Skip to content

Latest commit

 

History

History
142 lines (95 loc) · 20.9 KB

UnsolvedMysteries.md

File metadata and controls

142 lines (95 loc) · 20.9 KB

Unsolved Mysteries in AVRs

Solve these - and let me know if you do. You'll get credit and a link to your report, and (depending on the scale of the mystery) maybe even free hardware.

Useful undocumented features that I'm not aware of are a good way to get freebies.

Boring and Pragmatic

Can anyone tell me how to use the gcc "warning()" funtion attribute to produce a warning when a function is used? I've been trying to figure out how to do this for years, and it would let me give much more useful compiler output in the cases where user code is doing things that are clearly suspect, but not necessarily wrong? There are a bunch of places where I'd like to show a warning if they've used a function, or used it when certain compilation options were set? I'd like to be able to do what I do now with badCall() and badArg() in my cores, only with warnings instead of errors.

Modern AVRs - Undocumented Feature Search

There are known to be several undocumented features on the AVR Dx-series. Except for the internal oscillator going up to 32, these are all known from the original headers.

There are bound to be more. Let's find them - the initial (generating prospects) part of the search can be automated. Other parts can be done semi-automated. I want to have reasons to mail people free boards as rewards for finding cool undocumnented, likely not-entirely-functional features!

The only way I know to find these other than digging through headers hoping to find stuff accidentally left in there, is through the less-than-sophisticated process known in info-sec circles as "fuzzing". That consists of setting options that are not expected to be valid or do anything while watching for unexpected behavior, and checking whether the bits "stick" when set or not. (some bits will not even stick, implying that the hardware can't even store them. These would not be expected to be associated with undocumented features unless you were to come across a single register that contains some bits that stick and some that don't. That would indicate that some bit in the register does something if set to some value. Otherwise, we would expect an unused/inactive register to be made up of either 8 sticky bits or more likely 8 non-stick bits, but absolutly not a mix of the two.

Reserved fuse bits

Most of these fuse bits are known to do nothing. Microchip has shipped product with them set to 0 instead of 1 at the factory; many people including Microchip's own QA deparment (if it has one eyeroll) didn't notice. So most of them don't do anything.

Reserved values for fuse group codes have more of a chance of being interesting. Do any of the other BOD voltage possibilities on the Dx-series do anything? The tinyAVRs have a ton of "secret" BOD voltage options which work just fine but are not documented.

What's with OSCCFG on the Dx having 4 bits, but only 2 values that do anything??? (I'm going to guess that at least some of them will be used in the DU, but will do nothing until then. They will need to configure the clock source such that USB will work if you can upload via USB without relying on a user-supplied bootloader, which the pulled product brief could have been interpreted as implying will be a feature (it's hard to say what "Immutable Boot" means, but it's certainly suggestive, no?)

Reserved Group Codes in registers

The easiest one to test. Pull all the group masks and group codes from the headers. For each group mask, check if there are any possible values not covered by the group codes. Turn the peripheral on, in a mode where that group does something visible, and see what happens if you set it to the reseved codes. Read it back and see if the setting "stuck". If not, try disabling the peripheral, writing it, and turning it back on, just in case.

There's nothing interesting in TCD0.CTRLA, the reserved count prescaler option just makes the timer stop (well, not start, since you have to turn it off to change this)

Reserved bits in registers associated with peripherals present on the part

For all reserved bits, try to write a 1 to them if they read as 0 or a 0 if they read as 1. Any for which the bit written "sticks" should be recorded.

Reserved Registers

The first two categories described where you're most likely to find active registers. If those don't give up any interesting secrets on the parts you're looking at, there's only one place left. Unfortunately, it's a big place: Registers within the extended I/O address space, yet which do not have any documented function.

Some of these will be associated with specific peripherals, and you can safely assume that they are going to relate only to that peripheral - these are the reserved bytes from the struct definition. They are also more likely to hide an undocumented feature than an arbitrary unused register address within the extended I/O space.

