From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mga03.intel.com (mga03.intel.com [134.134.136.65]) by dpdk.org (Postfix) with ESMTP id 570162C57 for ; Wed, 1 Feb 2017 15:18:45 +0100 (CET) Received: from orsmga001.jf.intel.com ([10.7.209.18]) by orsmga103.jf.intel.com with ESMTP; 01 Feb 2017 06:18:44 -0800 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.33,320,1477983600"; d="scan'208";a="1089679112" Received: from bricha3-mobl3.ger.corp.intel.com ([10.237.221.61]) by orsmga001.jf.intel.com with SMTP; 01 Feb 2017 06:18:41 -0800 Received: by (sSMTP sendmail emulation); Wed, 01 Feb 2017 14:18:40 +0000 Date: Wed, 1 Feb 2017 14:18:40 +0000 From: Bruce Richardson To: Jan Blunck Cc: Adrien Mazarguil , Thomas Monjalon , Aaron Conole , dev@dpdk.org, Stephen Hemminger Message-ID: <20170201141839.GA162640@bricha3-MOBL3.ger.corp.intel.com> References: <1485529023-5486-1-git-send-email-aconole@redhat.com> <20170127093729.5cef9138@xeon-e3> <2880962.uYJx3WqeFl@xps13> <20170201105430.GQ10133@6wind.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: Organization: Intel Research and =?iso-8859-1?Q?De=ACvel?= =?iso-8859-1?Q?opment?= Ireland Ltd. User-Agent: Mutt/1.7.1 (2016-10-04) Subject: Re: [dpdk-dev] [PATCH 25/25] rte_eal_init: add info about rte_errno codes X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Wed, 01 Feb 2017 14:18:45 -0000 On Wed, Feb 01, 2017 at 01:06:03PM +0100, Jan Blunck wrote: > On Wed, Feb 1, 2017 at 11:54 AM, Adrien Mazarguil > wrote: > > On Mon, Jan 30, 2017 at 09:19:29PM +0100, Thomas Monjalon wrote: > >> 2017-01-30 13:38, Aaron Conole: > >> > Stephen Hemminger writes: > >> > > Bruce Richardson wrote: > >> > >> On Fri, Jan 27, 2017 at 08:33:46AM -0800, Stephen Hemminger wrote: > >> > >> > Why use rte_errno? > >> > >> > Most DPDK calls just return negative value on error which > >> > >> > corresponds to error number. > >> > >> > Are you trying to keep ABI compatibility? Doesn't make sense > >> > >> > because before all these > >> > >> > errors were panic's no working application is going to care. > >> > >> > >> > >> Either will work, but I actually prefer this way. I view using rte_errno > >> > >> to be better as it can work in just about all cases, including with > >> > >> functions which return pointers. This allows you to have a standard > >> > >> method across all functions for returning error codes, and it only > >> > >> requires a single sentinal value to indicate error, rather than using a > >> > >> whole range of values. > >> > > > >> > > The problem is DPDK is getting more inconsistent on how this is done. > >> > > As long as error returns are always same as kernel/glibc errno's it really doesn't > >> > > matter much which way the value is returned from a technical point of view > >> > > but the inconsistency is sure to be a usability problem and source of errors. > >> > > >> > I am using rte_errno here because I assumed it was the preferred > >> > method. In fact, looking at some recently contributed modules (for > >> > instance pdump), it seems that folks are using it. > >> > > >> > I'm not really sure the purpose of having rte_errno if it isn't used, so > >> > it'd be helpful to know if there's some consensus on reflecting errors > >> > via this variable, or on returning error codes. Whichever is the more > >> > consistent with the way the DPDK project does things, I'm game :). > >> > >> I think we can use both return value and rte_errno. > >> We could try to enforce rte_errno as mandatory everywhere. > >> > >> Adrien did the recent rte_flow API. > >> Please Adrien, could you give your thought? > > > > Sure, actually as already pointed out in this thread, both approaches have > > pros and cons depending on the use-case. > > > > Through return value: > > > > Pros > > ---- > > > > - Most common approach used in DPPK today. > > - Used internally by the Linux kernel (negative errno) and in the pthreads > > library (positive errno). > > - Avoids the need to access an external, global variable requiring its own > > thread-local storage. > > - Inherently thread-safe and reentrant (i.e. safe with signal handlers). > > - Returned value is also the error code, two facts reported at once. > > Caller can decide to ignore return value if no error handling is wanted. > Not always the case. In the case of a rx or tx burst call, if there is a negative error that must be checked for or assigned to zero in some cases to make other logic in the path work sanely, e.g. updating an array of stats using the return value. > > > > Cons > > ---- > > > > - Difficult to use with functions returning anything other than signed > > integers with negative values having no other meaning. > > - The returned value must be assigned to a local variable in order not to > > discard it and process it later most of the time. > > I believe this is Pro since the rte_errno even needs to assign to a > thread-local variable even. No, it's a con, since for errno the value will be preserved in the absense of other errors. The