mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
struct vmspace is freed from cpu_sched_exit() to pmap_release().
This has the advantage of being able to rely on MI code to decide
when a free should occur, instead of having to inspect the reference
count ourselves.
At the same time, turn the per-CPU vmspace pointer into a pmap pointer,
so that pmap_release() can deal with pmaps exclusively.
Reviewed (and embrassing bug spotted) by: jake
to <sys/gmon.h>. Cleaned them up a little by not attempting to ifdef
for incomplete and out of date support for GUPROF in userland, as in
the sparc64 version.
elf_reloc() backends for two reasons. First, to support the possibility
of there being two elf linkers in the kernel (eg: amd64), and second, to
pass the relocbase explicitly (for relocating .o format kld files).
"options OFW_NEWPCI").
This is a bit overdue, the new sparc64 OFW PCI code which is
meant to replace the old one is in place for 10 months and
enabled by default in GENERIC for 8 months. FreeBSD 5.2 and
5.2.1 also shipped with the new code enabled by default.
- Some minor clean-up, e.g. remove functions that encapsulated
the #ifdefs for OFW_NEWPCI, remove unused resp. no longer
required includes, etc.
Approved by: tmm, no objections on freebsd-sparc64
and type for printing info about the device that didn't probe from child, not
parent.
This fixes a panic on systems where not yet supported devices hang off of the
nexus, e.g. on E450.
Reported by: joerg
- Remove second license, the first was not that different and should be
fine.
- Add nexus_attach(), and do not perform its task in nexus_probe() any
more.
- Remove nexus_write_ivar(), since it was quite pointless.
- Remove superfluous devinfo members.
- Clean up some comments, minor style issues and prototypes.
bridges, the EBus bridge has resource ranges it claims exclusively to
map its children into in its BARs. Hence, we need to allocate these
completely and manage them for the children, instead of just passing
allocations through to the PCI layer as we did before.
While being there, split ebus_probe(), which did also contain code
normally belonging into the attach method, into ebus_probe() and
ebus_attach(), and perform some minor cleanups.
- Fix some comments; remove numerous superfluous or outdated ones.
- Correctly pass on the requesting device when handing requests up
to the parent bus.
- Use the complete device name, including unit number, to build the
IOMMU instance name.
- Inline a function that was only used once, and was trivial.
phandle_t. Since both are typedefed to unsigned int, this is more
or less cosmetic.
- Fix the code that determines whether a creator instance was used
for firmware output (and should not be blanked on initialization).
Since r1.2 of dev/fb/creator.c, this consisted comparing a handle of
an instance of a package with a handle of the package itself.
Use the test from r1.1, which utilizes OF_instance_to_package().
Submitted by: Marius Strobl <marius@alchemy.franken.de>
iommu_dvma_vallocseg(), which I botched in r1.32. This bug could
cause an endless loop when a map was loaded and DVMA was scarce,
or that map had a stringent alignment or boundary.
Report and additional testing: Marius Strobl <marius@alchemy.franken.de>
in cpu_fork(). This prevents the stack tracer from running past the
end of the stack (only the pc is checked in that case), which became
fatal when db_print_backtrace() was introduced and called outside
of ddb.
Additional testing: kris
implementation could be characterized as a hybrid of the amd64 and i386
implementations. Specifically, the direct virtual-to-physical mapping is
used if possible and sf_buf_alloc() is used if the direct map cannot.
level of abstraction for any and all CPU mask and CPU bitmap variables
so that platforms have the ability to break free from the hard limit
of 32 CPUs, simply because we don't have more bits in an u_int. Note
that the type is not supposed to solve massive parallelism, where
the number of CPUs can be larger than the width of the widest integral
type. As such, cpumask_t is not supposed to be a compound type. If
such would be necessary in the future, we can deal with the issues
then and there. For now, it can be assumed that the type is integral
and unsigned.
With this commit, all MD definitions start off as u_int. This allows
us to phase-in cpumask_t at our leasure without breaking anything.
Once cpumask_t is used consistently, platforms can switch to wider
(or smaller) types if such would be beneficial (or not; whatever :-)
Compile-tested on: i386
dependent function by the same name and a machine-independent function,
sf_buf_mext(). Aside from the virtue of making more of the code machine-
independent, this change also makes the interface more logical. Before,
sf_buf_free() did more than simply undo an sf_buf_alloc(); it also
unwired and if necessary freed the page. That is now the purpose of
sf_buf_mext(). Thus, sf_buf_alloc() and sf_buf_free() can now be used
as a general-purpose emphemeral map cache.
only. This is a MAJOR incompatible change for the sparc64 platform,
but will not effect FreeBSD on other architectures.
Reviewed by: imp for UPDATING, freebsd-sparc for the change itself.
ever since alpha/alpha/pmap.c revision 1.81 introduced the list allpmaps,
there has been no reason for having this function on Alpha. Briefly,
when pmap_growkernel() relied upon the list of all processes to find and
update the various pmaps to reflect a growth in the kernel's valid
address space, pmap_init2() served to avoid a race between pmap
initialization and pmap_growkernel(). Specifically, pmap_pinit2() was
responsible for initializing the kernel portions of the pmap and
pmap_pinit2() was called after the process structure contained a pointer
to the new pmap for use by pmap_growkernel(). Thus, an update to the
kernel's address space might be applied to the new pmap unnecessarily,
but an update would never be lost.
