with Some References to Logging.
(This is a low-level follow-up to theses posts:
- How to Update Your Bootstrap Code/BIOS
- Some Comments on Basic Requirements and Principles of Hardware/OS Security
- Some Theorizing about How the Init Process Should Work after the Kernel Is Up
This is just a rough model, a rough ideal, not a specification. Real devices will tend to vary from this model. It's just presented as a framework for discussion, and possibly as a model to refer to when documenting real hardware.
(1) Simple ALU/CPU test.
The first thing the CPU should do on restart is check the Arithmetic-Logic Unit, not in the grand sense, but in a limited sense.
Something like (assuming an 8 bit binary ALU) adding 165 to 90 and checking that the result comes out 255 (A5sixteen + 5Asixteen == FFsixteen), and then adding 1 to the result to see if the result is 0 with a carry, would be a good, quick check. This would be roughly equivalent to trying to remember what day it is when you wake up, then checking to see that you remember what the day before and the day after are.
It doesn't tell you much, but it at least tells you that your brains are trying to work.
* If the ALU appears to give the wrong result, there likely won't be much that can be done -- maybe set a diagnostic flag and halt safely.
* In some devices, halting itself is not safe, and an alternative to simply halting such as having the device securely self-destruct may be safer. Halting safely may have non-obvious meanings.
Now, it's very likely that this test can be made a part of the next step, but we need to be conscious of it.
(2) Initial boot ROM test.
There should be an initial boot ROM that brings the CPU up. The size should be in the range of 1,000 instructions to 32,768 instructions.
Ideally, I would strongly suggest that it contain a bare-metal Forth interpreter as a debugger/monitor, but it may contain some other kind of debug/monitor. It may just contain a collection of simple Basic Input-Output library functions, but I personally do not recommend that. It needs to have some ability to interact with a technician.
And, of course, it contains the machine instructions to carry out the first several steps of the boot-up process.
This second step would then be to perform a simple, non-cryptographic checksum of the initial boot ROM.
Which means that the ROM contains its own test routines. This is clearly an example of chicken-and-egg logical circularity. It is therefore not very meaningful.
This is not the time for cryptographic checksums.
* Success does not mean that the CPU is secure or safe. Failure, on the other hand, gives us another opportunity to set a diagnostic flag for a technician to find, and halt safely, whatever halting safely means.
On modern, highly integrated CPUs, this ROM is a part of the physical CPU package. It should not be re-programmable in the field.
(That's one reason it should be small -- making it small helps reduce the chance for serious bugs that can't be fixed. This smallest part of the boot process cannot be safely re-written and cannot safely be allowed to be overridden.)
For all that it should not be re-programmable in the field, the source should be available to the end-administrator, and there should be some means of verifying that the executable object in the initial boot ROM matches the source that the vendor says should be there.
(3) Internal RAM check.
Most modern CPUs will have some package internal RAM, distinct from CPU registers. It is a good idea to check these RAM locations at this point, to see that what is written out can be read back, using bit patterns that can catch short and open circuits in the RAM.
Just enough RAM should be tested to see that the initial boot-up ROM routines can run safely. If the debug/monitor is a Forth interpreter, it should have enough return stack for at least 8 levels of nested call, 16 native integers on the parameter stack, and 8 native integers of per-user variable space. That's 32 cells of RAM, or room for 32 full address words, in non-Forth terminology.
(I'm speaking roughly, more complex integrated packages will need more than that, much more in some cases. Very simple devices might actually need only half that. The engineers should be able to determine actual numbers from their spec. If they can't, they should raise a management diagnostic flag and put the project in a wait state.)
* Again, if there are errors, there is not much we can do but set a diagnostic flag and do its best to halt safely, whatever halting safely means.
(4) Lowest level diagnostic firmware.
At this point, we can be moderately confident that the debug/monitor can safely be entered, so it should be entered and initialize itself.
The next several steps should run under the control of the debug/monitor.
* Again, if the debug/monitor fails to came up in a stable state, the device should set a diagnostic flag and halt itself as safely as possible.
** This means that the debug/monitor needs a resident watchdog cycle that will operate at this level.
(5) First test/diagnostic device.
We want a low-level serial I/O (port) device of high reliability, through which the technician can read error messages and interact with the debug/monitor.
