In sbi_dbtr_read_trig(), the range check is:
if (trig_idx_base + trig_count >= hs->total_trigs)
When trig_idx_base and trig_count are both unsigned long values supplied
by S-mode, their sum can wrap past ULONG_MAX to a small value, making
the check pass. For example trig_idx_base=1, trig_count=ULONG_MAX wraps
to 0, which is less than total_trigs.
This allows the subsequent for_each_trig_entry loop to access trigger
entries far beyond the triggers[] array, corrupting M-mode heap memory
via CSR read-back writes and leaking M-mode internal state to S-mode
shared memory.
Rewrite the condition as trig_count >= total_trigs - trig_idx_base. The
subtraction is safe because the preceding check already guarantees
trig_idx_base < total_trigs.
Fixes: 97f234f15c ("lib: sbi: Introduce the SBI debug triggers extension support")
Signed-off-by: liutong <liutong@iscas.ac.cn>
Reviewed-by: Himanshu Chauhan <himanshu.chauhan@oss.qualcomm.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260910052629.3359493-2-liutong@iscas.ac.cn
Signed-off-by: Anup Patel <anup@brainfault.org>
The per-HART event list is kept sorted by expiry time: the head is the
event programmed into the timer device by __sbi_timer_update_device(),
and sbi_timer_process() stops at the first unexpired entry.
__sbi_timer_event_start() finds the first event that expires after the
new one, but then inserts the new event with sbi_list_add(), i.e. after
that event instead of before it. Starting an event at time 50 while an
event at time 1000 is pending yields the list [1000, 50]: the device is
programmed for 1000 and the earlier event fires 950 ticks late.
Today each HART only ever has its S-mode timer event on the list, so
the misordering needs a second timer event user to show up. Once there
is one, on HARTs without Sstc an S-mode timer programmed while a later
M-mode event is pending is delivered only when that later event
expires.
Use sbi_list_add_tail() so the new event goes in front of next_ev.
Fixes: b7fa8a246c ("lib: sbi_timer: Add support for timer events")
Signed-off-by: Nickolai Zeldovich <nickolai@csail.mit.edu>
Assisted-By: Claude Fable 5.1
Reviewe-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20261011222412.769530-1-nickolai@csail.mit.edu
Signed-off-by: Anup Patel <anup@brainfault.org>
sbi_pmu_event_get_info() maps the caller's shared memory into M-mode with
sbi_hart_protection_temp_map_range(). The must-be-zero check added inside the
event loop returns straight to the supervisor without reaching the matching
temp_unmap_range() at the end of the function, so the mapping is left behind.
This has a security impact beyond leaking a mapping. Under Smepmp the
temporary mapping is a single reserved PMP entry, SBI_SMEPMP_RESV_ENTRY.
sbi_hart_smepmp_temp_map_range() refuses to reprogram that entry while it is
already enabled:
if (sbi_hart_is_pmp_enabled(SBI_SMEPMP_RESV_ENTRY))
return SBI_ENOSPC;
and no caller in the tree checks that return value. sbi_pmu.c, sbi_sse.c,
sbi_mpxy.c, sbi_dbtr.c and sbi_ecall_dbcn.c all discard it. So once the entry
is left behind, every later shared memory ecall on that hart runs with PMP 0
still covering the previous region, and M-mode faults the first time it
dereferences the new buffer.
A supervisor triggers this with one ecall carrying a reserved bit in the first
event_idx. The next SBI_EXT_DBCN_CONSOLE_WRITE then takes an M-mode load
access fault on the console buffer and the firmware stops making forward
progress, taking DBCN, SSE, MPXY and DBTR down with it on that hart. It also
breaks the Smepmp invariant that M-mode does not retain standing access to
supervisor memory.
