On Thu, Nov 9, 2017 at 10:13 AM, Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote: > > > On Thu, 9 Nov 2017, Dan Williams wrote: > >> On Thu, Nov 9, 2017 at 8:37 AM, Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote: >> > >> > >> > On Wed, 8 Nov 2017, Dan Williams wrote: >> > >> >> On Wed, Nov 8, 2017 at 12:26 PM, Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote: >> >> > On Wed, 8 Nov 2017, Christoph Hellwig wrote: >> >> > >> >> >> Can you start by explaining what you actually need the vmap for? >> >> > >> >> > It is possible to use lvm on persistent memory. You can create linear or >> >> > striped logical volumes on persistent memory and these volumes still have >> >> > the direct_access method, so they can be mapped with the function >> >> > dax_direct_access(). >> >> > >> >> > If we create logical volumes on persistent memory, the method >> >> > dax_direct_access() won't return the whole device, it will return only a >> >> > part. When dax_direct_access() returns the whole device, my driver just >> >> > uses it without vmap. When dax_direct_access() return only a part of the >> >> > device, my driver calls it repeatedly to get all the parts and then >> >> > assembles the parts into a linear address space with vmap. >> >> >> >> I know I proposed "call dax_direct_access() once" as a strawman for an >> >> in-kernel driver user, but it's better to call it per access so you >> >> can better stay in sync with base driver events like new media errors >> >> and unplug / driver-unload. Either that, or at least have a plan how >> >> to handle those events. >> > >> > Calling it on every access would be inacceptable performance overkill. How >> > is it supposed to work anyway? - if something intends to move data on >> > persistent memory while some driver accesse it, then we need two functions >> > - dax_direct_access() and dax_relinquish_direct_access(). The current >> > kernel lacks a function dax_relinquish_direct_access() that would mark a >> > region of data as moveable, so we can't move the data anyway. >> >> We take a global reference on the hosting device while pages are >> registered, see the percpu_ref usage in kernel/memremap.c, and we hold >> the dax_read_lock() over calls to dax_direct_access() to temporarily >> hold the device alive for the duration of the call. > > If would be good if you provided some function that locks down persistent > memory in the long-term. Locking it on every access just kills performance > unacceptably. > > For changing mapping, you could provide a callback. When the callback is > called, the driver that uses persistent memory could quiesce itself, > release the long-term lock and let the system change the mapping. I'll take a look at this. It dovetails with some of the discussions we are having about how to support RDMA to persistent memory and notification/callback to tear down memory registrations. >> While pages are pinned for DMA the devm_memremap_pages() mapping is >> pinned. Otherwise, an error reading persistent memory is identical to >> an error reading DRAM. > > The question is if storage controllers and their drivers can react to this > in a sensible way. Did someone test it? The drivers don't need to react, once the pages are pinned for dma the hot-unplug will not progress until all those page references are dropped. -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
