We forgot a space in some log messages, so the output ended
up looking like
gicv3_dist_write: invalid guest write at offset 0000000000008000size 8
with a missing space before "size". Add the missing spaces.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220303202341.2232284-5-peter.maydell@linaro.org
For debugging guest use of the ITS, it can be helpful to trace
when the ITS reads and writes the in-memory tables.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220303202341.2232284-3-peter.maydell@linaro.org
When debugging code that's using the ITS, it's helpful to
see tracing of the ITS commands that the guest executes. Add
suitable trace events.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220303202341.2232284-2-peter.maydell@linaro.org
In most of the ITS command processing, we check different error
possibilities one at a time and log them appropriately. In
process_mapti() and process_mapd() we have code which checks
multiple error cases at once, which means the logging is less
specific than it could be. Split those cases up.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-14-peter.maydell@linaro.org
When handling MAPI/MAPTI, we allow the supplied interrupt ID to be
either 1023 or something in the valid LPI range. This is a mistake:
only a real valid LPI is allowed. (The general behaviour of the ITS
is that most interrupt ID fields require a value in the LPI range;
the exception is that fields specifying a doorbell value, which are
all in GICv4 commands, allow also 1023 to mean "no doorbell".)
Remove the condition that incorrectly allows 1023 here.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-13-peter.maydell@linaro.org
In the MAPC command, if V=0 this is a request to delete a collection
table entry and the rdbase field of the command packet will not be
used. In particular, the specification says that the "UNPREDICTABLE
if rdbase is not valid" only applies for V=1.
We were doing a check-and-log-guest-error on rdbase regardless of
whether the V bit was set, and also (harmlessly but confusingly)
storing the contents of the rdbase field into the updated collection
table entry. Update the code so that if V=0 we don't check or use
the rdbase field value.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-12-peter.maydell@linaro.org
Currently we track in the TableDesc and CmdQDesc structs the state of
the GITS_BASER<n> and GITS_CBASER Valid bits. However we aren't very
consistent abut checking the valid field: we test it in update_cte()
and update_dte(), but not anywhere else we look things up in tables.
The GIC specification says that it is UNPREDICTABLE if a guest fails
to set any of these Valid bits before enabling the ITS via
GITS_CTLR.Enabled. So we can choose to handle Valid == 0 as
equivalent to a zero-length table. This is in fact how we're already
catching this case in most of the table-access paths: when Valid is 0
we leave the num_entries fields in TableDesc or CmdQDesc set to zero,
and then the out-of-bounds check "index >= num_entries" that we have
to do anyway before doing any of these table lookups will always be
true, catching the no-valid-table case without any extra code.
So we can remove the checks on the valid field from update_cte()
and update_dte(): since these happen after the bounds check there
was never any case when the test could fail. That means the valid
fields would be entirely unused, so just remove them.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-11-peter.maydell@linaro.org
Make the update_ite() struct use the new ITEntry struct, so that
callers don't need to assemble the in-memory ITE data themselves, and
only get_ite() and update_ite() need to care about that in-memory
layout. We can then drop the no-longer-used IteEntry struct
definition.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-10-peter.maydell@linaro.org
In get_ite() we currently return the caller some of the fields of an
Interrupt Table Entry via a set of pointer arguments, and validate
some of them internally (interrupt type and valid bit) to return a
simple true/false 'valid' indication. Define a new ITEntry struct
which has all the fields that the in-memory ITE has, and bring the
get_ite() function in to line with get_dte() and get_cte().
This paves the way for handling virtual interrupts, which will want
a different subset of the fields in the ITE. Handling them under
the old "lots of pointer arguments" scheme would have meant a
confusingly large set of arguments for this function.
