without Giant held.
A quick outline of the locking strategy:
Since all IOMMUs are synchronized, there is a single lock, iommu_mtx,
which protects the hardware registers (where needed) and the global and
per-IOMMU software states. As soon as the IOMMUs are divorced, each struct
iommu_state will have its own mutex (and the remaining global state
will be moved into the struct).
The dvma rman has its own internal mutex; the TSB slots may only be
accessed by the owner of the corresponding resource, so neither needs
extra protection.
Since there is a second access path to maps via LRU queues, the consumer-
provided locking is not sufficient; therefore, each map which is on a
queue is additionally protected by iommu_mtx (in part, there is one
member which only the map owner may access). Each map on a queue may
be accessed and removed from or repositioned in a queue in any context as
long as the lock is held; only the owner may insert a map.
To reduce lock contention, some bus_dma functions remove the map from
the queue temporarily (on behalf of the map owner) for some operations and
reinsert it when they are done. Shorter operations and operations which are
not done on behalf of the lock owner are completely covered by the lock.
To facilitate the locking, reorganize the streaming buffer handling;
while being there, fix an old oversight which would cause the streaming
buffer to always be flushed, regardless of whether streaming was enabled
in the TSB entry. The streaming buffer is still disabled for now, since
there are a number of drivers which lack critical bus_dmamp_sync() calls.
Additional testing by: jake
Add two new arguments to bus_dma_tag_create(): lockfunc and lockfuncarg.
Lockfunc allows a driver to provide a function for managing its locking
semantics while using busdma. At the moment, this is used for the
asynchronous busdma_swi and callback mechanism. Two lockfunc implementations
are provided: busdma_lock_mutex() performs standard mutex operations on the
mutex that is specified from lockfuncarg. dftl_lock() is a panic
implementation and is defaulted to when NULL, NULL are passed to
bus_dma_tag_create(). The only time that NULL, NULL should ever be used is
when the driver ensures that bus_dmamap_load() will not be deferred.
Drivers that do not provide their own locking can pass
busdma_lock_mutex,&Giant args in order to preserve the former behaviour.
sparc64 and powerpc do not provide real busdma_swi functions, so this is
largely a noop on those platforms. The busdma_swi on is64 is not properly
locked yet, so warnings will be emitted on this platform when busdma
callback deferrals happen.
If anyone gets panics or warnings from dflt_lock() being called, please
let me know right away.
Reviewed by: tmm, gibbs
for now. It introduces a OFW PCI bus driver and a generic OFW PCI-PCI
bridge driver. By utilizing these, the PCI handling is much more elegant
now.
The advantages of the new approach are:
- Device enumeration should hopefully be more like on Solaris now,
so unit numbers should match what's printed on the box more
closely.
- Real interrupt routing is implemented now, so cardbus bridges
etc. have at least a chance to work.
- The quirk tables are gone and have been replaced by (hopefully
sufficient) heuristics.
- Much cleaner code.
There was also a report that previously bogus interrupt assignments
are fixed now, which can be attributed to the new heuristics.
A pitfall, and the reason why this is not the default yet, is that
it changes device enumeration, as mentioned above, which can make
it necessary to change the system configuration if more than one
unit of a device type is present (on a system with two hme cars,
for example, it is possible that hme0 becomes hme1 and vice versa
after enabling the option). Systems with multiple disk controllers
may need to be booted into single user (and require manual specification
of the root file system on boot) to adjust the fstab.
Nevertheless, I would like to encourage users to use this option,
so that it can be made the default soon.
In detail, the changes are:
- Introduce an OFW PCI bus driver; it inherits most methods from the
generic PCI bus driver, but uses the firmware for enumeration,
performs additional initialization for devices and firmware-specific
interrupt routing. It also implements an OFW-specific method to allow
child devices to get their firmware nodes.
- Introduce an OFW PCI-PCI bridge driver; again, it inherits most
of the generic PCI-PCI bridge driver; it has it's own method for
interrupt routing, as well as some sparc64-specific methods (one to
get the node again, and one to adjust the bridge bus range, since
we need to reenumerate all PCI buses).
- Convert the apb driver to the new way of handling things.
- Provide a common framework for OFW bridge drivers, used be the two
drivers above.
- Provide a small common framework for interrupt routing (for all
bridge types).
