[Fwd: Re: gdb/725: Crash using debug target and regcaches (in 5.3 branch?)]

[Andrew Cagney]

> On Thu, Nov 21, 2002 at 05:42:24PM -0500, Andrew Cagney wrote:
> 
>> FYI,
>> 
>> Too many memory reads/writes was one reason for a ptrace'd threaded 
>> shlib program running slow, I suspect this is the other.
> 
> 
> Maybe, maybe not... definitely needs to go though!  Thanks for such a
> thorough investigation, it gave me a good idea.

[snip]

>> currently:
>> runtest linux-dp.exp print-threads.exp  17.21s user 48.22s system 82% cpu 1:19.56 total
>> With change:
>> runtest linux-dp.exp print-threads.exp  16.67s user 45.35s system 82% cpu 1:15.27 total

Given that the numbers are being overwhelmed by all those memory read 
ptrace calls, a ~5% improvement is significant.

Try something simpler like running gdb under strace (tweak 
testsuite/lib/gdb.exp to run 'strace $GDB' instead of $GDB) and then 
count how many ptrace calls of each type occure.

>>  Briefly, the GNU/Linux thread code is giving regcache.c conflicting 
>>  stories over which inferior ptid should be in the register cache.  As a 
>>  consequence, every single register fetch leads to a regcache flush and 
>>  re-fetch.  Outch!
>>  
>>  
>>  Briefly,  core GDB tries to fetch a register.  This eventually leads to 
>>  the call:
>>  
>>  regcache_raw_read(REGNUM)
>>  
>>  	registers_tpid != inferior_tpid
>>  (gdb) print registers_ptid
>>  $6 = {pid = 31263, lwp = 0, tid = 0}
>>  (gdb) print inferior_ptid
>>  $7 = {pid = 31263, lwp = 31263, tid = 0}
>>  		-> flush regcache
>>  		-> registers_tpid = inferior_tpid
>>  	-- at this point regnum is invalid
>>  	target_fetch_registers (regnum)
>>  
>>  Since the inferior doesn't match the target, the cache is flushed, 
>>  inferior_ptid is updated, and the register is fetched.  The fetch flows 
>>  on down into the depths of the target and the call:
>>  
>>  Seen the problem yet?
> 
> 
> Yup.  Saw something else very interesting, too.
> 
> 
>>  The long term fix is to have per-thread register caches, that is 
>>  progressing.
>>  
>>  I don't know about a short term fix though.
> 
> 
> I was working on a short-term fix and discovered it was almost entirely
> in place already.  Look at a couple of random fetch_inferior_registers
> implementations; every one that a GNU/Linux platform uses already will
> fetch the LWP's registers if the LWP is non-zero.  So why not give that
> to 'em?  Leave the inferior_ptid as it is, and make
> fetch_inferior_registers honor the LWP id.

It feels right.  I'm hopeing that, eventually, the code will supply the 
registers direct to a (one of many) `struct thread_info' object.

> So, thoughts on the attached patch?

Thread maintainer question (not so sure about the #ifdef linux though :-).

Andrew