From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Simmons, James A." Date: Fri, 01 May 2015 20:18:56 +0000 Subject: RE: [HPDD-discuss] [PATCH 2/11] Staging: lustre: fld: Use kzalloc and kfree Message-Id: <524505df3433441494cf082a425f2ee7@EXCHCS32.ornl.gov> List-Id: References: <1430495482-933-1-git-send-email-Julia.Lawall@lip6.fr> <1430495482-933-11-git-send-email-Julia.Lawall@lip6.fr> In-Reply-To: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: 'Julia Lawall' Cc: Oleg Drokin , "devel@driverdev.osuosl.org" , Greg Kroah-Hartman , "kernel-janitors@vger.kernel.org" , "linux-kernel@vger.kernel.org" , "HPDD-discuss@lists.01.org" >> >From: Julia Lawall >> > >> >Replace OBD_ALLOC, OBD_ALLOC_WAIT, OBD_ALLOC_PTR, and OBD_ALLOC_PTR_WAIT by >> >kalloc/kcalloc, and OBD_FREE and OBD_FREE_PTR by kfree. >> >> Nak: James Simmons >> >> A simple replace will not work. The OBD_ALLOC and OBD_FREE functions allocate memory >> anywhere from one page to 4MB in size. You can't use kmalloc for the 4MB allocations. >> Currently lustre uses a 4 page water mark to determine if we allocate using vmalloc. Even >> using kmalloc for 4 pages has shown high failure rates on some systems. It gets even more >> messy with 64K page systems like ppc64 boxes. Now I'm not suggesting to port the larger >> allocations to vmalloc either since issues have been founded with using vmalloc. For example >> when using large stripe count files the MDS rpc generated crosses the 4 page line and vmalloc >> is used. Using vmalloc caused a global spinlock to be taken which causes meta data operations >> to serialized on the MDS servers. > >It's not the LARGE functions that do the switching? For example OBD_ALLOC >ends up at __OBD_MALLOC_VERBOSE, which as far as I can see calls kmalloc >(with __GFP_ZERO, and hance the use of kzalloc). Yes the LARGE functions do the switching. I was expecting also patches to remove the OBD_ALLOC_LARGE functions as well which is not the case here. I do have one question still. The macro __OBD_MALLOC_VERBOSE allowed the ability to simulate memory allocation failures at a certain percentage rate. Does something exist in the kernel to duplicate that functionality? Once these macros are gone we lose the ability to simulate high memory allocation failures.