715 useful
I've got keyboard input working on the 715/33 (should also work on 715/50 and 715/75). There's still some bits which I don't want to commit, so you can ftp those from puffin.external.hp.com/pub/parisc/src/patch-715-19991113
if you want to type amazing sash commands on your 715. try `help' as your first one.
For the machines which have ASP, am I right in saying they have no other sources of interrupts? ie what I want to do is:
if we detect an ASP, set the interrupt routine to load the IRR from ASP and then only make one pass through do_irq_mask instead of two.
This lets me not bother which irq line ASP is interrupting on, at the expense of not being able to take interrupts from any other source. And it'll reduce interrupt latency on those older machines with ASP, so it seems like a win, unless I'm wrong and it's possible to add other sources of interrupt to them.
Signature · 4 lines
Matthew Wilcox <[email protected]> "Windows and MacOS are products, contrived by engineers in the service of specific companies. Unix, by contrast, is not so much a product as it is a painstakingly compiled oral history of the hacker subculture." - N Stephenson
Quoted 2 lines
For the machines which have ASP, am I right in saying they have no other sources of interrupts? ie what I want to do is:
Under hpux 10.X, I was told that the NCR 710/720 controller devices are sources of interrupts. They do this by writing to the processor. Thus, it appears they bypass the ASP.
Signature · 2 lines
J. David Anglin [email protected] National Research Council of Canada (613) 990-0752 (FAX: 952-6605)
Quoted 6 lines · John David Anglin
Quoted 2 lines
For the machines which have ASP, am I right in saying they have no other sources of interrupts? ie what I want to do is:
Under hpux 10.X, I was told that the NCR 710/720 controller devices are sources of interrupts. They do this by writing to the processor. Thus, it appears they bypass the ASP.
I think those are routed through ASP just like they are on Lasi.
At some point we need to be concerned with Zalon interrupt handling too.
- Alex
Quoted 11 lines · John David Anglin
Quoted 6 lines
Quoted 2 lines
For the machines which have ASP, am I right in saying they have no other sources of interrupts? ie what I want to do is:
Under hpux 10.X, I was told that the NCR 710/720 controller devices are sources of interrupts. They do this by writing to the processor. Thus, it appears they bypass the ASP.
I think those are routed through ASP just like they are on Lasi.
At some point we need to be concerned with Zalon interrupt handling too.
I was having problems with SCSI timeouts and LPMCs on my 735 after upgrading
Quoted 3 lines · to 10.20. Grant Grundler wrote:
In looking at the PA7300LC ERS, I see that detection of out of range DMA addresses has to be enabled in the MIOC for LPMCs to occur as a result of these events.
This is really interesting.
Quoted 11 lines
Quoted 2 lines
From page 83 in the discussion on "System Start-UP",
"
MIOC_CONTROL(address 0xF###'F080). This register contains bits that control both the I/O system and the memory system. Refer to the memory controller section for information on setting the memory-related bits; the I/O-related bits should be set as follows:
* setdma_noeccif the PA7300LC should not flag bad data during DMA reads; normally, if a double-bit memory error is detected during an outbound DMA transaction, the bad data word(s) are marked with bad parity on the GSC bus;
* setlpmc_ento detect and generate an LPMC on DMA accesses to memory space that fall outside of the system's installed physical memory;
The SCSI driver will attempt to use "Transaction based Interrupts" only for GSC SCSI cards (use "Zalon" front end). The c720 driver programs the card to generate interrupts by writing to the Processor! (ie way above memory). I wonder if this is causing the LPMC's.
You need to patch out code in c720_init() which calls wsio_get_interrupts(). Don't patch out the call. Rather NOP out the STW's one at a time which should immediately follow that call until it works.
Should be two places the return value is stored and one of them enables transaction based interrupts.
Quoted 8 lines
* set
fast_memoryif dwmode (also inMIOC_CONTROL) is set;
* set lopowhilat if the PA7300LC's power consumption is critical, and you wish to slightly reduce it, even at the expense of increasing CPU to I/O late ncy;
* (setdma_nocache, pgape, and pgdpe only for diagnostic purposes)."If I knew where where
lpmc_enwas set in 10.20, maybe I could turn it off for testing.
Very likely early during processor initialization. I don't know that code.
grant
Quoted 6 lines
Dave
Signature · 3 lines
J. David Anglin [email protected] National Research Council of Canada (613) 990-0752 (FAX: 952-660 5)
Grant Grundler
Unix Developement Lab
+1.408.447.7253
Signature · 2 lines
J. David Anglin [email protected] National Research Council of Canada (613) 990-0752 (FAX: 952-6605)
Quoted 6 lines · John David Anglin
Quoted 2 lines
For the machines which have ASP, am I right in saying they have no other sources of interrupts? ie what I want to do is:
Under hpux 10.X, I was told that the NCR 710/720 controller devices are sources of interrupts. They do this by writing to the processor. Thus, it appears they bypass the ASP.
