> > >>> Last alloc bulk version have a bug, when page_owner is on, system maybe >>> crashed due to alloc bulk invoke prep_new_page in irq disabled context, >>> and set_page_owner use page_flag(may GFP_KERNEL) to get memory and save >>> stacktrace. >>> We fix it by a circumvention plan -- bandon alloc bulk feature when >>> page_owner is set. >>> I think both alloc_bulk and page_owner is valuable, so, it's worth to >>> find a way enable alloc bulk when page owner is on. >> >>Why do you think it's valuable? The point of alloc_bulk is speed. And ok. if you think in debug version use alloc bulk is not valuable. But I think alloc buk used prep_new_page's way reflect that debug feature shouldn't assume they can used the same way to get memory, just like the page it traced. If we not give debugger some info(alloc_gfp), they have no way, but to use page's gfp to get memory.(alloc bulk show it's wrong.) For now, only PAGE_OWNER get memory in prep_new_page, for future, may have more feature need memory. And may have more feature invoke prep_new_page in different context to the page. So, give a alloc_gfp to lead debugger to alloc memory is valuable. >>Doing the allocation of a stack for each page removes the speed. > >>Where's the value? >Without alloc bulk, we also need to alloc pages from loop invoke >alloc_page, and this also will invoke set_page_owner to save stack. >So, find a way to let alloc_bulk work, I think is valuable. >If you don't think so, I sorry for my disturb. > >>If you're really interested in making both these things work together, >>then I observed that all of these pages are going to have the same >>allocation stack. Allocating that stack once per call to alloc_bulk, >>reference counting it and having each page use the shared stack would > >>be a good way of combining the two (it'd even save memory!) >Yes, your idea is very good. > >But, how we let set_page_owner knew that recent pages are all in same stack? >Maybe save stack by caller, and passed handle in prep_new_page? >But, I think caller shouldn't care of page trace, it's page owner feature's duties. >Let same stack's page have same magic number maybe a good way, emm >But, even they have same stack, set_page_owner should know how to get memory >rightly (Unless you pass it to, so back to the duties). >In the past, we just pass page gfp to it, in alloc bulk, not right. > >>This patch, I think, does not make sense. >Well, this is a simple way that I can find, and not affect perfermance both in normal and >PAGE_OWNER is on, but can work alloc bulk and PAGE_OWNER. >If you don't think so, I'm sorry for that > >
