On 03/06/18 01:45, Brian J. Johnson wrote:
> On 03/05/2018 12:22 PM, Laszlo Ersek wrote:
>> PEIMs generally "execute in place" (XIP), i.e. they run from flash, not
>> RAM. In this status they use "temporary RAM" (e.g. CPU caches configured
>> like RAM) for stack & heap; whatever HOBs they produce are stored in
>> "temp RAM" as well. Then one of the PEIMs "discovers permanent RAM"
>> (basically it programs the memory controller and publishes the RAM
>> ranges). In turn the PEI core "migrates" PEIMs from temporary to
>> permanent RAM, including the HOB list.
>> Before the temporary RAM migration (when still executing in place from
>> flash), PEIMs cannot have writeable global variables. For example,
>> dynamic PCDs are also maintained in a HOB (the PCD HOB).
>> A PEIM normally cannot (and shouldn't) tell whether it is dispatched
>> before or after permanent RAM is published. If needed, a PEIM can
>> advertise that it depends on permanent RAM (for example because it needs
>> a lot of heap memory) by including "gEfiPeiMemoryDiscoveredPpiGuid" in
>> its DEPEX.
>> Finally, it seems like a PEIM can also express, "I'm fine with being
>> dispatched from both flash (XIP) vs. permanent RAM, just the PEI core
>> tell me whichever it is". Apparently, if the PEIM is dispatched from
>> flash (before permanent RAM is available), its call to
>> RegisterForShadow() returns EFI_SUCCESS (and then its entry point
>> function will be invoked for a 2nd time, after the temp RAM migration).
>> And when a PEIM is dispatched from RAM (either for the very first time,
>> or for the second time, after being dispatched from flash first), the
>> same call returns EFI_ALREADY_STARTED.
>> Honestly, I'm unsure what this is good for (both in general, and
>> specifically for Tcg2Pei). Apparently, Tcg2Pei needs permanent RAM for
>> doing the measurements (which makes sense); I just wonder what exactly
>> it does so that it cannot simply specify
>> "gEfiPeiMemoryDiscoveredPpiGuid" in its DEPEX.
> I haven't looked at this particular PEIM. But one case where
> registering for shadowing is useful is for improving performance when
> running from permanent RAM, where writable global variables are
> available. For instance, when running from flash, a ReportStatusCode
> PEIM may need to go through a slow process to locate an output hardware
> device on every PPI call. This may involve traversing the HOB list,
> consulting other PPIs, even probing hardware addresses. But once it's
> shadowed to RAM, it can locate the device once, then cache its address
> in a global. Not to mention that the code itself is far, far faster
> when run from RAM vs. flash. (That's probably a key difference between
> a real machine and a VM.)
Ah, this sounds like a great example. Status code reporting is useful /
important for debugging even before permanent RAM is installed, so the
PEIM providing that PPI should not have a depex on
gEfiPeiMemoryDiscoveredPpiGuid. However, once the permanent RAM is
published, it makes sense to speed up the implementation. Thanks for
> Also, I've personally written a PEIM which saves a bunch of internal
> state in a HOB, since that's the main writable storage when running from
> flash. That state includes pointers to other data (in flash.) Once the
> data is all shadowed to RAM, it updates the HOB to point to the data in
> RAM. That way it's a lot faster to access the data.
Another good example. (I think I've been generally blind to this perf
difference between flash and RAM, specifically concerning execution -- I
must have been spoiled by virt, as you say :) )
> I also have other PEIMs which are constrained (via DEPEX) to run *only*
> from RAM, since they have larger memory requirements than can be
> satisfied from temporary cache-as-RAM. That's certainly a valid
> technique as well.
Right, that's quite known to OVMF too, due to CpuMpPei being "hungry"
> RegisterForShadow() is a useful tool for making the most of the
> restricted PEI environment. And having it standardized like this is, as
> Andrew said, a lot better than the hacks people had to use beforehand.
I agree. I'm happy to have learned about this service!
edk2-devel mailing list