use sys use "die"use "extremum"use "memops"use "syswrap"use "threadhooks"use "types"use "units"use "result"use "slfill"use "backtrace" pkg std = const startalloctrace : (f : byte[:] -> void) const endalloctrace : (-> void) const bytealloc : (sz:size -> byte#) const zbytealloc : (sz:size -> byte#) const bytefree : (m:byte#, sz:size -> void) /* null pointers. only used internally. */ pkglocal const Zslicep = (0 : byte#) pkglocal const Align = 16 /* minimum allocation alignment */ pkglocal const align : (m : std.size, align : std.size -> std.size) pkglocal const allocsz : (sz : std.size -> std.size) ;; const Zslab = (0 : slab#)const Zchunk = (0 : chunk#)const Slabsz = 512*KiBconst Cachemax = 4const Bktmax = 128*KiB /* a balance between wasted space and falling back to mmap */const Pagesz = 4*KiB var buckets : bucket[32] /* excessive */var trace : boolvar tracefd : std.fdvar bigcache : cacheelt[32] type bucket = struct sz : size /* aligned size */ nper : size /* max number of elements per slab */ slabs : slab# /* partially filled or free slabs */ cache : slab# /* cache of empty slabs, to prevent thrashing */ ncache : size /* size of cache */;; type slab = struct head : byte# /* head of virtual addresses, so we don't leak address space */ next : slab# /* the next slab on the chain */ prev : slab# /* the prev slab on the chain */ freehd : chunk# /* the nodes we're allocating */ nfree : size /* the number of free nodes */ magic : size /* ensure we didn't index into the void */;; /* NB: must be smaller than sizeof(slab) */type chunk = struct next : chunk# /* the next chunk in the free list */;; type cacheelt = struct sz : std.size p : byte#;; const __init__ = { for var i = 0; i < buckets.len && (Align << i) <= Bktmax; i++ bktinit(&buckets[i], Align << i) ;;} const startalloctrace = {path match openmode(path, Owrite | Ocreat, 0o644) | `Ok fd: tracefd = fd | `Err e: -> void ;; trace = true} const endalloctrace = { std.close(tracefd) trace = false} const zbytealloc = {sz var p p = bytealloc(sz) memfill(p, 0, sz) -> p} const tracealloc = {p, sz var stk : void#[23] /* [type, addr, sz, 10 stack slots] */ slfill(stk[:], (0 : void#)) stk[0] = (0 : void#) stk[1] = (p : void#) stk[2] = (sz : void#) backtrace(stk[3:]) writealloctrace(stk[:])} const tracefree = {p, sz var stk : void#[3] stk[0] = (1 : void#) stk[1] = (p : void#) stk[2] = (sz : void#) writealloctrace(stk[:])} const writealloctrace = {sl var len, p len = sl.len * sizeof(void#) p = (sl : byte#) write(tracefd, p[:len])} /* Allocates a blob that is 'sz' bytes long. Dies if the allocation fails */const bytealloc = {sz var bkt, p if sz <= Bktmax bkt = &buckets[bktnum(sz)] lock(memlck) p = bktalloc(bkt) unlock(memlck) else p = bigalloc(sz) ;; if trace lock(memlck) tracealloc(p, sz) unlock(memlck) ;; -> p} /* frees a blob that is 'sz' bytes long. */const bytefree = {p, sz var bkt if p == (0 : byte#) -> void ;; if trace lock(memlck) tracefree(p, sz) unlock(memlck) ;; memfill(p, 0xa8, sz) if (sz < Bktmax) bkt = &buckets[bktnum(sz)] lock(memlck) bktfree(bkt, p) unlock(memlck) else bigfree(p, sz) ;;} const bigalloc = {sz var p p = Failmem /* We need to round up to the page size so that unmapping doesn't end up taking out the tail of another allocation. */ sz = align(sz, Pagesz) /* check our cache */ lock(memlck) for var i = 0; i < bigcache.len; i++ if sz < bigcache[i].sz /* While the allocator may never use a small chunk, we need to keep it together so when unmapping it in the free call, end up unmapping the whole thing. We