The new struct ITEntry is obviously confusably similar to the
existing IteEntry struct, whose fields are the raw 12 bytes
of the in-memory ITE. In the next commit we will make update_ite()
use ITEntry instead of IteEntry, which will allow us to delete
the IteEntry struct and remove the confusion.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-9-peter.maydell@linaro.org
The get_ite() code has some awkward nested if statements; clean
them up by returning early if the memory accesses fail.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-8-peter.maydell@linaro.org
In get_ite() and update_ite() we work with a 12-byte in-guest-memory
table entry, which we intend to handle as an 8-byte value followed by
a 4-byte value. Unfortunately the calculation of the address of the
4-byte value is wrong, because we write it as:
table_base_address + (index * entrysize) + 4
(obfuscated by the way the expression has been written)
when it should be + 8. This bug meant that we overwrote the top
bytes of the 8-byte value with the 4-byte value. There are no
guest-visible effects because the top half of the 8-byte value
contains only the doorbell interrupt field, which is used only in
GICv4, and the two bugs in the "write ITE" and "read ITE" codepaths
cancel each other out.
We can't simply change the calculation, because this would break
migration of a (TCG) guest from the old version of QEMU which had
in-guest-memory interrupt tables written using the buggy version of
update_ite(). We must also at the same time change the layout of the
fields within the ITE_L and ITE_H values so that the in-memory
locations of the fields we care about (VALID, INTTYPE, INTID and
ICID) stay the same.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-7-peter.maydell@linaro.org
Make update_cte() take a CTEntry struct rather than all the fields
of the new CTE as separate arguments.
This brings it into line with the update_dte() API.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-6-peter.maydell@linaro.org
In the ITS, a CTE is an entry in the collection table, which contains
multiple fields. Currently the function get_cte() which reads one
entry from the device table returns a success/failure boolean and
passes back the raw 64-bit integer CTE value via a pointer argument.
We then extract fields from the CTE as we need them.
Create a real C struct with the same fields as the CTE, and
populate it in get_cte(), so that that function and update_cte()
are the only ones which need to care about the in-guest-memory
format of the CTE.
This brings get_cte()'s API into line with get_dte().
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-5-peter.maydell@linaro.org
Make update_dte() take a DTEntry struct rather than all the fields of
the new DTE as separate arguments.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-4-peter.maydell@linaro.org
In the ITS, a DTE is an entry in the device table, which contains
multiple fields. Currently the function get_dte() which reads one
entry from the device table returns it as a raw 64-bit integer,
which we then pass around in that form, only extracting fields
from it as we need them.
Create a real C struct with the same fields as the DTE, and
populate it in get_dte(), so that that function and update_dte()
are the only ones that need to care about the in-guest-memory
format of the DTE.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-3-peter.maydell@linaro.org
Currently the ITS accesses each 8-byte doubleword in a 4-doubleword
command packet with a separate address_space_ldq_le() call. This is
awkward because the individual command processing functions have
ended up with code to handle "load more doublewords out of the
packet", which is both unwieldy and also a potential source of bugs
because it's not obvious when looking at a line that pulls a field
out of the 'value' variable which of the 4 doublewords that variable
currently holds.
Switch to using address_space_map() to map the whole command packet
at once and fish the four doublewords out of it. Then each process_*
function can start with a few lines of code that extract the fields
it cares about.
This requires us to split out the guts of process_its_cmd() into a
new do_process_its_cmd(), because we were previously overloading the
value and offset arguments as a backdoor way to directly pass the
devid and eventid from a write to GITS_TRANSLATER. The new
do_process_its_cmd() takes those arguments directly, and
process_its_cmd() is just a wrapper that does the "read fields from
command packet" part.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220201193207.2771604-2-peter.maydell@linaro.org
Implement the ITS MOVI command. This command specifies a (physical) LPI
by DeviceID and EventID and provides a new ICID for it. The ITS must
find the interrupt translation table entry for the LPI, which will
tell it the old ICID. It then moves the pending state of the LPI from
the old redistributor to the new one and updates the ICID field in
the translation table entry.
This is another GICv3 ITS command that we forgot to implement. Linux
does use this one, but only if the guest powers off one of its CPUs.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220122182444.724087-15-peter.maydell@linaro.org
Implement the ITS MOVALL command, which takes all the pending
interrupts on a source redistributor and makes the not-pending on
that source redistributor and pending on a destination redistributor.
