On 21/02/18 03:36, Dave Chinner wrote:
> On Tue, Feb 20, 2018 at 03:56:00PM -0800, Matthew Wilcox wrote:
>> On Wed, Feb 21, 2018 at 08:36:04AM +1100, Dave Chinner wrote:
>>> FWIW, I'm not wanting to use it to replace static variables. All the
>>> structures are dynamically allocated right now, and get assigned to
>>> other dynamically allocated pointers. I'd likely split the current
>>> structures into a "ro after init"
I would prefer to use a different terminology, because, if I have
understood the use case, this is not exactly the same as __ro_after_init
So, this is my understanding:
* "const" needs to be known at link time - there might be some
adjustments later on, ex: patching of "const" pointers, after relocation
has taken place - I am assuming we are not planning to patch const data
The compiler can perform whatever optimization it feels like and it is
allowed to do, on this.
* __ro_after_init is almost the same as a const, from a r/w perspective,
but it will become effectively read_only after the completion of the
init phase. The compiler cannot use it in any way to detect errors,
AFAIK. The system will just generate a runtime error is someone tries to
alter some __ro_after_init data, when it's read-only.
The only trick available is to use, after the protection, a different
type of handle, const.
* pmalloc pools can be protected (hence the "p") at any time, but they
start as r/w. Also, they cannot be declared statically.
* data which is either const or __ro_after_init is placed into specific
sections (on arm64 it's actually the same) and its pages are then marked
as read-only.
>>> structure and rw structure, so
>>> how does the "__ro_after_init" attribute work in that case? Is it
>>> something like this?
>>>
>>> struct xfs_mount {
>>> struct xfs_mount_ro{
>>> .......
>>> } *ro __ro_after_init;
> ^^^^^^^^
>
> pointer, not embedded structure....
I doubt this would work, because I think it's not possible to put a
field of a structure into a separate section, afaik.
__ro_after_init would refer to the ro field, not to the memory it refers to.
>>> ......
>>
>> No, you'd do:
>>
>> struct xfs_mount_ro {
>> [...]
>> };
is this something that is readonly from the beginning and then shared
among mount points or is it specific to each mount point?
>> struct xfs_mount {
>> const struct xfs_mount_ro *ro;
>> [...]
>> };
>
> .... so that's pretty much the same thing :P
The "const" modifier is a nice way to catch errors through the compiler,
iff the ro data will not be initialized through this handle, when it's
still writable.
>>> Also, what compile time checks are in place to catch writes to
>>> ro structure members? Is sparse going to be able to check this sort
>>> of thing, like is does with endian-specific variables?
>>
>> Just labelling the pointer const should be enough for the compiler to
>> catch unintended writes.
>
> Ok.
yes, anyway the first one trying to alter it at run time, is in for some
surprise.
>>>> I'd be interested to have your review of the pmalloc API, if you think
>>>> something is missing, once I send out the next revision.
>>>
>>> I'll look at it in more depth when it comes past again. :P
>>
>> I think the key question is whether you want a slab-style interface
>> or whether you want a kmalloc-style interface. I'd been assuming
>> the former, but Igor has implemented the latter already.
>
> Slabs are rally only useful when you have lots of a specific type of
> object. I'm concerned mostly about one-off per-mount point
> structures, of which there are relatively few. A heap-like pool per
> mount is fine for this.
That was my same sentiment.
Actually it would be even possible to simulate caches with pools: each
pool supports a granularity parameter, during creation. One could have
multiple pools, each with different granularity, but it would probably
lead to a proliferation of pools.
Instead, I preferred to have pmalloc as a drop-in replacement for the
variants of k/v/kv malloc.
The only real issue was the - previous - inability of tracking the size
of an allocation, given its address, but that is taken care of by the
patch for the genalloc bitmap.
If I could have a pointer to a good candidate for the pmalloc treatment,
I could come up with a patch, to show how it could be done.
Then it might be easier to discuss if the API needs to be modified
and/or extended somehow.
--
igor
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
