On Sun, 2008-03-09 at 18:42 +0300, Michael Tokarev wrote: > Michael Tokarev wrote: > > James Bottomley wrote: > >> On Sun, 2008-03-09 at 21:29 +0900, FUJITA Tomonori wrote: > >>> On Sun, 09 Mar 2008 14:23:13 +0300 > >>> Michael Tokarev <mjt@xxxxxxxxxx> wrote: > >>> > >>>> Just got quite.. bad situation on a production server > >>>> here. The machine locked up hard several times in a > >>>> row (required hard reboot). So I finally enabled watchdog > >>>> subsystem which helped. > >>>> > >>>> Now I see the following (over netconsole): > >>>> > >>>> DMA: Out of SW-IOMMU space for 65536 bytes at device 0000:08:07.0 > >>>> ------------[ cut here ]------------ > >>>> kernel BUG at drivers/scsi/aic7xxx/aic79xx_osm.c:1490! > >>> Seems that you was out of swiommu space (and aic79xx can't handle it > >>> though it should). This happened because: > >>> > >>> a) you produced more I/Os than swiommu can handle. > >>> > >>> b) swiommu space leaks due to bugs. > >>> > >>> If you hit this problem due to a), the following boot option might > >>> help: > >>> > >>> swiotlb=65536 > > > > Running with this parameter now - no lockups so far. > > > >> Actually, it's worse than this. The aic79xx is a fully 64 bit capable > >> PCI card, it shouldn't be using the iommu at all. However, it has three > >> DMA modes: 64 bit, 39 bit and 32 bit; with a corresponding resource > >> cost increasing with the number of bits. It employs special APIs to > >> size the masks according to the memory, in aic79xx_osm_pci.c: > > [] > >> Could you firstly tell me how much memory you have, and secondly > >> instrument this code with the patch below to see if we can work out what > >> it's doing? > > > > The memory map is below (6Gb total). The patch - kernel is being compiled > > right now. > > And here's the result (without swiotlb=65536): > > DEBUG: RETURNED REQUIRED MASK ffffffff > DEBUG: SET 32 BIT ADDRESSING > > (which doesn't look like a good thing, provided this > machine has 6Gb of memory...) That's the root cause then. There's a bug in the generic implementation of dma_get_required_mask(), a fix for which is below, if you could try it (still with the debugging patches to make sure it's working). James --- diff --git a/drivers/base/platform.c b/drivers/base/platform.c index efaf282..911ec60 100644 --- a/drivers/base/platform.c +++ b/drivers/base/platform.c @@ -648,7 +648,7 @@ u64 dma_get_required_mask(struct device *dev) high_totalram += high_totalram - 1; mask = (((u64)high_totalram) << 32) + 0xffffffff; } - return mask & *dev->dma_mask; + return mask; } EXPORT_SYMBOL_GPL(dma_get_required_mask); #endif -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