This is a GICv3 ITS command which we forgot to implement. (It is
not used by Linux guests.)
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220122182444.724087-14-peter.maydell@linaro.org
Currently when we fill in a TableDesc based on the value the guest
has written to the GITS_BASER<n> register, we calculate both:
* num_entries : the number of entries in the table, constrained
by the amount of memory the guest has given it
* num_ids : the number of IDs we support for this table,
constrained by the implementation choices and the architecture
(eg DeviceIDs are 16 bits, so num_ids is 1 << 16)
When validating ITS commands, however, we check only num_ids,
thus allowing a broken guest to specify table entries that
index off the end of it. This will only corrupt guest memory,
but the ITS is supposed to reject such commands as invalid.
Instead of calculating both num_entries and num_ids, set
num_entries to the minimum of the two limits, and check that.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220122182444.724087-13-peter.maydell@linaro.org
The ITS has a bank of 8 GITS_BASER<n> registers, which allow the
guest to specify the base address of various data tables. Each
register has a read-only type field indicating which table it is for
and a read-write field where the guest can write in the base address
(among other things). We currently allow the guest to write the
writeable fields for all eight registers, even if the type field is 0
indicating "Unimplemented". This means the guest can provoke QEMU
into asserting by writing an address into one of these unimplemented
base registers, which bypasses the "if (!value) continue" check in
extract_table_params() and lets us hit the assertion that the type
field is one of the permitted table types.
Prevent the assertion by not allowing the guest to write to the
unimplemented base registers. This means their value will remain 0
and extract_table_params() will ignore them.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220122182444.724087-12-peter.maydell@linaro.org
The MemoryRegionOps gicv3_its_translation_ops currently provides only
a .write_with_attrs function, because the only register in this
region is the write-only GITS_TRANSLATER. However, if you don't
provide a read function and the guest tries reading from this memory
region, QEMU will crash because
memory_region_read_with_attrs_accessor() calls a NULL pointer.
Add a read function which always returns 0, to cover both bogus
attempts to read GITS_TRANSLATER and also reads from the rest of the
region, which is documented to be reserved, RES0.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220122182444.724087-11-peter.maydell@linaro.org
The ITS specification says that when the guest writes to GITS_CBASER
this causes GITS_CREADR to be cleared. However it does not have an
equivalent clause for GITS_CWRITER. (This is because GITS_CREADR is
read-only, but GITS_CWRITER is writable and the guest can initialize
it.) Remove the code that clears GITS_CWRITER on GITS_CBASER writes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220122182444.724087-6-peter.maydell@linaro.org
The current ITS code clears GITS_CREADR when GITS_CTLR.ENABLED is set.
This is not correct -- guest code can validly clear ENABLED and then
set it again and expect the ITS to continue processing where it left
off. Remove the erroneous assignment.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220122182444.724087-5-peter.maydell@linaro.org
In our implementation, all ITSes connected to a GIC share a single
AddressSpace, which we keep in the GICv3State::dma_as field and
initialized based on the GIC's 'sysmem' property. The right place
to set it up by calling address_space_init() is therefore in the
GIC's realize method, not the ITS's realize.
This fixes a theoretical bug where QEMU hangs on startup if the board
model creates two ITSes connected to the same GIC -- we would call
address_space_init() twice on the same AddressSpace*, which creates
an infinite loop in the QTAILQ that softmmu/memory.c uses to store
its list of AddressSpaces and causes any subsequent attempt to
iterate through that list to loop forever. There aren't any board
models like that in the tree at the moment, though.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220122182444.724087-4-peter.maydell@linaro.org
The ITS currently has no tracepoints; add a minimal set
that allows basic monitoring of guest register accesses and
reading of commands from the command queue.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220122182444.724087-3-peter.maydell@linaro.org
In process_its_cmd(), we read an ICID out of the interrupt table
entry, and then use it as an index into the collection table. Add a
check that it is within range for the collection table first.
