I'd be much more comfortable doing this once direct-style IO lands in Mirage. mmap is a _really_ POSIX-oriented abstraction, and exposing it in all Mirage_block.S will add a fair bit of complexity to any implementations. It's true that it's simpler if you just restrict it to read-only, but there is still the problem of when to unmap something.
Note that access to the data is still determined by an underlying scheduler -- but instead of Lwt or the OCaml runtime, you are now depending on the page table subsystem to do this decision making for you.
Do you have an example of a usecase you want to solve here -- is it for fairly small pieces of encrypted data?
Anil
On 7 Dec 2022, at 14:39, Romain Calascibetta <romain.calascibetta@xxxxxxxxx> wrote:
Hi,
It's not a real `mmap` but more a `read()` without `Lwt`. In the case when we limit the access to the block (only for reading), it's fine to provide a `read()` without a scheduling idea mainly because whatever what we do with the block, it's a read-only block and data will be the same all the time. The idea behind that is to unlock the ability to create a read-only file-system and where the access of datas will not be determined by a underlying scheduler. A new signature like: ```ocaml module type Mirage_block.RD = sig type t
val read : offset:int64 -> Cstruct.t end ```
Will allow us to make a `Mirage_kv.RO` and be able to compose such layout (the file-system layout) with something else (like `ccm_block`).
Best,
On 30 Nov 2022, at 15:13, Hannes Mehnert < hannes@xxxxxxxxxxx> wrote:
## `mmap` available on `Mirage_block.S` (dinosaure, https://github.com/mirage/mirage-block/issues/53)- dinosaure has an implementation to get a part of the block (similar to mmap), without being in the Lwt monad- at the moment, read is in Lwt.t, i.e. does not block, but returns the filled page(s)- dinosaure needs a blocking function that returns the data- the solo5 interface is already blocking (and synchronous), mirage-block-solo5 adds the asynchronous stuff- christiano mentions that it could be done with locking- maybe develop a block read-only interface with a synchronous read
In general, having "automatic" scheduling via mmap is a bad idea for anything non-trivial, since you slow to a crawl when under memory pressure and having a lot of page faults. There's no way a caller can determine whether or not a set of accesses will result in a blocking fetch or not.
It may be workable for a read-only mmap, but... why do you want it? To get out of Lwt allocations?
Anil
--
|
