Re: [RFC 3/3] Save/restore LWP state in context switches.

[Brian Gerst]

On Tue, Oct 5, 2010 at 2:30 PM, Hans Rosenfeld <hans.rosenfeld@xxxxxxx> wrote:
> LWP (Light-Weight Profiling) is a new per-thread profiling mechanism
> that can be enabled by any thread at any time if the OS claims to
> support it (by setting a bit in XCR0). A threads LWP state
> (configuration & unsaved collected data) is supposed to be saved and
> restored with xsave and xrstor by the OS.
>
> Unfortunately, LWP does not support any kind of lazy switching, nor does
> it use the TS bit in CR0. Since any thread can enable LWP at any time
> without the kernel knowing, the context switch code is supposed to
> save/restore LWP context unconditionally. This would require a valid
> xsave state area for all threads, whether or not they use any FPU or LWP
> functionality. It would also make the already complex lazy switching
> code more complicated.
>
> To avoid this memory overhead, especially for systems not supporting
> LWP, and also to avoid more intrusive changes to the code that handles
> FPU state, this patch handles LWP separately from the FPU. Only if a
> system supports LWP, the context switch code checks whether LWP has been
> used by the thread that is being taken off the CPU by reading the
> LWP_CBADDR MSR, which is nonzero if LWP has been used by the thread.
> Only in that case the LWP state is saved to the common xsave area in the
> threads FPU context. This means, of course, that an FPU context has to
> be allocated and initialized when a thread first uses LWP before using
> the FPU.
>
> Similarly, restoring the LWP state is only done when an FPU context
> exists and the LWP bit in the xstate header is set.
>
> To make things a little more complicated, xsave and xrstor _do_ use the
> TS bit and trap when it is set. To avoid unwanted traps, the TS bit has
> to be cleared before and restored after doing xsave or xrstor for LWP.
>
> Signed-off-by: Hans Rosenfeld <hans.rosenfeld@xxxxxxx>

I would prefer to see the xsave code refactored so that you would only
need one xsave/xrstor call per context switch.  We are currently
treating xsave as an extension to the FPU state.  But it would be
better to treat FPU as a subset of the extended state.  That way more
state can be added without touching the FPU code..

--
Brian Gerst
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/