can't return a larger size here than requested, because we don't track the size. Eviction takes care of freeing it. */ p = bigcache[i].p bigcache[i].sz -= sz bigcache[i].p = ((p : intptr) + (sz : intptr) : byte#) break ;; ;; unlock(memlck) if p != Failmem -> p ;; /* ok, lets give up and get memory from the os */ p = getmem(sz) if p != Failmem -> p ;; die("could not get memory\n")} const bigfree = {p, sz var evictsz, evictp, evictidx var endp, endblk sz = align(sz, Pagesz) evictp = p evictidx = -1 evictsz = sz endp = ((p : intptr) + (sz : intptr) : byte#) lock(memlck) for var i = 0; i < bigcache.len; i++ endblk = ((bigcache[i].p : intptr) + (bigcache[i].sz : intptr) : byte#) /* merge in front of existing block */ if bigcache[i].p == endp bigcache[i].sz += sz bigcache[i].p = p evictidx = -1 evictsz = 0 break /* merge in behind existing block */ elif endblk == p bigcache[i].sz += sz evictidx = -1 evictsz = 0 break ;; /* evict */ if bigcache[i].sz < evictsz evictidx = i evictsz = bigcache[i].sz evictp = bigcache[i].p ;; ;; if evictidx != -1 bigcache[evictidx].p = p bigcache[evictidx].sz = sz ;; unlock(memlck) /* Now that we've removed it, we can free it. It's not in the cache, so we don't need the lock held. */ if evictsz > 0 freemem(evictp, evictsz) ;;} /* Sets up a single empty bucket */const bktinit = {b, sz b.sz = align(sz, Align) b.nper = (Slabsz - sizeof(slab))/b.sz b.slabs = Zslab b.cache = Zslab b.ncache = 0} /* Creates a slab for bucket 'bkt', and fills the chunk list */const mkslab = {bkt var p, s var b, bnext var off /* offset of chunk head */ if bkt.ncache > 0 s = bkt.cache bkt.cache = s.next bkt.ncache-- s.next = Zslab -> s ;; /* tricky: we need power of two alignment, so we allocate double the needed size, chop off the unaligned ends, and waste the address space. Since the OS is "smart enough", this shouldn't actually cost us memory, and 64 bits of address space means that we're not going to have issues with running out of address space for a while. On a 32 bit system this would be a bad idea. */ p = getmem(Slabsz*2) if p == Failmem die("Unable to get memory") ;; b = Zchunk s = (align((p : size), Slabsz) : slab#) s.head = p s.nfree = bkt.nper s.next = Zslab s.prev = Zslab s.magic = 0xfee1baff1ed /* skip past the slab header */ off = align(sizeof(slab), Align) bnext = nextchunk((s : chunk#), off) s.freehd = bnext for var i = 0; i < bkt.nper; i++ b = bnext bnext = nextchunk(b, bkt.sz) b.next = bnext ;; b.next = Zchunk -> s} /* Allocates a node from bucket 'bkt', crashing if theallocation cannot be satisfied. Will create a new slabif there are no slabs on the freelist.*/const bktalloc = {bkt var s, c /* find a slab */ s = bkt.slabs if s == Zslab s = mkslab(bkt) bkt.slabs = s if s == Zslab die("No memory left") ;; ;; /* grab the first chunk on the slab */ c = s.freehd s.freehd = c.next s.nfree-- if s.freehd == Zchunk bkt.slabs = s.next if s.next != Zslab s.next.prev = Zslab ;; ;; -> (c : byte#)} /*Frees a chunk of memory 'm' into bucket 'bkt'.Assumes that the memory already came from a slabthat was created for bucket 'bkt'. Will crashif this is not the case.