Introduce d_version field in struct cdevsw, this must always be
initialized to D_VERSION.
Flip sense of D_NOGIANT flag to D_NEEDGIANT, this involves removing
four D_NOGIANT flags and adding 145 D_NEEDGIANT flags.
at it, use the ANSI C generic pointer type for the second argument,
thus matching the documentation.
Remove the now extraneous (and now conflicting) function declarations
in various libc sources. Remove now unnecessary casts.
Reviewed by: bde
MAC address in the EEPROM, and we need to get it from OpenFirmware.
This isn't very pretty but time is lacking to do this in a better
way this near 5.2-RELEASE. This is a RELENG_5_2 candidate.
Original version by: Marius Strobl <marius@alchemy.franken.de>
Tested by: Pete Bentley <pete@sorted.org>
Reviewed by: jake
is useless for threaded programs, multiple threads can not share same
stack.
The alternative signal stack is private for thread, no lock is needed,
the orignal P_ALTSTACK is now moved into td_pflags and renamed to
TDP_ALTSTACK.
For single thread or Linux clone() based threaded program, there is no
semantic changed, because those programs only have one kernel thread
in every process.
Reviewed by: deischen, dfr
1) mp_maxid is a valid FreeBSD CPU ID in the range 0 .. MAXCPU - 1.
2) For all active CPUs in the system, PCPU_GET(cpuid) <= mp_maxid.
Approved by: re (scottl)
Tested on: i386, amd64, alpha
very early (SI_SUB_TUNABLES - 1) and is responsible for setting mp_maxid.
cpu_mp_probe() is now called at SI_SUB_CPU and determines if SMP is
actually present and sets mp_ncpus and all_cpus. Splitting these up
allows an architecture to probe CPUs later than SI_SUB_TUNABLES by just
setting mp_maxid to MAXCPU in cpu_mp_setmaxid(). This could allow the
CPU probing code to live in a module, for example, since modules
sysinit's in modules cannot be invoked prior to SI_SUB_KLD. This is
needed to re-enable the ACPI module on i386.
- For the alpha SMP probing code, use LOCATE_PCS() instead of duplicating
its contents in a few places. Also, add a smp_cpu_enabled() function
to avoid duplicating some code. There is room for further code
reduction later since much of this code is also present in cpu_mp_start().
- All archs besides i386 still set mp_maxid to the same values they set it
to before this change. i386 now sets mp_maxid to MAXCPU.
Tested on: alpha, amd64, i386, ia64, sparc64
Approved by: re (scottl)
physical mapping.
- Move the sf_buf API to its own header file; make struct sf_buf's
definition machine dependent. In this commit, we remove an
unnecessary field from struct sf_buf on the alpha, amd64, and ia64.
Ultimately, we may eliminate struct sf_buf on those architecures
except as an opaque pointer that references a vm page.
cache after a data access error we must discard all cache lines. When
disabled existing cache lines are not invalidated by stores to memory, so
we risk reading stale data that was cached before the data access error if
we don't flush them. This is especially fatal when the memory involved
is the active part of the kernel or user stack. For good measure we also
flush the instruction cache.
This fixes random crashes when the X server probes the PCI bus through
/dev/pci.
Since all callers either passed 0 or 1 for clear_ret, define bit 0 in
the flags for use as clear_ret. Reserve bits 1, 2 and 3 for use by MI
code for possible (but unlikely) future use. The remaining bits are for
use by MD code.
This change is triggered by a need on ia64 to have another knob for
get_mcontext().
A small helper function pmap_is_prefaultable() is added. This function
encapsulate the few lines of pmap_prefault() that actually vary from
machine to machine. Note: pmap_is_prefaultable() and pmap_mincore() have
much in common. Going forward, it's worth considering their merger.
quantities on every other architecture.) This change is required in order
to move pmap_prefault() out of the pmap and into the machine-independent
layer.
evaluating them at compile time rather than at run time. As for x86
and amd64, this requires GCC and it's enabled only if __OPTIMIZE__ is
defined (ie, if at least -O is used).
Reviewed by: jake
systems where the data/stack/etc limits are too big for a 32 bit process.
Move the 5 or so identical instances of ELF_RTLD_ADDR() into imgact_elf.c.
Supply an ia32_fixlimits function. Export the clip/default values to
sysctl under the compat.ia32 heirarchy.
Have mmap(0, ...) respect the current p->p_limits[RLIMIT_DATA].rlim_max
value rather than the sysctl tweakable variable. This allows mmap to
place mappings at sensible locations when limits have been reduced.
Have the imgact_elf.c ld-elf.so.1 placement algorithm use the same
method as mmap(0, ...) now does.
