Re: Firefox still crashes

[Venkatesh Srinivas <me@xxxxxxxxxxxxxxxxxxx>]

> > What's a "magazine"?
> >
> > Pierre

In libc, nmalloc (lib/libc/stdlib/nmalloc.c) provides malloc() ( from 
malus locus, 'bad place' ) and free() for single-threaded and 
multithreaded applications. In the DragonFly 2.4 release cycle, the 
original allocator (phkmalloc, inherited from FreeBSD) was replaced with a 
port of the kernel slab allocator; in the 2.8 release cycle I committed 
some work to change the multithreaded strategy.

The DragonFly 2.4 and 2.6 libc allocator had two strategies, one for small 
requests, one for large requests; large requests are served directly via 
mmap. Small requests are served from 64k, 64k-aligned regions of memory 
called 'slabs'. Each slab only services requests of a given size, 
minimizing fragmentation. (The DragonFly libc slab allocator was fairly 
different from the original Sun design -- the Sun design has variable-size 
slabs and a hash table for block->slab mappings). For multithreaded 
applications, the allocator kept track of four sets of slab structures; 
threads would attempt to use the set they'd used most recently. If they 
failed to lock that said, they'd move on to the next set.

The 2.7/2.8 allocator has a new structure -- a magazine. A magazine is a 
fixed-size array of blocks of the same size. Each thread carries a pair of 
magazines; when a thread tries to allocate something, it first checks its 
magazines for a buffer _without any locking_. If the magazines are not 
able to support an allocation, a central collection of magazines, called 
the 'depot', is locked and a magazine is retrieved. If the depot is empty, 
we fall back to the slab allocator. The design is also from Sun -- see a 
paper called 'Magazines and Vmem: Extending the Slab Allocator to Many 
CPUs and Multiple Resources'. When I last measured, in the 2.6 release 
cycle, the magazine layer sped up sysbench OLTP / MySQL by approximately 
20%.

The 2.7/2.8 allocator also has work to reduce the number of mmap/munmap 
system calls relative to the earlier version of the allocator; rather than 
immediately unmapping a slab when it has no outstanding allocations, we 
keep around up to 64 old slabs and we attempt to allocate slabs in bursts 
from the system. When I last measured, the reduction in mmaps/munmaps was 
fairly dramatic.

The latest bug was a fairly unfortunate one -- most of the locks in 
nmalloc use libc's spinlocks. The depot locks, however, used 
pthread_spinlocks; when nmalloc was linked against libc, it used the stub 
pthread_spinlocks in libc, rather than the versions in libthread. This 
meant that accesses to the depot magazine lists were not synchronized at 
all and the magazine lists were getting corrupted. Oops...

-- vs