Older blog entries for mjg59 (starting at number 325)

Why UEFI secure boot is difficult for Linux

I wrote about the technical details of supporting the UEFI secure boot specification with Linux. Despite me pretty clearly saying that this was ignoring issues of licensing and key distribution and the like, people are now using it to claim that Linux could support secure boot with minimal effort. In a sense, they're right. The technical implementation details are fairly straightforward. But they're not the difficult bit.

Secure boot requires that all code that can touch hardware be trusted

Right now, if you can run unstrusted code before the OS then you can subvert the OS. Secure boot gives you a mechanism for making sure you only run trusted code, which protects against that. So your UEFI drivers have to be signed, your bootloader has to be signed, and your bootloader must only load a signed kernel. If you've only booted trusted code then you know that your OS is safe. But, unlike trusted boot, secure boot provides no way for you to know that only trusted code was executed. That has to be ensured by OS policy.

This doesn't sound like too much of a problem. But it is. Imagine we have a signed Linux bootloader and a signed Linux kernel, and that these signatures are made with a globally trusted key. These will boot on any hardware using secure boot. Now imagine that an attacker writes a kernel module that sets up a fake UEFI environment, stops the kernel from running code and then executes the Windows bootloader - kind of like kexec, but a little more fiddly. The Windows bootloader believes that it's performing a secure boot, but in fact everything that it believes is trustworthy is the attacker's fake UEFI code. The user's encryption passphrase is logged, the Windows kernel is modified to hide the Linux code and despite everything looking fine your credit card details are on their way to China. In this scenario, the signed Linux kernel is simply used as a malware loader. The only sign that anything is wrong is that boot will be slightly slowed down.

Signing the kernel isn't enough. Signed Linux kernels must refuse to load any unsigned kernel modules. Virtualbox on Linux? Dead. Nvidia binary driver on Linux? Dead. All out of tree kernel modules? Utterly, utterly dead. Building an updated driver locally? Not going to happen. That's going to make some people fairly unhappy.

(The same applies to Windows, of course. Windows 7 allows you to disable driver integrity checks. Windows 8 will have to forbid that when the system's using secure boot)

Licensing

GPLv3 has various requirements for signing keys to be available. Microsoft's new requirement that systems support the installation of user keys would let users boot their own modified bootloaders, so that may end up being sufficient to satisfy the license. But we're then beholden on Microsoft - if they remove that requirement then users lose that freedom, and suddenly we're in an awkward licensing situation. There are ongoing conversations about exactly what we're able to do here, but it's not a solved problem.

Key distribution

The UEFI spec doesn't describe or mandate a central certifying authority. Microsoft require that everyone carry their key. We could generate our own, but we have much less sway with vendors. There's no way to guarantee that all hardware vendors will generate our key. And, obviously, if we generate a key, we can't just hand the private half out to others. That means that it becomes impossible for people to produce derivative versions of Linux distributions without getting their own key. The kind of identity verification that would be required for getting such a key is likely to be expensive, and also fairly likely to require that the distribution have a legally registered company in order to facilitate the identity verification. Think Extended Validation certificates, not Startssl Free. Hobbyist Linux distributions will be a thing of the past.

Doesn't custom mode fix this?

Microsoft's certification requirements now state that all systems must support a custom mode, implying that it will be possible for a user to install their own keys. Linux vendors would then be able to ship with their own keys on the install media and impose their own policies. Everyone's happy. It's not really good enough, though. People have spent incredible amounts of time and effort making it easy to install Linux by doing little more than putting a CD in a drive. Asking them to go into the firmware and reconfigure things adds an extra barrier that restricts the ability to install Linux to more technically skilled users. And it's even worse than that. This is the full description of the requirement for custom mode:

1. It shall be possible for a physically present user to use the Custom Mode firmware setup option to modify the contents of the Secure Boot signature databases and the PK.
   
2. If the user ends up deleting the PK then, upon exiting the Custom Mode firmware setup, the system will be operating in Setup Mode with Secure Boot turned off.
   
3. The firmware setup shall indicate if Secure Boot is turned on, and if it is operated in Standard or Custom Mode. The firmware setup must provide an option to return from Custom to Standard Mode which restores the factory defaults.

