On Wed, May 11, 2022 at 02:37:24PM +0200, Christoph Hellwig wrote: > On Wed, May 11, 2022 at 10:10:40AM +0000, Conor.Dooley@xxxxxxxxxxxxx wrote: > > Without even trying the patch, I double checked the boot log from > > 3f70356edf56 and I get a "software IO TLB: Cannot allocate buffer" > > With the patch its a "software IO TLB: swiotlb_init_remap: failed > > to allocate tlb structure". So spot on & I feel like an idiot for > > not spotting that before! > > > > Is failing being fatal valid, or should it fail gracefully like it > > used to do? To me, blissfully unaware about swiotlb, the "current" > > behaviour of failing gracefully makes more sense. > > Given that we're at -rc6 I think the most important thing for now is to > avoid a regression and restore the old behavior. I'll send out a > series with this and the nslab related fixes for Xen today. > > But we should look into why allocating the memory fails for your > plaforms. Does it have very little memory? I can't really think > of why else the memblock allocation for swiotlb would fail. I guess the default to use memblock_alloc_low() backfires on system with physical memory living at 0x1000200000: [ 0.000000] Early memory node ranges [ 0.000000] node 0: [mem 0x0000001000200000-0x000000103fffffff] The default limit for "low" memory is 0xffffffff and there is simply no memory there. -- Sincerely yours, Mike.