Re: [Crash-utility] Adding a new command rbtree

----- Original Message ----- Hello Qiao, We're getting closer, but there is still a problem with red-black tree dumps. For this command to be useful, the "tree" command should do what the "list" command does when dumping a list of structures that have embedded list_head members -- which is to dump the addresses of the data structures in the list. In the case of radix tree lists, your patch does the right thing, which is to dump the addresses of the items found in the lowest radix_tree_node slots[] array. Taking an address_space address of ffff8805f5a73580, the address_space.page_tree radix tree dump shows the page struct addresses, and therefore the "-s page.mapping" option works -- confirming that each page struct has a reference back to the address_space: crash> tree -t radix -r address_space.page_tree ffff8805f5a73580 -s page.mapping ffffea00178472c0 mapping = 0xffff8805f5a73580 ffffea0017869240 mapping = 0xffff8805f5a73580 ffffea0017869280 mapping = 0xffff8805f5a73580 ffffea00178692c0 mapping = 0xffff8805f5a73580 ffffea0017869300 mapping = 0xffff8805f5a73580 ffffea0017869340 mapping = 0xffff8805f5a73580 ffffea0017869380 ... But in the case of red-black trees, your patch is not consistent, because it dumps the addresses of the embedded rb_node structure in each data structure in the list. A user would typically want to see the addresses of the containing data structure, *not* the embedded rb_node addresses. For example, suppose we want to see the list of vm_area_structs that are members of the red-black tree whose root is in an mm_struct. Using your latest patch, this is what I see, where the mm_struct address is ffff880828163180: crash> vm | head -3 PID: 8426 TASK: ffff88042bd9cd00 CPU: 3 COMMAND: "crash" MM PGD RSS TOTAL_VM ffff880828163180 ffff880828f7a000 193048k 344404k crash> Therefore, this command should dump the address of each vm_area_struct in the tree: crash> tree -t rb -r mm_struct.mm_rb -o vm_area_struct.vm_rb ffff880828163180 ffff880828a82d80 ffff880828a82798 ffff880828a83758 ffff880828a81bc8 ffff880828a83560 ffff880828a821b0 ffff88081c5a57d8 ffff88081c5a7170 ffff880828a80038 ffff880828a80230 ... However, the addresses above are the addresses of the vm_area_struct.vm_rb rb_node members -- not the addresses of the containing vm_area_structs. Confusing the issue even more, if I remove the "-o vm_area_struct.vm_rb" option from the command line, I see the exact same output as above: crash> tree -t rb -r mm_struct.mm_rb ffff880828163180 ffff880828a82d80 ffff880828a82798 ffff880828a83758 ffff880828a81bc8 ffff880828a83560 ffff880828a821b0 ffff88081c5a57d8 ffff88081c5a7170 ffff880828a80038 ffff880828a80230 ... For red-black trees, the "-o offset" option is always required unless the rb_node member in the containing data structure has a member offset of 0. And presuming that I want a list of vm_area_struct addresses, your patch doesn't do that. I do note that "-s struct[.member,member]" will work correctly if the "-o vm_area_struct.vm_rb" is applied, so you're on the right track. But again, the red-black tree dump should be similar to the radix tree dump, and both of them should be similar to the "list" command. Another minor nit -- you should disallow the usage of "-o offset" when a radix tree is being dumped. Although your patch apparently ignores it, it should display an error message and fail the command. Thanks, Dave