On Sun, Mar 24, 2019 at 03:29:04PM +0100, Thomas Gleixner wrote: >Wei, > >On Tue, 12 Feb 2019, Wei Yang wrote: >> >> This patch changes the implementation from the first perception to the >> second to reduce one different handling on end_pfn. After doing so, the >> code is easier to read. > >It's maybe slightly easier to read, but it's still completely unreadable >garbage. > > Not your fault, it was garbage before. > >But refining garbage still results in garbage. Just the smell is slightly >different. > >Why? > > 1) Doing all the calculations PFN based is just a pointless > indirection. Just look at all the rounding magic and back and forth > conversions. > > All of that can be done purely address/size based which makes the code > truly readable. > > 2) The 5(3) sections are more or less copied code with a few changes of > constants, except for the first section (A) which has an extra quirk > for 32bit. Plus all the 64bit vs. 32bit #ifdeffery which is not making > it any better. > > This copied mess can be avoided by using helper functions and proper > loops. > > 3) During the bootmem phase the code tries to preserve large mappings > _AFTER_ splitting them up and then it tries to merge the resulting > overlaps. > > This is completely backwards because the expansion of the range can be > tried right away when then mapping of a large page is attempted. Surely > not with the current mess, but with a proper loop based approach it can > be done nicely. > > Btw, there is a bug in that expansion code which could result in > undoing the enforced 4K mapping of the lower 2M/4M range on 32bit. It's > probably not an issue in practice because the low range is usually not > contiguous. But it works by chance not by design.
Hi, Thomas I want to confirm with you, this bug is in adjust_range_page_size_mask(), right? -- Wei Yang Help you, Help me