On 2019/11/21 12:52, Hugh Dickins wrote: > On Thu, 21 Nov 2019, zhengbin (A) wrote: >> On 2019/11/20 23:45, Matthew Wilcox wrote: >>> On Wed, Nov 20, 2019 at 10:23:18PM +0800, zhengbin wrote: >>>> I have tried to change last_ino type to unsigned long, while this was >>>> rejected, see details on https://patchwork.kernel.org/patch/11023915. >>> Did you end up trying sbitmap? >> Maybe sbitmap is not a good solution, max_inodes of tmpfs are controlled by mount options--nrinodes, >> >> which can be modified by remountfs(bigger or smaller), as the comment of function sbitmap_resize says: >> >> * Doesn't reallocate anything. It's up to the caller to ensure that the new >> * depth doesn't exceed the depth that the sb was initialized with. >> >> We can modify this to meet the growing requirements, there will still be questions as follows: >> >> 1. tmpfs is a ram filesystem, we need to allocate sbitmap memory for sbinfo->max_inodes(while this maybe huge) >> >> 2.If remountfs changes max_inode, we have to deal with it, while this may take a long time >> >> (bigger: we need to free the old sbitmap memory, allocate new memory, copy the old sbitmap to new sbitmap >> >> smaller: How do we deal with it?ie: we use sb->map[inode number/8] to find the sbitmap, we need to change the exist >> >> inode numbers?while this maybe used by userspace application.) >> >>> What I think is fundamentally wrong with this patch is that you've found a >>> problem in get_next_ino() and decided to use a different scheme for this >>> one filesystem, leaving every other filesystem which uses get_next_ino() >>> facing the same problem. >>> >>> That could be acceptable if you explained why tmpfs is fundamentally >>> different from all the other filesystems that use get_next_ino(), but >>> you haven't (and I don't think there is such a difference. eg pipes, >>> autofs and ipc mqueue could all have the same problem. >> tmpfs is same with all the other filesystems that use get_next_ino(), but we need to solve this problem one by one. >> >> If tmpfs is ok, we can modify the other filesystems too. Besides, I do not recommend all file systems share the same >> >> global variable, for performance impact consideration. >> >>> There are some other problems I noticed, but they're not worth bringing >>> up until this fundamental design choice is justified. >> Agree, thanks. > Just a rushed FYI without looking at your patch or comments. > > Internally (in Google) we do rely on good tmpfs inode numbers more > than on those of other get_next_ino() filesystems, and carry a patch > to mm/shmem.c for it to use 64-bit inode numbers (and separate inode > number space for each superblock) - essentially, > > ino = sbinfo->next_ino++; > /* Avoid 0 in the low 32 bits: might appear deleted */ > if (unlikely((unsigned int)ino == 0)) > ino = sbinfo->next_ino++; > > Which I think would be faster, and need less memory, than IDA. > But whether that is of general interest, or of interest to you, > depends upon how prevalent 32-bit executables built without > __FILE_OFFSET_BITS=64 still are these days. So how google think about this? inode number > 32-bit, but 32-bit executables cat not handle this? "separate inode number space for each superblock" can reduce the probability, but still can not solve it. > > Hugh
