885
edits
Changes
Add kernel diff
The anti-downgrade fuses were [[Fuses#Anti-downgrade|updated]].
The anti-downgrade fuses were [[Fuses#Anti-downgrade|updated]].
==== Kernel ====
* Compiler/libc changes:
** The kernel is now linked using RELR for relocations instead of RELA (see compiler support in lld for .relr.dyn).
** This greatly reduces the relocations segment size; it has decreased from 0x3A50 bytes in 16.0.0 to 0x90 in 17.0.0.
** Many minor optimization changes, e.g. mul+add -> madd, madd -> smaddl/umaddl, (a + b - 1) >> 36 is now (a + b) > 0x1000000000, various reordering.
* crt0 changes:
** crt0 is no longer located at _start, instead _start is `b crt0` followed by 0x7FC of zeroes.
** crt0 is now located at the start of .rodata.
*** The crt0 page is now identity-mapped R-X in .rodata instead of RWX at start-of-text.
** Many system registers which were previously set from KInitArguments are now set using a register constants table in the crt0 .rodata segment.
*** These are ttbr0_el1, ttbr1_el1, tcr_el1, mair_el1, and sctlr_el1.
*** This table is initially zeroes, and is initialized to the correct values by KernelLdr before returning to Kernel/setting permissions.
** Kernel Map now stores offsets relative to itself rather than relative to _start.
*** Kernel map also now stores an additional offset (to the "register constants").
** The big idea here is to make the crt0 page no longer executable after init.
*** This mitigates the ability to execute gadgets (via ROP/etc) to set TTBR1_EL1 (and other important registers) to user-controlled values.
**** The *only* ttbr1_el1 gadget in all of kernel now sets it to the constant in .rodata, which can't be modified after KernelLdr finishes.
*** This also enables setting the WXN bit while still identity-mapped, instead of having to do it later in boot.
* KernelLdr changes:
** INI1 is now used in-place, if KSystemControl does not have a preferred layout.
* Initialize0 changes:
** Initialize0 now receives the initial process binary size from KernelLdr and stores it in a global.
*** Initialize1 forwards this to the rest of the kernel as with the address.
** Initialize0 no longer memsets the slab region to zero before calling the ifdef'd out function for the unknown debug region.
*** This is now done by InitializeSlabHeaps().
* All exception returns now migitate post-eret speculative execution.
** All "eret" instructions are now "eret; dsb nsh; isb;"
* KInitialPageAllocator::Allocate(Aligned) now memsets the pages to zero before returning them to the caller.
** Correspondingly, KInitialPageTable no longer memsets those pages to zero after allocating them.
* KInitialProcessReader::CreateProcessParameter now ands sizes with 0x1FFFFF000 before overflow checking.
** This may actually just be compiler-garbage due to the types being u32-cast-to-higher-width?
* CreateAndStartInitialProcesses changes:
** A difference check is now an != when allocating page group.
** Segment loading/uncompressing has now been refactored:
*** The entire page group is no longer mapped while loading the segments.
*** KInitialProcessReader::Load is now responsible; it now takes the page group as argument, clears bss (using linear map), and then calls a helper to load each segment.
**** This helper creates a page group for just the pages relevant to the segment, copies the data (using linear map), and then if compressed maps the page group, uncompresses, and unmaps.
* KMemoryRegionType had a number of large changes:
** A new memory type is now inserted after the SecureAppletMemory region (id is 0xC200028E).
** Low 0x2 ID derivations changed to accommodate this.
** As a knock-on effect(?) type IDs for pool partitions changed substantially (likely due to derivation changes elsewhere).
* New KProcess field ("has application system resource").
** This is set to 1 when initializing a KProcess with CreateProcessFlag_IsApplication and system_resource_num_pages == 0.
** When this is true, svc::GetInfo() always returns 0 for InfoType_SystemResourceSizeTotal and InfoType_SystemResourceSizeUsed.
*** This also modifies the calculations for various SystemResourceSize calculations.
*** MapPhysicalMemory() and UnmapPhysicalMemory() will also now return svc::ResultInvalidState().
* The KProcess::Initialize() overload used by initial processes now supports system_resource_num_pages != 0 (and allocates a system resource in this case).
** NOTE: KInitialProcessReader::CreateProcessParameter still hardcodes param->system_resource_num_pages = 0 for all KIPs.