Confirmed on qemu virt with a bare metal S-mode payload, QEMU 8.2.2:
build -cpu protection result
master 6ad246a1 rv64 pmp completes
master 6ad246a1 rv64,smepmp=true epmp firmware dies
master + this patch rv64,smepmp=true epmp completes
The middle row leaves this as the last entry in the QEMU exception trace,
where the epc is inside nputs() in M-mode and the tval is the supervisor
console buffer:
cause:0000000000000005, epc:0x000000008000204c,
tval:0x0000000080500000, desc=fault_load
Without Smepmp the map and unmap pair compiles down to nothing, which is why
the first row survives.
This only affects master. v1.9 and earlier have the same map and unmap pair
but no return between them, so no released version needs a backport.
Route the early exit through the existing unmap.
Fixes: d97cfb33ae ("lib: sbi_pmu: Return invalid param error for reserved event_idx bits")
Signed-off-by: Sebastian Alba Vives <sebasjosue84@gmail.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20261012022033.1165060-1-sebasjosue84@gmail.com
Signed-off-by: Anup Patel <anup@brainfault.org>
mask_len counts mask bits: it is derived from vlenb * 8, and the mask
indexing below divides by 8 to turn a bit index into a byte index.
get_vreg() counts bytes: its size argument feeds "vsetvli x0, %0, e8,
m8" and a vse8.v, both of which operate on 8-bit elements.
Both emulators pass mask_len directly as get_vreg()'s size, so the
firmware writes eight times more than intended into mask[], a
MASK_BUFFLEN / 8 == 128 byte stack buffer. The write stays in bounds
only while vlenb * 8 <= 128, that is VLEN <= 128. Above that it runs
past the end of the buffer and corrupts the M-mode stack, including the
saved return address.
The pre-rework code passed vlenb, a byte count, which was correct; the
chunked version replaced it with a bit count.
Reachable from S-mode or U-mode with a single masked, misaligned vector
load or store on a hart whose VLEN exceeds 128 bits. Observed on
qemu-system-riscv64 with -cpu rv64,v=true,vlen=256,zicclsm=false: the
firmware takes an instruction access fault in M-mode with mepc and ra
both zero. The same payload at vlen=128 enters the emulator and
completes normally.
Convert pos and size to bytes at both call sites. mask_len is a power of
two and at least 8, and pos is already a multiple of mask_len, so both
divisions are exact.
Fixes: a8be5e94 ("lib: sbi: Rework misaligned vector load/store")
Signed-off-by: Sebastian Alba Vives <sebasjosue84@gmail.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260922030633.10886-1-sebasjosue84@gmail.com
Signed-off-by: Anup Patel <anup@brainfault.org>
SKIP_MATCH selects the first counter in the caller-provided set, but the
current path checks only that the index exists and is active. A supervisor
can select an active hardware counter for a firmware event. CLEAR_VALUE or
AUTO_START then indexes firmware-counter state with
ctr_idx - num_hw_ctrs, which underflows. The inverse mismatch sends a
firmware-counter index to hardware-counter operations.
Reject the selected counter when its hardware/firmware class cannot monitor
the requested event. Return SBI_ENOTSUPP, matching the PMU specification's
case where the selected counter cannot monitor that event.
Fixes: 13d40f21d5 ("lib: sbi: Add PMU support")
Signed-off-by: Pengpeng Hou <pengpeng@iscas.ac.cn>
Tested-by: Sebastian Alba Vives <sebasjosue84@gmail.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260827115718.50124-1-pengpeng@iscas.ac.cn
Signed-off-by: Anup Patel <anup@brainfault.org>
sbi_hsm_hart_start() acquires the target hart's start ticket, then
writes next_arg1, next_addr and next_mode before checking whether the
target is stopped.
A start request for a non-retentively suspended hart is rejected, but
those writes have already replaced the resume context consumed by
sbi_hsm_hart_resume_finish(). The same ordering also lets other rejected
start requests modify target scratch state.
Check the state immediately after acquiring the start ticket and before
updating the scratch context. The ticket serializes start requests for
the target; retain the cmpxchg as the publication barrier and defensive
state check.