application can delay handling the error as long as it wants, in the absense of causes of subsequent errors. > > > - All function calls must be tested for errors. > > The rte_errno needs to do this too to decide if it needs to assign a > value to rte_errno. > Thats inside the called function, not the application. See my earlier comment above about having to check your return value is in the valid "logical range" expected from the call. Having a negative number of packets received does not make logical sense, so you have to check the return value when updating stats etc. > > > > Through rte_errno: > > > > Pros > > ---- > > > > - errno-like, well known behavior defined by the C standard and used > > everywhere in the C library. > > - Testing return values is not mandatory, e.g. rte_errno can be initialized > > to zero before calling a group of functions and checking its value > > afterward (rte_errno is only updated in case of error). > > - Assigning a local variable to store its value is not necessary as long as > > another function that may affect rte_errno is not called. > > > > Cons > > ---- > > > > - Not fully reentrant, thread-safety is fine for most purposes but signal > > handlers affecting it still cause undefined behavior (they must at least > > save and restore its value in case they modify it). > > - Accessing non-local storage may affect CPU cycle-sensitive functions such > > as TX/RX burst. > > Actually testing for errors mean you also have to reset the rte_errno > variable before. That also means you have to access thread-local > storage twice. > Not true. Your return value still indicates an error via a single sentinal value. Only in that case do you (the app) access the global value, to find out the exact error reason. > Besides that the problem of rte_errno is that you do error handling > twice because the implementation still needs to check for the error > condition before assigning a meaningful error value to rte_errno. > After that again the user code needs to check for the return value to > decide if looking at rte_errno makes any sense. > Yes, in the case of an error occuring there will be an extra write to a global variable, and a subsequent read from that value (which should not be a problem, as the write will have occurred in the same thread). However, this is irrelevant to normal path processing. Error should be the exception not the rule. > > > My opinion is that rte_errno is best for control path operations while using > > the return value makes more sense in the data path. The major issue being > > that function returning anything other than int (e.g. TX/RX burst) cannot > > describe any kind of error to the application. > > > > I went with both in rte_flow (return + rte_errno) mostly due to the return > > type of a few functions (e.g. rte_flow_create()) and wanted to keep the API > > consistent while maintaining compatibility with other DPDK APIs. Note there > > is little overhead for API functions to set rte_errno _and_ return its > > value, it's mostly free +1, and error cases should be rare, even if there is a small cost. . > > > > I think using both is best also because it leaves applications the choice of > > error-handling method, however if I had to pick one I'd go with rte_errno > > and standardize on -1 as the default error value (as in the C library). > > +1 though I think the sentinal value will vary depending on each case. I would look to keep the standard packet rx/tx functions and ones like them returning a zero on any error, to simplify programming logic, and also because in many cases the only real causes of error they can produce is from bad parameters. Functions returning pointers obviously will use NULL as error value. > > Below are a bunch of use-case examples to illustrate how rte_errno could > > be convenient to applications. > > > > Easily creating many flow rules during init in a all-or-nothing fashion: > > > > rte_errno = 0; > > for (i = 0; i != num; ++i) > > rule[i] = rte_flow_create(port, ...); > > if (unlikely(rte_errno)) { > > rte_flow_flush(port); > > return -1; > > } > > > > Complete TX packet burst failure with explanation (could also detect partial > > failures by initializing rte_errno to 0): > > > > sent = rte_eth_tx_burst(...); > > if (unlikely(!sent)) { > > switch (rte_errno) { > > case E2BIG: > > // too many packets in burst > > ... > > case EMSGSIZE: > > // first packet is too large > > ... > > case ENOBUFS: > > // TX queue is full > > ... > > } > > } > > > > TX burst functions in PMDs could be modified as follows with minimal impact > > on their performance and no ABI change: > > > > uint16_t sent = 0; > > int error; // new variable > > > > [process burst] > > if (unlikely([something went wrong])) { // this check already exists > > error = EPROBLEM; // new assignment > > goto error; // instead of "return sent" > > } > > [process burst] > > return sent; > > error: > > rte_errno = error; > > return sent; > > > > -- > > Adrien Mazarguil > > 6WIND