From mboxrd@z Thu Jan 1 00:00:00 1970 Message-ID: <47DD5A2C.3050606@domain.hid> Date: Sun, 16 Mar 2008 18:34:36 +0100 From: Philippe Gerum MIME-Version: 1.0 References: <47DD3BF9.2@domain.hid> <47DD3FB4.5080305@domain.hid> <47DD424A.7040006@domain.hid> <47DD47F0.5020003@domain.hid> <47DD4C70.2040208@domain.hid> <47DD4E1F.9050400@domain.hid> <47DD4FDA.8020105@domain.hid> <47DD53EF.9070302@domain.hid> <47DD55CF.80903@domain.hid> In-Reply-To: <47DD55CF.80903@domain.hid> Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Sender: Philippe Gerum Subject: Re: [Xenomai-core] pgprot_noncached for io-remapping? Reply-To: rpm@xenomai.org List-Id: "Xenomai life and development \(bug reports, patches, discussions\)" List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Jan Kiszka Cc: Xenomai-core@domain.hid Jan Kiszka wrote: > Philippe Gerum wrote: >> Jan Kiszka wrote: >>> Philippe Gerum wrote: >>>> Jan Kiszka wrote: >>>>> Philippe Gerum wrote: >>>>>> Jan Kiszka wrote: >>>>>>> Philippe Gerum wrote: >>>>>>>> Jan Kiszka wrote: >>>>>>>>> Hi, >>>>>>>>> >>>>>>>>> doesn't this patch [1] have some relevance for us as well? As we use >>>>>>>>> xnarch_remap_io_page_range also for non-IO memory, I'm hesitating to >>>>>>>>> suggest that we apply this unconditionally at xnarch level. Ideas welcome. >>>>>>>>> >>>>>>>> Yes, I think it makes a lot of sense on powerpc at least, since doing so will >>>>>>>> set the PAGE_GUARDED bit as well, and we obviously want to avoid any >>>>>>>> out-of-order access of I/O memory. >>>>>>>> >>>>>>>> (I don't see the reason to force the VM_RESERVED and VM_IO on the vma though, >>>>>>>> since remap_pfn_range will do it anyway.) >>>>>>> No, I was talking about cases where we may pass kmalloc'ed memory to >>>>>>> xnarch_remap_io_page_range. In that case, caching and out-of-order >>>>>>> access may be desired for performance reasons. >>>>>>> >>>>>> xnarch_remap_io_page_range is intended for I/O memory only, some assumptions are >>>>>> made on this. rtdm_mmap_buffer() should be fixed; it would be much better to >>>>>> define another internal interface at xnarch level to specifically perform >>>>>> kmalloc mapping. >>>>> Yeah, probably. But I think the issue is not just limited to RTDM. The >>>>> xnheap can be kmalloc-hosted as well. >>>>> >>>> This one is used with DMA memory. What I would suggest, is something like this: >>>> >>>> --- ksrc/skins/rtdm/drvlib.c (revision 3590) >>>> +++ ksrc/skins/rtdm/drvlib.c (working copy) >>>> @@ -1738,9 +1738,12 @@ >>>> } >>>> return 0; >>>> } else >>>> -#endif /* CONFIG_MMU */ >>>> return xnarch_remap_io_page_range(vma, maddr, paddr, >>>> size, PAGE_SHARED); >>>> +#else >>>> + return xnarch_remap_kmem_page_range(vma, maddr, paddr, >>>> + size, PAGE_SHARED); >>>> +#endif /* CONFIG_MMU */ >>>> } >>>> >>>> static struct file_operations rtdm_mmap_fops = { >>>> >>>> >>>> I.e. split the cases where MMU is absent from the one where MMU is there but we >>>> come from rtdm_iomap_to_user. >>> Makes no sense to me yet. With CONFIG_MMU we have 3 cases, not just two, >>> no? We have to use mmap_data->src_paddr to tell kmem_page apart from >>> io_page. >> That's what the patch I sent right after this one does. > > Nope, vaddr can be NULL in both cases, only paddr is differentiating. > > This one should do what we need: > > Index: ksrc/skins/rtdm/drvlib.c > =================================================================== > --- ksrc/skins/rtdm/drvlib.c (Revision 3594) > +++ ksrc/skins/rtdm/drvlib.c (Arbeitskopie) > @@ -1739,8 +1739,12 @@ static int rtdm_mmap_buffer(struct file > return 0; > } else > #endif /* CONFIG_MMU */ > + if (mmap_data->src_paddr) > return xnarch_remap_io_page_range(vma, maddr, paddr, > size, PAGE_SHARED); > + else > + return xnarch_remap_kmem_page_range(vma, maddr, paddr, > + size, PAGE_SHARED); > } > > static struct file_operations rtdm_mmap_fops = { > > Ok. So if that's fine with you, I think we should merge that because the current situation is error-prone. For the time being, I've defined a remap_kmem wrapper which mirrors the previous actions of the remap_io one, before I fixed the latter with the noncache protection bits. At least we should get the same behaviour than previously wrt to rtdm_mmap_buffer. This leaves some time to think about the kmem mapping modes without breaking the current situation, but they should be correct now. -- Philippe.