As noted later, I was the source of John's information. More notes:
o This is a the HP-UX c720 driver taking advantage of the architecture. Linux SCSI drivers won't do that (yet).
o Alex misunderstood - the HP-UX c720 driver *can* bypass all line based interrupt support (ASP/LASI/Dino/...). It only does it for a card which already uses transaction based interupts (TBIs) (HSC F/W SCSI). TBIs are required to work on HSC bus to meet PA I/O Arch and the c720 driver was coded with this safe assumption. TBIs sadly do NOT work under PCI on all platforms.
o Someone also already pointed out the timer uses the processor interrupts. I don't know if WAX (EISA) has it's own IAR but would expect it too. My advice: boot HP-UX 10.20 or 11.X, run Q4 and include pa_intr.pl, "run DumpEIRR" and you'll see exactly which EIR bits are used.
o and lastly, I too am skeptical (like Philipp) how useful this optimization is. I can help measure it once the box is running. I have my own set of optimizations I *think* we need. Philipp agreed we would put them in once I had measurements showing the improvement.
cheers,
grant
Grant Grundler
Unix Developement Lab
+1.408.447.7253
Quoted 6 lines
o Alex misunderstood - the HP-UX c720 driver *can* bypass all line based interrupt support (ASP/LASI/Dino/...). It only does it for a card which already uses transaction based interupts (TBIs) (HSC F/W SCSI). TBIs are required to work on HSC bus to meet PA I/O Arch and the c720 driver was coded with this safe assumption. TBIs sadly do NOT work under PCI on all platforms.
In fact, to the best of my knowledge, software-visible TBIs are unique to PA-RISC among all common architectures. Am I wrong ?
Quoted 5 lines
o and lastly, I too am skeptical (like Philipp) how useful this optimization is. I can help measure it once the box is running. I have my own set of optimizations I *think* we need. Philipp agreed we would put them in once I had measurements showing the improvement.
measurements showing the improvement to be significant enough to put what otherwise has all the features of a nasty hack in the API. What am I afraid is going to happen is:
1. we find out there is an interesting, though not essential, performance win
2. we implement the necessary API
3. the API gets used in exactly one driver
4. the only two developers running machines making use of the driver disappear for a week
5. we break the driver
6. we have a long discussion, and fix the API
goto step 4.
So the possibility of just removing the interface support for using TBIs on busses for which TBI is not the classical interrupt mechanism doesn't sound all that bad compared to leaving it in, and if someone is really convinced TBIs over PCI are worth implementing, I'd suggest to do some benchmarks as soon as possible (i.e. _now_ for PA1.1 boxes).
Philipp Rumpf
Quoted 2 lines · Matthew Wilcox
For the machines which have ASP, am I right in saying they have no other sources of interrupts? ie what I want to do is:
I'm interpretting this question as:
"For machines which have ASP, am I right in saying that no other IO devices apart from ASP generate interrupts to the CPU?"
I think the answer to that question is no, because:
- they have GSC slots, so you could have a card mode lasi or dino device
- doesn't WAX have it's own interrupts?
However, you might be able to set it up so that you still get the optimization if at runtime it's found that the ASP is the only detected device.
Did I interpret your question properly?
- Alex
Quoted 11 lines
For the machines which have ASP, am I right in saying they have no other sources of interrupts? ie what I want to do is:
if we detect an ASP, set the interrupt routine to load the IRR from ASP and then only make one pass through
do_irq_maskinstead of two.This lets me not bother which irq line ASP is interrupting on, at the expense of not being able to take interrupts from any other source. And it'll reduce interrupt latency on those older machines with ASP, so it seems like a win, unless I'm wrong and it's possible to add other sources of interrupt to them.
Matthew, I think you are overestimating the negative impact this code has on interrupt latency. Furthermore, I do not believe solving a problem that is a limitation of the interface used - our current driver table device management system does not allow offsetting IRQ numbers in the same way a linked list would (and does, for PCI).
The very same problem you want to avoid has to be solved in a clean way for Lasi, and handling it different on otherwise very similar machines (712, 715) does more to confuse developers and users (for whom we cannot totally avoid contact with virtual IRQ numbers) than it does to improve the actual performance of the kernel.
If you should disagree on the order of performance loss we are talking about here, please provide benchmarks - should the delay turn out to be unacceptable, we need to find a way of lowering it, even if it does mean to obfuscate code.
Philipp Rumpf