Fixes: e8e9ed3790 ("lib: sbi: Set the state of a hart to START_PENDING after the hart is ready")
Signed-off-by: Pengpeng Hou <pengpeng@iscas.ac.cn>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260831130817.54625-1-pengpeng@iscas.ac.cn
Signed-off-by: Anup Patel <anup@brainfault.org>
Timer callbacks can request that an expired event be restarted. Such
events are moved to a local restart_list, but the restart loop tests that
list while reading its first entry from tstate->event_list.
If the main event list is empty, the loop treats its list head as a timer
event. Otherwise it repeatedly removes and reinserts the wrong event
while restart_list remains nonempty, so timer processing cannot
complete.
No in-tree callback currently requests a restart, but the timer-event
callback interface exposes this path to OpenSBI library and platform
users.
Read the event from restart_list, matching the loop condition and the
list that owns the entry.
Fixes: b7fa8a246c ("lib: sbi_timer: Add support for timer events")
Signed-off-by: Pengpeng Hou <pengpeng@iscas.ac.cn>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260831130641.54100-1-pengpeng@iscas.ac.cn
Signed-off-by: Anup Patel <anup@brainfault.org>
The RPMI shared-memory helpers validate a queue index before using it
to select queue_ctx_tbl, but the current comparison accepts queue_id
equal to queue_count.
Only entries 0 through queue_count - 1 are initialized from the device
tree. Accepting queue_count can therefore use an uninitialized queue
context when a protocol operation selects a queue omitted by the
description.
Reject queue_id greater than or equal to queue_count in both transfer
directions.
Fixes: 91f46fb47e ("lib/utils: Add RPMI messaging protocol and shared memory transport support")
Signed-off-by: Pengpeng Hou <pengpeng@iscas.ac.cn>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260831130444.53629-1-pengpeng@iscas.ac.cn
Signed-off-by: Anup Patel <anup@brainfault.org>
The RPMI shared-memory mailbox transport derives qctx->num_slots from the
device-tree queue window size and the riscv,slot-size property. The current
initialization only checks that the queue window is nonzero before
subtracting the two RPMI header slots.
Reject queue windows that cannot provide the two header slots plus at least
two data ring slots. A zero data-slot count would later be used as a modulo
divisor, and a one-slot ring is also unusable with the current full/empty
convention because one empty slot is reserved to distinguish the two
states.
Also reject data-slot counts that cannot be represented by qctx->num_slots.
Fixes: 91f46fb47e ("lib/utils: Add RPMI messaging protocol and shared memory transport support")
Signed-off-by: Pengpeng Hou <pengpeng@iscas.ac.cn>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260828142521.14267-1-pengpeng@iscas.ac.cn
Signed-off-by: Anup Patel <anup@brainfault.org>
The HPM feature detector and CSR emulator build mhpm_mask bits with a
plain signed 1. Both macro families expand through counter 31, so the
last expansion evaluates 1 << 31 and shifts a signed int into its sign
bit.
Use BIT() at both sites. The mask is unsigned and bit 31 is part of the
represented HPM counter set.
Fixes: c744ed77b1 ("lib: sbi_pmu: Enable noncontigous hpm event and counters")
Fixes: ee1f83ca84 ("lib: sbi_pmu: remove mhpm_count field in hart feature")
Signed-off-by: Pengpeng Hou <pengpeng@iscas.ac.cn>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260828142354.13035-1-pengpeng@iscas.ac.cn
Signed-off-by: Anup Patel <anup@brainfault.org>
The sbi_mpxy_register_channel() currently performs ceiling division by
PAGE_SIZE but stores the resulting page count in mpxy_shmem_size. All
consumers interpret mpxy_shmem_size as a byte count, so channels requiring
more than one page get an undersized shared memory value.
Address the above issue by using ROUNDUP() to keep the rounded value in
bytes.
Fixes: ec09918426 ("lib: sbi: Update MPXY framework and SBI extension as per latest spec")
Signed-off-by: Shibo Zhu <3499129952@qq.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/tencent_14AA3BA0083001905C262A5CD64E442FC205@qq.com
Signed-off-by: Anup Patel <anup@brainfault.org>
fdt_cpu_fixup() reserves a hardcoded 32 extra bytes, plus 16 more per
HART inside the "zicntr" loop. Those 32 bytes also have to cover every
"status" property rewritten to "disabled", so once more than a handful
of HARTs are disabled the fixups fail with FDT_ERR_NOSPACE (-3).
