On Thu, Jun 25, 2020 at 10:24:33AM +0200, Peter Zijlstra wrote: > On Thu, Jun 25, 2020 at 10:03:13AM +0200, Peter Zijlstra wrote: > > I'm sure Will will respond, but the basic issue is the trainwreck C11 > > made of dependent loads. > > > > Anyway, here's a link to the last time this came up: > > > > https://lore.kernel.org/linux-arm-kernel/20171116174830.GX3624@xxxxxxxxxxxxxxxxxx/ > > Another good read: > > https://lore.kernel.org/lkml/20150520005510.GA23559@xxxxxxxxxxxxxxxxxx/ > > and having (partially) re-read that, I now worry intensily about things > like latch_tree_find(), cyc2ns_read_begin, __ktime_get_fast_ns(). > > It looks like kernel/time/sched_clock.c uses raw_read_seqcount() which > deviates from the above patterns by, for some reason, using a primitive > that includes an extra smp_rmb(). > > And this is just the few things I could remember off the top of my head, > who knows what else is out there. As an example, let us consider __ktime_get_fast_ns(), the critical bit is: seq = raw_read_seqcount_latch(&tkf->seq); tkr = tkf->base + (seq & 0x01); now = tkr->base; And we hard rely on that being a dependent load, so: LOAD seq, (tkf->seq) LOAD tkr, tkf->base AND seq, 1 MUL seq, sizeof(tk_read_base) ADD tkr, seq LOAD now, (tkr->base) Such that we obtain 'now' as a direct dependency on 'seq'. This ensures the loads are ordered. A compiler can wreck this by translating it into something like: LOAD seq, (tkf->seq) LOAD tkr, tkf->base AND seq, 1 CMP seq, 0 JE 1f ADD tkr, sizeof(tk_read_base) 1: LOAD now, (tkr->base) Because now the machine can speculate and load now before seq, breaking the ordering.