So, we can separate these into four categiories:

  1. Registers associated with specific peripherals yet marked reserved.
  2. Registers below 0x1000 that are associated with a peripheral on some parts, but this part doesn't have that peripheral. These are not very likely to be fruitful - though if you find a bonus instance of a peripheral, that was certainly worth it - if you know what might have gone there, this also makes figuring out how to use it really easy compared to a mystery register.
  3. Registers between 0x0040 and 0x1000, yet which are NOT associated with a peripheral on this or similar parts,
  4. Registers between 0x1000 and 0x1040 - the NVMCTRL peripheral (we also know 2 of the registers here on the EA-series - EECALB and EECALC - and what the bitfields in them do vaguely speaking - though not the group codes - though we don't know whether production parts actually have anything here)
  5. Registers 0x1040 and the start of SRAM not including those containing the EEPROM, USERROW or SIGROW. These are less likely than other regions to have interesting things hiding

Preparation

Generate 5 lists corresponding to the 5 categories listed above for the parts you are looking at.

Be sure that none of the registers associated with the serial port you using show up in any of the lists.

For each of these categories, you should do the following:

For a reference state - which will be the state your test sketches start in (likely nothing more than a UART enabled to output serial at some baud rate) loop over all the addresses from 0x040 (0x0000 to 0x003F is the I/O space, and we pretty much know what that's about: the low half is 7 VPORTs and 4 GPIO registers, and the high half has the CCP, SPL, SPH and SREG in it, and none of the registers there are writable if they're not one of those (I already checked for unused-yet-present I/O registers ;-) - An unused I/O register that worked would for sure be worth free hardware (I mean, I want it so badly that I checked for them myself - and we all know that one of my guiding principles is "Never do yourself what you can trick someone else into doing for you") - if there were any to be found. But there aren't, which is a shame.) to the start of the EEPROM section. Format that into a uint8_t array and store it in progmem (if it fits). That's your reference. Most of the reserved addresses will read 0x00.

The Serial.printHex() methods are recommended here.

Reserved registers - data gathering

From your reference known state - well... I guess I'd take the one's complement of the value Then read it back. If you see a non-0xFF value in the address you wrote, that's a class A prospect. If you see 0xFF, that's a class B prospect. And if you see 0x00, there's no reason to belive it is writable... however, your code should also after each address, read every register in the 0x0040 to 0x1000 range mentioned above and compare it to the reference. any register that now holds a new value should be recorded (to serial). After encountering that, write the most recent address you were working with where you saw that artifact to EEPROM, do a softeware reset, and pick up where you left off. It may be that this never happens. It may be that there's some address or range of addresses that always gets stuff written to reserved registers, I suspect that this test won't flag any, but it's a good thing to check.

Once youve got a list of prospects

Sort them by peripheral, and see if you can change the behavior of the peripheral in a manner other than "stops working entirely" by writing invalid values.

Modern AVRs - Specific

AVR DA: OSCTEST???

The AVR128DA64 initial header release contained this line in the EVSYS struct:

USEROSCTEST;  /* User 43 - OSCTEST */

Someone should really try it out and figure out what it does! I'd output PWM from a timer clocked from TCA0, and monitor it on the scope while triggering the event. Have the chip running something simple like blink, so if the OSCTEST event is an HCF so you know if the event straight up freezes the chip or something. My guess is that it's and event so that they could fire it while executing code so they can see if whatever they're testing breaks stuff.

tinyAVR 2-series: NVMREFEN?

VREF.CTRLB has a bit called NVMREFEN The description makes very clear that the person writing the documentation didn't know what it did, and probably couldn't find anyone who did.

Bit 2 – NVMREFEN NVM Reference Force Enable
This bit field controls whether the NVMREFEN reference is always on or not

Wonder what the NVM Reference is actually used for? Can you get it to a point where, measuring the current with sufficient precision, that toggling that bit changes the current drawn? If so, can you find any reserved bits in VREF that a) aren't read-only and b) that impact by how much it changes? (I'd try writing all 1's to the reserved bits, and seeing if any stick). What about other parts? Is there an extra reserved bit that "sticks" in an analogous location?

A red herring (mostly)

Don't bother with the BIST stuff from early 2-series io headers, it's a test mode for memory. Unlikly to be of any value. Not to say it is not of interest by virtue of it's existence, but only in the "hmmm looks like they were having memory problems on the 2-series. It'd certainly be interesting to discover that the same registers and fnctionality were presemt on other series, or conversely, that they'd been removed entirely from the silicon, not just the headers and datasheet.