(Parallel port could work, but it would usually be a waste of I/O pins for no real gain.)
* This is the last point where we want to just set a diagnostic flag and halt as safely as possible on error. Any dangerous side-effects of having started the debug port should be addressed before halting safely at this stage.
(6) Full test of CPU internal devices.
This step can be performed somewhat in parallel with the next step. Details are determined by the internal devices and the interface devices. Conceptually, however, this is a separate step.
All internal registers should be tested to the extent that it is safe to test them without starting external devices. This includes being able to write and read any segment base and limit/extent registers, but not does not actually include testing their functionality.
If the CPU provides automatic testing, this is probably the stage where it should be performed (which may require suspending or shutting down, then restarting the monitor/debug processes).
Watchdog timers should be checked to the extent possible and started during this step.
If there is internal low-level ROM that remains to be tested, or if management requires cryptographic checksum checks on the initial boot ROM, this is the stage to do those.
Note that the keys used here are not, repeat, not the manufacturer's update keys. Those are separate.
However, for all that management might require cryptographic self-checks at this stage, engineers should consider such checks to be exercising the CPU and looking for broken hardware, and not related to security. There should be a manufacturer's boot key, and the checksums should be performed with the manufacturer's boot key, since the initial boot ROM is the manufacturer's code.
How to hide the manufacturer's boot key should be specified in the design, but, if the test port enabled in step (5) allows technician input at this step, such efforts to hide the manufacturer's key can't really prevent attack, only discourage attack.
Even if the device has a proper system/user separation, the device is in system state right now, and the key has to be readable to be used.
The key could be encrypted and hidden, spread out in odd corners of the ROM. There could be two routines to read it, and the one generally accessible through the test port could be protected by security switch/strap and/or extra password. But the supervisor, by definition, allows the contents of ROM to be read and dumped through the test port at this stage. A determined engineer would be able to analyze the code and find the internal routine, and jump to it. Therefore, this raises the bar, but does not prevent access.
Another approach to raising the bar is the provision of a boundary between system/supervisor mode and key-access mode. The supervisor could use hardware to protect the key except when in key-access mode, and could use software to shut down the test port when key-access mode is entered. This would make it much more difficult to get access to supervisor commands while the key is readable, but there are probably going to be errors in the construction that allow some windows of opportunity. It is not guaranteed that every design will be able to close off all windows of opportunity.
Such efforts to protect the boot key may be useful. They do raise the bar. But they do not really protect the boot key, only discourage access.
And legal proscriptions such as that epitome of legal irony called DMCA do not prevent people who ignore the law from getting over the bar.
Thus, the key used to checksum the initial boot ROM must not be assumed to be unknown to attackers. (And, really, we really don't need to assume it is unknown, if we don't believe in fairy tales about protecting intellectual property at a distance. As long as this initial boot ROM can't be re-written. As long as the update keys are separate.)
The extra ROM, if it exists, should not be loaded yet, only tested.
If extra RAM is required to do the checksums, the RAM should be checked first, enough to perform the checksums
All remaining internal RAM should be checked at this stage.
(7) Low-level I/O subsystems.
Finally, the CPU package is ready to check its own fundamental address decode, data and address buffers, and so forth. Not regular I/O devices, but the devices that give it access to low-level flash ROM, cache, working RAM, and the I/O space, in that order.
They should be powered up and given rudimentary tests.
Note that the flash ROM, cache, working RAM, and I/O devices themselves should not yet be powered up, much less tested.
Only the interfaces are powered up and tested at this step, and they must be powered up in a state that keeps the devices on the other side powered down.
* On errors here, any devices enabled to the point of error should be powered down in whatever order is safe (often in reverse order of power-up), diagnostic messages should be sent through the diagnostic port, and the device should set a diagnostic flag and enter as safe a wait state as possible.
** It may be desirable to enter a loop that repeats the diagnostic messages.
It would seem to be desirable to provide some way for a technician to interrogate the device for active diagnostic messages.
** But security will usually demand that input on the diagnostic port be shut down unless a protected hardware switch or strap for this function is set to the enabled position/state. This is one of several such security switch/straps, and the diagnostic message will reflect the straps state to some extent.
This kind of security switch or strap is not perfect protection, but it is often sufficient, and is usually better than nothing. (Better than nothing if all involved understand it is not perfect, anyway.)
