On Mon, Oct 23, 2017 at 5:06 AM, David Laight <David.Laight@xxxxxxxxxx> wrote: > From: Jim Quinlan >> Sent: 20 October 2017 16:28 >> On Fri, Oct 20, 2017 at 10:57 AM, Christoph Hellwig <hch@xxxxxx> wrote: >> > On Fri, Oct 20, 2017 at 10:41:56AM -0400, Jim Quinlan wrote: >> >> I am not sure I understand your comment -- the size of the request >> >> shouldn't be a factor. Let's look at your example of the DMA request >> >> of 3fffff00 to 4000000f (physical memory). Lets say it is for 15 >> >> pages. If we block out the last page [0x3ffff000..0x3fffffff] from >> >> what is available, there is no 15 page span that can happen across the >> >> 0x40000000 boundary. For SG, there can be no merge that connects a >> >> page from one region to another region. Can you give an example of >> >> the scenario you are thinking of? >> > >> > What prevents a merge from say the regions of >> > 0....3fffffff and 40000000....7fffffff? >> >> Huh? [0x3ffff000...x3ffffff] is not available to be used. Drawing from >> the original example, we now have to tell Linux that these are now our >> effective memory regions: >> >> memc0-a@[ 0....3fffefff] <=> pci@[ 0....3fffefff] >> memc0-b@[100000000...13fffefff] <=> pci@[ 40000000....7fffefff] >> memc1-a@[ 40000000....7fffefff] <=> pci@[ 80000000....bfffefff] >> memc1-b@[300000000...33fffefff] <=> pci@[ c0000000....ffffefff] >> memc2-a@[ 80000000....bfffefff] <=> pci@[100000000...13fffefff] >> memc2-b@[c00000000...c3fffffff] <=> pci@[140000000...17fffffff] >> >> This leaves a one-page gap between phsyical memory regions which would >> normally be contiguous. One cannot have a dma alloc that spans any two >> regions. This is a drastic step, but I don't see an alternative. >> Perhaps I may be missing what you are saying... > > Isn't this all unnecessary? > Both kmalloc() and dma_alloc() are constrained to allocate memory > that doesn't cross an address boundary that is larger than the size. > So if you allocate 16k it won't cross a 16k physical address boundary. > > David > Hi David, Christoph was also concerned about this: "For the block world take a look at __blk_segment_map_sg which does the merging of contiguous pages into a single SG segment. You'd have to override BIOVEC_PHYS_MERGEABLE to prevent this from happening in your supported architectures for the block layer." Do you consider this a non-issue as well or can this happen and span memory regions?
