Shrink measured size #15

krystian-hebel · 2024-02-22T21:33:07Z

This set of changes makes the measured size of SKL smaller by making all the non-const data built at runtime and moving sections around. Size changes, calculated as offset to bootloader_data symbol:

for 64b build, change from 0xe058 to 0xd4d8, new measured size: 0x2f4a
for 32b build, change from 0x705a to 0x7480, new measured size: 0x30de

Increase in size for 32b is caused mostly due to measured size being just above multiply of 4K in size, and next variable is 4K-aligned.

Not tested on hardware, I'm planning to do so after switching from current tags format to SLRT.

At least GNU ld 2.30 and earlier fail to discard the generic part of .got.plt when no actual entries were allocated. As this section can't be discarded, linker script checks if section's size is what we expect it to be (either empty or 3 * sizeof(long)) and fails otherwise. .got.plt is removed by objcopy when final binary is built. Signed-off-by: Krystian Hebel <[email protected]>

andyhhp · 2024-02-23T02:09:06Z

head.S

+	add	%edx, %ebx
+	add	$MAX_STACK_SIZE, %ebx
+	cmp	%ebx, %esp
+	jge	.Lenough_stack


So you can actually use the fact that the SKL size is 64k to your advantage here:

movzwl booloader_data+2(%ebp), %eax add $MAX_STACK_SIZE, %ax jno .Lenough_stack

ought to be equivalent, I think ?

booloader_data+2(%ebp) has just the size from booloader_data to the end of that data, not from start of SLB, so booloader_data offset would still have to be added to that.

Using halfword trick is nice, but it won't matter for SLRT because it's size field is u32.

Modified in 65e9d93 in next PR

dpsmith · 2024-02-23T12:23:40Z

link.lds

+
+	/*
+	 * This section is expected to be empty. To keep amount of data sent to the
+	 * TPM by SINIT instruction low, all data is either read-only (.rodata), or


nit: s/SINIT/SKINIT/

Fixed in https://github.com/TrenchBoot/secure-kernel-loader/compare/412bd740fd2d1b043ae6ad8a47726fdb7362848f..602004f3fab90a9502d60ef7b0ccae528c20b30f

iommu.c

link.lds

event_log.c

tpmlib/crb.c

Big, sparsely filled structures don't have to be in measured part of SLB. This change moves them to .bss where they are filled at runtime, after .bss is cleared. Note that pagetables are not moved yet, it will be done in following commit. Signed-off-by: Krystian Hebel <[email protected]>

Pagetables take 7*4K = 28K of space. Building them at runtime means that they won't have to be measured, as long as the code creating them is measured and proper memory protections are in place. Signed-off-by: Krystian Hebel <[email protected]>

To keep amount of data sent to the TPM by SKINIT instruction low, .data section is made empty by moving its contents which is either turned read-only (.rodata), or uninitialized and built at run time (.bss). Linker script asserts that there is no leftover .data. Signed-off-by: Krystian Hebel <[email protected]>

Stack area doesn't need to be measured by SKINIT. Initial %esp always points to the end of SLB, there is no reason to move it from there. Minimal stack size is checked by the code as the amount of available space depends on size of SLRT passed by the bootloader. Signed-off-by: Krystian Hebel <[email protected]>

krystian-hebel · 2024-04-09T15:24:57Z

Loosely related to the commit that moves non-RO data to .bss: https://nvd.nist.gov/vuln/detail/CVE-2017-16837. To be fair, similar vulnerability doesn't exist in the current code as all data was measured, including .bss, stack etc.

With reordered sections and most of the data structures built at runtime, size of actual code that has to be measured by TPM can be reduced. As that size field used to specify offset to bootloader data, a new field was added to the header for that purpose. Signed-off-by: Krystian Hebel <[email protected]>

Far return is used, so segment selector and relocated instruction pointer are build on the stack, instead of modifying the code. Signed-off-by: Krystian Hebel <[email protected]>

pm_kernel_entry and zero_page were renamed to dlme_entry and dlme_arg, respectively. Their previous names were valid only for Linux. Signed-off-by: Krystian Hebel <[email protected]>

Linux, Multiboot2 and simple payload don't have overlapping uses for values passed through registers or stack. All of those can be set at the same time. Signed-off-by: Krystian Hebel <[email protected]>

Signed-off-by: Krystian Hebel <[email protected]>

SKL hashes are no longer passed from the bootloader. Measured code and read-only data are not changed, and all other data is created at runtime so SKL is able to measure itself, which it now does. Those hashes are only used for event log entries, PCR was extended as a result of SKINIT instruction on Dynamic Launch Event. Signed-off-by: Krystian Hebel <[email protected]>

This removes different handling between boot protocols. Everything is reduced to one range of consecutive memory for DLME, as well as pointer to the entry point, which is measured to PCR 17 as an offset from DLME base to avoid attacks based on jumping to different parts of DLME. Signed-off-by: Krystian Hebel <[email protected]>

