Secure Kernel

From Vita Development Wiki

Entrypoint

The entrypoint of secure_kernel is +0x100. First thing it does is disable irq. Then it zeroes the bss segment 8 bytes at a time.

It sets up $sp to 0x808FF0, and $gp to 0x80F8A8. Then it sets dbg::TraceEnableFlag to *0xE005003C.

In the dump trace was enabled, and buf_ptr was 0x4002C160.

Then it sets up cfg:

    * Clears bit4 (EVM), this moves exception vectors to 0x00800000.
    * Sets bit3 (IVM), this enables interrupt table at 0x00800030.

Then it copies a jmp instruction to the reset-vector, unsure why.

Then it calls a function that sets up icache: If icache size is 0, function just returns without doing anything. It also bails if icache line width is < 2 or >= 5 (reserved values).

Then it calls a function at 0x400B0, which is an uncached code that does:

  • Enable icache.
  • ORs 0x400 into CFG (this is "reserved" according to datasheet!).

Then it calls a function that just iterates through an empty table of function pointers. C++ object initialization?

Then it calls main().

Then it calls a function that just iterates through an empty table of function pointers, again. C++ object dtors.

Finally goes into death-mode:

Writes a jump instruction to inf-loop to the reset-vector. This works because jmp instruction encoding contains an abs address.

Sets $lp = inf_sleep_loop.

Then it sets up args and jumps to a small stub at 0x008000E0. This small stub clears everything in f00d-mem in the region 0x00800100-0x8080FF0. Unknown why last 0x10 bytes are not cleared.


F00D messages

These are sent to ARM using the lower 16-bits at 0xE0000000. When ARM has read it, it is set to 0.

ARM can write a 32-bit response to 0xE0000010. For ARM->F00D, bit0 is used to indictate the message was written by ARM.

     1 = Request succeded
     4 = Debug string
 0x101 = Main init started
 0x102 = Sm can be loaded/resumed
 0x103 = Sm resumed successfully
 0x104 = Sm was shut down
 0x106 = Main shutting down
 0x107 = Suspend beginning
 0x108 = Ready for suspending, when using the async version.
0x8016 = Error: Invalid address range
0x802F = Error: Failed to init E003, E006.


Debug prints

secure_kernel supports tracing to a buffer.

Enabled by -7FF8h($gp) being non-zero. Out-buf address is stored in -7FF4($gp). It writes in a loop, 16 bytes at a time, inserting a null-terminator at buf[15] each "line".

After out-buf is written, writes 0x20000 to 0xE0000000. This will either signal ARM or disable ARM communications. Then inf-loop, this is a panic function.

Print types

Addresses are xored with a stack cookie that's fixed for all functions.

SuspendEncrypt BadAddr:                        00000003
SWI 1 BadAddr:                                 00000004
SWI 1 UnreachableCode:                         00000006
SWI 7 UnreachableCode:                         00000007
Addrcheck IntegerOverflow Food:                0000000B
Addrcheck IntegerOverflow Kernel:              0000000B
Addrcheck IntegerOverflow Module:              0000000C
IRQ register func w irq enabled:               0000000E
Force exit dmac w irq enabled:                 0000000F
Crypto irq enabled:                            00000010
Resuming suspendbuf w irq enabled:             00000012
Creating suspendbuf w irq enabled:             00000013
Reset:                                         00000014 <xored-exception-lr> <xored-exception-pc> <exception-lr>
RI:                                            00000015 <xored-exception-pc>
ZDIV:                                          00000016 <xored-exception-pc>
Trace:                                         00000017 <func-addr>
SWI 7:                                         00000018 <xored-r1>
DMAC when updating suspendbuf key:             00000019 <ret-val>
DMAC cmac bad enum:                            0000001A <enum-value>
Bad enum to suspend AES-CBC function:          0000001B <enum-value>
Generating new suspendbuf key failed:          0000001C
Creating suspendbuf w irq8 module-registered:  0000001D
Addrcheck IntegerOverflow Tz:                  0000001E
Addrcheck IntegerOverflow Tz2:                 0000001F
Bad ptr to suspend AES-CBC function:           00000020
IRQ register func w bad irq number:            00000022
Recieved unknown IRQ:                          00000023

SWI

When swi-handler is entered stack is set to 0x00807CC0. It stores context (all regs) on this stack. Then stack is set to 0x00807C40 from which it loads SPRs $tp, $gp, $sp. After SWI has been executed, it restores context from first stack.

in_r4 = Syscall number, starting with 1.

