[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20220924085509.zftbujx224aam5hd@gmail.com>
Date: Sat, 24 Sep 2022 01:55:09 -0700
From: Fangrui Song <i@...kray.me>
To: James Y Knight <jyknight@...gle.com>, musl@...ts.openwall.com,
Florian Weimer <fweimer@...hat.com>, Rich Felker <dalias@...c.org>,
baiyang <baiyang@...il.com>
Subject: Re: The heap memory performance (malloc/free/realloc) is
significantly degraded in musl 1.2 (compared to 1.1)
On 2022-09-20, Szabolcs Nagy wrote:
>* James Y Knight <jyknight@...gle.com> [2022-09-20 12:59:00 -0400]:
>
>> On Tue, Sep 20, 2022 at 9:58 AM Siddhesh Poyarekar <siddhesh@...hat.com>
>> wrote:
>>
>> > Adding support for something that's already declared as bad
>> > programming practice seems like a step backwards. Instead, I hope we
>> > find a way to discourage active use of malloc_usable_size more
>> > strongly.
>>
>>
>> BTW, if folks aren't aware, there is already work on the C++ side to expose
>> an API which lets you request a heap allocation of _at least_ the given
>> size, which rounds the actual size up in whatever way the allocator likes,
>> and returns the pointer and actual size allocated. With this API, you
>> declare an explicit intent that all of the memory -- up to the returned
>> size -- is valid to use without needing to go back to the allocator to ask
>> for more.
>>
>> The proposal is still making its way through the standardization process,
>> but hopefully it'll make it into the next version of C++ after C++23. (Of
>> course, that's not a sure thing until it happens.) Here's the doc, with
>> more rationale/etc:
>> https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p0901r9.html
>
>this does not seem to discuss how existing applications
>that override new() would cope with this.
>
>nor how existing implementations on top of c allocators
>would implement it (given that we just agreed that
>malloc_usable_size is not suitable for such use).
>
>nor how existing allocator tooling (interposers, profilers)
>would handle the new interface.
>
>>
>> Also, as noted in the doc, jemalloc experimentally implemented this
>> functionality in its non-standard API, via a function it called "smallocx"
>> -- though jemalloc hides the API so it can't be used by default. The API is
>> effectively:
>> typedef struct { void *ptr; size_t size; } smallocx_return_t;
>> smallocx_return_t smallocx(size_t size, int flags);
>> https://github.com/jemalloc/jemalloc/blob/a0734fd6ee326cd2059edbe4bca7092988a63684/src/jemalloc.c#L3414
>> (That's consistent with jemalloc's other non-standard APIs, which stick
>> alignment/etc into a "flags" argument, but probably not suitable for a
>> more-standardized cross-implementation API)
>>
>> tcmalloc implements similar functionality, as well, with family of
>> functions named "tcmalloc_size_returning_operator_new":
>
>so there are already incompatible c apis, which means this
>should not be considered a viable proposal at this point.
Small addition: https://wg21.link/P0401R6 (allocate_at_least) has made it into C++23.
https://reviews.llvm.org/D122877 libc++ has implemented it in the
trivial way that just returns the user-requested size.
>> https://github.com/google/tcmalloc/blob/267aa2ec2817ab9d09b3fbb65ecb90193dd4348e/tcmalloc/malloc_extension.h#L549
>> which of course also isn't a suitable API to support cross-implementation.
>>
>> If someone wants to push forward this area, IMO, it would be really great
>> to have an API exposing this functionality designed to be implemented in a
>> common way across libc malloc implementations -- and eventually added to
>> POSIX or C.
>
>this is done the wrong way around.
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