There's a few things missing from this, namely:

• Any description of the UI. It's effectively impossible to document Linux installation when the first step becomes (a) complicated and (b) vendor specific. Vendors are using the UEFI transition to differentiate themselves by coming up with their own unique firmware interfaces. Custom mode is going to look different everywhere.
• Any description of the key format. A raw binary representation of the key? An EFI_SIGNATURE_DATA struct? A base64 encoding of one, further protected with ROT13? We just don't know.
• Any way to use custom mode for unattended installs. It's a firmware interface that requires a physically present user. Want to install a few thousand machines over the network? This isn't a scalable approach
• …and this one's nitpicking, but there's not actually any requirement that the user be able to add keys - a vendor could conform to this language by only letting users delete keys. This is actually ok as long as the user deletes Pk, because then we'll effectively be back in setup mode and can install our own keys from the installer, but it still results in some practical problems

So no, custom mode doesn't make everything ok. Custom mode with a mandated UI and a documented key format would be much closer, but it wouldn't solve the problem of unattended automated installs.

Summary

We can write the code required to support secure boot on Linux in a minimal amount of time - in fact, most of it's now done. But significant practical problems remain, and so far we have no workable solutions for any of them.

comments

Syndicated 2012-01-18 00:55:05 from Matthew Garrett

Firmware bugs considered enraging

Part of our work to make it possible to use UEFI Secure Boot on Linux has been to improve our EFI variable support code. Right now this has a hardcoded assumption that variables are 1024 bytes or smaller, which was true in pre-1.0 versions of the EFI spec. Modern implementations allow the maximum variable size to be determined by the hardware, and with implementations using large key sizes and hashes 1024 bytes isn't really going to cut it. My first attempt at this was a little ugly but also fell foul of the fact that sysfs only allows writes of up to the size of a page - so 4KB on most of the platforms we're caring about. So I've now reimplemented it as a filesystem[1], which is trickier but avoids this problem nicely.

Things were almost working fine - I could read variables of arbitrary sizes, and I could write to existing variables. I was just finishing hooking up new variable creation, but in the process accidentally set the contents of the Boot0002 variable to 0xffffffff 0xffffffff 0x00000000. Boot* variables provide the UEFI firmware with the different configured boot devices on the system - they can point either at a raw device or at a bootloader on a device, and they can do so using various different namespaces. They have a defined format, as documented in chapter 9 of the UEFI spec. At boot time the boot manager reads the variables and attempts to boot from them in a configured order as found in the BootOrder variable.

Now, obviously, 0xffffffff 0x00000000 is unlikely to conform to the specification. And when I rebooted the machine, it gave me a flashing cursor and did nothing. Fair enough - I should be able to choose another boot path from the boot manager. Except the boot manager behaves identically, and I get a flashing cursor and nothing else.

I reported this to the EDK2 development list, and Andrew Fish (who invented EFI back in the 90s) pointed me at the code that's probably responsible. It's in the BDS (Boot Device Selection) library that's part of the UEFI reference implementation from Intel, and you can find it here. The relevant function is BdsLibVariableToOption, which is as follows (with irrelevant bits elided):

BdsLibVariableToOption (
  IN OUT LIST_ENTRY                   *BdsCommonOptionList,
  IN  CHAR16                          *VariableName
  )
{
  UINT16                    FilePathSize;
  UINT8                     *Variable;
  UINT8                     *TempPtr;
  UINTN                     VariableSize;
  VOID                      *LoadOptions;
  UINT32                    LoadOptionsSize;
  CHAR16                    *Description;

  //
  // Read the variable. We will never free this data.
  //
  Variable = BdsLibGetVariableAndSize (
              VariableName,
              &gEfiGlobalVariableGuid,
              &VariableSize
              );
  if (Variable == NULL) {
    return NULL;
  }

So so far so good - we read the variable from flash and put it in Variable, Variable is now 0xffffffff 0xffffffff 0x00000000. If it hadn't existed we'd have skipped over and continued. VariableSize is 12.

  //
  // Get the option attribute
  //
  TempPtr   =  Variable;
  Attribute =  *(UINT32 *) Variable;
  TempPtr   += sizeof (UINT32);

Attribute is now 0xffffffff and TempPtr points to Variable + 4.