There are only 8 syscalls. Syscall numbers >= 8 are ignored by the handler and returns error 0x800F032C.


IRQ

When irq-handler is entered stack is set to 0x00807B40. Then it reads irq number from control bus space using the ldcb instruction. Irq number must be < 12, otherwise panic.

Then it sets up a special "user irq" context:

 sp = 0x0080B000
 tp = userctx.tp
 gp = userctx.gp 
 <rest same as kernel irq context>

And calls the function with r1=irq_id. The table is empty in the dump.

Above holds for all irqs except 8, which is special, see below.

Just before calling the callback, interrupts are enabled again. This leads to reentracy vulnerabilities.

IRQ8

IRQ8 is the 8th interrupt. This is the one triggered when ARM sends a request to F00D using 0xE0000010. And secure_kernel has a handler for this one.

IRQ9

This interrupt is sent when ARM writes to sm cmd registers (0xE0000014 and maybe more?). It's handled inside every sm module.


Secure Kernel Commands

Handler starts with a switch statement that handles reset commands.

For a command 0x100401 0x10 is size of shared buffer that's used by f00d handler, 0x4 is command ID, 0x1 is validity flag(?).

None of the following 5 commands are enabled/disabled depending on f00d-state. These are all unconditional.

0xB01

 Reset ScePervasiveResetReg +0x190 to enable re-writing the mask for the protected memory (keyring)
 Set mask to 0x10 for slots 0x20E and 0x20F (set in 0xE0070008)and trigger reset with 0 and 1 to device control 0xE0070000

0xC01

  Reads protected mem/keyring slot 0x50C.
  If lower 4 bits are nonzero:
    Reset ScePervasiveResetReg +0x190 to enable re-writing the mask for the protected memory (keyring)
    Set mask to 0x10 for slots 0x20E and 0x20F (set in 0xE0070008)and trigger reset with 0 (not 1 this time) to device control 0xE0070000

0xD01

  Reads protected mem/keyring slot 0x50C.
  If lower 4 bits are nonzero:
    Does same thing as 0xB01.

0xE01

 Does same thing as 0xB01.
 Then reads 3 times from ScePervasiveMisc (0xE3100000).
 It reads three times, if first read != 0x20, second != 0x30, third != 0x31 then it writes 6 to 0xE0070014.

0xF01: GetEncryptedInfoBlk

 It encrypts a block of size 0x80 with key=eeprom_blk_515, and hardcoded iv from .data.
 Block looks like this:
   +0x00: Magic (0xACB4ACB1)
   +0x04: One
   +0x08: Random (read from 0xE005003C)
   +0x0C: Zero
   +0x10: EEPROM sector 0x511
   +0x30: EEPROM sector 0x512
   +0x50: EEPROM sector 0x517
   +0x70: AES-256-CMAC using key from EEPROM sector 0x514.
 It memcpys this encrypted info-blk size 0x80 to 0x4001FF00.
 Then it programs (u32)1, followed by zeroes, to EEPROM sector 0x516.

After processing, 0xFFFFFFFF is written to 0xE0000010. Then comes the real switch. This one gives different func-ptrs. But before func-ptr is called there's a check in a table-lookup based on f00d-state. So not all cmds are allowed in all states. If not allowed error 0x8029 is sent.

Then it reads lower u16 of f00d mailbox. If all zeroes then it returns 0x802D. Then it reads lower u16 and now wants it to be 0, if not it returns 0x802D.

After this it finally calls the funcptr for cmdhandler. For unknown cmdid it returns 0x802A.

The allowed commands are as follows:

  State: 0   Allowed: 0
  State: 1   Allowed: 0
  State: 2   Allowed: 2
  State: 3   Allowed: 78E
  State: 4   Allowed: 2
  State: 5   Allowed: 72
  State: 6   Allowed: 42
  State: 7   Allowed: 2
  State: 8   Allowed: 2
  State: 9   Allowed: 0
  State: 10  Allowed: 0

todo: Decode these.

0x101: ArmPanic

 Sets food-state to 9.
 This will later cause main() to return, triggering memclr + infloop.
 Then it writes 0xF to control bus addr 0.
 Control bus addr 0 is interrupt controller, but bits don't match architecture doc.
 Corrupt code?
 Followed by 3 NOPs. Wat.

