functions are implemented with CMPXCHG8B instruction where it is available,
i. e., all Pentium-class and later processors. Note this instruction is
also used for atomic_store_rel_64() because a simple XCHG-like instruction
for 64-bit memory access does not exist, unfortunately. If the processor
lacks the instruction, i. e., 80486-class CPUs, two 32-bit load/store are
performed with interrupt temporarily disabled, assuming it does not support
SMP. Although this assumption may be little naive, it is true in reality.
This implementation is inspired by Linux.
specify their own version of atomic_cmpset_* which could have been
different than the membar version.
Right now, however, FreeBSD is bound mostly to GCC-like compilers and
it is desired to add new support and compat shim mostly when there is
a real necessity, in order to avoid too much compatibility bloats.
In this optic, bring back atomic_cmpset_{acq, rel}_* to be the same as
atomic_cmpset_* and unwind the atomic_cmpset_barr_* introduction.
Requested by: jhb
Reviewed by: jhb
Tested by: Giovanni Trematerra <giovanni dot trematerra at
gmail dot com>
not defined through macros or similar) in order to be later compiled in
the kernel and offer this way the support for modules (and
compatibility among the UP case and SMP case).
Fix this for the newly introduced atomic_cmpset_barr_* cases by defining
and specifying a template. Note that the new DEFINE_CMPSET_GEN()
template save more typing on amd64 than the current code. [1]
- Fix the style for memory barriers on amd64.
[1] Reported by: Paul B. Mahol <onemda at gmail dot com>
memory barriers should also ensure that the compiler doesn't reorder paths
where they are used. GCC, however, does that aggressively, even in
presence of volatile operands. The most reliable way GCC offers for avoid
instructions reordering is clobbering "memory" even if that is
theoretically an heavy-weight operation, flushing the content of all
the registers and forcing reload of them (We could rely, however, on
gcc DTRT by just understanding the purpose as this is a well-known
pattern for many modern operating-systems).
Not all our memory barriers, right now, clobber memory for GCC-like
compilers. The most notable cases are IA32 and amd64 where the memory
barrier are treacted the same as normal atomic instructions.
Fix this by offering the possibility to implement atomic instructions
with memory barriers separately from the normal version and implement
the GCC-like specific one using memory clobbering.
Thanks to Chris Lattner (@apple) for his discussion on llvm specifics.
Reported by: jhb
Reviewed by: jhb
Tested by: rdivacky, Giovanni Trematerra
<giovanni dot trematerra at gmail dot com>
memory barriers on i386. It works as a serialization instruction on
all IA32 CPUs.
Alternative solution of using {s,l,}fence requires run-time checking
of the presense of the corresponding SSE or SSE2 extensions, and
possible boot-time patching of the kernel text.
Suggested by: many
and ifnet functions
- add memory barriers to <machine/atomic.h>
- update drivers to only conditionally define their own
- add lockless producer / consumer ring buffer
- remove ring buffer implementation from cxgb and update its callers
- add if_transmit(struct ifnet *ifp, struct mbuf *m) to ifnet to
allow drivers to efficiently manage multiple hardware queues
(i.e. not serialize all packets through one ifq)
- expose if_qflush to allow drivers to flush any driver managed queues
This work was supported by Bitgravity Inc. and Chelsio Inc.
MPLOCKED. The cleaning in rev.1.25 was supposed to have been undone
by rev.1.26, but 1.26 could never have actually affected asm files
since atomic.h is full of C declarations so including it in asm files
would just give syntax errors. The asm MPLOCKED is even less needed
than when misplaced definitions of it were first removed, and is now
unused in any asm file in the src tree except in anachronismns in
sys/i386/i386/support.s.
These functions are used a lot for mutexes, so this reduces the text
size of an average kernel by about 0.75%. This wasn't intended to
be a significant optimization, but it somehow increased the maximum
number of packets per second that can be transmitted by my bge hardware
from 320000 to 460000 (this benchmark is CPU-bound and remarkably
sensitive to changes in the text section).
Details: we would prefer to leave the result of the cmpxchg in %al,
but cannot tell gcc that it is there, so we have to convert it to an
integer register. We converted to %al, then to %[re]ax, but the
latter step is usually wasted since gcc usually only wants the condition
code and can recover it from %al just as easily as from %[re]ax. Let
gcc promote %al in the few cases where this is needed.
Nearby style fixes;
- let gcc manage the load of `res', and don't abuse `res' for a copy of `exp'
- don't echo `res's name in comments
- consistently spell the condition code as 'e' after comparison for equality
- don't hard-code %al anywhere except in constraints
- for the version that doesn't use cmpxchg, there is no requirement to use
%al anywhere, so don't hard-code it in the constraints either.