  //
  // Get the option's device path size
  //
  FilePathSize =  *(UINT16 *) TempPtr;
  TempPtr      += sizeof (UINT16);

FilePathSize is 0xffff, TempPtr points to Variable + 6.

  //
  // Get the option's description string size
  //
  TempPtr     += StrSize ((CHAR16 *) TempPtr);

TempPtr points to 0xffff 0x0000, so StrSize (which is basically strlen) will be 4. TempPtr now points to Variable + 10.

  //
  // Get the option's device path
  //
  DevicePath =  (EFI_DEVICE_PATH_PROTOCOL *) TempPtr;
  TempPtr    += FilePathSize;

TempPtr now points to Variable + 65545 (FilePathSize is 0xffff).

  LoadOptions     = TempPtr;
  LoadOptionsSize = (UINT32) (VariableSize - (UINTN) (TempPtr - Variable));

LoadOptionsSize is now 12 - (Variable + 65545 - Variable), or 12 - 65545, or -65533. But it's cast to an unsigned 32 bit integer, so it's actually 4294901763.

  Option->LoadOptions = AllocateZeroPool (LoadOptionsSize);
  ASSERT(Option->LoadOptions != NULL);

We attempt to allocate just under 4GB of RAM. This probably fails - if it does the boot manager exits. This probably means game over. But if it somehow succeeds:

CopyMem (Option->LoadOptions, LoadOptions, LoadOptionsSize);

we then proceed to read almost 4GB of content from uninitialised addresses, and since Variable was probably allocated below 4GB that almost certainly includes all of your PCI space (which is typically still below 4GB) and bits of your hardware probably generate very unhappy signals on the bus and you lose anyway.

So now I have a machine that won't boot, and the clear CMOS jumper doesn't clear the flash contents so I have no idea how to recover from it. And because this code is present in the Intel reference implementation, doing the same thing on most other UEFI machines would probably have the same outcome. Thankfully, it's not something people are likely to do by accident - using any of the standard interfaces will always generate a valid entry, so you could only trigger this when modifying variables by hand. But now I need to set up another test machine.

[1] All code in Linux will evolve until the point where it's implemented as a filesystem.

comments

Syndicated 2012-01-06 20:17:56 from Matthew Garrett

The economic incentive to violate the GPL

My post yesterday on how Google gains financial benefit from vendor GPL violations contained an assertion that some people have questioned - namely, "unscrupulous hardware vendors save money by ignoring their GPL obligations". And, to be fair, as written it's true but not entirely convincing. So instead, let's consider "unscrupulous hardware vendors have economic incentives to ignore their GPL obligations".

The direct act of compliance costs money

Complying with the GPL means having the source code that built the binaries you ship. This is easy if your workflow involves putting source in at one end and getting binaries out at the other, but getting to that workflow means having a certain degree of engineering rigour. If your current build process involves mixing a bunch of known good binaries you got from somewhere but you can't remember where with a hacked up source tree that exists on someone's hard drive and then pushing all of these into a tool that only runs on Windows ME, before taking the resulting image and replacing chunks of it by hand, compliance is effectively impossible.

We all know that this is against all kinds of best practices and probably causes so many problems that it's more expensive in the long term, but retooling and hiring someone to oversee all of this takes time and money, and given the margins on many of these devices that's probably enough to make you uncompetitive for a couple of product cycles. Maybe you'll be in a better position afterwards, but you don't know that there'll be an afterwards.

Suppliers who don't provide you with the source code may be cheaper than those who do

You can't be in compliance if you don't have the source code in the first place. The same arguments that apply to the hardware vendors also apply to the people selling you your chips, so there's also an economic incentive for them to avoid complying. And there's an obvious incentive for you to choose the cheaper chipset, even if they don't comply.

Getting the source may cost money

Buying a chipset doesn't necessarily get you the software that makes it work - several silicon vendors will charge you for the SDK. But many of these devices are effectively reference platforms, so are basically identical from a hardware perspective. So if one of your competitors paid for the SDK, you can just dump the binaries off their machine, flash them onto your own boards and save yourself a decent amount of money. You obviously don't get the source, and nor do you have the standing to insist that the vendor whose binaries you misappropriated give you the source.

