pmu_ctr_write_hw() writes CSR_MCYCLE + cidx with no check on cidx,
while its sibling pmu_ctr_start_hw() guards the same index space:
if (cidx >= num_hw_ctrs || cidx == 1)
return SBI_EINVAL;
A cidx at or above num_hw_ctrs addresses an mhpmcounter the platform
need not implement, and cidx == 1 addresses 0xb01, which is not
assigned. Either one traps in M-mode instead of returning an error.
Give the function the same guard and a return value. The guard is
copied from pmu_ctr_start_hw() rather than limited to the upper
bound, so that the two functions accept exactly the same index space.
Only the call site in sbi_pmu_ctr_cfg_match() propagates the error.
The two call sites inside pmu_ctr_start_hw() run after the identical
guard in that function, so they cannot fail and are left unchanged.
Fixes: 13d40f21d5 ("lib: sbi: Add PMU support")
Signed-off-by: Sebastian Alba Vives <sebasjosue84@gmail.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20261025145338.2569358-3-sebasjosue84@gmail.com
Signed-off-by: Anup Patel <anup@brainfault.org>
sbi_pmu_ctr_cfg_match() calls pmu_ctr_start_hw() for the AUTO_START
flag and drops the result. pmu_ctr_start_hw() carries its own index
guard
if (cidx >= num_hw_ctrs || cidx == 1)
return SBI_EINVAL;
and can also return SBI_EALREADY_STARTED when the counter is already
running. Both are reported to the caller as a successful configure.
Neither is benign. A rejected index means nothing was configured, and
an already running counter may have been configured for a different
event previously, so the supervisor must not be told its configure
succeeded.
This is also the only one of the three start paths in the file that
discards the error. sbi_pmu_ctr_start() propagates it:
ret = pmu_ctr_start_hw(cidx, ival, bUpdate);
if (ret)
return ret;
and the firmware branch directly below this call site propagates the
error from pmu_dev->fw_counter_start() as well.
Fixes: 13d40f21d5 ("lib: sbi: Add PMU support")
Signed-off-by: Sebastian Alba Vives <sebasjosue84@gmail.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20261025145338.2569358-2-sebasjosue84@gmail.com
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>