Subversion Repositories HelenOS

Rev

Go to most recent revision | Show changed files | Directory listing | RSS feed

Filtering Options

Rev Age Author Path Log message Diff
2208 6260 d 22 h decky /trunk/kernel/generic/ use spinlock only on console output, not other print functions
cleanup
 
2193 6262 d 1 h decky /trunk/uspace/tester/ reimplement all tests from original ipcc
(cleanup needed)
 
2192 6262 d 1 h decky /trunk/uspace/libc/arch/sparc64/include/ fix uspace atomic_add according kernel code  
2188 6262 d 2 h decky /trunk/ new user space testing framework (unfinished)  
2187 6262 d 3 h decky /trunk/kernel/generic/ add syscall for getting current thread ID  
2186 6262 d 3 h decky /trunk/uspace/tester/ rename ipcc.c to tester.c  
2185 6262 d 3 h decky /trunk/ rename ipcc to tester  
2184 6262 d 5 h decky /trunk/kernel/generic/src/lib/ add zetta (sextillion) prefix
make constants explicitly unsigned
 
2183 6262 d 23 h jermar /trunk/kernel/generic/ Continue to de-oversynchronize the kernel.
- replace as->refcount with an atomic counter; accesses to this
reference counter are not to be done when the as->lock mutex is held;
this gets us rid of mutex_lock_active();

Remove the possibility of a deadlock between TLB shootdown and asidlock.
- get rid of mutex_lock_active() on as->lock
- when locking the asidlock spinlock, always do it conditionally and with
preemption disabled; in the unsuccessful case, enable interrupts and try again
- there should be no deadlock between TLB shootdown and the as->lock mutexes
- PLEASE REVIEW !!!

Add DEADLOCK_PROBE's to places where we have spinlock_trylock() loops.
 
2170 6267 d 17 h jermar /trunk/kernel/ Simplify synchronization in as_switch().
The function was oversynchronized, which
was causing deadlocks on the address
space mutex.

Now, address spaces can only be switched
when the asidlock is held. This also protects
stealing of ASIDs. No other synchronization
is necessary.
 
2161 6269 d 20 h jermar /trunk/kernel/arch/sparc64/src/mm/ Fix TSB bug during TSB refill.
When one wants to enable a TSB entry, he or she should set the
entry invalid bit to false, as opposed to setting it to true.
 
2160 6269 d 20 h decky /trunk/boot/arch/ remove duplicate dependency  
2144 6270 d 18 h jermar /trunk/kernel/arch/sparc64/src/mm/ Fix TSB size.  
2143 6271 d 0 h jermar /trunk/kernel/arch/sparc64/include/ Remove unneeded enum member.  
2142 6271 d 0 h jermar /trunk/kernel/generic/include/mm/ Remove out of date comment.  
2141 6271 d 16 h jermar /trunk/ The Ultimate Solution To Illegal Virtual Aliases.
It is better to avoid them completely than to fight them.
Switch the sparc64 port to 16K pages. The TLBs and TSBs
continue to operate with 8K pages only. Page tables and
other generic parts operate with 16K pages.

Because the MMU doesn't support 16K directly, each 16K
page is emulated by a pair of 8K pages. With 16K pages,
illegal aliases cannot be created in 16K D-cache.
 
2139 6271 d 17 h decky /trunk/contrib/toolchain/ move to GCC 4.1.2, make scripts really host platform independent
remove orphaned libobjc
 
2138 6271 d 19 h decky /trunk/contrib/toolchain/ revert previous change (libobjc is actually HOST platform dependent)  
2134 6272 d 20 h jermar /trunk/kernel/ Reworked handling of illegal virtual aliases caused by frame reuse.
We moved the incomplete handling from backend's frame method to
backend's page_fault method. The page_fault method is the one that
can create an illegal alias if it writes the userspace frame using
kernel address with a different page color than the page to which is
this frame mapped in userspace. When we detect this, we do D-cache
shootdown on all processors (!!!).

If we add code that accesses userspace memory from kernel address
space, we will have to check for illegal virtual aliases at all such
places.

I tested this on a 4-way simulated E6500 and a real-world Ultra 5,
which has unfortunatelly only one processor.

This solves ticket #26.
 
2133 6272 d 20 h jermar /trunk/kernel/generic/src/mm/ Coding style and indentation fixes.  

Show All