In the absence of enforcement, GPL compliance only works if it's the norm

Let's imagine two companies, A and B. Both build a tablet device, and buy the full SDK including source code. Both find a bunch of bugs in the vendor SDK and fix a different subset of them. They ship. A provides source code. B doesn't. B can now take A's bugfixes and incorporate them, resulting in a more compelling product without any significant extra cost. You now have two products that can sell for the same price, but B's is better. A would need to prove that B copied their bugfixes rather than simply fixing them themselves , which probably isn't going to happen.

In a larger market, if B is the only vendor who does this then their advantage isn't large - some of A's work is misappropriated by B, but A does benefit from the engineering work contributed by C, D, E, F and G. A combination of social pressure and legal threats may bring B into compliance. But if infringement is the norm, A has no incentive at all to release the source - by doing so they'll be helping not only B, but also C, D, E, F and G. Everyone undercuts A and they go out of business quite quickly.

Moral: In the absence of enforcement, if everyone else is infringing, a single company who complies is at a disadvantage.

If compliance cost nothing then everyone would do it

You can argue that cheap tablets from China are infringing simply because nobody knows better. But what's HTC's excuse? They've clearly decided that there's a benefit in holding back their source code releases[1], balancing this against the risk of being sued. They know full well what they're doing. If compliance was free they'd ship the source at the same time as they shipped the binaries. Other significant vendors are also fully aware of their obligations but choose to ignore them anyway.

Summary

There are economic incentives to infringe the GPL, and therefore (all else being equal) an infringing device can be sold for less money. All else being equal, a cheaper device will sell more units. More sales means more devices selling adverts for Google. Google makes more money because Android vendors infringe the GPL.

[1] The usual argument is "We will release the source code within 120 days", implying that it's a process that takes time and we should just be patient. Every single time I've started making threatening noises, the source has appeared within a week.

comments

Syndicated 2012-01-04 15:08:34 from Matthew Garrett

Android, GPL violations and Google

A bit over a year ago, I wrote about how an incredible number of Android tablets on the market were in violation of the terms of the GPL. I've had rather a lot else to do since then so it's now awfully out of date - but taking a quick browse through the current stack of cheaper devices indicates that things aren't all that much better. We've got source code for some chipsets that were missing it before, but to compensate we've got a whole bunch of new hardware that's entirely lacking. It's all pretty poor, really.

At the time, I wrote the following:

"(Side note: People sometimes ask why Google aren't doing more to prevent infringing devices. For the vast majority of these cases, Google's sole contribution has been to put Android source code on a public website. Red Hat own more of the infringing code than Google do. There's no real reason why Google should be the ones taking the lead role here, and there's fairly sound business reasons why it's not in their interest to do so)"

Factually speaking, nothing's changed. Each of these devices contains code owned by Google, and Google could absolutely take legal action against the vendors. Equally, so could Red Hat, Intel, Nokia and dozens of other companies who hold copyright on portions of the code carried on these devices, and so could thousands of individuals around the world. Nobody's obliged to enforce their copyrights, and in the absence of anyone else doing so it's unreasonable to insist that Google should do it.

However.

Google gives Android away. This seems like an odd thing for them to do, given that it's a significant engineering effort and costs a lot of money to produce. But remember what Android brings to Google - it's a platform with a well-integrated mechanism for distributing advertising to users. Scanning the market shows a huge number of ad-supported apps, and Google's getting money for every single one of those that gets shown. The more Android devices, the bigger the market for apps - and the wider their advertising reach.

In other words: unscrupulous hardware vendors save money by ignoring their GPL obligations. This lets them appeal on price, increasing the number of Android devices in use and increasing Google's profits. Google makes money off other people's violation of the GPL.

Could Google do anything to stop this? Yes. They could sue for copyright infringement, but that kind of thing's time consuming and awkward and any argument about the GPL always seems to end up as a big argument involving conspiracy theories. Instead, Google could attach some extra conditions to the Android trademark. Requiring that the trademark only be attached to GPL-compliant products ought to allow Google to take advantage of the existing well-tested mechanisms for seizing counterfeit goods, providing a direct economic incentive for companies to come into compliance. For added marks, they could restrict the adwords code to devices that use the trademark - if the vendor removes the trademark, applications depending on the adwords functionality would refuse to run and Google wouldn't make money off the infringing hardware.

