avoid exceptions in copy_from_user?
This is probably a flawed idea but I was just curious: could the kernel use speculative loads to avoid taking exceptions when a copy_from_user or related function faults? (Something like the routine below.)
I thought that if the chk.s instructions could fit into otherwise unused slots, then failing copies would avoid needing to take an interrupt and search the fixup table, and could instead branch directly to the code to clean up and return.
I suspect the answer is that it's not worthwhile, because if the copy fails then the kernel is about to go and do something expensive like paging or killing the task, and so a little more overhead is not important, and anyhow adding instructions to the non-failing case is undesirable.
// int readit(int *a)
// return the contents of a, or -1 if a is not a valid pointer
.text
.align 16 .global readit# .proc readit# readit: .prologue // one input, no locals alloc r2 = ar.pfs, 1, 0, 0, 0 .body // Try to read the thing ld8.s ret0 = [in0] ;; // Did we break? chk.s ret0, .recovery // Return br.ret.sptk.many b0
.recovery:
mov ret0 = -1 ;; br.ret.sptk.many b0 .end readit#
Signature · 1 line
Martin
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Martin Pool
Martin> This is probably a flawed idea but I was just curious: could Martin> the kernel use speculative loads to avoid taking exceptions Martin> when a copy_from_user or related function faults? Martin> (Something like the routine below.)
You're right, it's flawed. ;-)
First problem: why optimize the _exceptional_ case? Show me one application which suffers from the cost of the exception handling and I'll buy you a beer (or two or three). OTOH, almost every application benefits from optimizing the common case of exception-free execution.
Second (and perhaps more important) problem: kernel code must not assume that a speculative load fails only if the load cannot possibly execute successfully (in ia64 lingo, this is called the "recovery" model, as opposed to the "no-recovery" model, which I'm not a fan of). For example, depending on the DCR settings, a speculative load may also fail on a TLB miss. Of course, you don't want the code to fail just because of a TLB miss.
Martin> I suspect the answer is that it's not worthwhile, because if Martin> the copy fails then the kernel is about to go and do Martin> something expensive like paging or killing the task, and so Martin> a little more overhead is not important, and anyhow adding Martin> instructions to the non-failing case is undesirable.
Yup.
My experience so far is that speculative loads are best left to the compiler. Human beings are just not very good at writing recovery code and keeping it in sync with the main code (e.g., we had some painful experiences with the strncpy() routine).
--david