*/const bktfree = {bkt, m var s, c s = (mtrunc(m, Slabsz) : slab#) iassert(s.magic == 0xfee1baff1ed, "bad free") c = (m : chunk#) if s.nfree == 0 if bkt.slabs != Zslab bkt.slabs.prev = s ;; s.next = bkt.slabs s.prev = Zslab bkt.slabs = s elif s.nfree == bkt.nper - 1 /* unlink the slab from the list */ if s.next != Zslab s.next.prev = s.prev ;; if s.prev != Zslab s.prev.next = s.next ;; if bkt.slabs == s bkt.slabs = s.next ;; /* HACK HACK HACK: if we can't unmap, keep an infinite cache per slab size. We should solve this better somehow. */ if bkt.ncache < Cachemax || !Canunmap s.next = bkt.cache s.prev = Zslab bkt.cache = s bkt.ncache++ else /* we mapped 2*Slabsz so we could align it, so we need to unmap the same */ freemem(s.head, Slabsz*2) ;; -> void ;; s.nfree++ c.next = s.freehd s.freehd = c} /*Finds the correct bucket index to allocate fromfor allocations of size 'sz'*/const bitpos : byte[32] = [ 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31]const bktnum = {sz var n, v v = (sz >> 3 : uint32) v |= v >> 1 v |= v >> 2 v |= v >> 4 v |= v >> 8 v |= v >> 16 n = bitpos[((v * 0x07c4acdd) & 0xffff_ffffui) >> 27] -> (n : size)} /*returns the actual size we allocated for a givensize request*/const allocsz = {sz var bktsz if sz <= Bktmax bktsz = Align for var i = 0; bktsz <= Bktmax; i++ if bktsz >= sz -> bktsz ;; bktsz *= 2 ;; else -> align(sz, Pagesz) ;; die("Size does not match any buckets")} /*aligns a size to a requested alignment.'align' must be a power of two*/const align = {v, align -> (v + align - 1) & ~(align - 1)} /*chunks are variable sizes, so we can't justindex to get to the next one*/const nextchunk = {b, sz : size -> ((b : intptr) + (sz : intptr) : chunk#)} /*truncates a pointer to 'align'. 'align' mustbe a power of two.*/const mtrunc = {m, align -> ((m : intptr) & ~((align : intptr) - 1) : byte#)}
use sys
use "die"
use "extremum"
use "memops"
use "syswrap"
use "threadhooks"
use "types"
use "units"
use "result"
use "slfill"
use "backtrace"
pkg std =
const startalloctrace : (f : byte[:] -> void)
const endalloctrace : (-> void)
const bytealloc : (sz:size -> byte#)
const zbytealloc : (sz:size -> byte#)
const bytefree : (m:byte#, sz:size -> void)
/* null pointers. only used internally. */
pkglocal const Zslicep = (0 : byte#)
pkglocal const Align = 16 /* minimum allocation alignment */
pkglocal const align : (m : std.size, align : std.size -> std.size)
pkglocal const allocsz : (sz : std.size -> std.size)
;;
const Zslab = (0 : slab#)
const Zchunk = (0 : chunk#)
const Slabsz = 512*KiB
const Cachemax = 4
const Bktmax = 128*KiB /* a balance between wasted space and falling back to mmap */
const Pagesz = 4*KiB
var buckets : bucket[32] /* excessive */
var trace : bool
var tracefd : std.fd
var bigcache : cacheelt[32]
type bucket = struct
sz : size /* aligned size */
nper : size /* max number of elements per slab */
slabs : slab# /* partially filled or free slabs */
cache : slab# /* cache of empty slabs, to prevent thrashing */
ncache : size /* size of cache */
;;
type slab = struct
head : byte# /* head of virtual addresses, so we don't leak address space */
next : slab# /* the next slab on the chain */
prev : slab# /* the prev slab on the chain */
freehd : chunk# /* the nodes we're allocating */
nfree : size /* the number of free nodes */
magic : size /* ensure we didn't index into the void */
;;
/* NB: must be smaller than sizeof(slab) */
type chunk = struct
next : chunk# /* the next chunk in the free list */
;;
type cacheelt = struct
sz : std.size
p : byte#
;;
const __init__ = {
for var i = 0; i < buckets.len && (Align << i) <= Bktmax; i++
bktinit(&buckets[i], Align << i)
;;
}
const startalloctrace = {path
match openmode(path, Owrite | Ocreat, 0o644)
| `Ok fd: tracefd = fd
| `Err e: -> void
;;
trace = true
}
const endalloctrace = {
std.close(tracefd)
trace = false
}
const zbytealloc = {sz
var p
p = bytealloc(sz)
memfill(p, 0, sz)
-> p
}