Signed-off-by: Krystian Hebel <[email protected]>

MULTIBOOT2_BOOTLOADER_MAGIC has been moved to defs.h as it is still used in head.S. Signed-off-by: Krystian Hebel <[email protected]>

It used to be defined in boot.h, even though it isn't expected to be ever needed anywhere but in code for handling event log internally. Signed-off-by: Krystian Hebel <[email protected]>

Test for INVALIDATE_ALL completion is now done inside the initialization function, instead of main.c. Function names were changed to better describe what the function does. Comment describing window in anti-DMA security was updated with new finds. Some bugs were fixed: - EventLogInt wasn't properly cleared (it is write-1-to-clear bit) - CmdBufEn was set on transition to the kernel, which didn't clear it before setting up head/tail pointers, which in turn lead to hang (it is listed as undefined behavior in IOMMU specification) - Command buffer tail/head pointer registers were incorrectly masked Signed-off-by: Krystian Hebel <[email protected]>

Release versions of SKL hanged in tis_send(), while versions with serial output worked, most likely thanks to increased delay between DRTM sequence and next TPM command. Apparently, some TPMs don't properly handle setting STS.commandReady when all of the following conditions are met: - TPM has recently finished DRTM sequence (internal work may still be happening at that point, there is no way to be sure), - TPM just transitioned to Idle state (e.g. by changing locality), - STS.commandReady is written periodically before TPM reports it is in Ready state. When all of the above applies, STS.commandReady is always read as 0, as if the TPM restarted transition from Idle to Ready each time it is asked to do so. To work around this, set this bit once and keep checking in a loop until TIMEOUT_B (2 seconds). Well behaving TPM must be able to enter Ready state before that time, if it doesn't, error is returned. Signed-off-by: Krystian Hebel <[email protected]>

This enables SHA tests to check if different compilers produce the same results. Definitions in string.h were moved outside of __STDC_HOSTED__ to avoid compiler warnings (promoted to errors because of -Werror) due to incompatible implicit declaration of built-in functions. Signed-off-by: Krystian Hebel <[email protected]>

-fstrict-aliasing is enabled by default when using optimization levels higher than 1, including -Os. With that option, compiler may assume that object of one type never resides at the same address as object of a different type. Both sha1_final() and sha256_final() used to write message length by casting a pointer to buffer into a pointer to u64, while surrounding code operated on the buffer directly. The problem manifests in GCC 11 and later versions. The commit fixes this issue and another UB caused by unaligned access at the beginning of transformation step by using memcpy() in both cases. Signed-off-by: Krystian Hebel <[email protected]> Signed-off-by: Sergii Dmytruk <[email protected]>

macpijan · 2024-04-17T17:49:46Z

Integrated both #17 and #19 here for AEM milestone release.

The review can still happen in individual (even closed) MRs if that makes it easier to review, and we will happily consider any comments.

…ssors" This reverts commit 335c103. Without -march=btver2 SKL can be used on AMD 15h or earlier and now that SKL more than fits into 64KiB limit, no need for such options to save a bit of space. Signed-off-by: Sergii Dmytruk <[email protected]>

DEV is specific to a northbridge, so there are multiple of them on systems with several CPU sockets. 0x18 is the device number of the first one with the rest using successive numbers (0x19, 0x1A, etc.). Signed-off-by: Sergii Dmytruk <[email protected]>

rossphilipson · 2024-04-18T17:55:28Z

Just FYI, I have modified the SLRT to make everything 8b aligned in the SLRT in Linux. The PR to do this was pulled into the linux-sl-6.7 working branch. I am also adding a TXT heap address to the Intel info table. This should go up as a PR soon.

krystian-hebel · 2024-08-07T14:45:57Z

Just FYI, I have modified the SLRT to make everything 8b aligned in the SLRT in Linux. The PR to do this was pulled into the linux-sl-6.7 working branch. I am also adding a TXT heap address to the Intel info table. This should go up as a PR soon.

Should be addressed by #21

This results in 64-bit alignment of 64-bit fields. Signed-off-by: Sergii Dmytruk <[email protected]>

rossphilipson

It makes sense to move the work from shrink_measured_size into master. We used the shrink_measured_size branch when working on the latest rebase of the AMD work on the internal PSP work we did. There are changes on top of these that will need to be reconciled once we can merge the PSP/DRTM bits for AMD. But for now, I think this LGTM:

Reviewed-by: Ross Philipson [email protected]

krystian-hebel requested review from andyhhp, rossphilipson and SergiiDmytruk February 22, 2024 21:33

krystian-hebel mentioned this pull request Feb 22, 2024

Update TrenchBoot boot protocol for AMD in Secure Kernel Loader TrenchBoot/trenchboot-issues#22

Closed

andyhhp reviewed Feb 23, 2024

View reviewed changes

krystian-hebel force-pushed the shrink_measured_size branch from 3b4f2ca to e2f93ca Compare February 23, 2024 11:39

dpsmith reviewed Feb 23, 2024

View reviewed changes

krystian-hebel force-pushed the shrink_measured_size branch 2 times, most recently from 602004f to 1019f11 Compare February 27, 2024 19:12