Or, of course, they could just carry on making extra money as a result of vendors denying users the freedoms granted by the copyright holders. Although that sounds kind of evil to me.

comments

Syndicated 2012-01-04 03:11:38 from Matthew Garrett

TVs are all awful

A discussion a couple of days ago about DPI detection (which is best summarised by this and this and I am not having this discussion again) made me remember a chain of other awful things about consumer displays and EDID and there not being enough gin in the world, and reading various bits of the internet and wikipedia seemed to indicate that almost everybody who's written about this has issues with either (a) technology or (b) English, so I might as well write something.

The first problem is unique (I hope) to 720p LCD TVs. 720p is an HD broadcast standard that's defined as having a resolution of 1280x720. A 720p TV is able to display that image without any downscaling. So, naively, you'd expect them to have 1280x720 displays. Now obviously I wouldn't bother mentioning this unless there was some kind of hilarious insanity involved, so you'll be entirely unsurprised when I tell you that most actually have 1366x768 displays. So your 720p content has to be upscaled to fill the screen anyway, but given that you'd have to do the same for displaying 720p content on a 1920x1080 device this isn't the worst thing ever in the world. No, it's more subtle than that.

EDID is a standard for a blob of data that allows a display device to express its capabilities to a video source in order to ensure that an appropriate mode is negotiated. It allows resolutions to be expressed in a bunch of ways - you can set a bunch of bits to indicate which standard modes you support (1366x768 is not one of these standard modes), you can express the standard timing resolution (the horizontal resolution divided by 8, followed by an aspect ratio) and you can express a detailed timing block (a full description of a supported resolution).

1366/8 = 170.75. Hm.

Ok, so 1366x768 can't be expressed in the standard timing resolution block. The closest you can provide for the horizontal resolution is either 1360 or 1368. You also can't supply a vertical resolution - all you can do is say that it's a 16:9 mode. For 1360, that ends up being 765. For 1368, that ends up being 769.

It's ok, though, because you can just put this in the detailed timing block, except it turns out that basically no TVs do, probably because the people making them are the ones who've taken all the gin.

So what we end up with is a bunch of hardware that people assume is 1280x720, but is actually 1366x768, except they're telling your computer that they're either 1360x765 or 1368x769. And you're probably running an OS that's doing sub-pixel anti-aliasing, which requires that the hardware be able to address the pixels directly which is obviously difficult if you think the screen is one size and actually it's another. Thankfully Linux takes care of you here, and this code makes everything ok. Phew, eh?

But ha ha, no, it's worse than that. And the rest applies to 1080p ones as well.

Back in the old days when TV signals were analogue and got turned into a picture by a bunch of magnets waving a beam of electrons about all over the place, it was impossible to guarantee that all TV sets were adjusted correctly and so you couldn't assume that the edges of a picture would actually be visible to the viewer. In order to put text on screen without risking bits of it being lost, you had to steer clear of the edges. Over time this became roughly standardised and the areas of the signal that weren't expected to be displayed were called overscan. Now, of course, we're in a mostly digital world and such things can be ignored, except that when digital TVs first appeared they were mostly used to watch analogue signals so still needed to overscan because otherwise you'd have the titles floating weirdly in the middle of the screen rather than towards the edges, and so because it's never possible to kill technology that's escaped into the wild we're stuck with it.

tl;dr - Your 1920x1080 TV takes a 1920x1080 signal, chops the edges off it and then stretches the rest to fit the screen because of decisions made in the 1930s.

So you plug your computer into a TV and even though you know what the resolution really is you still don't get to address the individual pixels. Even worse, the edges of your screen are missing.

The best thing about overscan is that it's not rigorously standardised - different broadcast bodies have different recommendations, but you're then still at the mercy of what your TV vendor decided to implement. So what usually happens is that graphics vendors have some way in their drivers to compensate for overscan, which involves you manually setting the degree of overscan that your TV provides. This works very simply - you take your 1920x1080 framebuffer and draw different sized black borders until the edge of your desktop lines up with the edge of your TV. The best bit about this is that while you're still scanning out a 1920x1080 mode, your desktop has now shrunk to something more like 1728x972 and your TV is then scaling it back up to 1920x1080. Once again, you lose.

