Creport

Revision as of 03:27, 12 November 2018 by SciresM (talk | contribs) (→‎6.1.0: details on changes)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

This process is launched by NS when PM signals that there is a crashing process.

Creport takes a string containing a pid formatted in base10 as input, and generates an error report. This error report can later be sent to the cloud server by Eupld services.

[2.1.0+]: An additional input argument string is now used. Only the first byte is used: inarg_flag = argv[1][0]; This is compared with '1' only with the below titleID code, and near the end of nnMain().

Crash dumping

It uses the svcDebugActiveProcess to start a debug-session for the pid. It loops svcGetDebugEvent to fetch all debug events.

It has a event buffer of 128 u64's that starts with "CREP\x01\x00\x00\x00".

It only cares about type EXCEPTION and ATTACH_PROCESS events.

For ATTACH_PROCESS events it saves the title-id of the crashed process (this is the first u64 of the type-specific data), and a flag whether to dump stack.

For EXCEPTION events, it adds records to the buffer.

Each exception record starts with 2x u64's: exception type, and exception address.

Then depending on exception type it stores:

  • UNDEFINED_INSTRUCTION (0): Always (u64) 0.
  • PREFETCH_ABORT (1): Always (u64) 0.
  • DATA_ABORT (2): (u64) Fault register.
  • UNALIGNED_ACCESS (3): (u64) Fault register.
  • UNDEFINED_SYSCALL (8): (u64) Syscall id.
  • ? (9): Always (u64) 0.

For all exceptions, it then adds more data from svcGetDebugThreadParam/svcGetDebugThreadContext. "Usually" (when is this not the case?), this data includes a size u64 for the register dump region + the values of all general purpose registers+SP and PC, as well as 0x80 of stack. This reads the flag from ATTACH_PROCESS to determine whether to read 0x10 bytes using svcReadDebugProcessMemory.

All other events (USER_BREAK, etc) don't store any extra data except type and address.

After it has fetched all events, if it didn't encounter USER_BREAK it constructs an error report:

  • Field10 "ErrorCode": (String) Error-code string formatted with "%04d-%04d".
  • Field115 "ProgramId": (String) Title-id snprintf'ed as "%08llx".
  • Field116 "AbortFlag": (Bool) 0.

It does *not* add the event buffer to the report(blacklisted) if title-id is any of the following(swkbd and all web-applets except offline-applet):

  • 0100000000001008
  • 010000000000100A
  • 010000000000100B
  • 0100000000001010
  • 0100000000001011

[2.1.0+] It also blacklists the creport-sysmodule title-id. Then, if inarg_flag(see above) is set to '1', all title-ids are blacklisted except for the following whitelist:

  • 0100704000B3A000 "Snipperclips" (Game)
  • 01007EF00011E000 "The Legend of Zelda: Breath of the Wild"
  • 01009B500007C000 "ARMS"
  • [2.3.0+] 0100C5E003B40000 "ARMS Global Testpunch"
  • 0100D87002EE0000 "Snipperclips - Cut it out, together!"
  • 0100F8F0000A2000 "Splatoon 2" (EUR)
  • 010000A00218E000 "Splatoon 2 Global Testfire"
  • 01000320000CC000 "1-2 Switch"
  • 0100152000022000 "Mario Kart 8 Deluxe"
  • 01003BC0000A0000 "Splatoon 2" (USA)
  • 01003C700009C000 "Splatoon 2" (JPN)

This is probably because of privacy concerns (software keyboard + browser could contain passwords and personal info).

The above whitelist handling is probably so that only Nintendo (published) applications get full exception info reported, since crash-reports for other applications probably(?) wouldn't be shared with third-parties.

For all other title-ids, it generates a random AES-128 key and CTR using csrng.

It encrypts the entire event buffer with this AES key and CTR. Then it encrypts the key-iv-pair using RSA-PSS with a hardcoded pubkey and exponent 0x10001.

These are added to the error report:

  • Field206 "EncryptionKey": (Raw) RSA-encrypted AES-key.
  • Field207 "EncryptedExceptionInfo": (Raw) Encrypted crash-info.

Thus you need the private key to decrypt the crash dump.

Finally, the error report is sent to "erpt:c" cmd1.

It uses "ns:dev" to terminate the pid. If certain checks pass, then it throws the following fatal-err depending on exception type:

  • 0 -> 0xA8
  • 1 -> 0x2A8
  • 2 -> 0x4A8,
  • 3 -> 0x6A8,
  • 6 -> No fatal-err
  • 8 -> 0x10A8
  • Default: 0x4A2

2.1.0

Exactly the following code was changed:

  • nnMain(): Updated.
  • L_1ee0(prev ver L_1c90): Updated.
  • L_ad34: New func, svcQueryDebugProcessMemory.
  • L_e290: New func. "L_12cf0(<inparams>); return inx0;"
  • L_12cf0: New func. Only called by L_e290.

nnMain:

  • Two input arguments are now used+required, see above.
  • During init near the start of this func, u64 val0 is now written to x24+32.
  • A lot of new code was added.
  • TID handling block was updated, see above.
  • The check for <is_blacklisted> was changed from "if(val<=0)<branch>" to "if(val<1)<branch>".
  • The code block at the very end of this func was updated.
    • Prev code: if(<loadedval> != 1)return; New code: if(<loadedval> != 1)<jump over the two following func calls which are the same as prev ver>
    • Following the two funcs mentioned above, prev code: if(<loadedval> == 1 && (u8 *(ptr+1) & 1) == 0)<call func>; return; New code: if(inarg_flag != '1' && (u8 *(ptr+1) & 1) == 0 && ((u8 *(ptr+0) ^ 0x1) & 0x1) == 0)<call func>; return;
    • <call func> here is throw_fatalerr(ptr+4). The above second block basically changed the conditions required for throwing fatal-error. For example, fatal-error is no longer thrown when applications crash.

2.3.0

Exactly the following code was changed:

Only change was adding a titleID to the above whitelist.

5.0.0

Many changes were made to add more detail to reports. In particular:

  • The second input flag is no longer actually used, instead whether the process is an application is parsed from the ATTACH_PROCESS debug info.
  • Support was added for reading a custom user error code from process memory in the UserBdsak case.
  • Support was added for reading a custom user "Dying Message" of up to 0x1000 bytes from process memory if the crashes process was an application.
  • All reports now have additional info in their crash reports:
    • A list of up to 0x60 threads is retrieved via svcGetThreadList, and each thread has a full register dump + stacktrace added to the report.
    • The crashing thread's PC and LR are used to try to locate the base executable region that caused the crash -- if found, it and up to 15 code regions with higher virtual addresses have their start and end addresses saved, and their executable name and GNU build IDs read out of .rodata and added to the report. This fixes the problem of crash reports in previous versions not including information on ASLR.

6.1.0

Support was improved for detecting code regions. In particular:

  • The number of processable code regions was increased from 16 to 96.
  • Instead of processing the crashing thread's PC or LR, now both are processed, and additionally every address in the thread's stacktrace are processed.
    • If the crashed module is an application, this is further done for all threads.