Messages in this thread | | | Date | Thu, 27 Nov 2014 17:39:28 +0100 | From | Richard Weinberger <> | Subject | Re: [PATCH 2/6] UBI: Fastmap: Ensure that only one fastmap work is scheduled |
| |
Am 27.11.2014 um 17:35 schrieb Artem Bityutskiy: > On Thu, 2014-11-27 at 17:13 +0100, Richard Weinberger wrote: >> Am 27.11.2014 um 16:27 schrieb Artem Bityutskiy: >>> On Mon, 2014-11-24 at 14:20 +0100, Richard Weinberger wrote: >>>> If the WL pool runs out of PEBs we schedule a fastmap write >>>> to refill it as soon as possible. >>>> Ensure that only one at a time is scheduled otherwise we might end in >>>> a fastmap write storm because writing the fastmap can schedule another >>>> write if bitflips are detected. >>> >>> Could you please provide some more details about the "write storm". Does >>> it happen when there are 2 fastmap works in the queue? Or if they run >>> simultaneously? Why the storm happens and white kind of "writes" it >>> consists of? >> >> If UBI needs to write a new fastmap while wear leveling (by using get_peb_for_wl()) >> a fastmap work is scheduled. >> We cannot write a fastmap in this context because we're in atomic context. >> At this point one fastmap write is scheduled. If now get_peb_for_wl() is executed >> a second time it will schedule another fastmap work because the pools are still not refilled. > > Sounds like just you do not need any works and any queues at all. All > you need is a "please, fastmap me!" flag. > > Then this flag should be checked every time we enter the background > thread or the fastmap code, and be acted upon. > > So the background thread would first check this flag, and if it is set - > call the fastmap stuff. The go do the WL works. > > Just off-the top of my head, take with grain of salt.
So you want me to redesign it? IMHO this is just a matter of taste.
Face it, my brain does not work like yours. I design things differently.
Thanks, //richard
| |