- Convert the psycho driver to the new framework; this gets rid of a
bunch of old kludges in pci_read_config(), and the whole
preinitialization (ofw_pci_init()).
- Convert the ISA MD part and the EBus driver to the new way
interrupts and nodes are handled.
- Introduce types for firmware interrupt properties.
- Rename the old sparcbus_if to ofw_pci_if by repo copy (it is only
required for PCI), and move it to a more correct location (new
support methodsx were also added, and an old one was deprecated).
- Fix a bunch of minor bugs, perform some cleanups.
In some cases, I introduced some minor code duplication to keep the
new code clean, in hopes that the old code will be unifdef'ed soon.
Reviewed in part by: imp
Tested by: jake, Marius Strobl <marius@alchemy.franken.de>,
Sergey Mokryshev <mokr@mokr.net>,
Chris Jackman <cjackNOSPAM@klatsch.org>
Info on u30 firmware provided by: kris
data access errors when trying to read/write to non-existant PCI devices.
fix the psycho bridge to use peek for probing devices. This no longer
fakes it if the OFW node doesn't exist (and the reg == 0).
Reviewed by: jake, tmm
- Move prototypes for sparc64-specific helper functions from bus.h to
bus_private.h
- Move the method pointers from struct bus_dma_tag into a separate
structure; this saves some memory, and allows to use a single method
table for each busdma backend, so that the bus drivers need no longer
be changed if the methods tables need to be modified.
- Remove the hierarchical tag method lookup. It was never really useful,
since the layering is fixed, and the current implementations do not
need to call into parent implementations anyway. Each tag inherits
its method table pointer and cookie from the parent (or the root tag)
now, and the method wrapper macros directly use the method table
of the tag.
- Add a method table to the non-IOMMU backend, remove unnecessary
prototypes, remove the extra parent tag argument.
- Rename sparc64_dmamem_alloc_map() and sparc64_dmamem_free_map() to
sparc64_dma_alloc_map() and sparc64_dma_free_map(), move them to a
better place and use them for all map allocations and deallocations.
- Add a method table to the iommu backend, and staticize functions,
remove the extra parent tag argument.
- Change the psycho and sbus drivers to just set cookie and method table
in the root tag.
- Miscellaneous small fixes.
to the machine-independent parts of the VM. At the same time, this
introduces vm object locking for the non-i386 platforms.
Two details:
1. KSTACK_GUARD has been removed in favor of KSTACK_GUARD_PAGES. The
different machine-dependent implementations used various combinations
of KSTACK_GUARD and KSTACK_GUARD_PAGES. To disable guard page, set
KSTACK_GUARD_PAGES to 0.
2. Remove the (unnecessary) clearing of PG_ZERO in vm_thread_new. In
5.x, (but not 4.x,) PG_ZERO can only be set if VM_ALLOC_ZERO is passed
to vm_page_alloc() or vm_page_grab().
This machine uses a non-standard scheme to specify the interrupts to
be assigned for devices in PCI slots; instead of giving the INO
or full interrupt number (which is done for the other devices in this
box), the firmware interrupt properties contain intpin numbers, which
have to be swizzled as usual on PCI-PCI bridges; however, the PCI host
bridge nodes have no interrupt map, so we need to guess the
correct INO by slot number of the device or the closest PCI-PCI
bridge leading to it, and the intpin.
To do this, this fix makes the following changes:
- Add a newbus method for sparc64 PCI host bridges to guess
the INO, and glue code in ofw_pci_orb_callback() to invoke it based
on a new quirk entry. The guessing is only done for interrupt numbers
too low to contain any IGN found on e450s.
- Create another new quirk entry was created to prevent mapping of EBus
interrupts at PCI level; the e450 has full INOs in the interrupt
properties of EBus devices, so trying to remap them could cause
problems.
- Set both quirk entries for e450s; remove the no-swizzle entry.
- Determine the psycho half (bus A or B) a driver instance manages
in psycho_attach()
- Implement the new guessing method for psycho, using the slot number,
psycho half and property value (intpin).
Thanks go to the testers, especially Brian Denehy, who tested many kernels
for me until I had found the right workaround.