This check is not strictly necessary, because:
* we range check the ICID from the guest before writing it into
the interrupt table entry, so the the only way to get an
out of range ICID in process_its_cmd() is if a badly-behaved
guest is writing directly to the interrupt table memory
* the collection table is in guest memory, so QEMU won't fall
over if we read off the end of it
However, it seems clearer to include the check.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-id: 20220111171048.3545974-14-peter.maydell@linaro.org
In a few places in the ITS command handling functions, we were
doing the range-check of an event ID or device ID only after using
it as a table index; move the checks to before the uses.
This misordering wouldn't have very bad effects because the
tables are in guest memory anyway.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-id: 20220111171048.3545974-13-peter.maydell@linaro.org
The ITS has several tables which all share a similar format,
described by the TableDesc struct: the guest may configure them
to be a single-level table or a two-level table. Currently we
open-code the process of finding the table entry in all the
functions which read or write the device table or the collection
table. Factor out the "get the address of the table entry"
logic into a new function, so that the code which needs to
read or write a table entry only needs to call table_entry_addr()
and then perform a suitable load or store to that address.
Note that the error handling is slightly complicated because
we want to handle two cases differently:
* failure to read the L1 table entry should end up causing
a command stall, like other kinds of DMA error
* an L1 table entry that says there is no L2 table for this
index (ie whose valid bit is 0) must result in us treating
the table entry as not-valid on read, and discarding
writes (this is mandated by the spec)
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-id: 20220111171048.3545974-12-peter.maydell@linaro.org
Fix process_mapd() to consistently return CMD_STALL for memory
errors and CMD_CONTINUE for parameter errors, as we claim in the
comments that we do.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220111171048.3545974-11-peter.maydell@linaro.org
Fix process_mapc() to consistently return CMD_STALL for memory
errors and CMD_CONTINUE for parameter errors, as we claim in the
comments that we do.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220111171048.3545974-10-peter.maydell@linaro.org
Fix process_mapti() to consistently return CMD_STALL for memory
errors and CMD_CONTINUE for parameter errors, as we claim in the
comments that we do.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220111171048.3545974-9-peter.maydell@linaro.org
Refactor process_its_cmd() so that it consistently uses
the structure
do thing;
if (error condition) {
return early;
}
do next thing;
rather than doing some of the work nested inside if (not error)
code blocks.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220111171048.3545974-8-peter.maydell@linaro.org
Fix process_its_cmd() to consistently return CMD_STALL for
memory errors and CMD_CONTINUE for parameter errors, as
we claim in the comments that we do.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220111171048.3545974-7-peter.maydell@linaro.org
When an ITS detects an error in a command, it has an
implementation-defined (CONSTRAINED UNPREDICTABLE) choice of whether
to ignore the command, proceeding to the next one in the queue, or to
stall the ITS command queue, processing nothing further. The
behaviour required when the read of the command packet from memory
fails is less clearly documented, but the same set of choices as for
command errors seem reasonable.
The intention of the QEMU implementation, as documented in the
comments, is that if we encounter a memory error reading the command
packet or one of the various data tables then we should stall, but
for command parameter errors we should ignore the queue and continue.
However, we don't actually do this. To get the desired behaviour,
the various process_* functions need to return true to cause
process_cmdq() to advance to the next command and keep processing,
and false to stall command processing. What they mostly do is return
false for any kind of error.
To make the code clearer, replace the 'bool' return from the process_
functions with an enum which may be either CMD_STALL or CMD_CONTINUE.
In this commit no behaviour changes; in subsequent commits we will
adjust the error-return paths for the process_ functions one by one.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Message-id: 20220111171048.3545974-6-peter.maydell@linaro.org
In process_cmdq(), we read 64 bits of the command packet, which
contain the command identifier, which we then switch() on to dispatch
to an appropriate sub-function. However, if address_space_ldq_le()
reports a memory transaction failure, we still read the command
identifier out of the data and switch() on it. Restructure the code
so that we stop immediately (stalling the command queue) in this
case.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220111171048.3545974-5-peter.maydell@linaro.org
process_its_cmd() returns a bool, like all the other process_ functions.
