Add ImagePropertiesRecordLib and Fix MAT Bugs

[TaylorBeebe]

The UEFI and SMM MAT logic contains duplicate logic for manipulating image properties records which is used to track runtime images. This patch series adds a new library, ImagePropertiesRecordLib, which consolidates this logic and fixes the bugs which currently exist in the MAT logic.

A writeup of the bug which initiated this patch series is below.

Bugzilla: https://bugzilla.tianocore.org/show_bug.cgi?id=4492

To create the MAT, we call into a function CoreGetMemoryMapWithSeparatedImageSection() which will get the memory map (and thus its required size) but update that size to be the size of the map plus the amount of extra space necessary to break up the memory map. The extra space in this case is indicated by a field called CodeSegmentMax which is set to be the most number of code segments found in a single image loaded. Let’s say we found an image with 2 code sections, making CodeSegmentMax = 2. With two code segments, here are a couple of ways an image could be laid out:

Image 1: Data | Code | Data | Code | Data
Image 2: Data | Code | Code | Data

So, if a single memory map descriptor is the same address range as a singe image, that one map descriptor could become at most 5, leading to the calculation (2 * CodeSegmentMax + 1) * NumberOfImageRecords.

After returning the required pool size (MemoryMapSize) the next call will be with a pool of that size. When we call into get the memory map, MemoryMapSize will be updated to be the actual size of the memory map and we make sure that size is what we expect before calling into SplitTable(). Within SplitTable(), MemoryMapSize is equal to the size of the actual memory map fetched, but the buffer containing the memory map is much larger as described above. We set IndexOld to be the last descriptor in the memory map fetched from CoreGetMemoryMap() and IndexNew to the last descriptor in the entire buffer. We work our way backwards through the descriptors generated from CoreGetMemoryMap() and write new descriptors to the end of the entire buffer. We call into GetMaxSplitRecord() which calculates how many images might be contained in a single memory map descriptor. Let’s say one descriptor range maps to one loaded image which has 2 code sections, so GetMaxSplitRecord() will return 4 (we subtract 1 because IndexNew is already pointing to a valid descriptor).

Let’s zoom into those 5 currently empty descriptors:

| | | | | |
^
IndexNew

Let’s say the code/data sections are laid out like Image 2 shown above. We call into SplitRecord() which will get the image covering the original descriptor range then call into SetNewRecord() to start populating the descriptors. There’s a stray portion of SplitRecord() which makes sure there isn’t any lingering memory that hasn’t been covered by one of the new descriptors which we can ignore in this case. In the case of microsoft/mu_basecore#2 above, we will write 4 descriptors but only return the value 3.

When we get back to SplitTable(), MaxSplitRecordCount = 4 and RealSplitRecordCount = 3 and those 5 descriptors look like this:

| d1 | c1 | c2 | d2 | |
^
IndexNew

We then do a mem copy with a destination of IndexNew + MaxSplitRecordCount – RealSplitRecordCount = IndexNew + 1 and a source of IndexNew and a length of 3 which would result in the following:

| d1 | d1 | c1 | c2 | |
^
IndexNew

And this is clearly not right because we have a duplicated the first data section and overwritten the last data section. Index 0 will be overwritten, but Index 4 is now a garbage memory descriptor which will be captured when we copy all the new records to the front of the buffer and update memory map size to be the new number of entries.