0x500201: LoadModule

 Reads 0x50 bytes from ArmBuffer into buf.
 If *(buf+4) and *(buf+8) are not aligned to 4 it fails.
 Then it TZ-checks addr *(buf+4) size 4.
 Then it TZ-checks addr *(buf+4) size 8.
 If any of these 3 checks fail, return code is 0x8016 or panic.
 Then it sets food-state to 4.
 memcpy(sp+0x68, buf+0x30, 0x20);
 *(sp+0x58)  = *(buf+0x28)
 *(sp+0x5C)  = *(buf+0x2C)
 *(sp+0x178) = *(buf+0x24)
 *(sp+0x54)  = *(buf+0x20)
 Then it calls the big function to load the SM with args (sp, sp+0x50).
 If it fails then either panics or returns 0x800000FF.
 It saves the address for "0x40-buf", without checking it to be in Tz.
 This is okay because they check before writing to it.
 Then it gets the address for the "shared buf", and loads arg0-arg3 from this ptr.
 Then it sets epc to 0x0080B000, sp to 0, gp to 0.
 And uses reti to jump to it.

0x100301: RestoreModule

  Reads 0x10 bytes from shared buf.
  Checks alignment on stuff also.
  TZ-check on *(buf+4) size 4, *(buf+8) size 8, *(buf+12) size 0x18. On fail sends 0x8016.
  After that it sets state to 4.
  Then it calls the big restore function.
  If it fails it sends zeroes module region, sends reply 0x8024, sets state to 3, and sends 0x102.
  If success, it sets the "0x40-buf" to *(buf+8), sets state to 5, and sends back 0x103.

It checks the suspendbuf-state already in f00d-memory. Checks magic, version, and size. On fail 0x800F0324. Then it uses the easter value to get the key into keyslot 0x22 and 0x33 somehow. Called function does this.

Then it calculates CMAC of the header that's already in f00d-memory into a buf on stack.

Then it calls a corrupted function that probably is this:

   SuspendBufReadPalist(
       dst=suspendContextBuf, size=0x140, skip=0x40, shared_buf, sp0=4, sp4=&some_stack_buf)

Then it continues the CMAC calculation on both contexts just read from ARM.

Then it calls a corrupted function that probably is this:

   SuspendBufReadPalist(
       dst=module_base, size=module_size+0x1C0, skip=0x180, shared_buf, sp0=4, sp4=&some_stack_buf)

Note that size here is read by the f00d header before the ARM one is copied in. So we can't modify it in this call unfortunately.

Then it continues the CMAC calculation on the module just read.

Then it calls the update keyslot 0x22 and 0x23 function. This is to prevent next suspendbuf from having the same key?

After that it does a timing-safe memcmp loop to verify the CMAC on stack against the header. Returns 0x800F0324 on CMAC cmpfail.

If hash checks out, it memcopies the contexts to whereever they were saved from. It loads all func-ptrs for IRQ listeners that were saved. It resets the F00D io-state that was saved. It sets MeP irq mask, but it makes sure to first mask away non-allowed ones.

Then it calls a function that calls the easteregg function. Probably to generate a new key.

0x100401: RequestModuleSuspend

 Reads 0x10 bytes from shared buf.
 Verifies alignment on *(buf+4), *(buf+8), *(buf+12).
 Verifies the following tz-ptrs:
   *(buf+8) size 4
   *(buf+4) size 8
   *(buf+12) size 0x18
 If bad it returns 8016.
 Sets f00d-state to 6.
 If sm is not ready for suspend, it saves the buf in bss.
 Also sets the flag that suspend has been requested then returns.
 If sm is ready to suspend, it calls the suspend function.

0x501: SubscribeSuspendAsyncEvent

 Sets f00d-state to 6.
 If sm is ready to be suspended, sends 0x108 and move into state 7.

0x601: ForceExitModule

 Sets f00d state to 6.
 Then it calls a function that:
    * Calls a function that appears to stop DMAC.
    * Zeroes the module region.
    * Flushes cache.
    * Writes some unknown DMAC registers.
    * Does a soft-reset.

0x80901: SetTraceBuffer

 Shared buf:
   +0: Addr
   +4: Size
 Checks that region is valid TZ, then sets them in the state.
 On bad addr, it sends error 0x8016, otherwise it sends 1.
 It also prints "rev %s\n" with "5679", however that only happens for some OTP configs.

0x80A01: SetRevocationList

 Reads 8 bytes from "shared buf".
 Calls the function to set the revocation list.
 If that function returns error x, the error that gets sent back is x&0xFF.
 On success, 1 is sent back.

Debug prints

secure_kernel supports tracing to a buffer.

Enabled by -7FF8h($gp) being non-zero. Out-buf address is stored in -7FF4($gp). It writes in a loop, 16 bytes at a time, inserting a null-terminator at buf[15] each "line".

