?

Log in

No account? Create an account
Pleased to flash you, hope you change my name... - LIBV Intentionally Breaks Videodrivers [entries|archive|friends|userinfo]
Luc Verhaegen

[ userinfo | livejournal userinfo ]
[ archive | journal archive ]

Pleased to flash you, hope you change my name... [Oct. 17th, 2018|12:49 am]
Luc Verhaegen
[Tags|, , , , , , , , , , , , , ]
[Current Location |Rumpelkammer]
[music |The Rolling Stones - Picture Disc 08 - Sympathy for the devil (1979)]

Remember that time when ATI employees tried to re-market their atomBIOS bytecode as scripts?

You probably don't, it was a decade ago.

It was in the middle of the RadeonHD versus -ati shitstorm. One was the original, written in actual C poking display registers directly, and depending on ATIs atomBIOS as little as practical. The other was the fork, implementing everything the fglrx way, and they stopped at nothing to smear the real code (including vandalizing the radeonhd repo, _after_ radeonhd had died). It was AMD teamed with SUSE versus ATI teamed with redhat. From a software and community point of view, it was real open source code and the goal of technical excellence, versus spite and grandstanding. From an AMD versus ATI point of view, it was trying to gain control over and clean up a broken company, and limiting the damage to server sales, versus fighting the new "corporate" overlord and people fighting to keep their old (ostensibly wrong) ways of working.

The RadeonHD project started april 2007, when we started working on a proposal for an open source driver. AMD management loved it, and supported innovations such as fully free docs (the first time since 3dfx went bust in 1999), and we started coding in July 2007. This is also when we were introduced to John Bridgman. At SUSE, we were told that John Bridgman was there to provide us with the information we needed, and to make sure that the required information would go through legal and be documented in public register documents. As an ATI employee, he had previously been tasked by AMD to provide working documentation infrastructure inside ATI (or bring one into existence?). From the very start, John Bridgman was underhandedly working on slowly killing the RadeonHD project. First by endlessly stalling and telling a different lie every week about why he did not manage to get us information this time round either. Later, when the RadeonHD driver did make it out to the public, by playing a very clear double game, specifically by supporting a competing driver project (which did things the ATI way) and publicly deriding or understating the role of the AMD backed SUSE project.

In November 2007, John Bridgman hired Alex Deucher, a supposed open source developer and x.org community member. While the level of support Bridgman had from his own ATI management is unclear to me, to AMD management he claimed that Alex was only there to help out the AMD sponsored project (again, the one with real code, and public docs), and that Alex was only working on the competing driver in his spare time (yeah right!). Let's just say that John slowly "softened" his claim there over time, as this is how one cooks a frog, and Mr Bridgman is an expert on cooking frogs.

One particularly curious instance occurred in January 2008, when John and Alex started to "communicate" differently about atomBIOS, specifically by consistently referring to it as a set of "scripts". You can see one shameful display here on alex his (now defunct) blog. He did this as part of a half-arsed series of write-ups trying to educate everyone about graphics drivers... Starting of course with... those "scripts" called atomBIOS...

I of course responded with a blog entry myself. Here is a quote from it: "At no point do AtomBIOS functions come close to fitting the definition of script, at least not as we get them. It might start life as "scripts", but what we get is the bytecode, stuck into the ROM of our graphics cards or our mainboard." (libv, 2008-01-28)

At the same time, Alex and John were busy renaming the C code for r100-r4xx to "legacy", inferring both that these old graphics cards were too old to support, and that actual C code is the legacy way of doing things. "The warping of these two words make it impossible to deny: it is something wrong, that somehow has to be made to appear right." (libv, 2008-01-29) Rewriting an obvious truth... Amazing behaviour from someone who was supposed to be part of the open source community.

At no point did ATI provide us with any tools to alter these so-called scripts. They provided only the interpreter for the bytecode, the bare minimum of what it took for the rest of the world to actually use atomBIOS. There never was atomBIOS language documentation. There was no tooling for converting to and from the bytecode. There never was tooling to alter or create PCI BIOS images from said atomBIOS scripts. And no open source tool to flash said BIOSes to the hardware was available. There was an obscure atiflash utility doing the rounds on the internet, and that tool still exits today, but it is not even clear who the author is. It has to be ATI, but it is all very clandestine, i think it is safe to assume that some individuals at some card makers sometimes break their NDAs and release this.

The only tool for looking at atomBIOS is Matthias Hopfs excellent atomdis. He wrote this in the first few weeks of the RadeonHD project. This became a central tool for RadeonHD development, as this gave us the insight into how things fit together. Yes, we did have register documentation, but Mr. Bridgman had given us twice 500 pages of "this bit does that" (the same ones made public a few months later), in the hope that we would not see the forest through the trees. Atomdis, register docs, and the (then) most experienced display driver developer on the planet (by quite a margin) made the RadeonHD a viable driver in record time, and it gave ATI no way back from an open source driver. When we showed Mr. Bridgman the output of atomdis in september 2007, he was amazed just how readable its output was compared to ATI internal tools. So much for scripts eh?

I have to confess though that i convinced Matthias to hold off making atomdis public, as i knew that people like Dave Airlie would use it against us otherwise (as they ended up doing with everything else, they just did not get to use atomdis for this purpose as well). In Q2 2009. after the RadeonHD project was well and truly dead, Matthias brought up the topic again, and i wholeheartedly agreed to throw it out. ATI and the forkers had succeeded anyway, and this way we would give others a tool to potentially help them move on from atomBIOS. Sadly, not much has happened with this code since.