TCA frequency mode

The frequency mode of a type A timer is inconsistent with the description of the timer. In frequency mode, the period is determined by the compare match value. It's specified generically, implying that all three WO channels should be avbailable, so what happens if you try to get different frequencies out of them? It seems like it's inevitably going to come out broken on one pin or the other, it's not clear how this would behave - either that, or the timer would need to keep three counts. If so, 1: that's ingenious however they did it, and 2: Well then, what do the other registers do when it's working this magic? TCA0.CNT does what? What about PER and PERBUF, which are theoretically not used here (it's conceivable that, if it does work, it's keeping three counts. That strikes me as extremely improbable, however.

TCA "restart" errata

Wait, why is this in the errata? It says that the problem is that it resets the direction? The datasheet says it is supposed to! Did they not read their own datasheet? Were major customers gathering outside Microchip HQ with pitchforks demanding this not be how it works?

USART RS485 = 0b10 mode

The tinyAVR 0/1-series parts have two RS485 bits, 0b01 gives the behavior we know, where it drives XDIR to control an external line driver. Early versions of the DA-series headers had it, (the same version with the 4x PLL option (which is in the released silicon) and the 4th power profile for the analog comparators). Since none of the bugs in the USART were fixed on the Dx-series, it stands to reason that this feature was also removed only on paper.

I can't make sense of what the second mode is supposed to do... the documentation on the parts that admit to having this feature is a bit terse:

Either an external line driver is supported as shown in the figure below (RS485=0x1 in USART.CTRLA), or control of
the transmitter driving the TxD pin is provided (RS485=0x2)

What does that mode do? Is whatever it does something that is plausibly what was intended, or did it not really work right ever and get dropped for that reason? Is it secretly present on Dx-series parts? Does it have any practical use?

Why was the 4th power profile for the analog comparators dropped from the DA datasheet and headers at the last minute?

The functionality is almost certainly still in the silicon. One imagines that it must have been broken, and they decided to drop it entirely, instead of of adding to the pile of errata on these parts. This would have had the slowest response time and the lowest power consumption. It may be returning on the EA-series.

What happens if you try to use it on a Dx=seres? How badly busted is it?

tinyAVR 0/1 CCL inputs: SPI MISO, MOSI, and SCK?

On everything else, the CCL can only get SPI SCK and MOSI, and supposedly only in Master mode. I'm not really sure what the point of this functionality is. The Master mode only caveat is not present in the 0/1-series datasheets - AND all three lines are available! Do they all work? Can you use them in slave mode? For that matter, are they really disabled in SPI slave mode on anything that's not a tinyAVR 0/1?

EEPROM cal registers

Check out the NVMCTRL section of the EA-series. Surely those won't be like that on release. Are any of those registers readable as non 0x00 or 0xFF values on Dx or tnyAVR part? Are thwy writable? If yes, cam you write crap to them such that the EEPROM ceases to function corectly. An investigation here that had any substantive findings, particularly if they were useful, is a good way to get a freebie from me. Obviously sommeting useful is worth more than just a series of registers you can write to to break EEPROM functionality.

classic AVRs

The 828 and the mystery of the ADC

The 828 is a sorry, sorry piece of hardware. It is clear reading the datasheet that it was supposed to have a differential ADC, and at the last minute, someone scrubbed out mention of it (they refer to single ended channels, normally they only call them that when they might be a differential channel. The register layout looks like that of the killer ADC on the 841 (one of the best on any classic AVR) - except that almost every bit is marked reserved instead. Rather, it has the lamest ADC possible, and is notable only in that it can use all 28 I/O pins. It has the one basic reference option (REFS0 is all alone), and the high MUX bit (normaly what selects differential or single ended channels on classic AVRs) is always supposed to be set to 0.

If we accept the hypothesis that the differential ADC was planned, but got scrubbed at the last minute, without them being able to swap back to a normal register layout, what does that imply? The fact that the register layout looks like it does, and that the changes to the datasheet look to have been made in such a hurried and slapdash manner strongly suggests that it must have happened very late in the process. I would venture a guess that there were no changes to the die between after QA found that the differential ADC had a major defect. Most likely, management refused to let them respin to fix it, and they didn't want to deal with it via Errata because 1) they knew very well that the chip would never get a fix, and 2) it was so severely broken that it would be embarassing to publically disclose what went wrong and expect customers to navigate the complicated workaround that might be needed to use it, if there was one.

