Jump to content

Switch - Atmosphère 1.2.2


Recommended Posts

1.2.2 is Atmosphère's fifty-ninth official release.

Please be sure to update fusee when upgrading to 1.2.2. fusee-primary no longer exists, and will not work any more.

With thanks to the @switchbrew team, Atmosphère 1.2.2 is bundled with hbl 2.4.1, and hbmenu 3.5.0.

The following was changed since the last release:

  • A number of fixes were made to Atmosphère's implementation of the new "sprofile" service added in 13.0.0.
    • Nintendo is finally transmitting data over the internet to certain consoles, which has allowed for validating our service implementation.
      • Unfortunately, there were several problems, and if your console began trying to use the new services atmosphère would show a fatal error with code 0xCAF6 (sprofile::ResultInvalidState()).
    • With actual test data in hand, a test program was written and it was verified that our implementation can successfully import/access profile data now.
      • Hopefully there are no more issues, and I sincerely apologize to anyone who got an 0xCAF6 fatal due to this.
  • A number of minor improvements were made to mesosphère, including:
    • KThread::GetContextForSchedulerLoop was implemented in assembly (using static assertions to verify offset-of-context-in-struct is correct).
      • This saves an unnecessary function call in the middle of the scheduler hot loop, replacing it with an addition instruction, which should improve microperformance.
    • Mesosphère's hardware maintenance instructions were audited via a script and now directly match Nintendo's kernels.
      • Notably, this inserts a missing instruction synchronization barrier when validating that slab heaps may be constructed.
      • This missing ISB could cause an abort on certain (see: particularly sensitive) hardware on boot if the relevant codepath was speculatively executed (it normally only executes on game launch...)
    • The SVC handlers for performing light IPC (normally unused) from 32-bit process were fixed.
    • A bug was fixed that would cause the register x27 to be overwritten with the contents of x26 when returning from a user exception handler.
    • A bug was fixed that would cause the kernel to use the userland stack pointer instead of the kernel stack pointer while generating an error report for a kernel abort.
  • General system stability improvements to enhance the user's experience.

And the following was changed in 1.2.1:

  • Support was implemented for 13.1.0.
    • mesosphère was updated to reflect the kernel behavioral changes made in 13.1.0.
      • KScheduler now issues a data memory barrier when unlocking the scheduler lock and when early-returning due to top-thread-is-current during scheduling.
    • erpt was updated to reflect the latest official behaviors.
      • The new service added in 13.0.0 ("sprofile") was revised, and the data formats it expects was changed.
        • This still appears to be (possibly(?)) untestable due to data not being transmitted yet, but I have greater confidence things will go smoothly than I did when 1.1.0 released.
  • A number of improvements were made to mesosphère, including:
    • A build target was created to build targeting the qemu virt board.
      • This facilitates writing unit tests for the kernel (and other atmosphère components) and running them under PC.
        • Please Note: Official system software will not work at all under this, and the Atmosphère project has zero interest in attempting to run official software of any kind. This is unit testing machinery, and explicitly not more than that.
      • This should hopefully allow us to have greater confidence that all of atmosphère's components work the way they're theoretically supposed to in the future.
      • Please Note: If you are a developer who is familiar with the Horizon operating system (or capable of becoming familiar), I would greatly appreciate help writing tests and improving the testing framework.
        • Please contact SciresM#0524 if you are capable and interested.
          • Really, if you are actually a developer who would like to help me get this off the ground, I would deeply appreciate it.
          • That said, if you are not a developer but want to be one, this probably isn't the best opportunity; I expect it to be highly technical.
            • Consider the ReSwitched discord's #hack-n-all channel for your educational purposes.
        • We are (at least for now) using catch2 for unit tests.
    • Almost all virtual calls in the kernel are now resolved statically.
      • This eliminates substantial virtual call overhead, and should lead to improved kernel microperformance in pretty much every function.
    • The remaining red black tree find operations which weren't using the optimized "find key" variant are now using the optimized version.
    • Custom assembly was written to improve tick-to-timespan conversion.
      • This works around gcc emitting suboptimal assembly at -Os (it emits good assembly at -O3, clang is fine at both -O3 and -Os).
    • KThread and KSession structures were updated to optimize member layout, saving 0x10 bytes per KThread/KSession object.
    • Rather than unnecessarily zero-ing all data in kernel objects only to overwrite members later, we now only initialize the members we need to in kernel object constructors.
      • This is what Nintendo was doing already.
    • A set of custom optimized atomic primitives were implemented and are used in place of std::atomic<>
      • This works around a gcc bug which downgrades specified memory order to seq_cst, and introduces clrex in places where it is appropriate.
      • This should strictly improve microperformance of many system calls.
    • An compile-time toggleable extension was added to support 40-bit physical addresses in MapRange capabilities (using currently reserved bits).
    • A number of minor bugs were fixed, including:
      • Initial cache management now better reflects official behavior.
        • This fixes an issue that caused certain hardware with cache sensitivity to produce cryptic kernel panics during boot.
      • Incorrect logic when checking thread priority capabilities was fixed to reflect official behavior.
      • The scheduler was updated to reflect latest official behavior, and a number of minor bugs involving clz/ctz were fixed.
      • Accesses to the processes local region were fixed to properly use kernel linear region, not userland pointers.
      • The cache SVCs exposed for 32-bit processes now better reflect official core mask request semantics.
      • A bug was fixed that could cause a kernel panic if SvcArbitrateLock was called on a thread with exactly one reference in the middle of handling a user-mode exception.
  • General system stability improvements to enhance the user's experience.

For information on the featureset supported by 1.2, please see the official release notes.

View the full article

Link to comment
Share on other sites

  • Create New...