Re: [PATCH] trace-cmd: fix kvm_mmu_prepare_zap_page even name and kvm_mmu_get_page event output in kvm plugin

[Gleb Natapov <gleb@redhat.com>]

On Wed, Dec 26, 2012 at 09:22:32AM -0500, Steven Rostedt wrote:
> On Wed, 2012-12-26 at 16:13 +0200, Gleb Natapov wrote:
> > On Wed, Dec 26, 2012 at 08:58:30AM -0500, Steven Rostedt wrote:
> > > On Tue, 2012-12-25 at 13:46 +0200, Gleb Natapov wrote:
> > > > kvm_mmu_zap_page even was renamed to kvm_mmu_prepare_zap_page.
> > > > Print out created field for kvm_mmu_get_page event.
> > > 
> > > trace-cmd needs to be backward compatible with older kernels. If older
> > > kernels used kvm_mmu_zap_page, then please add a check for that and have
> > > the plugin cope with either one.
> > > 
> > Something like this?
> > 
> >         if (pevent_find_event_by_name(pevent, "kvmmmu", "kvm_mmu_prepare_zap_page"))
> >                 pevent_register_event_handler(pevent, -1, "kvmmmu",
> >                                 "kvm_mmu_prepare_zap_page", kvm_mmu_print_role,
> >                                 NULL);
> >         else
> >                 pevent_register_event_handler(pevent, -1, "kvmmmu",
> >                                 "kvm_mmu_zap_page", kvm_mmu_print_role, NULL);
> 
> Sure, if it works.
> 
Well, it isn't and if it were it would have prevented tarce-cmd running
on new kernel from reading traces generated on older kernels. I will
register both.

> > 
> > Also when trace-cmd encounters an event it does not recognize it prints
> > mysterious message:
> > 
> > trace-cmd: No such file or directory
> >   bad op token {
> > 
> > Is this a bug?
> > 
> 
> Yes and no ;-)
> 
> I need to get rid of the "perror" part, for errors that don't set errno.
> That's been on my todo list for a long time. Maybe when I come back to
> work next week I'll fix that.
> 
> The 'bad op token {' happens when it tries to parse an event and it
> comes across something that it doesn't recognize. In this case a '{'.
> 
> The kvm events are notorious with having extremely complex "print_fmt"
> fields in their event format files. When no event handler is registered
> for an event, trace-cmd uses the print_fmt to figure out how to print
> it. This was never meant to be too complex of a parser. For example,
> looking at my kvm_mmu_prepare_zap_page, we have:
> 
> print fmt: "%s", ({ const char *ret = p->buffer + p->len; static const char *access_str[] = 
> { "---", "--x", "w--", "w-x", "-u-", "-ux", "wu-", "wux" }; union kvm_mmu_page_role role;
>  role.word = REC->role; trace_seq_printf(p, "sp gfn %llx %u%s q%u%s %s%s" " %snxe root %u
>  %s%c", REC->gfn, role.level, role.cr4_pae ? " pae" : "", role.quadrant, role.direct ? " 
> direct" : "", access_str[role.access], role.invalid ? " invalid" : "", role.nxe ? "" : "!",
>  REC->root_count, REC->unsync ? "unsync" : "sync", 0); ret; })
> 
> trace-cmd has no idea on how to parse "({ const char *ret ..." in fact
> it dies on that first "{" because trace-cmd is not a full C parser.
> 
> This was why the plugins were created in the first place. To handle
> various events that have too complex print_fmts for trace-cmd to
> understand.
> 
> Most events are simple print_fmts and do not need plugins. Like
> sock_rcvqueue_full:
> 
> print fmt: "rmem_alloc=%d truesize=%u sk_rcvbuf=%d", REC->rmem_alloc,
> REC->truesize, REC->sk_rcvbuf
> 
> trace-cmd has no problem parsing events like that.
> 
Got it.

--
			Gleb.