Tested by: Brian Denehy <B.Denehy@90east.com>, jake, fenner,
Marius Strobl <marius@alchemy.franken.de>,
Marian Dobre <mari@onix.ro>
Approved by: re (scottl)
The current name is confusing, because it indicates to
the client that a bus_dmamap_sync() operation is not
necessary when the flag is specified, which is wrong.
The main purpose of this flag is to hint the underlying
architecture that DMA memory should be mapped in a coherent
way, but the architecture can ignore it. But if the
architecture does supports coherent mapping of memory, then
it makes bus_dmamap_sync() calls cheap.
This flag is the same as the one in NetBSD's Bus DMA.
Reviewed by: gibbs, scottl, des (implicitly)
Approved by: re@ (jhb)
BUS_DMASYNC_ definitions remain as before. The does not change the ABI,
and reverts the API to be a bit more compatible and flexible. This has
survived a full 'make universe'.
Approved by: re (bmah)
- Fix visibilty test for LONG_BIT and WORD_BIT. `#if defined(__FOO_VISIBLE)'
is alays wrong because __FOO_VISIBLE is always defined (to 0 for
invisibility).
sys/<arch>/include/limits.h
sys/<arch>/include/_limits.h:
- Style fixes.
Submitted by: bde
Reviewed by: bsdmike
Approved by: re (scottl)
Remove DBL_DIG, DBL_MIN, DBL_MAX and their FLT_ counterparts, they
were marked for deprecation ever since SUSv1 at least.
Only define ULLONG_MIN/MAX and LLONG_MAX if long long type is
supported.
Restore a lost comment in MI _limits.h file and remove it from
sys/limits.h where it does not belong.
that were added to sparc64 and later powerpc, really should have been in
the MI area. But changing that now with insufficient preperation will
just cause too much pain.
Move MD_FETCH() to the MI sys/linker.h file to avoid another two copies
of it.
to get actual constant values. This is in preparation for machine/limits.h
retirement.
Discussed on: standards@
Submitted by: Craig Rodrigues <rodrigc@attbi.com> (*)
Modified by: kan
the cpu dependent files. It will need to be done differently for USIII.
- Simplify the logic for detecting context rollovers. Instead of dealing
with it when the next context switch would cause the context numbers to
rollover, deal with it when they actually do rollover.
- Move some things around in cpu_switch so that we only do 1 membar #Sync
when switching address space, instead of 2.
- Detect kernel threads by comparing the new vm space to vmspace0, instead
if checking if the tlb context is 0.
- Removed some debug code.
enum to an int and redefine the BUS_DMASYNC_* constants as
flags. This allows us to specify several operations in one
call to bus_dmamap_sync() as in NetBSD.
These are called through function pointers so that different implementations
can be provided for cheetah, where the block load instructions may or may
not be a win, and so they can be disabled with the machdep.use_vis tunable.
In terms of raw bandwidth the integer versions are faster, but not allocating
lines in the L2 cache for useless data gives a measurable improvement in user
time for the benchmarks I tested (mostly buildworld with -j8).
As far as I can tell the instructions used are implemented on everything
back to UltraSPARC I, so there should not be a problem with different cpu
types.
can do 64 bytes at a time and don't allocate lines in the L2 cache. These
assume that everything is 64 byte aligned, and that there's more than 128
bytes of data (best for whole pages). The block load and store instructions
don't follow normal memory ordering rules and require either a memory barrier
or move between registers before the data can actually be used. This
implementation correctly shuffles around 3 out of the 4 sets of registers
in order to avoid memory barriers expect for the last 2 blocks.
used to support block copy and zero operations in the kernel which use the
floating point registers.
- While I'm changing the size, improve the layout of struct pcb, sort by size,
then alphabetical etc.
- Add some assertions to validate assumptions made about how the pcb is
allocated.
pages which represent actual physical memory we must strip off the fake
page in order to allow illegal aliases to be detected. Otherwise we map
uncacheable in the virtual and physical caches and set the side effect bit,
as is required for mapping device memory.
This fixes gstat on sparc64, which wants to mmap kernel memory through a
character device.
where physical addresses larger than virtual addresses, such as i386s
with PAE.
- Use this to represent physical addresses in the MI vm system and in the
i386 pmap code. This also changes the paddr parameter to d_mmap_t.
- Fix printf formats to handle physical addresses >4G in the i386 memory
detection code, and due to kvtop returning vm_paddr_t instead of u_long.