Bumping the reservation to a fixed worst case is not safe either:
fdt_open_into() cannot know how much space the caller's buffer really
has, it just trusts the requested size. As OpenSBI does not own the
memory following a device tree passed in by the previous booting stage,
over-declaring the blob risks corrupting whatever follows it.
Walk the HART nodes once without modifying them instead, and reserve
exactly what the fixups need. The read-only decision logic is factored
into fdt_cpu_fixup_needed() so the sizing and fixup passes cannot drift
apart, which also makes the per-HART fdt_open_into() redundant. Check
the return values of fdt_setprop_string() and fdt_appendprop_string()
as well, as running out of space used to fail silently.
Fixes: dd9439fbac ("lib: utils: Add a fdt_cpu_fixup() helper")
Signed-off-by: Chen Pei <cp0613@linux.alibaba.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260817084303.1232-1-cp0613@linux.alibaba.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Now that the lock relationship between an event and its owner hart is
explicit, add TSA annotations for the SSE handling code. This includes
indicating which fields are protected by a spinlock, as well as
annotating which functions need to run under a spinlock.
Exclude sse_local_init(), which initializes the spinlock to be acquired,
and sse_event_get() / sse_event_put(), which perform conditional locking
that TSA is not able to follow properly.
Signed-off-by: Carlos López <carlos.lopezr4096@gmail.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260909162322.29778-14-carlos.lopezr4096@gmail.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Add Thread Safety Analysis (TSA) annotations for the HTIF driver
implementation.
This requires annotating the tohost/fromhost and htif_console_buf
globals to indicate that they are protected by htif_lock.
htif_system_reset() can be excluded since it is only called under
special circumstances from sbi_system_reset(), when all other harts
have been stopped.
Signed-off-by: Carlos López <carlos.lopezr4096@gmail.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260909162322.29778-10-carlos.lopezr4096@gmail.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Add Thread Safety Analysis (TSA) annotations for spinlock APIs, allowing
the compiler to reason about lock acquisition and release.
Annotations have a different meaning if used in header declarations vs C
file implementations. In header declarations, they specify the semantics
of the given function, meaning that we must specify that the spinlock
functions acquire and release a lock ("capability" in clang terms).
In function implementations, attributes affect the function body and
callers in the same translation unit; since the spinlock functions
contain inline assembly, they must be excluded from analysis.
Signed-off-by: Carlos López <carlos.lopezr4096@gmail.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260909162322.29778-4-carlos.lopezr4096@gmail.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Currently, the domain initialization order is implied by the order
in which domains are populated by sbi_platform_domains_init() from
sbi_domain_finalize(). This is not documented anywhere and forces
unecessary ordering between domain DT nodes.
To address the above, introduce per-domain 32-bit integer to represent
intialization order (aka "init_order") where domain with a lower
initialization order will be booted first and two domains must not
have same initialization order. For DT based domain creation, new
"init-order" DT property can be used in domain DT node to specify
the initialization order. The ROOT domain is assumed to have lowest
initialization order (aka 0xffffffff).
Signed-off-by: Anup Patel <anup.patel@oss.qualcomm.com>
Reviewed-by: Pawandeep Oza <pawandeep.oza@oss.qualcomm.com>
Tested-by: Pawandeep Oza <pawandeep.oza@oss.qualcomm.com>
Reviewed-by: Yu-Chien Peter Lin <peter.lin@sifive.com>
Link: https://lore.kernel.org/r/20260905131748.922920-4-anup.patel@oss.qualcomm.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Align the PMU extension implementation with the error codes required
by the SBI v3.0 specification, chapter 11:
- sbi_pmu_counter_start and sbi_pmu_counter_stop (secs 11.9-11.10,
tables 39-42): the start_flags/stop_flags bits 2:(XLEN-1) are
reserved and must be zero, so return SBI_ERR_INVALID_PARAM when any
reserved flag bit is set. Introduce SBI_PMU_START_FLAGS_MASK and
SBI_PMU_STOP_FLAGS_MASK for the valid bits of each function.