One major advantage of full openness is the ability to support future use cases that were unforeseeable at the time of hardware, code or documentation release. One such use-case is the today rather common (still) use of GPUs for cryptocurrency "mining". This was not a thing back in 2007-2009 when we were doing RadeonHD. It was not a thing when AMD created a GPGPU department out of a mix of ATI and new employees (this is the department that has been pushing out 3D ISA information ever since). It was also not considered when ATI stopped the flow of (non shader ISA) documentation back in Q1 2009 (coinciding with radeonHD dying). We only just got to the point of providing a 3d driver then, and never got near considering pushing for open firmware for Radeon GPUs. Fully open power management and fully open 3d engine firmware could mean that both could be optimised to provide a measurable boost to a very specific use-case, which cryptocurrency mining usually is. By retreating in its proprietary shell with the death of the RadeonHD driver, and by working on the blob based fig-leaf driver to keep the public from hating the fglrx driver as much, ATI has denied us the ability to gain those few 10ths of percents or even whole percents that there are likely to be gained from optimised support.

Today though, miners are mostly just altering the gpu and memory frequencies and voltages in their atomBIOS based data tables (not touching the function tables). They tend to use a binary only gui tool to edit those values. Miners also extensively use the rather horrible atiflash. That all seems very limited compared to what could have been if AMD had not lost the internal battle.

So much for history, as giving ATI the two fingered salute is actually more of an added bonus for me. I primarily wanted to play around with some newer ideas for tracing registers, and i wanted to see how much more effective working with capstone would make me. Since SPI chips are simple hardware, and SPI engines are just as trivial, ati spi engines seemed like an easy target. The fact that there is one version of atiflash (that made it out) that runs under linux, made this an obvious and fun project. So i spent the last month, on and off, instrumenting this binary, flexing my REing muscle.

The information i gleaned is now stuck into flashrom, a tool to which i have been contributing since 2007, from right before i joined SUSE. I even held a coreboot devroom at FOSDEM in 2010, and had a talk on how to RE BIOSes to figure out board specific flash enables. I then was on a long flashrom hiatus from 2010 til earlier this year, when i was solely focused on ARM. But a request for a quick board enable got me into flashrom again. Artificial barriers are made to be breached, and it is fun and rewarding to breach them, and board enables always were a quick fix.

The current changes are still in review at review.coreboot.org, but i have a github clone for those who want immediate and simple access to the complete tree.

To use the ati_spi programmer in flashrom you need to use:
./flashrom -pati_spi
and then specify the usual flashrom arguments and commands.

Current support is limited to the hardware i have in my hands today, namely Rx6xx, and only spans to that hardware that was obviously directly compatible to the Rx6xx SPI engine and GPIO block. This list of a whopping 182 devices includes:
* rx6xx: which are the radeon HD2xxx and HD3xxx series, released april 2007
* rx7xx: or HD4xxx, released june 2008.
* evergreen: or HD5xxx, released september 2009.
* northern island: HD6xxx series, released oktober 2010
* Lombok, part of the Southern Island family, release in january 2012.

The other Southern Islanders have some weird IO read hack that i need to go test.

I have just ordered 250eur worth of used cards that should extend support across all PCIE devices all the way through to Polaris. Vega is still out of reach, as that is prohibitively expensive for a project that is not going to generate any revenue for yours truly anyway (FYI, i am self-employed these days, and i now need to find a balance between fun, progress and actual revenue, and i can only do this sort of thing during downtime between customer projects). According to pci.ids, Vega 12 and 20 are not out yet, but they too will need specific changes when they come out. Having spent all that time instrumenting atiflash, I do have enough info to quickly cover the entire range.

One good thing about targetting AMD/ATI again is that I have been blackballed there since 2007 anyway, so i am not burning any bridges that were not already burned a decade ago \o/

If anyone wants to support this or related work, either as a donation, or as actually invoiced time, drop me an email. If any miners are interested in specifically targetting Vega, soon, or doing other interesting things with ATI hardware, I will be happy to set up a coinbase wallet. Get in touch.

Oh, and test out flashrom, and report back so we can update the device table to "Tested, OK" status, as there are currently 180 entries ranked as "Not-Tested", a number that is probably going to grow quite a bit in the next few weeks as hardware trickles in :)
linkReply

Comments:
From: (Anonymous)
2018-10-18 12:26 pm (UTC)

Mod

Great news! As far as I researched if there will be a technique to wrap DEV_ID that is being read from the AMD chips, we will be able to easily turn any Radeon graphics card to a FirePro/FireGL without need to hack the driver. AMD chips are not limited in its professional features, they only have internal dev_id that is being read by the vBIOS and driver.

(Reply) (Thread)
From: (Anonymous)
2018-10-23 07:55 am (UTC)

Please explain

Hello Luc,

thank you for doing this work. It is much appreciated.

From what I understand it will be possible to load new firmware on those amd graphics cards to free the blobs out of the graphics cards. But what will I flash on them? I am running coreboot on my devices, but having one amd rx480 on one computer, so this has a non-free firmware. Are there free firmwares available for those devices; since you wrote in your next entry that flashrom supports those cards now?

Greetings,
foobar
(Reply) (Thread)
[User Picture]From: libv
2018-10-23 11:19 am (UTC)

Re: Please explain

There are no such images available today. ATI never wanted us to have to tools to do so, so they never happened.

This work removes a big barrier to making free, or at least freer, firmware available though.

The first stage will probably be to fix known issues in atombios code. Such as the lack of HPD pin for DVI in R500, and the bad crtc sync bit in r500 and most of R600.
(Reply) (Parent) (Thread)
From: (Anonymous)
2018-10-23 11:32 am (UTC)

Re: Please explain

Hello Luc,

thank you for the explanation. And thank you again for your work for a more open world. There are so many firmwares in our devices which we do not control like ssds, graphics, etc. Thank you for making them accessible

Greetings,
foobar.
(Reply) (Parent) (Thread)