** In some cases, the security switch/straps should not exist at all, and attempts to find or force them should be met with the device's self-destruction. In other cases, lock and key are sufficient. In yet other cases, such as in-home appliance controllers, a screw panel may be sufficient, and the desired level of protection.
Straps are generally preferred to switches, to discourage uninformed users from playing with them.
*** However, attempts to protect the device from access by the device's legal owner or lawfully designated system administrator should always be considered highly suspect, and require a much higher level of engineering safety assurance. If the owner/end-admin user must be prevented from maintenance access, it should be assumed that the device simply cannot be maintained -- thus, quite possibly should self-destruct on failure.
(8) Supervisor, extended ROM, internal parameter store.
The initial boot ROM may actually be the bottom of a larger boot ROM, or there may be a separate boot ROM containing more program functions, such as low-level supervisor functions, to be loaded and used during initial boot up. This additional ROM firmware, if it exists, should be constructed to extend, but not replace the functionality in the initial boot ROM.
This extra initial boot ROM was tested in step (6), it should be possible to begin loading and executing things from it now. It would contain the extensions in stepped modules, starting modules necessary to support the bootstrap process as it proceeds.
Considering the early (classic) Macintosh, a megabyte of ROM should be able to provide a significant level of GUI interface for the supervisor, giving end-admins with primarily visual orientation improved ability to handle low-level administration. But we don't have display output at this point, such functionality should be oriented toward the technician's serial port at this stage.
This supervisor would also contain the basic input/output functionality, so it could be called, really, a true "Basic Input/Output Operating System" -- BIOOS. But that would be confusing, so let's not do that. Let's just call it a supervisor.
It could also contain "advanced" hooks and virtual OS support such as a "hypervisor" has, but we won't give in to the temptation to hype it. It's just a supervisor. And most of it will not be running yet.
This remaining initial boot ROM is not an extension boot ROM such as I describe below, but considered part of the initial boot ROM.
There should be internal persistent store that is separate from the extension boot (flash) ROM, to keep track of boot parameters such as the owner's cryptographic keys and the manufacturer's update cryptographic keys for checksumming the extension flash ROM, passwords, high-level boot device specification, etc. It should all be write protected under normal operation. The part containing the true cryptographic keys for the device and such must be both read- and write-protected under normal operation, preferably requiring a security switch/strap to enable write access.
Techniques for protecting these keys have been partially discussed above. The difference is that these are the owner's keys and update keys, and those are the manufacturer's boot keys.
This parameter store should be tested and brought up at this point.
Details such as how to protect it, how to enable access, and what to do on errors are determined by the engineers' design specification.
In the extreme analysis, physical access to a device means that anything it contains can be read and used. The engineering problem is the question of what kinds of cryptological attacks are expected, and how much effort should be expended to defend the device from unauthorized access.
Sales literature and such should never attempt to hide this fact, only assert the level to which they are attempting to raise the bar.
Again, attempts to protect the device from access by the legitimate owner/end-admin should be considered detrimental to the security of the device.
* At this point, reading the owner's keys and update keys from the test port should be protected by security switch/strap and password. But, again, until the boot process has proceeded far enough to be able to switch between system and user mode, the protections have to be assumed to be imperfect.
Providing a key-access mode such as described above for the manufacturer's key should mitigate the dangers and raise the bar to something reasonable for some applications, but not for all.
Some existing applications really should never be produced and sold as products.
(As an example, consider the "portable digital purse" in many cell phones. That is an abomination. Separated from the cell phone, it might be workable, but only with specially designed integrated packages, and only if the bank always keeps a copy of the real data. Full discussion of that is well beyond the scope of this rant.)
(9) Private cache.
If there is private cache RAM local to the first boot CPU, separate from the internal RAM, it should be tested now. Or it could be schedule and set to run mostly in a lesser privileged mode after lesser privileged modes are available.
If there are segment base and limit/extent registers, their functionality may be testable against the local cache.
In particular, if the stack register(s) have segment base and limit, and can be pointed into cache, it might be possible to test them and initialize the stacks into such cache here, providing some early stack separation.
If dedicated stack caches are provided in the hardware, they should be tested here. If they can be used in locked mode (no spills, deep enough), the supervisor should switch to them now.