SergiiDmytruk reviewed Mar 11, 2024

View reviewed changes

iommu.c Outdated Show resolved Hide resolved

link.lds Outdated Show resolved Hide resolved

event_log.c Show resolved Hide resolved

tpmlib/crb.c Show resolved Hide resolved

krystian-hebel force-pushed the shrink_measured_size branch from 1019f11 to cb2b08a Compare March 26, 2024 15:19

krystian-hebel added 4 commits April 8, 2024 18:24

head: build pagetables at runtime

bac993a

Pagetables take 7*4K = 28K of space. Building them at runtime means that they won't have to be measured, as long as the code creating them is measured and proper memory protections are in place. Signed-off-by: Krystian Hebel <[email protected]>

krystian-hebel force-pushed the shrink_measured_size branch from cb2b08a to d8ac246 Compare April 8, 2024 16:24

SergiiDmytruk approved these changes Apr 8, 2024

View reviewed changes

krystian-hebel added 2 commits April 10, 2024 19:09

head: don't use self-modifying code to switch to 64b mode

67853db

Far return is used, so segment selector and relocated instruction pointer are build on the stack, instead of modifying the code. Signed-off-by: Krystian Hebel <[email protected]>

krystian-hebel force-pushed the shrink_measured_size branch from d8ac246 to 67853db Compare April 10, 2024 17:09

krystian-hebel added 8 commits April 10, 2024 20:58

main: rename variables

729d480

pm_kernel_entry and zero_page were renamed to dlme_entry and dlme_arg, respectively. Their previous names were valid only for Linux. Signed-off-by: Krystian Hebel <[email protected]>

head: don't test for boot protocol, use hand-off that works for all

08ddd83

Linux, Multiboot2 and simple payload don't have overlapping uses for values passed through registers or stack. All of those can be set at the same time. Signed-off-by: Krystian Hebel <[email protected]>

include/slrt.h: new file with definitions for SLRT

e36b5ae

Signed-off-by: Krystian Hebel <[email protected]>

main: remove leftover functions specific to Linux and Multiboot2

a74b5ab

Signed-off-by: Krystian Hebel <[email protected]>

include: remove obsolete files (linux-bootparams, multiboot2, tags)

fc97d74

MULTIBOOT2_BOOTLOADER_MAGIC has been moved to defs.h as it is still used in head.S. Signed-off-by: Krystian Hebel <[email protected]>

event_log: move definition of ev_log_hash_t to code

8c4fea6

It used to be defined in boot.h, even though it isn't expected to be ever needed anywhere but in code for handling event log internally. Signed-off-by: Krystian Hebel <[email protected]>

krystian-hebel added 4 commits April 16, 2024 14:45

SergiiDmytruk added 2 commits April 17, 2024 21:13

SergiiDmytruk mentioned this pull request Aug 7, 2024

Update GRUB for SL specification and push changes to TrenchBoot GRUB repository TrenchBoot/trenchboot-issues#32

Closed

include/slrt.h: update sizes and order of fields

ac26ede

This results in 64-bit alignment of 64-bit fields. Signed-off-by: Sergii Dmytruk <[email protected]>

rossphilipson approved these changes Aug 12, 2024

View reviewed changes

rossphilipson merged commit 1cc0a75 into master Aug 12, 2024
56 of 57 checks passed

SergiiDmytruk deleted the shrink_measured_size branch August 12, 2024 17:58

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Shrink measured size #15

Shrink measured size #15

krystian-hebel commented Feb 22, 2024 •

edited

Loading

andyhhp Feb 23, 2024 •

edited

Loading

krystian-hebel Feb 23, 2024

krystian-hebel Feb 28, 2024 •

edited

Loading

dpsmith Feb 23, 2024

krystian-hebel Feb 28, 2024

krystian-hebel commented Apr 9, 2024

macpijan commented Apr 17, 2024 •

edited

Loading

rossphilipson commented Apr 18, 2024

krystian-hebel commented Aug 7, 2024

rossphilipson left a comment

Shrink measured size #15

Shrink measured size #15

Conversation

krystian-hebel commented Feb 22, 2024 • edited Loading

andyhhp Feb 23, 2024 • edited Loading

Choose a reason for hiding this comment

krystian-hebel Feb 23, 2024

Choose a reason for hiding this comment

krystian-hebel Feb 28, 2024 • edited Loading

Choose a reason for hiding this comment

dpsmith Feb 23, 2024

Choose a reason for hiding this comment

krystian-hebel Feb 28, 2024

Choose a reason for hiding this comment

krystian-hebel commented Apr 9, 2024

macpijan commented Apr 17, 2024 • edited Loading

rossphilipson commented Apr 18, 2024

krystian-hebel commented Aug 7, 2024

rossphilipson left a comment

Choose a reason for hiding this comment

krystian-hebel commented Feb 22, 2024 •

edited

Loading

andyhhp Feb 23, 2024 •

edited

Loading

krystian-hebel Feb 28, 2024 •

edited

Loading

macpijan commented Apr 17, 2024 •

edited

Loading