Style non-fix:
- for the versions that use cmpxchg, keep using "a" (was %[re]ax, now %al)
for the main output operand, although this is not required. The input
and output operands that use the "a" constraint are now decoupled, and
this makes things clearer except for the reason that the output register
is hard-coded. It is now just a hack to tell gcc that the input "a" has
been clobbered without increasing the number of operands.
variable and returns the previous value of the variable.
Tested on: i386, alpha, sparc64, arm (cognet)
Reviewed by: arch@
Submitted by: cognet (arm)
MFC after: 1 week
constraint is actually only allowed for register operands. Instead, use
separate input and output memory constraints.
Education from: alc
Reviewed by: alc
Tested on: i386, alpha
MFC after: 1 week
variables rather than void * variables. This makes it easier and simpler
to get asm constraints and volatile keywords correct.
MFC after: 3 days
Tested on: i386, alpha, sparc64
Compiled on: ia64, powerpc, amd64
Kernel toolchain busted on: arm
further changes and fixes in the future:
- Use aliases via macros rather than duplicated inlines wherever possible.
- Move all the aliases to the bottom of these files and the inline
functions to the top.
- Add various comments.
- On alpha, drop atomic_{load_acq,store_rel}_{8,char,16,short}().
- On i386 and amd64, don't duplicate the extern declarations for functions
in the two non-inline cases (KLD_MODULE and compiler doesn't do inlines),
instead, consolidate those two cases.
- Some whitespace fixes.
Approved by: re (scottl)
place.
This moves the dependency on GCC's and other compiler's features into
the central sys/cdefs.h file, while the individual source files can
then refer to #ifdef __COMPILER_FEATURE_FOO where they by now used to
refer to #if __GNUC__ > 3.1415 && __BARC__ <= 42.
By now, GCC and ICC (the Intel compiler) have been actively tested on
IA32 platforms by netchild. Extension to other compilers is supposed
to be possible, of course.
Submitted by: netchild
Reviewed by: various developers on arch@, some time ago
of atomic_store_rel().
- Use the 80386 versions of atomic_load_acq() and atomic_store_rel() that
do not use serializing instructions on all UP kernels since a UP machine
does need to synchronize with other CPUs. This trims lots of cycles from
spin locks on UP kernels among other things.
Benchmarked by: rwatson
to build the kernel. It doesn't affect the operation if gcc.
Most of the changes are just adding __INTEL_COMPILER to #ifdef's, as
icc v8 may define __GNUC__ some parts may look strange but are
necessary.
Additional changes:
- in_cksum.[ch]:
* use a generic C version instead of the assembly version in the !gcc
case (ASM code breaks with the optimizations icc does)
-> no bad checksums with an icc compiled kernel
Help from: andre, grehan, das
Stolen from: alpha version via ppc version
The entire checksum code should IMHO be replaced with the DragonFly
version (because it isn't guaranteed future revisions of gcc will
include similar optimizations) as in:
---snip---
Revision Changes Path
1.12 +1 -0 src/sys/conf/files.i386
1.4 +142 -558 src/sys/i386/i386/in_cksum.c
1.5 +33 -69 src/sys/i386/include/in_cksum.h
1.5 +2 -0 src/sys/netinet/igmp.c
1.6 +0 -1 src/sys/netinet/in.h
1.6 +2 -0 src/sys/netinet/ip_icmp.c
1.4 +3 -4 src/contrib/ipfilter/ip_compat.h
1.3 +1 -2 src/sbin/natd/icmp.c
1.4 +0 -1 src/sbin/natd/natd.c
1.48 +1 -0 src/sys/conf/files
1.2 +0 -1 src/sys/conf/files.amd64
1.13 +0 -1 src/sys/conf/files.i386
1.5 +0 -1 src/sys/conf/files.pc98
1.7 +1 -1 src/sys/contrib/ipfilter/netinet/fil.c
1.10 +2 -3 src/sys/contrib/ipfilter/netinet/ip_compat.h
1.10 +1 -1 src/sys/contrib/ipfilter/netinet/ip_fil.c
1.7 +1 -1 src/sys/dev/netif/txp/if_txp.c
1.7 +1 -1 src/sys/net/ip_mroute/ip_mroute.c
1.7 +1 -2 src/sys/net/ipfw/ip_fw2.c
1.6 +1 -2 src/sys/netinet/igmp.c
1.4 +158 -116 src/sys/netinet/in_cksum.c
1.6 +1 -1 src/sys/netinet/ip_gre.c
1.7 +1 -2 src/sys/netinet/ip_icmp.c
1.10 +1 -1 src/sys/netinet/ip_input.c
1.10 +1 -2 src/sys/netinet/ip_output.c
1.13 +1 -2 src/sys/netinet/tcp_input.c
1.9 +1 -2 src/sys/netinet/tcp_output.c
1.10 +1 -1 src/sys/netinet/tcp_subr.c
1.10 +1 -1 src/sys/netinet/tcp_syncache.c
1.9 +1 -2 src/sys/netinet/udp_usrreq.c
1.5 +1 -2 src/sys/netinet6/ipsec.c
1.5 +1 -2 src/sys/netproto/ipsec/ipsec.c
1.5 +1 -1 src/sys/netproto/ipsec/ipsec_input.c
1.4 +1 -2 src/sys/netproto/ipsec/ipsec_output.c
and finally remove
sys/i386/i386 in_cksum.c
sys/i386/include in_cksum.h
---snip---
- endian.h:
* DTRT in C++ mode
- quad.h:
* we don't use gcc v1 anymore, remove support for it
Suggested by: bde (long ago)
- assym.h:
* avoid zero-length arrays (remove dependency on a gcc specific
feature)
This change changes the contents of the object file, but as it's
only used to generate some values for a header, and the generator
knows how to handle this, there's no impact in the gcc case.
