is the
problem of the core.
An example of such
kind of mechanism can be found in spi-alloc.c file in our core. The
lightweightness is achieved by the knowledge that all the blocks to be
allocated are of the same size.
I'd have said that since this increases the per-transfer costs (to set
up and manage the list memberships) you want to increase the weight of
that part of the API, not decrease it. ;)
Disagree. Let's look deeper. kmalloc itself is more heavyweght than
setting up list memberships.
The list setting commands are pretty essential and will not add a lot to
the assembly code.
Note that your current API maps to mine roughly by equating
allocate your spi_msg
allocate my { spi_message + one spi_transfer }
So if you're doing one allocation anyway, you already have the relevant
linked list (in spi_message) and pre-known size. So this patch wouldn't
improve any direct translation of your driver stack.
This is the patch to your API, so I don't see why you're mentioning it here.
And your understanding is not quite correct. What if I'm going to send a
chain of 5 messages? I'll allocate 5 spi_msg's in my case which all are
gonna be of the same size -- thus the technique described above is
applicable. In case of your core it's not.
And I'm not worrying that much about direct translation, I'm just trying
to make your core acceptabe for us so that we could give up ours and
work with yours.
So... it's just one more step towards convergence you are not likely to
accept.
We'd like to use this allocation
technique for both message structure and transfer structure The problem
with the current code is that transfers are represnted as an array so it
can be of any size effectively.
Could you elaborate on this problem you perceive? This isn't the only
driver API in Linux to talk in terms of arrays describing transfers,
even potentially large arrays.
The problem is: we're using real-time enhancements patch developed by
Ingo/Sven/Daniel etc. You cannot call kmalloc from the interrupt
context if you're using this patch. Yeah, yeah -- the interrupt
handlers are in threads by default, but we can't go for that since we
want immediate acknowledgement from the interrupt context, and that
implies spi_message/spi_transfer allocation.
Consider how "struct scatterlist" is used, and how USB manages the
descriptors for isochronous transfers. They don't use linked lists
there, and haven't seemed to suffer from it.
Not sure if I understand why it's relevant to what we're discussing.
Vitaly
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