const tracealloc = {p, sz
var stk : void#[23] /* [type, addr, sz, 10 stack slots] */
slfill(stk[:], (0 : void#))
stk[0] = (0 : void#)
stk[1] = (p : void#)
stk[2] = (sz : void#)
backtrace(stk[3:])
writealloctrace(stk[:])
}
const tracefree = {p, sz
var stk : void#[3]
stk[0] = (1 : void#)
stk[1] = (p : void#)
stk[2] = (sz : void#)
writealloctrace(stk[:])
}
const writealloctrace = {sl
var len, p
len = sl.len * sizeof(void#)
p = (sl : byte#)
write(tracefd, p[:len])
}
/* Allocates a blob that is 'sz' bytes long. Dies if the allocation fails */
const bytealloc = {sz
var bkt, p
if sz <= Bktmax
bkt = &buckets[bktnum(sz)]
lock(memlck)
p = bktalloc(bkt)
unlock(memlck)
else
p = bigalloc(sz)
;;
if trace
lock(memlck)
tracealloc(p, sz)
unlock(memlck)
;;
-> p
}
/* frees a blob that is 'sz' bytes long. */
const bytefree = {p, sz
var bkt
if p == (0 : byte#)
-> void
;;
if trace
lock(memlck)
tracefree(p, sz)
unlock(memlck)
;;
memfill(p, 0xa8, sz)
if (sz < Bktmax)
bkt = &buckets[bktnum(sz)]
lock(memlck)
bktfree(bkt, p)
unlock(memlck)
else
bigfree(p, sz)
;;
}
const bigalloc = {sz
var p
p = Failmem
/*
We need to round up to the page size so that unmapping
doesn't end up taking out the tail of another allocation.
*/
sz = align(sz, Pagesz)
/* check our cache */
lock(memlck)
for var i = 0; i < bigcache.len; i++
if sz < bigcache[i].sz
/*
While the allocator may never use a small chunk, we need
to keep it together so when unmapping it in the free call,
end up unmapping the whole thing. We can't return a larger
size here than requested, because we don't track the size.
Eviction takes care of freeing it.
*/
p = bigcache[i].p
bigcache[i].sz -= sz
bigcache[i].p = ((p : intptr) + (sz : intptr) : byte#)
break
;;
;;
unlock(memlck)
if p != Failmem
-> p
;;
/* ok, lets give up and get memory from the os */
p = getmem(sz)
if p != Failmem
-> p
;;
die("could not get memory\n")
}
const bigfree = {p, sz
var evictsz, evictp, evictidx
var endp, endblk
sz = align(sz, Pagesz)
evictp = p
evictidx = -1
evictsz = sz
endp = ((p : intptr) + (sz : intptr) : byte#)
lock(memlck)
for var i = 0; i < bigcache.len; i++
endblk = ((bigcache[i].p : intptr) + (bigcache[i].sz : intptr) : byte#)
/* merge in front of existing block */
if bigcache[i].p == endp
bigcache[i].sz += sz
bigcache[i].p = p
evictidx = -1
evictsz = 0
break
/* merge in behind existing block */
elif endblk == p
bigcache[i].sz += sz
evictidx = -1
evictsz = 0
break
;;
/* evict */
if bigcache[i].sz < evictsz
evictidx = i
evictsz = bigcache[i].sz
evictp = bigcache[i].p
;;
;;
if evictidx != -1
bigcache[evictidx].p = p
bigcache[evictidx].sz = sz
;;
unlock(memlck)
/*
Now that we've removed it, we can
free it. It's not in the cache, so
we don't need the lock held.
*/
if evictsz > 0
freemem(evictp, evictsz)
;;
}
/* Sets up a single empty bucket */
const bktinit = {b, sz
b.sz = align(sz, Align)
b.nper = (Slabsz - sizeof(slab))/b.sz
b.slabs = Zslab
b.cache = Zslab
b.ncache = 0
}
/* Creates a slab for bucket 'bkt', and fills the chunk list */
const mkslab = {bkt
var p, s
var b, bnext
var off /* offset of chunk head */
if bkt.ncache > 0
s = bkt.cache
bkt.cache = s.next
bkt.ncache--
s.next = Zslab
-> s
;;
/*
tricky: we need power of two alignment, so we allocate double the
needed size, chop off the unaligned ends, and waste the address
space. Since the OS is "smart enough", this shouldn't actually
cost us memory, and 64 bits of address space means that we're not
going to have issues with running out of address space for a
while. On a 32 bit system this would be a bad idea.