Explained by: bde
Submitted by: Marius Strobl <marius@alchemy.franken.de>
- aicasm.c:
* minor change to teach it about the way icc spells "-nostdinc"
Not approved by: gibbs (no reply to my mail)
- bump __FreeBSD_version (lang/icc needs to know about the changes)
Incarnations of this patch survive gcc compiles since a loooong time,
I use it on my desktop. An icc compiled kernel works since Nov. 2003
(exceptions: snd_* if used as modules), it survives a build of the
entire ports collection with icc.
Parts of this commit contains suggestions or submissions from
Marius Strobl <marius@alchemy.franken.de>.
Reviewed by: -arch
Submitted by: netchild
is the warning that points to the bug in `(char *)malloc(...)' where
malloc() is implicitly declared as returning int. We do similar things
here, but they work because u_int is the same as uintptr_t on i386's.)
This is most beneficial for vmware client os installs.
Reviewed by: jmallet, iedowse, tlambert2@mindspring.com
MFC After: never, -STABLE does not currently use this instruction
Some things needed bits of <i386/include/lock.h> - cy.c now has its
own (only) copy of the COM_(UN)LOCK() macros, and IMASK_(UN)LOCK()
has been moved to <i386/include/apic.h> (AKA <machine/apic.h>).
Reviewed by: jhb
non-386 atomic_load_acq(). %eax is an input since its value is used in
the cmpxchg instruction, but we don't care what value it is, so setting
it to a specific value is just wasteful. Thus, it is being used without
being initialized as the warning stated, but it is ok for it to be used
because its value isn't important. Thus, we are only sort of lying when
we say it is an output only operand.
- Add "cc" to the clobber list for atomic_load_acq() since the cmpxchgl
changes ZF.
that modules can call.
- Remove the old gcc <= 2.8 versions of the atomic ops.
- Resort the order of some things in the file so that there is only
one #ifdef for KLD_MODULE, and so that all WANT_FUNCTIONS stuff is
moved to the bottom of the file.
- Remove ATOMIC_ACQ_REL() and just use explicit macros instead.
implement memory fences for the 486+. The 386 still uses versions w/o
memory fences as all operations on the 386 are not program ordered.
The 386 versions are not MP safe.
in most of the atomic operations. Now for these operations, you can
use the normal atomic operation, you can use the operation with a read
barrier, or you can use the operation with a write barrier. The function
names follow the same semantics used in the ia64 instruction set. An
atomic operation with a read barrier has the extra suffix 'acq', due to
it having "acquire" semantics. An atomic operation with a write barrier
has the extra suffix 'rel'. These suffixes are inserted between the
name of the operation to perform and the typename. For example, the
atomic_add_int() function now has 3 variants:
- atomic_add_int() - this is the same as the previous function
- atomic_add_acq_int() - this function combines the add operation with a
read memory barrier
- atomic_add_rel_int() - this function combines the add operation with a
write memory barrier
- Add 'ptr' to the list of types that we can perform atomic operations
on. This allows one to do atomic operations on uintptr_t's. This is
useful in the mutex code, for example, because the actual mutex lock is
a pointer.
- Add two new operations for doing loads and stores with memory barriers.
The new load operations use a read barrier before the load, and the
new store operations use a write barrier after the load. For example,
atomic_load_acq_int() will atomically load an integer as well as
enforcing a read barrier.
specifies the instruction's operation size. GCC will default to 32-bit
operands reguardless of the prototype (ie, formal parameters' type)
of an inline function.