Note that we cannot remove all references to the sysctl tweakable
maxdsiz etc variables because /etc/login.conf specifies a datasize
of 'unlimited'. And that causes exec etc to fail since it can no
longer find space to mmap things.
atomically extracts and holds the physical page that is associated with the
given pmap and virtual address. Such a function is needed to make the
memory mapping optimizations used by, for example, pipes and raw disk I/O
MP-safe.
Reviewed by: tegge
rl(4) driver and put it in a new re(4) driver. The re(4) driver shares
the if_rlreg.h file with rl(4) but is a separate module. (Ultimately
I may change this. For now, it's convenient.)
rl(4) has been modified so that it will never attach to an 8139C+
chip, leaving it to re(4) instead. Only re(4) has the PCI IDs to
match the 8169/8169S/8110S gigE chips. if_re.c contains the same
basic code that was originally bolted onto if_rl.c, with the
following updates:
- Added support for jumbo frames. Currently, there seems to be
a limit of approximately 6200 bytes for jumbo frames on transmit.
(This was determined via experimentation.) The 8169S/8110S chips
apparently are limited to 7.5K frames on transmit. This may require
some more work, though the framework to handle jumbo frames on RX
is in place: the re_rxeof() routine will gather up frames than span
multiple 2K clusters into a single mbuf list.
- Fixed bug in re_txeof(): if we reap some of the TX buffers,
but there are still some pending, re-arm the timer before exiting
re_txeof() so that another timeout interrupt will be generated, just
in case re_start() doesn't do it for us.
- Handle the 'link state changed' interrupt
- Fix a detach bug. If re(4) is loaded as a module, and you do
tcpdump -i re0, then you do 'kldunload if_re,' the system will
panic after a few seconds. This happens because ether_ifdetach()
ends up calling the BPF detach code, which notices the interface
is in promiscuous mode and tries to switch promisc mode off while
detaching the BPF listner. This ultimately results in a call
to re_ioctl() (due to SIOCSIFFLAGS), which in turn calls re_init()
to handle the IFF_PROMISC flag change. Unfortunately, calling re_init()
here turns the chip back on and restarts the 1-second timeout loop
that drives re_tick(). By the time the timeout fires, if_re.ko
has been unloaded, which results in a call to invalid code and
blows up the system.
To fix this, I cleared the IFF_UP flag before calling ether_ifdetach(),
which stops the ioctl routine from trying to reset the chip.
- Modified comments in re_rxeof() relating to the difference in
RX descriptor status bit layout between the 8139C+ and the gigE
chips. The layout is different because the frame length field
was expanded from 12 bits to 13, and they got rid of one of the
status bits to make room.
- Add diagnostic code (re_diag()) to test for the case where a user
has installed a broken 32-bit 8169 PCI NIC in a 64-bit slot. Some
NICs have the REQ64# and ACK64# lines connected even though the
board is 32-bit only (in this case, they should be pulled high).
This fools the chip into doing 64-bit DMA transfers even though
there is no 64-bit data path. To detect this, re_diag() puts the
chip into digital loopback mode and sets the receiver to promiscuous
mode, then initiates a single 64-byte packet transmission. The
frame is echoed back to the host, and if the frame contents are
intact, we know DMA is working correctly, otherwise we complain
loudly on the console and abort the device attach. (At the moment,
I don't know of any way to work around the problem other than
physically modifying the board, so until/unless I can think of a
software workaround, this will have do to.)
- Created re(4) man page
- Modified rlphy.c to allow re(4) to attach as well as rl(4).
Note that this code works for the sample 8169/Marvell 88E1000 NIC
that I have, but probably won't work for the 8169S/8110S chips.
RealTek has sent me some sample NICs, but they haven't arrived yet.
I will probably need to add an rlgphy driver to handle the on-board
PHY in the 8169S/8110S (it needs special DSP initialization).
a correctable DMA error. Failing to do so can cause the error interrupt
to be triggered over and over again.
- Clean up the comments for UEAFSR_* constants, fix a typo (UEAFSR_BLK is
(1 << 23), not (1 << 22)), and add two more. Also, add similar constants
for the CE AFSR bits.
address of the device identified by its phandle_t by traversing OFW's
device tree. The space and address returned by this function can
subsequently be passed to sparc64_fake_bustag() to construct a valid
tag and handle for use by the newbus I/O functions.
Use of this function is expected to be limited to pre-newbus access to
devices, such as consoles and keyboards.
Partially obtained from: tmm
Reviewed by: jake, jmg, tmm
SBus testing made possible by: jake
Tested with: LINT
sockets into machine-dependent files. The rationale for this
migration is illustrated by the modified amd64 allocator. It uses the
amd64's direct map to avoid emphemeral mappings in the kernel's
address space. On an SMP, the emphemeral mappings result in an IPI
for TLB shootdown for each transmitted page. Yuck.
Maintainers of other 64-bit platforms with direct maps should be able
to use the amd64 allocator as a reference implementation.
bus tag is to allow bus space accesses prior to having newbus
fully initialized, such as would be the case for console drivers.
Since barriers are a fundamental part of bus space accesses, not
allowing them on fake tags would defeat the purpose of these tags.
We use the barrier function normally associated with nexus. This
is the barrier used when subordinates haven't defined a barrier
themselves.
