[Devel] Re: BC: resource beancounters (v6) (with userpages reclamation + configfs)

[Herbert Poetzl]

On Thu, Nov 09, 2006 at 07:49:28PM +0300, Kirill Korotaev wrote:
> MAJOR CHANGES in v6 (see details below):
> - configfs interface instead of syscalls (as wanted by CKRM people...)
> - added numfiles resource accounting
> - added numtasks resource accounting
> 
> numfiles and numtasks controllers demonstrate how
> clean and simple BC interface is.
> 
> Patch set is applicable to 2.6.19-rc5-mm1
> 
> -----------------------------------------------------
> 
> Resource BeanCounters (BC).
> 
> BC allows to account and control consumption
> of kernel resources used by *group* of processes
> (users, containers, ...).
> 
> Draft BC description on OpenVZ wiki can be found at
> http://wiki.openvz.org/UBC_parameters
> 
> The full BC patch set allows to control:
> - kernel memory. All the kernel objects allocatable
> on user demand and not reclaimable should be accounted and
> limited for DoS protection.
> e.g. page tables, task structs, vmas etc.

kernel memory is not accounted on a linux system
right now, and user can probably quite easily use
that for DoS ... shouldn't that become a separate
user limit in the first place?

> - virtual memory pages. BCs allow to
> limit a container to some amount of memory and
> introduces 2-level OOM killer taking into account
> container's consumption.
> pages shared between containers are correctly
> charged as fractions (tunable).

how much overhead does this add to the memory/page
management? do we really want to account shared pages
at all, if so, why is accounting a fraction to each
group the 'proper' way to do so? (would expect to
account it to all of them equally)

IMHO it would be sufficient to extend existing page
accounting to a 'memory' namespace isntead of adding
another complex mechanism like the beancounters

> - network buffers. These includes TCP/IP rcv/snd
> buffers, dgram snd buffers, unix, netlinks and
> other buffers.
> 
> - minor resources accounted/limited by number:
> tasks, files, flocks, ptys, siginfo, pinned dcache
> mem, sockets, iptentries (for containers with
> virtualized networking)
> 
> Summary of changes from v5 patch set:
> * configfs interface instead of syscalls (as wanted by CKRM people)
> * added numfiles resource accounting
> * added numtasks resource accounting
> * introduced dummy_resource to handle case when
>   no resource registered
> * calls to rss accounting are integrated to rmap calls

again, how much overhead does this add?
can you provide some numbers/tests here?

TIA,
Herbert

> Summary of changes from v4 patch set:
> * changed set of resources - kmemsize, privvmpages, physpages
> * added event hooks for resources (init, limit hit etc)
> * added user pages reclamation (bc_try_to_free_pages)
> * removed pages sharing accounting - charge to first user
> * task now carries only one BC pointer, simplified
> * make set_bcid syscall move arbitrary task into BC
> * resources are not recharged when task moves
> * each vm_area_struct carries a BC pointer
> 
> Summary of changes from v3 patch set:
> 
> * Added basic user pages accounting (lockedpages/privvmpages)
> * spell in Kconfig
> * Makefile reworked
> * EXPORT_SYMBOL_GPL
> * union w/o name in struct page
> * bc_task_charge is void now
> * adjust minheld/maxheld splitted
> 
> Summary of changes from v2 patch set:
> 
> * introduced atomic_dec_and_lock_irqsave()
> * bc_adjust_held_minmax comment
> * added __must_check for bc_*charge* funcs
> * use hash_long() instead of own one
> * bc/Kconfig is sourced from init/Kconfig now
> * introduced bcid_t type with comment from Alan Cox
> * check for barrier <= limit in sys_set_bclimit()
> * removed (bc == NULL) checks
> * replaced memcpy in beancounter_findcrate with assignment
> * moved check 'if (mask & BC_ALLOC)' out of the lock
> * removed unnecessary memset()
> 
> Summary of changes from v1 patch set:
> 
> * CONFIG_BEANCOUNTERS is 'n' by default
> * fixed Kconfig includes in arches
> * removed hierarchical beancounters to simplify first patchset
> * removed unused 'private' pointer
> * removed unused EXPORTS
> * MAXVALUE redeclared as LONG_MAX
> * beancounter_findcreate clarification
> * renamed UBC -> BC, ub -> bc etc.
> * moved BC inheritance into copy_process
> * introduced reset_exec_bc() with proposed BUG_ON
> * removed task_bc beancounter (not used yet, for numproc)
> * fixed syscalls for sparc
> * added sys_get_bcstat(): return info that was in /proc
> * cond_syscall instead of #ifdefs
> 
> Many thanks to Oleg Nesterov, Alan Cox, Matt Helsley and others
> for patch review and comments.
> 
> Thanks,
> Kirill
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo at vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/