However we were putting its return value into 'res', not 'result',
which meant we would ignore it when deciding whether to continue
or stall the command queue. Fix the typo.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Message-id: 20220111171048.3545974-4-peter.maydell@linaro.org
The bounds check on the number of interrupt IDs is correct, but
doesn't match our convention; change the variable name, initialize it
to the 2^n value rather than (2^n)-1, and use >= instead of > in the
comparison.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-id: 20220111171048.3545974-3-peter.maydell@linaro.org
In process_its_cmd() and process_mapti() we must check the
event ID against a limit defined by the size field in the DTE,
which specifies the number of ID bits minus one. Convert
this code to our num_foo convention:
* change the variable names
* use uint64_t and 1ULL when calculating the number
of valid event IDs, because DTE.SIZE is 5 bits and
so num_eventids may be up to 2^32
* fix the off-by-one error in the comparison
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-id: 20220111171048.3545974-2-peter.maydell@linaro.org
In several places we have a local variable max_l2_entries which is
the number of entries which will fit in a level 2 table. The
calculations done on this value are correct; rename it to
num_l2_entries to fit the convention we're using in this code.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
The ITS code has to check whether various parameters passed in
commands are in-bounds, where the limit is defined in terms of the
number of bits that are available for the parameter. (For example,
the GITS_TYPER.Devbits ID register field specifies the number of
DeviceID bits minus 1, and device IDs passed in the MAPTI and MAPD
command packets must fit in that many bits.)
Currently we have off-by-one bugs in many of these bounds checks.
The typical problem is that we define a max_foo as 1 << n. In
the Devbits example, we set
s->dt.max_ids = 1UL << (GITS_TYPER.Devbits + 1).
However later when we do the bounds check we write
if (devid > s->dt.max_ids) { /* command error */ }
which incorrectly permits a devid of 1 << n.
These bugs will not cause QEMU crashes because the ID values being
checked are only used for accesses into tables held in guest memory
which we access with address_space_*() functions, but they are
incorrect behaviour of our emulation.
Fix them by standardizing on this pattern:
* bounds limits are named num_foos and are the 2^n value
(equal to the number of valid foo values)
* bounds checks are either
if (fooid < num_foos) { good }
or
if (fooid >= num_foos) { bad }
In this commit we fix the handling of the number of IDs
in the device table and the collection table, and the number
of commands that will fit in the command queue.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Use FIELD macros to handle CTEs, rather than ad-hoc mask-and-shift.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Currently the ITS code that reads and writes DTEs uses open-coded
shift-and-mask to assemble the various fields into the 64-bit DTE
word. The names of the macros used for mask and shift values are
also somewhat inconsistent, and don't follow our usual convention
that a MASK macro should specify the bits in their place in the word.
Replace all these with use of the FIELD macro.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The MAPI command takes arguments DeviceID, EventID, ICID, and is
defined to be equivalent to MAPTI DeviceID, EventID, EventID, ICID.
(That is, where MAPTI takes an explicit pINTID, MAPI uses the EventID
as the pINTID.)
We didn't quite get this right. In particular the error checks for
MAPI include "EventID does not specify a valid LPI identifier", which
is the same as MAPTI's error check for the pINTID field. QEMU's code
skips the pINTID error check entirely in the MAPI case.
We can fix this bug and in the process simplify the code by switching
to the obvious implementation of setting pIntid = eventid early
if ignore_pInt is true.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The GITS_TYPE_PHYSICAL define is the value we set the
GITS_TYPER.Physical field to -- this is 1 to indicate that we support
physical LPIs. (Support for virtual LPIs is the GITS_TYPER.Virtual
field.) We also use this define as the *value* that we write into an
interrupt translation table entry's INTTYPE field, which should be 1
for a physical interrupt and 0 for a virtual interrupt. Finally, we
use it as a *mask* when we read the interrupt translation table entry
INTTYPE field.