The HDMI spec actually defines an extension block for EDID that indicates whether the display will overscan or not, but doesn't provide any way to work out how much it'll overscan. We haven't seen many of those in the wild. It's also possible to send an HDMI information frame that indicates whether or not the video source is expecting to be overscanned or not, but (a) we don't do that and (b) it'll probably be ignored even if we did, because who ever tests this stuff. The HDMI spec also says that the default behaviour for 1920x1080 (but not 1366x768) should be to assume overscan. Charming.

The best thing about all of this is that the same TV will often have different behaviour depending on whether you connect via DVI or HDMI, but some TVs will still overscan DVI. Some TVs have options in the menu to disable overscan and others don't. Some monitors will overscan if you feed them an HD resolution over HDMI, so if you have HD content and don't want to lose the edges then your hardware needs to scale it down and let the display scale it back up again. It's all awful. I recommend you drink until everything's already blurry and then none of this will matter.

comments

Syndicated 2012-01-03 17:46:40 from Matthew Garrett

Clarifying the "secure boot attack"

Yesterday I wrote about an alleged attack on the Windows 8 secure boot implementation. As I later clarified, it turns out that the story was, to put it charitably, entirely wrong. The attack is a boot kit targeted towards BIOS-based boots. It lives in the MBR. It'll never be executed on any UEFI systems, let alone secure boot ones. In fact, this is precisely the kind of attack that secure boot is intended to protect against. So, context.

The MBR contains code that's executed by the BIOS at boot time. This code is unverifiable - it's permitted to have arbitrary functionality. There's only 440 bytes, but that's enough to jump to somewhere else and read code from elsewhere. There's no way for the BIOS to know that this code is malicious. And one thing this code can obviously do is load the normal boot code and modify it to behave differently. Any self-validation code in the loader can be patched out at this point. The modified loader will then load the kernel, and potentially also modify it. At this point, you've lost. Any attempts to validate the code can be redirected to the original code and so everything will look fine, up until the point where the user runs a specific application and suddenly your kernel is sending all your keystrokes over UDP to someone in Nigeria.

These attacks exist now. They're in the wild. In a normal UEFI world you'd do the same thing by just replacing the UEFI bootloader. But with secure boot you'll be able to validate that the bootloader is appropriately signed and if someone's modified it you'll drop into some remediation mode that recovers your files, from install media if necessary.

Obviously, this protection is based on all the components of secure boot (ie, everything that runs before ExitBootServices() is called) being perfect. As I said, if any of them accept untrusted input and misinterpret it in such a way that they can be tricked into running arbitrary code, you'll still have problems. But when discussing the pros and cons of secure boot, it's important to make sure that we're talking about reality rather than making provably false assertions.

comments

Syndicated 2011-11-18 21:03:08 from Matthew Garrett

Attacks on secure boot

This is interesting. It's obviously lacking in details yet, but it does highlight one weakness of secure boot. The security for secure boot is all rooted in the firmware - there's no external measurement to validate that everything functioned as expected. That means that if you can cause any trusted component to execute arbitrary code then you've won. So, what reads arbitrary user data? The most obvious components are any driver that binds to user-controlled hardware, any filesystem driver that reads user-provided filesystems and any signed bootloader that reads user-configured data. A USB drive could potentially trigger a bug in the USB stack and run arbitrary code. A malformed FAT filesystem could potentially trigger a bug in the FAT driver and run arbitrary code. A malformed bootloader configuration file or kernel could potentially trigger a bug in the bootloader and run arbitrary code. It may even be possible to find bugs in the PE-COFF binary loader. And once you have the ability to run arbitrary code, you can replace all the EFI entry points and convince the OS that everything is fine anyway.

None of this should be surprising. Secure boot is predicated upon the firmware only executing trusted material until the OS handoff. If you can sidestep that restriction then the entire chain of trust falls down. We're talking about a large body of code that was written without the assumption that it would have to be resistant to sustained attack, and which has now been put in a context where people are actively trying to break it. Bugs are pretty inevitable. I'd expect a lot of work to be done on firmware implementations between now and Windows 8 ship date.

comments

Syndicated 2011-11-17 16:52:03 from Matthew Garrett

GPT disks in a BIOS world

Starting with Fedora 16 we're installing using GPT disklabels by default, even on BIOS-based systems. This is worth noting because most BIOSes have absolutely no idea what GPT is, which you'd think would create some problems. And, unsurprisingly, it does. Shock. But let's have an overview.

