Re: [patch 1/5] Staging: VME Framework for the Linux Kernel

From: Martyn Welch
Date: Mon Aug 10 2009 - 11:53:04 EST


Emilio G. Cota wrote:
Martyn Welch wrote:
Emilio G. Cota wrote:
- semaphores? isn't it screaming for mutexes?
The semaphores are initialized in mutex mode.

I know, but then please use mutexes.

I'm clearly missing something here - can you provide me with an example?
- some function signatures (get_whatever) pass too many
pointers; it's rather ugly. Passing a pointer to an
appropriate struct seems a better solution.
When implementing these functions it seemed like a simpler option to pass the parameters rather than defining yet another structure and using it for a single call. If the *_get functions are changed then the *_set functions should also be changed for consistency. I'm not overly bothered either way - if the consensus is that a structure would be better then we can go with a structure.

Yeh, In my opinion a structure there would be prettier. However
let's put this off a bit, I'm currently re-working this a bit.

- vme_alloc_consistent looks pretty fishy; why don't you pass
that responsibility to the specific master? There you obviously
know if you're bridging over PCI or whatever. Or even better;
why is this needed at all?
In the model I have chosen it is up to the VME device driver to create buffers rather than the VME bridge drivers. After all it is the device drivers for the specific hardware found on the VME bus that knows what it is going to do with these buffers. This method is provided as a helper. It ensures that a VME device driver (not to be confused with the VME bridge driver) is able to allocate a buffer suitable for utilizing the DMA engines found in the bridge or indeed to be mapped onto VME though a slave window without explicitly knowing which VME bridge is being used.

hmm I see. The way you provide coherent DMA mappings for VME is through
providing coherent (consistent) mappings for PCI, and then mapping the
VME space onto those. I haven't tested if this works, but sounds
reasonable.
Master windows don't require a contiguous buffer, are you referring to something else when you say master?
I was thinking of the hypothetical case where you'd have a bridge not
over PCI; but anyway since quite a few things seem tied to PCI this
is fine.

That's one reason for the helper function. If VME bridges are added which sit of on other buses then vme_alloc_consistent can be extended to support this without requiring VME device drivers to be altered to reflect the fact that pci_alloc_consistent might not work for all VME bridges.

- Please explain me what all the DMA functions do; are they
meant to be used by master or slaves?
The TODO file (drivers/staging/vme/TODO) contains a description of the current API including the DMA functions, does that provide enough of a description?

I've had a closer look; it seems to me that most of it is unnecessary;
there's no show those lists to a driver. I'd just provide a single
'do_dma(attributes)' call that sleeps until it's done (or similar).

The DMA controller in the tsi-148 can do PCI to PCI; PCI to VME; VME to PCI; VME to VME; Patterns to VME and Patterns to PCI transfers. How do you specify all those options without providing a structure where over 50% of the fields aren't used for any given transfer?

Every bridge I have seen is capable of link-list execution. The API provides the ability to do scatter-gather style DMA transfers, this could be used as part of a "zero-copy" type driver for high speed VME capture devices. How would you support the link-list execution with a single call?

I was also looking at the ability to queue DMA transfers if the controller was currently busy - if the consensus is that this is overkill I will happily remove this.

Have a look at the interface for slaves we've got for our tsi148
driver, available at:
http://repo.or.cz/w/tsi148vmebridge.git

/* API for new drivers */
extern int vme_request_irq(unsigned int, int (*)(void *),
void *, const char *);
[snip]
extern int vme_bus_error_check(int);

That's pretty thin and it covers our slaves' needs. Do you see
anything missing there?
This interface, especially struct vme_mapping seems to have been written solely for the tsi-148 bridge. The API I have defined aims to provide a consistent API across multiple VME bridges. Whilst this API clearly works for you with the tsi-148 I am unsure how suitable it will be for other bridge chips.

The API I have proposed is designed to support more than just the tsi148 chipset. Have you thought about how the above API will be supported on other VME bridges?

We only have access to tsi148 chips. Since apparently there aren't
many bridges around, I think is saner to just provide an interface
that works well with the devices we have access to (i.e. tsi148 and
Tundra Universe in your case), and when new chips come along, we simply
modify the interface as needed. Of course this doesn't mean we shouldn't
try to make it generic, but within reason.

Agreed - which I believe is the approach I've taken.
For masters there's no interface there because it was the
master's driver who directly provided these calls to slaves.
I had it in my to-do list to split that from the tsi148, in
the same fashion as you've done with this work.

- Note above the interrupt handler; simply needs the cookie. Also,
shouldn't your vme_request_irq() just require the IRQ vector?
No. I believe that you are using the term IRQ vector where we would use the term status ID, the value which is returned from an IACK cycle. Your interrupt handling code assigns a single interrupt handler to all interrupt levels, purely using the interrupt vector/status ID to determine which interrupt handler will be used. This adds an artificial limitation and would not work in some instances that we have seen. Our framework provides the ability to attach an interrupt handler to each combination of IRQ level and Status ID/vector.
Fair enough. For sanity we tend not to share IRQ vectors among
different modules, but yes, the interface should know about this.

- I'd like to see the whole picture, or 'vertical slice', i.e.
the bus interface + a master + a slave driver. How would
the slave's driver get the addresses and sizes of the mappings,
interrupt lines, etc. for each of the devices it controls?
For the time being we quickly hacked an xml-based scheme to get
this info upon installation, but it's clearly not suitable
for mainline.
I have written a test driver for a very old slave card we had lying around. In that case we used module parameters to determine the location of the device on the bus (it used a fixed VME address space). In this instance the interrupt level and ID could be configured in the registers available in the VME address space, hence I added module parameters to allow these to be configured. In this respect configuration of VME devices is very similar to ISA devices - neither of the buses has supported discovery mechanism from the outset and thus old cards. I therefore implemented a mechanism similar to how I believe ISA approaches this.

The framework that I have proposed aims to provide a consistent API to manage allowing the resources provided by the VME bridges to be managed in as a consistent a manner as possible. I believe it is up to the device drivers for the devices found on the VME bus to determine the best way to configure this as it is not provided by the VME specifications.

The problem is how to manage several (say 10) devices that have to be
controlled with the same driver; passing parameters upon loading the
driver's module doesn't seem to scale beyond one device..
I implemented it using param array, I agree that the arguments might get quite long with 10 devices, especially if multiple parameters are required, but I don't see why it shouldn't work.

At the moment I see no clean way of doing this; right now we're doing
it through an IOCTL, copying from user-space a tree that describes the
devices to install, namely mappings and IRQs for each of them.

Greg, any ideas on how to deal with this?

For using the VME bus for communication between two SBCs, I think we could use the CRCSR space to provide some information about the resources used on each board for say, a virtual network driver like the rapidIO subsystem provides. Possibly using the "device tree" stuff used on PowerPC, Blackfin and Sparc archs (I believe) for passing device layout between the firmware and kernel at boot.

hmm I don't see how we could implement this common consistent view of
the bus in a way that could work with every bridge. For the time
being I'd defer this and simply pass that responsibility to user-space.

That's the conclusion I sort of came to. I guess how the VME address space(s) are used is a system integration issue.
Sure, np,

Martyn

Thanks,
E.


--
Martyn Welch MEng MPhil MIET (Principal Software Engineer) T:+44(0)1327322748
GE Fanuc Intelligent Platforms Ltd, |Registered in England and Wales
Tove Valley Business Park, Towcester, |(3828642) at 100 Barbirolli Square,
Northants, NN12 6PF, UK T:+44(0)1327359444 |Manchester,M2 3AB VAT:GB 927559189
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/