Untangle this confusion: define an ITE_INTTYPE_VIRTUAL and
ITE_INTTYPE_PHYSICAL to be the valid values of the ITE INTTYPE
field, and replace the ad-hoc collection of ITE_ENTRY_* defines with
use of the FIELD() macro to define the fields of an ITE and the
FIELD_EX64() and FIELD_DP64() macros to read and write them.
We use ITE in the new setup, rather than ITE_ENTRY, because
ITE stands for "Interrupt translation entry" and so the extra
"entry" would be redundant.
We take the opportunity to correct the name of the field that holds
the GICv4 'doorbell' interrupt ID (this is always the value 1023 in a
GICv3, which is why we were calling it the 'spurious' field).
The GITS_TYPE_PHYSICAL define is then used in only one place, where
we set the initial GITS_TYPER value. Since GITS_TYPER.Physical is
essentially a boolean, hiding the '1' value behind a macro is more
confusing than helpful, so expand out the macro there and remove the
define entirely.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
We set the TableDesc entry_sz field from the appropriate
GITS_BASER.ENTRYSIZE field. That ID register field specifies the
number of bytes per table entry minus one. However when we use
td->entry_sz we assume it to be the number of bytes per table entry
(for instance we calculate the number of entries in a page by
dividing the page size by the entry size).
The effects of this bug are:
* we miscalculate the maximum number of entries in the table,
so our checks on guest index values are wrong (too lax)
* when looking up an entry in the second level of an indirect
table, we calculate an incorrect index into the L2 table.
Because we make the same incorrect calculation on both
reads and writes of the L2 table, the guest won't notice
unless it's unlucky enough to use an index value that
causes us to index off the end of the L2 table page and
cause guest memory corruption in whatever follows
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The extract_table_params() decodes the fields in the GITS_BASER<n>
registers into TableDesc structs. Since the fields are the same for
all the GITS_BASER<n> registers, there is currently a lot of code
duplication within the switch (type) statement. Refactor so that the
cases include only what is genuinely different for each type:
the calculation of the number of bits in the ID value that indexes
into the table.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
In extract_table_params() we process each GITS_BASER<n> register. If
the register's Valid bit is not set, this means there is no
in-guest-memory table and so we should not try to interpret the other
fields in the register. This was incorrectly coded as a 'return'
rather than a 'break', so instead of looping round to process the
next GITS_BASER<n> we would stop entirely, treating any later tables
as being not valid also.
This has no real guest-visible effects because (since we don't have
GITS_TYPER.HCC != 0) the guest must in any case set up all the
GITS_BASER<n> to point to valid tables, so this only happens in an
odd misbehaving-guest corner case.
Fix the check to 'break', so that we leave the case statement and
loop back around to the next GITS_BASER<n>.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The TableDesc struct defines properties of the in-guest-memory tables
which the guest tells us about by writing to the GITS_BASER<n>
registers. This struct currently has a union 'maxids', but all the
fields of the union have the same type (uint32_t) and do the same
thing (record one-greater-than the maximum ID value that can be used
as an index into the table).
We're about to add another table type (the GICv4 vPE table); rather
than adding another specifically-named union field for that table
type with the same type as the other union fields, remove the union
entirely and just have a 'uint32_t max_ids' struct field.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
We currently define a bitmask for the GITS_CTLR ENABLED bit in
two ways: as ITS_CTLR_ENABLED, and via the FIELD() macro as
R_GITS_CTLR_ENABLED_MASK. Consistently use the FIELD macro version
everywhere and remove the redundant ITS_CTLR_ENABLED define.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
The checks in the ITS on the rdbase values in guest commands are
off-by-one: they permit the guest to pass us a value equal to
s->gicv3->num_cpu, but the valid values are 0...num_cpu-1. This
meant the guest could cause us to index off the end of the
s->gicv3->cpu[] array when calling gicv3_redist_process_lpi(), and we
would probably crash.
(This is not a security bug, because this code is only usable
with emulation, not with KVM.)
Cc: qemu-stable@nongnu.org
Fixes: 17fb5e36aa ("hw/intc: GICv3 redistributor ITS processing")
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