* Changes to KPageTable(Base) around KMemoryState:
** There is no longer a bijective mapping between svc::MemoryState and kern::KMemoryState.
** In particular, KMemoryState_Io has been split into two memory states:
*** KMemoryState_Io(Register) no longer has bit 13 (0x2000) set (new value is 0x180001).
*** For memory mapped with SvcMapIoRegion called with svc::MemoryMapping_Memory, KMemoryState_Io(Memory) retains that bit set (value is 0x182001).
*** KPageTableBase functions dealing with Io mappings now take in MemoryState arguments, and/or MemoryMapping arguments (for the IoRegion functions).
** KMemoryState_ThreadLocal no longer has bit 13 (0x2000) set (new value is 0x400000C).
** KMemoryState_Kernel no longer has bit 13 (0x2000) set (new value is 0x13).
** KMemoryState_Static no longer has bit 13 (0x2000) set (new value is 0x40002).
** KMemoryState_Insecure now supports FlagCanQueryPhysical (new value is 0x55C3817).
** To accommodate this, KPageTableBase::QueryMapping/Contains/GetRegionAddress/GetRegionSize now take an svc::MemoryState (u8) instead of the full KMemoryState.
*** In a (presumably) happy accident, this produces much, much better assembly for the switch statement.
** KPageTableBase::CheckMemoryState was made ALWAYS_INLINE and now calls an impl-func which takes KMemoryBlock * as argument.
* KPageTableBase::MapPageGroup no longer sets the io bit in page properties.
** This is the overload used by process creation.
* KMemoryBlockManager::UpdateIfMatch now takes set_disable_attr, clear_disable_attr.
** KPageTableBase::MapPhysicalMemory passes true for set_disable_attr if the address is exactly the start of the alias region.
* KPageTableBase::UnmapPhysicalMemory now passes clear_disable_attr = 1 to KMemoryManager::Update if the address is exactly the start of the alias region.
* KProcessPageTable::Initialize no longer has an unused truncated-process-id argument.
* Changes to KPageTable(Base) mapping for first-reference:
** KPageTable::Operate is no longer allowed to take MapFirst as operation.
** KPageTable::MapContiguousWithBaseAttribute no longer supports not_first argument, always calls OpenAdditionalReference.
** KPageTable::OperateOnPageGroup is now allowed to take MapFirst as operation, and MapWithPageGroup can now call OpenFirst or OpenAdditional for page group references.
** KPageTableBase::AllocatePageGroupAndOperate now passes MapFirst.
* Miscellaneous page table changes:
** KSupervisorPageTable::Initialize now checks that the WXN bit is set in sctlr_el1 instead of setting it.
** KPageTable::Finalize now calls a second OnFinalize() stub after NoteUpdated().
* KPageTableBase::MapStatic alignment checks were loosened/changed.
* New KMemoryAttribute bit 0x10 ("PermissionLocked").
** This can be set via SvcSetMemoryAttribute.
*** NOTE: Once set, this bit is irrevocable and can never be unset.
**** This is to enable relro (read only relocations).
*** This requires a new KMemoryStateFlag (bit 27) "FlagCanPermissionLock", which is set only on CodeData and AliasCodeData.
** KPageTable::SetMemoryAttribute now calls a new KMemoryBlockManager::UpdateAttributes function specifically for updating the attributes.
** This bit is allowed to be set when unmapped CodeMemory (as it can be set on (Alias)CodeData).
* HandleException now uses UserspaceAccess functions to retrieve the instruction when EsrEc is Unknown, IllegalState, Bkpt, or Brk.
* InvalidateProcessDataCache now special-cases being called on the current process, with a simpler (new) KPageTableBase function.
* Changes around signaling/thread termination.
** KThread::BeginTerminate no longer calls NotifyAvailable on the thread.
** KThread::DoWorkerTask now acquires the scheduler lock and calls NotifyAvailable on the thread.
** KThread and KProcess exit now use separate KWorkerTaskManagers (0 = Thread, 1 = Process).
*** Main() now initializes the two KWorkerTaskManagers, and now aborts if their priorities (both constant 11) are zero.
* KSleepManager's no longer saves and restores tcr_el1 when saving/restoring system registers.
==See Also==
==See Also==