Date: Wed, 4 Nov 2015 23:17:30 +0100 From: Svatopluk Kraus <onwahe@gmail.com> To: Jason Harmening <jason.harmening@gmail.com> Cc: Ian Lepore <ian@freebsd.org>, Ganbold Tsagaankhuu <ganbold@gmail.com>, src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: Re: svn commit: r289759 - in head/sys/arm: arm include Message-ID: <CAFHCsPWE3YLcO_GKL7pGkB5wU342WDZ5T695S_DPG7eBQNmqaQ@mail.gmail.com> In-Reply-To: <CAM=8qa=GX3yZmcXmMZz-nfmwOFhXyyKh_6%2BLVR=y7P5pURYvpQ@mail.gmail.com> References: <201510221638.t9MGc1cc053885@repo.freebsd.org> <CAGtf9xOx_iSNqsPT7OdTmAy-ER_wFzOC9YxFigpQ%2B0wLj7=ytQ@mail.gmail.com> <56348FF8.3010602@gmail.com> <CAM=8qa=4EqPvRoUnMytQBe32GtvPRZRE_s34tOzmVNCLe8c=Fw@mail.gmail.com> <1446394311.91534.236.camel@freebsd.org> <CAM=8qa=GX3yZmcXmMZz-nfmwOFhXyyKh_6%2BLVR=y7P5pURYvpQ@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, Nov 3, 2015 at 8:45 AM, Jason Harmening <jason.harmening@gmail.com> wrote: > > On Sun, Nov 1, 2015 at 8:11 AM, Ian Lepore <ian@freebsd.org> wrote: >> >> >> It's almost certainly not related to sysinit ordering. This exception >> is happening during mmc probing after interrupts are enabled. >> >> It appears that the problem is the faulting code is running on one of >> the very early pre-allocated kernel stacks (perhaps in an interrupt >> handler on an idle thread stack), and these stacks are not in memory >> represented in the vm page arrays. The mmc code is using a 64-byte >> buffer on the stack and mapping it for DMA. Normally that causes a >> bounce for cacheline alignment, but unluckily in this case that buffer >> on the stack just happened to be aligned to a cacheline boundary and a >> multiple of the cacheline size, so no bounce. That causes the new sync >> logic that is based on keeping vm_page_t pointers and offsets to get a >> NULL pointer back from PHYS_TO_VM_PAGE when mapping, then it dies at >> sync time trying to dereference that. It used to work because the sync >> logic used to use the vaddr, not a page pointer. >> >> Michal was working on a patch yesterday. >> > > Ah, thanks for pointing that out Ian. I was left scratching my head > (admittedly on the road and w/o easy access to the code) wondering what on > earth would be trying to do DMA during SI_SUB_CPU. > Using of fictitious pages is not so easy here as in case pointed by kib@ where they are allocated and freed inside one function. For sync list sake, they must be allocated when a buffer is loaded and freed when is unloaded. Michal uses pmap_kextract() in case when KVA of buffer is not zero in his proof-of-concept patch: https://gist.github.com/strejda/d5ca3ed202427a2e0557 When KVA of buffer is not zero, temporary mapping is not used at all, so vm_page_t array is not needed too. IMO, buffer's physical address can be saved in sync list to be used when its KVA is not zero. Thus pmap_kextract() won't be called in dma_dcache_sync().
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAFHCsPWE3YLcO_GKL7pGkB5wU342WDZ5T695S_DPG7eBQNmqaQ>