GPT, or GUID Partition Table, is part of the UEFI specification. It defines a partition table format that allows up to 128 partitions per disk, with 64 bit start and end values allowing partitions up to 9.4ZB (assuming 512 byte blocks). This is great, because the existing MBR partitioning format only allows up to 2.2TB when using 512 byte blocks. But most BIOSes (and most older operating systems) don't understand GPT, so plugging in a GPT-partitioned disk would result in the system believing that the drive was uninitialised. This is avoided by specifying a protective MBR. This is a valid MBR partition table with a single partition covering the entire disk (or the first 2.2TB of the disk if it's larger than that) and the partition type set to 0xee ("GPT Protective"). GPT-unaware BIOSes and operating systems will see a partition they don't understand and simply ignore it.

But how do we boot a GPT-labelled disk with a protective MBR on a system that doesn't understand GPT? The key here is that BIOS is pretty dumb. Typically a BIOS will see a disk and just attempt to execute the code in the first sector. This MBR code knows how to do the rest of the boot, including parsing the partition table if necessary. The BIOS doesn't need to care at all.

Of course, some BIOSes choose to care. We've seen a small number of machines that, when exposed to a GPT disk, refuse to boot because they parse the MBR partition map and don't like what they see. This is typically accompanied by a message along the lines of "No operating system found". What we've found is that they're looking for a partition marked with the bootable flag, and if no partitions are marked bootable they assume that there's no OS. This is in contrast to the traditional use of the flag, which is merely a hint to the MBR as to which partition boot code it should execute.

So, should we set that flag? The UEFI specification specifically forbids it - table 15 states that the BootIndicator byte must be set to 0. Once again we're left in an unfortunate position where the specification and reality collide in an awkward way.

If this happens to you after a Fedora 16 install, you have two choices. The first is to reinstall with the "nogpt" boot argument. The installer will then set up a traditional MBR partition table. The second is to boot off a live CD and run fdisk against the boot disk. It'll give a bunch of scary warnings. Ignore them. Hit "a", then "1", then "w" to write it to disk. Things ought to work then. We'll figure out something better for F17.

comments

Syndicated 2011-11-17 15:54:30 from Matthew Garrett

Making timeouts work with suspend

A reasonably common design for applications that want to run code at a specific time is something like:

time_t wakeup_time = get_next_event_time();
time_t now = time(NULL);
sleep(wakeup_time-now);

This works absolutely fine, except that sleep() ignores time spent with a suspended system. If you sleep(3600) and then immediately suspend for 45 minutes you'll wake up after 105 minutes, not 60. Which probably isn't what you want. If you want a timer that'll expire at a specific time (or immediately after resume if that time passed during suspend), use the POSIX timer family (timer_create, timer_settime and friends) with CLOCK_REALTIME. It's a signal-driven interface rather than a blocking one, so implementation may be a little more complicated, but it has the advantage of actually working.

comments

Syndicated 2011-11-17 13:51:47 from Matthew Garrett

Properly booting a Mac

This is mostly for my own reference, but since it might be useful to others:

By "Properly booting" I mean "Integrating into the boot system as well as Mac OS X does". The device should be visible from the boot picker menu and should be selectable as a startup disk. For this to happen the boot should be in HFS+ format and have the following files:

• /mach_kernel (can be empty)
• /System/Library/CoreServices/boot.efi (may be booted, if so should be a symlink to the actual bootloader)
• /System/Library/CoreServices/SystemVersion.plist which should look something like

  <xml version="1.0" encoding="UTF-8"?>
  <plist version="1.0">
  <dict>
          <key>ProductBuildVersion</key>
          <string></string>
          <key>ProductName</key>
          <string>Linux</string>
          <key>ProductVersion</key>
          <string>Fedora 16</string>
  </dict>
  </plist>

That's enough to get it to appear in the startup disk boot pane. Getting it in the boot picker requires it to be blessed. You probably also want a .VolumeIcon.icns in / in order to get an appropriate icon.

Now all I need is an aesthetically appealing boot loader.

comments

Syndicated 2011-11-09 22:06:07 from Matthew Garrett