Note that this is a name change only; vm_paddr_t is still the same as
vm_offset_t on all currently supported platforms.
Sponsored by: DARPA, Network Associates Laboratories
Discussed with: re, phk (cdevsw change)
on future UltraSPARC cpus for which the data cache is not direct mapped.
- Move UltraSPARC I and II (spitfire, blackbird, sapphire, sabre) specific
functions to spitfire.c, and add cheetah.c for UltraSPARC III specific
functions. Initially just cache flushing, but there are a few other
functions that will need to move here.
- Add an ipi handler for data cache flushing on UltraSPARC III.
- Use function pointers to select the right cache flushing functions based
on cpu_impl.
With this it is possible to boot single user from an mfs root on UltraSPARC
III systems, including spinning up secondary processors. There is currently
no support for the host to pci bridge, and no documentation for it is
publically available.
Thanks to Oleg Derevenetz for providing access to a system with UltraSPARC
III+ cpus.
are machine dependent because they are not required to update the tlb when
mappings are added or removed, and doing so is machine dependent.
In addition, an implementation may require that pages mapped with pmap_kenter
have a backing vm_page_t, which is not necessarily true of all physical
pages, and so may choose to pass the vm_page_t to pmap_kenter instead of the
physical address in order to make this requirement clear.
o Add a MD header private to libc called _fpmath.h; this header
contains bitfield layouts of MD floating-point types.
o Add a MI header private to libc called fpmath.h; this header
contains bitfield layouts of MI floating-point types.
o Add private libc variables to lib/libc/$arch/gen/infinity.c for
storing NaN values.
o Add __double_t and __float_t to <machine/_types.h>, and provide
double_t and float_t typedefs in <math.h>.
o Add some C99 manifest constants (FP_ILOGB0, FP_ILOGBNAN, HUGE_VALF,
HUGE_VALL, INFINITY, NAN, and return values for fpclassify()) to
<math.h> and others (FLT_EVAL_METHOD, DECIMAL_DIG) to <float.h> via
<machine/float.h>.
o Add C99 macro fpclassify() which calls __fpclassify{d,f,l}() based
on the size of its argument. __fpclassifyl() is never called on
alpha because (sizeof(long double) == sizeof(double)), which is good
since __fpclassifyl() can't deal with such a small `long double'.
This was developed by David Schultz and myself with input from bde and
fenner.
PR: 23103
Submitted by: David Schultz <dschultz@uclink.Berkeley.EDU>
(significant portions)
Reviewed by: bde, fenner (earlier versions)
counterparts to bus_dmamem_alloc() and bus_dmamem_free(). This allows
the caller to specify the size of the allocation instead of it defaulting
to the max_size field of the busdma tag.
This is intended to aid in converting drivers to busdma. Lots of
hardware cannot understand scatter/gather lists, which forces the
driver to copy the i/o buffers to a single contiguous region
before sending it to the hardware. Without these new methods, this
would require a new busdma tag for each operation, or a complex
internal allocator/cache for each driver.
Allocations greater than PAGE_SIZE are rounded up to the next
PAGE_SIZE by contigmalloc(), so this is not suitable for multiple
static allocations that would be better served by a single
fixed-length subdivided allocation.
Reviewed by: jake (sparc64)
1.) Fix an off-by-one in the DVMA space handling, which would make it
possible to allocate one page beyond the end of the DVMA area.
This page was aliased to the first page. Apparently, this bug was
responsible for the trashed nvram/eeprom some people were reporting,
in conjunction with a number of unfortunate coincidences.
2.) Fix broken boundary and and lowaddr calculations.
3.) Fix a memory leak on an error path.
4.) Update a outdated comment to reflect the introduction of IOMMU_MAX_PRE,
make the usage of IOMMU_MAX_PRE more consistent and KASSERT that the
preallocation size is not 0.
5.) Fix a case where an error return was lost.
6.) When signalling an error to the caller by invoking the callback, do
not use a segment pointer of NULL for compatability with existing
drivers.
Also, increase the maximum segment number to 64; it is rather arbitrary,
with the exception of the of the stack space consumed by the segment
array.
Special thanks go to Harti Brandt <brandt@fokus.fraunhofer.de> for
spotting 4 and 5, and testing many iterations of patches.
Pointy hats to: tmm