*/
p = getmem(Slabsz*2)
if p == Failmem
die("Unable to get memory")
;;
b = Zchunk
s = (align((p : size), Slabsz) : slab#)
s.head = p
s.nfree = bkt.nper
s.next = Zslab
s.prev = Zslab
s.magic = 0xfee1baff1ed
/* skip past the slab header */
off = align(sizeof(slab), Align)
bnext = nextchunk((s : chunk#), off)
s.freehd = bnext
for var i = 0; i < bkt.nper; i++
b = bnext
bnext = nextchunk(b, bkt.sz)
b.next = bnext
;;
b.next = Zchunk
-> s
}
/*
Allocates a node from bucket 'bkt', crashing if the
allocation cannot be satisfied. Will create a new slab
if there are no slabs on the freelist.
*/
const bktalloc = {bkt
var s, c
/* find a slab */
s = bkt.slabs
if s == Zslab
s = mkslab(bkt)
bkt.slabs = s
if s == Zslab
die("No memory left")
;;
;;
/* grab the first chunk on the slab */
c = s.freehd
s.freehd = c.next
s.nfree--
if s.freehd == Zchunk
bkt.slabs = s.next
if s.next != Zslab
s.next.prev = Zslab
;;
;;
-> (c : byte#)
}
/*
Frees a chunk of memory 'm' into bucket 'bkt'.
Assumes that the memory already came from a slab
that was created for bucket 'bkt'. Will crash
if this is not the case.
*/
const bktfree = {bkt, m
var s, c
s = (mtrunc(m, Slabsz) : slab#)
iassert(s.magic == 0xfee1baff1ed, "bad free")
c = (m : chunk#)
if s.nfree == 0
if bkt.slabs != Zslab
bkt.slabs.prev = s
;;
s.next = bkt.slabs
s.prev = Zslab
bkt.slabs = s
elif s.nfree == bkt.nper - 1
/* unlink the slab from the list */
if s.next != Zslab
s.next.prev = s.prev
;;
if s.prev != Zslab
s.prev.next = s.next
;;
if bkt.slabs == s
bkt.slabs = s.next
;;
/*
HACK HACK HACK: if we can't unmap, keep an infinite cache per slab size.
We should solve this better somehow.
*/
if bkt.ncache < Cachemax || !Canunmap
s.next = bkt.cache
s.prev = Zslab
bkt.cache = s
bkt.ncache++
else
/* we mapped 2*Slabsz so we could align it,
so we need to unmap the same */
freemem(s.head, Slabsz*2)
;;
-> void
;;
s.nfree++
c.next = s.freehd
s.freehd = c
}
/*
Finds the correct bucket index to allocate from
for allocations of size 'sz'
*/
const bitpos : byte[32] = [
0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30,
8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31
]
const bktnum = {sz
var n, v
v = (sz >> 3 : uint32)
v |= v >> 1
v |= v >> 2
v |= v >> 4
v |= v >> 8
v |= v >> 16
n = bitpos[((v * 0x07c4acdd) & 0xffff_ffffui) >> 27]
-> (n : size)
}
/*
returns the actual size we allocated for a given
size request
*/
const allocsz = {sz
var bktsz
if sz <= Bktmax
bktsz = Align
for var i = 0; bktsz <= Bktmax; i++
if bktsz >= sz
-> bktsz
;;
bktsz *= 2
;;
else
-> align(sz, Pagesz)
;;
die("Size does not match any buckets")
}
/*
aligns a size to a requested alignment.
'align' must be a power of two
*/
const align = {v, align
-> (v + align - 1) & ~(align - 1)
}
/*
chunks are variable sizes, so we can't just
index to get to the next one
*/
const nextchunk = {b, sz : size
-> ((b : intptr) + (sz : intptr) : chunk#)
}
/*
truncates a pointer to 'align'. 'align' must
be a power of two.
*/
const mtrunc = {m, align
-> ((m : intptr) & ~((align : intptr) - 1) : byte#)
}