So they were left with a ATtiny88-like chip, with it's only advantage being 1 hardware serial port. Everything else about it is worse than the much older, much cheaper 88: The USI clock pin is pulled low if the WDT isn't in use (per errata, same bug as the 1634 has) The TWI is slave only. You get remappable PWM pins within 1 port, but you still only have 2 timers to generate the PWM with.... Was there really a market for that? Who's gonna buy that? If that's what they had, I'd have expected them to take the design (and possibly some design team members) behind the woodshed. But they didn't. Tey're still selling them 8 years later. Someone must be buying enough of them to keep it profitable to make these. Why? Someone picked it because they wanted lots of ADC mux options? Because they couldn't make a software USART work?

As I pointed out, it looks like they still have the register layout for the fancy ADC that it didn't get... and if they had time to do a die revision to remove the busted ADC, they would have been able to fix it, and the fact that they considered it completely unusable suggests that the mistake was not subtle at all. That all implies that the tiny828's in fact do have a differential ADC under the hood, it's just got a really nasty bit of silicon errata. That would also explain why they are still in production: A few key customers, ones that Atmel worked closely with, know what happened, might have had access to the silicon pre-release, and were made aware of the nature of the defect and some workaround that made it usable.

The first test, would be to make some crude resistor ladder contraption, put one on a breakout board, and feed a bunch of the pins each with a different voltage.Say you do the first 16. Then you can increment through the mux values and see that you're getting a different value every time. Then do the same, only with the high bit of the mux set. If you get the same values, that suggests that this theory is wrong. Treat the value you read as a signed 16-bit integer. 0's or 1023's could mean anything, so you'd need to dig deeper to see if there's anything to this... seemingly random numbers are a good sign. If swapping the ends of your resistor ladder turns positive numbers into negative ones, you've hit pay-dirt!

Then you just have to piece together how it was supposed to work (841's ADC is the reference to uee), and then figure out what's wrong with it and how to work around the bug....

I would consider the possibility that the broken pin could be related somehow. You might need the WDT on to make it not-pull-low for things to work. Or the opposite. Or for it to be pulled high externally. The observed behavior of it though looks almost like a prtotection diode is pulling it low, suggesting that they connected it to the nearest power rail, but that was for some reason not actually powered when the WDT/ULP isn't running. We do know that a hidden ADC is not required for this though, so it may be unrelated, as the 1634 has no indications of a fancier ADC, but still has a bugged pin. On the other hand, the 1634 may have been slated to get the same fancy ADC, had the same problem, and had only slightly more time to deal with it by switching back to a normal register layout.

An alternate way to gain insight would be to find a piece of commercial hardware with an ATtiny828 in it. No commercial manufacturer (ie, big company, not maker-startup-type) would pick the t828 without a good reason. It's obvious that it's crown jewel is missing, and it otherwise lacks compelling advantages. Something that is likely to be measuring analog voltages, and has a tiny828 will provide two clues (the chip is probably locked - if it's not, and you can get a hex file out of it, even if you can't learn anything out of the hex file, I can, and would love to see it, and would likely qualify for freebies: It will likely solve the mystery straight up): First, in the form of the hardware connections: trace out the connections, look for anything odd - is anything other than the power pin tied to Vcc, particularly the broken pin? Second, when power is applied, check what that broken pin is doing. If it's pulled down per the bug, then the bugged pin is unrelated. If it's not - that gets interesting. And if it's connected in a way that doesn't make sense for an I/O pin, that's even more interesting. This all depends on finding such hardware in the wild though...

Hell, for that matter, I'd give freebies for any commercial board that has a tiny828 on it unless it it was being used simply as a 32-pin tiny with a hardware USART and did nothing analog whatsoever (in which case the company's choice of processor might reflect poorly enough on them that it would be blackmail material ;-) )

If you manage that, it still wouldn't be better than the modern AVRs, but it would be a demonstration of persistence and clever low-level work that would win you the admiration of your peers (who you just demonstrated were actually not peers: you're better than they are, because you solved this and they didn't), and you'll get a few boards of your choice from among the ones I sell to replace the poor 828's (or more 828's, if your investigation found that they were actually pretty cool.

Related: the bugged pin on he 1634 and 828

Both of thee have a bugged pin if the WDT isn't enabled. Can anything be determined from investigation of this?