After out-buf is written, writes 0x20000 to 0xE0000000. This will either signal ARM or disable ARM communications. Then inf-loop, this is a panic function.

Print types

Addresses are xored with a stack cookie that's fixed for all functions.

SuspendEncrypt BadAddr:                        00000003
SWI 1 BadAddr:                                 00000004
SWI 1 UnreachableCode:                         00000006
SWI 7 UnreachableCode:                         00000007
Addrcheck IntegerOverflow Food:                0000000B
Addrcheck IntegerOverflow Kernel:              0000000B
Addrcheck IntegerOverflow Module:              0000000C
IRQ register func w irq enabled:               0000000E
Force exit dmac w irq enabled:                 0000000F
Crypto irq enabled:                            00000010
Resuming suspendbuf w irq enabled:             00000012
Creating suspendbuf w irq enabled:             00000013
Reset:                                         00000014 <xored-exception-lr> <xored-exception-pc> <exception-lr>
RI:                                            00000015 <xored-exception-pc>
ZDIV:                                          00000016 <xored-exception-pc>
Trace:                                         00000017 <func-addr>
SWI 7:                                         00000018 <xored-r1>
DMAC when updating suspendbuf key:             00000019 <ret-val>
DMAC cmac bad enum:                            0000001A <enum-value>
Bad enum to suspend AES-CBC function:          0000001B <enum-value>
Generating new suspendbuf key failed:          0000001C
Creating suspendbuf w irq8 module-registered:  0000001D
Addrcheck IntegerOverflow Tz:                  0000001E
Addrcheck IntegerOverflow Tz2:                 0000001F
Bad ptr to suspend AES-CBC function:           00000020
IRQ register func w bad irq number:            00000022
Recieved unknown IRQ:                          00000023

Syscalls

1: Unload

  Sets f00d-state to 8.
  Writes r1 to "the shared2-buf".
  Sends msg 1 to ARM.
  Then it uses dmac with operation 0xC to memset entire module region to 0,
  Finally it jumps to secure_kernel entrypoint, triggering a softreboot.

2: ReadyToSuspend

  This just writes 1 to a state-field and returns 1,
  or returns 0x800F0329 if it's already non-zero.
  This flag tells kernel whether or not module ready to be suspended.

3: SuspendSelfIfRequested

  Returns 0x800F0329 if SuspendReady wasn't called prior to this.
  It checks a flag if 0x501 cmd has subscribed to the event.
  Then it will send msg 0x108 to ARM and goes into f00d-state 7.
  Regardless of that flag it also checks whether or not ARM has requested suspend.
  If so then it calls the suspend function.
  Otherwise it just returns 1.

4: RegisterIrqHandler

  r1 = irq-number
  r2 = func-ptr (or NULL to deregister)
  r3 = old-func-ptr-out (or NULL to ignore)
  Checks that r1 is < 12, and it also checks it against a mask of allowed sm IRQs:
    if (r1*2) & 0xE7E == 0: fail
  On either failure returns 0x800F0316.
  Checks that r2 is inside the module range or NULL, returns 0x800F030E on fail.
  Checks that r3 is inside the module range or NULL, returns 0x800F030E on fail.
  Returns 0x800F0316 if r2 not aligned to 2, or if (r3 & 2) != 0.
  Calls a function to get current function ptr in the irq-table entry for irq-number r1.
  Then it calls a function to register irq listener.
  This function does the following:
     Check addr, check irq number, panic on fail.
     Write r2 to IRQ func-ptr table.
     It reads the MeP interrupt mask register from control bus.
     Then depending on r2 was NULL or not, it sets n:th bit in the mask, where n is the IRQ number.
     Then it writes it back to MeP to enable the interrupt.
  Then it will write the old-func-ptr (before was overwritten) to r3, if r3 not set to NULL.

5: TracePrintf

  This just printfs with "%s" to the emit buffer, and returns 1.

6: CheckRvkList

  If r1 or r2 is 0 it returns 0x800F0B16.
  Then it checks that r1 size 0x130 is in module region.
  And that r2 size 0x80 also, same error.
  It returns 0x800F0326 if no rvk list has been inited.
  After that it returns result of the revocation checking function
  (same as used for sm normally).

7: UnloadPanic

  This one trace-printfs.
  Then it sets f00d-state to 8.
  Then it writes 0x800F033B to the "0x40 buffer".
  Then corrupted insn?
  Then it zeroes the module region and soft-reboots.

8: ReadEepromFlag

  This reads a bool from eeprom block 0x510 byte16 bit31.