- sbi_pmu_counter_start and sbi_pmu_counter_stop (tables 40 and 42):
return SBI_ERR_ALREADY_STARTED / SBI_ERR_ALREADY_STOPPED when the
set of counters includes a counter which is already started or
stopped, instead of ignoring the error returned for each counter.
- sbi_pmu_event_get_info (sec 11.14, table 47): the output word must
indicate whether the event is supported, but firmware events were
only matched against the hardware event map and were always
reported as unsupported. Report a validated firmware event as
supported.
Signed-off-by: David E. Garcia Porras <david.garcia@aheadcomputing.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260818210023.466462-4-david.garcia@aheadcomputing.com
Signed-off-by: Anup Patel <anup@brainfault.org>
sbi_pmu_event_get_info() walks hw_event_map[] to decide whether a
requested event is supported. For SBI_PMU_EVENT_RAW_IDX and
SBI_PMU_EVENT_RAW_V2_IDX it compares the requested event_data against
temp->select / temp->select_mask without first checking that the map
entry being examined is itself a raw event entry.
Non-raw hardware event entries are added via sbi_pmu_add_hw_event_counter_map(),
leave select and select_mask at zero, hence they satisfy:
temp->select == (event_data & temp->select_mask)
so the first non-raw entry visited will always match.
The issue's observability depends purely on the ordering of hw_event_map[]:
if the platform registers its raw events last, every raw event query, including
unsupported ones, will be reported as supported.
Fix it by checking event_idx against temp->start_idx and temp->end_idx before
comparing select/select_mask.
Fixes: e434584216 ("lib: sbi_pmu: Implement SBI PMU event info function")
Signed-off-by: David E. Garcia Porras <david.garcia@aheadcomputing.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260818210023.466462-3-david.garcia@aheadcomputing.com
Signed-off-by: Anup Patel <anup@brainfault.org>
As per section 11.14 of the SBI specification (Function: Get PMU Event
Info, FID #8), Table 47, the event_idx word of an event info entry
only uses BIT[0:19]; BIT[20:31] are reserved for the future purpose
and must be zero. Table 48 further requires the SBI implementation to
return SBI_ERR_INVALID_PARAM if any reserved bit in an event_idx word
is set.
sbi_pmu_event_get_info() does not check the reserved bits, so a
malformed event_idx is silently passed on to pmu_event_validate()
instead of failing the call. Add SBI_PMU_EVENT_IDX_MBZ_MASK covering
the must-be-zero bits and return SBI_ERR_INVALID_PARAM when any of
them are set.
Fixes: e434584216 ("lib: sbi_pmu: Implement SBI PMU event info function")
Signed-off-by: David E. Garcia Porras <david.garcia@aheadcomputing.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260818210023.466462-2-david.garcia@aheadcomputing.com
Signed-off-by: Anup Patel <anup@brainfault.org>
SSE injection sets hstatus.SPV to the virtualization state of the
interrupted context before entering the supervisor handler. This patch
completes that existing requirement and does not change the SSE ABI.
The existing code sets SPV when the interrupted context is virtualized,
but leaves it unchanged otherwise. A stale SPV value can therefore make
an event that interrupted host execution appear to have interrupted a
guest. Event completion can then resume with virtualization enabled.
Clear SPV when the interrupted context is not virtualized so the
handler-visible state matches the interrupted context. This also lets
supervisor software, such as the Linux PMU and perf code, reliably tell
whether an SSE interrupted host or virtualized execution.
Fixes: c8cdf01d8f ("lib: sbi: Add support for Supervisor Software Events extension")
Signed-off-by: Zhanpeng Zhang <zhangzhanpeng.jasper@bytedance.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20260731071820.7318-1-zhangzhanpeng.jasper@bytedance.com
Signed-off-by: Anup Patel <anup@brainfault.org>