From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753050AbZBZRnl (ORCPT ); Thu, 26 Feb 2009 12:43:41 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1751626AbZBZRnd (ORCPT ); Thu, 26 Feb 2009 12:43:33 -0500 Received: from mx3.mail.elte.hu ([157.181.1.138]:35337 "EHLO mx3.mail.elte.hu" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751048AbZBZRnc (ORCPT ); Thu, 26 Feb 2009 12:43:32 -0500 Date: Thu, 26 Feb 2009 18:43:03 +0100 From: Ingo Molnar To: Frederic Weisbecker Cc: Linus Torvalds , Andrew Morton , linux-kernel@vger.kernel.org, Steven Rostedt , Lai Jiangshan , Peter Zijlstra Subject: Re: [PATCH 1/3] add binary printf Message-ID: <20090226174303.GC29439@elte.hu> References: <49a38304.0506d00a.1f4b.406d@mx.google.com> <20090226130243.GA22460@elte.hu> <20090226170524.GB5889@nowhere> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20090226170524.GB5889@nowhere> User-Agent: Mutt/1.5.18 (2008-05-17) X-ELTE-VirusStatus: clean X-ELTE-SpamScore: -1.5 X-ELTE-SpamLevel: X-ELTE-SpamCheck: no X-ELTE-SpamVersion: ELTE 2.0 X-ELTE-SpamCheck-Details: score=-1.5 required=5.9 tests=BAYES_00 autolearn=no SpamAssassin version=3.2.3 -1.5 BAYES_00 BODY: Bayesian spam probability is 0 to 1% [score: 0.0000] Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org * Frederic Weisbecker wrote: > On Thu, Feb 26, 2009 at 02:02:43PM +0100, Ingo Molnar wrote: > > > > * Frederic Weisbecker wrote: > > > > > From: Lai Jiangshan > > > > > > Impact: Add APIs for binary trace printk infrastructure > > > > > > vbin_printf(): write args to binary buffer, string is copied > > > when "%s" is occurred. > > > > > > bstr_printf(): read from binary buffer for args and format a string > > > > > > [fweisbec@gmail.com: ported to latest -tip] > > > > > > Signed-off-by: Lai Jiangshan > > > Signed-off-by: Frederic Weisbecker > > > Cc: Steven Rostedt > > > --- > > > include/linux/string.h | 7 + > > > lib/Kconfig | 3 + > > > lib/vsprintf.c | 442 ++++++++++++++++++++++++++++++++++++++++++++++++ > > > 3 files changed, 452 insertions(+), 0 deletions(-) > > > > OK, it's a nice idea and speedup for printf based tracing - > > which is common and convenient. Would you mind to post the > > performance measurements you've done using the new bstr_printf() > > facility? (the nanoseconds latency figures you did in the timer > > irq in a system under load and on a system that is idle) > > > Ok. > > > > The new printf code itself should be done cleaner i think and is > > not acceptable in its current form. > > > > These two new functions: > > > > > +#ifdef CONFIG_BINARY_PRINTF > > > +/* > > > + * bprintf service: > > > + * vbin_printf() - VA arguments to binary data > > > + * bstr_printf() - Binary data to text string > > > + */ > > > > Duplicate hundreds of lines of code into three large functions > > (vsnprintf, vbin_printf, bstr_printf). These functions only have > > a difference in the way the argument list is iterated and the > > way the parsed result is stored: > > > > vsnprintf: iterates va_list, stores into string > > bstr_printf: iterates bin_buf, stores into string > > vbin_printf: iterates va_list, stores into bin_buf > > > > We should try _much_ harder at unifying these functions before > > giving up and duplicating them... > > > > An opaque in_buf/out_buf handle plus two helper function > > pointers passed in would be an obvious implementation. > > > > That way we'd have a single generic (inline) function that knows > > about the printf format itself: > > > > __generic_printf(void *in_buf, > > void *out_buf, > > void * (*read_in_buf)(void **), > > void * (*store_out_buf)(void **)); > > > > And we'd have various variants for read_in_buf and > > store_out_buf. The generic function iterates the following way: > > > > in_val = read_in_buf(&in_buf); > > ... > > store_out_buf(&out_buf, in_val); > > > > (where in_val is wide enough to store a single argument.) The > > iterators modify the in_buf / out_buf pointers. Argument > > skipping can be done by reading the in-buf and not using it. I > > think we can do it with just two iterator methods. > > > > Or something like that - you get the idea. It can all be inlined > > so that we'd end up with essentially the same vsnprint() > > instruction sequence we have today. > > > > Ingo > > > Ok, I just looked deeply inside vsnprintf, and I don't think > such a generic interface would allow that. We need to know the > size of the argument, it's precision, width and flags.... And > we need to know if we want to skip the non format char. > > > What do you think of the following: > > __ generic_printf(void *in, > void *out, > void *(*read_in)(void **buf, int size), > void *(store_char)(char *dst, char *end, char val, int field_width, int flags), > void *(*store_string)(char *dst, char *end, char *val, int field_width, int precision, int flags), > void *(*store_pointer)(char type, char *dst, char *end, void *val, > int field_width, int precision, int flags), > void *(*store_number)(char *dst, char *size, int base,int field_width, int precision, int flags), > bool skip_non_format > ) > > > Well, something like that... > > read_in can advance the pointer to the buffer itself (buf can > be a va_args or u32 *) and it returns a value, void * is > generic for the type. > > The storage functions are more specialized because of the > interpretation of flags, precision... So we can easily pass > the usual string(), pointer(), .... that are already present > in vsnprintf.c or use custom ones. They return the advanced > dst pointer. > > And at last, skip_non_format will decide if we want to > consider non-format characters from fmt to be copied as common > %c characters or if we want to ignore them (useful for > vbin_printf()). hm, that indeed looks very wide - storing into a binary buffer does complicate the iterator interface significantly. But at least vsnprintf() and bstr_printf() could be unified - they both output into a string buffer, just have different methods to iterate arguments. Ingo