* Errors at this point will be treated similarly to errors in step 7.
(10) Exit low-level boot and enter intermediate level boot process.
At this point, all resources owned by the boot-up CPU should have been tested.
Also, at this point, much of the work can and should be done in less secure modes of operation. The less time spent in system/supervisor mode, the better.
(10.1) Testing other CPUs.
If there are multiple CPUs, this is the step where they should be tested. The approach to testing the CPUs depends on their design, whether they share initial boot ROMs or are under management of the initial boot CPUs, etc.
From a functional point of view, it is useful if the first boot CPU can check the initial boot ROMs of the other CPUs before powering them up, if those ROMs are not shared. It may also be useful for the first boot CPU to initiate internal test routines on the others, and monitor their states as they complete.
At any rate, as much as possible should be done in parallel here, but care should be exercised to avoid one CPU invalidating the results of another.
* Again, errors at this point will be treated similarly to errors in step 7.
(10.2) Testing shared memory management hardware access, if it exists.
While waiting for the other CPUs to come up, any true memory management hardware should be tested and partially initialized.
At this point, only writing and reading registers should be tested, and enough initialization to allow un-mapped access.
* Again, errors at this point will be treated similarly to errors in step 7. MMU is pretty much vital, if it exists.
(10.3) Finding and testing shared RAM.
Shared main RAM should be searched for before shared cache.
As other CPUs come up, they can be allocated to test shared main RAM. (Really, modern designs should go to multiple CPUs before going to larger address space or faster CPUs, any more.) If there are multiple CPUs, testing RAM should be delegated to CPUs other than the first boot CPU.
This also gets tangled up in testing MMU.
Tests should be run first without address translation, then spot-checked with address translation.
As soon as enough good RAM has been found to support the return address stack and local variable store (one stack in the common case now, but preferably two in the future, a thread heap and a process heap) the supervisor OS, to the extent it exists, should be started now if it has not already been started. (See next step.)
Otherwise, parallel checks on RAM should proceed without OS support.
Either way, the boot ROM should support checking RAM in the background as long as the device is operational. RAM which is currently allocated would be left alone, and RAM which is not currently allocated would have test patterns written to them and read, helping erase data that programs leave behind.
Such concurrent RAM testing would be provided in the supervisor in the initial boot up ROM, but should run in a privilege-reduced state (user mode instead of system/supervisor).
* Usually, errors in RAM can be treated by slowing physical banks down until they work without errors, or by mapping physical banks out. Again, a log of such errors must be kept, and any errors in RAM should initiate a RAM checking process that will continue in the background as long as the device is running.
** If there are too many errors at this point, they may be treated similarly to errors in step 7.
*** Any logs kept in local RAM should be transferred to main RAM once enough main RAM is available (and known good).
(10.4) Testing shared cache.
As other CPUs come up, they can also be allocating to testing shared cache. As with testing main RAM, testing cache should be delegated to CPUs other than the first boot CPU. Also, main RAM comes before cache until there is enough known-good RAM to properly support multiple supervisor processes.
And this also gets tangled up in testing MMU.
Tests should be run first without being assigned to RAM, then again with RAM assigned.
* If there are errors in the cache, it might be okay to disable or partially disable the cache. Engineers must make such decisions.
** Errors at this point errors may still be treated similarly to errors in step 7, depending on engineering decisions. If it is acceptable to run with limited cache, or without cache, some logging mechanism that details the availability of cache must be set up. Such logging would be temporarily kept in internal RAM.
*** The decision about when to enable cache is something of an engineering decision, but, in many cases, once cache is known to be functional, and main RAM has also been verified, the cache can be put into operation.
In some designs, caches should not be assigned to RAM that is still being tested.
(11) Fully operational supervisor.
At this point, most of the remaining functionality of the supervisor (other than GUI and other high-level I/O) should be made available. Multi-tasking and multi-processing would both be supported (started in the previous step), with process management and memory allocation.
One additional function may become available at this point -- extending the supervisor via ROM or flash ROM.
If there is an extension ROM, the initial boot ROM knows where it is. If it is supposed to exist, the checksum should be calculated and confirmed at this point.
The key to use depends on whether the extension has been provided by the manufacturer or the end-user/owner. Manufacturer's updates should be checked with the update key (not the boot key), and owner's extensions should be checked with the owner's key.
Failure would result in a state such as in step (7).
Testing the extension proceeds as follows:
There are at least two banks of flash ROM. In the two bank configuration, one is a shadow bank and the other is an operational bank.
If the checksum of the operational bank is the same as the unwritable extension ROM, the contents are compared. If they are different, the operational bank is not loaded, and the error is logged and potentially displayed on console.
If the checksum of the operational bank is different from the unwritable ROM, it is checked against the shadow bank. If the shadow bank and the operational bank have the same checksum, the contents of the two are compared. If the contents are different, the operational bank is not loaded and the error is logged and potentially displayed on console.
If the contents are identical, the cryptographic checksum is checked for validity. If it is not valid, the operational bank is not loaded, and the error is logged and potentially displayed on console.
* If the operational bank verifies, it is loaded and boot proceeds.
** If the operational bank fails to verify, a flag in the boot parameters determines whether to continue or to drop into to a maintenance mode.
If the device drops into a maintenance mode, the test port becomes active, and a request for admin password is sent out it. A flag is set, and boot proceeds in a safe mode, to bring up I/O devices safely.
(When the operational bank is updated, the checksum checked and verified, and committed, the operational bank is copied directly onto the shadow bank. But that discussion is not part of this rant.)
Other approaches can be taken to maintain a valid supervisor. For instance, two shadow copies can be kept to avoid having to restore the factory extensions and go through the update process again from scratch.
The extensions can override much of the initial boot ROM, but the monitor/debugger must never be overridden. It can be extended in some ways, but it must not be overridden.
There should be no way to write to this flash ROM except by setting another protected hardware switch or strap which physically controls the write-protect circuit for the flash. This switch or strap should not be the same as mentioned in step (7), but may be physically adjacent to it, depending on the engineers' assessment of threat.
*** The initial boot ROM should not proceed to the flash ROM extensions unless said switches or straps are unset.
(12) I/O devices.
(12.1) Locating and testing normal I/O device controllers.
As known good main RAM becomes available, the boot process can shift to locating the controllers for normal I/O devices such as network controllers, rotating disk controllers, flash RAM controllers, keyboards, printers, etc.
There may be some priority to be observed when testing normal I/O device controllers, as to which to initiate first.
It also may be possible to initiate controller self-tests or allocate another CPU to test the controllers, so that locating the controllers and testing them can be done somewhat it parallel.
Timers and other such hardware resources would be more fully enabled at this point.
* Errors for most controllers should be logged, and should not cause the processor to halt.
(12.2) Identifying and testing devices.
As controllers become available and known good, the devices attached to them should be identified, initialized, and tested.
This might also occur in parallel with finding and testing other controllers.
* Errors for most devices should be logged, and should not cause the processor to halt.
** Some intelligence about the form and number of logs taken at this point can and should be exercised. We don't want RAM filled with messages that, for example, the network is unavailable. One message showing when problems began, and a count of error events, with a record of the last error, should be sufficient for most such errors.
(12.3) Low-level boot logging.
As video output and persistent store become available, error events should be displayed on screen and recorded in an error message partition. Again, there should be a strategy to avoid filling the error message partition, and to allow as many error notifications as possible to remain on screen.
If the device is booting to maintenance mode, and an admin has not logged in via the test port at this point, the video device may present a console login prompt/window, as well. Or it may present one for other reasons, such as a console request from the keyboard.
The video display could also have scrolling windows showing current system logs.
Also, parameter RAM flags may prevent console login to a local video device/keyboard pair, requiring admin login at the test port via some serial terminal device.
(12.4) High-level boot.
The supervisor would have hooks and APIs to present walled virtual run-time sessions to high-level OSses, including walled instances of itself and walled instances of Linux or BSD OSses, or Plan Nine, etc., to the extent the CPU can support such things, and to the extent the device is designed to support such things.
And parameter RAM would have flags to indicate whether a boot menu should be provided, or which high-level OSses available should boot.
If walled instances are not supported, only a single high level OS would be allowed to boot, and the supervisor would still map system calls from the high-level OS into device resources.
This is my idea of what should happen in the boot-up process. Unfortunately, most computers I am familiar with do a lot of other stuff and not enough of this.