| Age | Commit message (Collapse) | Author |
|---|
|
Register the VRAM manager with the dmem cgroup reclaim infrastructure
so that lowering dmem.max below current VRAM usage triggers TTM
eviction rather than failing with -EBUSY.
Guard place->flags in amdgpu_ttm_bo_eviction_valuable() against NULL,
as the TTM reclaim path passes a NULL place in cgroup drain mode.
Use drmm_cgroup_register_region() so that the region is automatically
unregistered at DRM device release, after drm_dev_unplug() has already
made drm_dev_enter() return false. The drm_dev_enter/exit guard in the
reclaim callback ensures no reclaim work touches the TTM manager after
driver unbind, closing the window between vram_mgr_fini() (called from
drm_driver.release) and the drmm cleanup that unregisters the region.
v3:
- Rebased on fix for uninitialized list and buddy allocator on the
drmm_cgroup_register_region() error path.
v5:
- Rebased on the introduction of struct dmem_cgroup_init.
- Clear the reclaim callback in amdgpu_vram_mgr_fini() to prevent
use-after-free if cgroup reclaim is triggered after driver unbind
while userspace holds an open DRM file descriptor. (Sashiko-bot)
- Switch from drmm_cgroup_register_region() to the raw
dmem_cgroup_register_region() and store the region in
amdgpu_vram_mgr.cg_region. Call dmem_cgroup_unregister_region()
in amdgpu_vram_mgr_fini() after ttm_resource_manager_evict_all()
to drain in-flight reclaim callbacks, and clear man->cg afterwards.
This is required because amdgpu's vram manager fini is called
explicitly during driver unbind, which may precede the DRM device
release and thus precede any drmm-based cleanup. (Sashiko-bot)
v6:
- Fix mgr->cg_region never being assigned, so
dmem_cgroup_unregister_region() in fini silently no-ops on NULL
and leaks the region. (Sashiko-bot)
- Reorder fini to call set_used(false) and evict_all() before
dmem_cgroup_unregister_region(), so ttm_resource_free() can
uncharge via man->cg during eviction; clear man->cg after
unregister. (Sashiko-bot)
v7:
- Move dmem_cgroup_unregister_region() before the early return on
evict_all() failure; not doing so leaves a dangling reclaim callback
pointing to the partially-torn-down VRAM manager, causing a
use-after-free when the cgroup later triggers reclaim. (Sashiko-bot)
- Switch back to drmm_cgroup_register_region() with a drm_dev_enter/
exit guard in the reclaim callback (matching xe), rather than manual
register/unregister. drm_dev_unplug() fires before vram_mgr_fini(),
so drm_dev_enter() returning false prevents any reclaim from touching
the manager during teardown. This also fixes the "vram" name
collision on multi-GPU systems, since drmm_cgroup_register_region()
automatically prefixes with "drm/<pci-addr>/". (Sashiko-bot)
v8:
- Move the dmem cgroup region teardown back into
amdgpu_vram_mgr_fini(): register the region with
dmem_cgroup_register_region() (keeping the "drm/<unique>/vram"
prefix), store it in amdgpu_vram_mgr.cg_region, and unregister it
explicitly after ttm_resource_manager_evict_all(). The v7
drmm-only approach left a use-after-free on the probe-error
teardown path: drm_dev_unplug() is not called there, so the
drm_dev_enter() guard stays open while drm_driver.release destroys
the manager before the drmm action unregisters the region. The
explicit unregister drains in-flight reclaim on every teardown
path; the drm_dev_enter() guard is kept as defense against reclaim
after unplug. (Sashiko-bot)
v9:
- Don't leak a gpu buddy manager on cgroup init failure. (Sashiko-bot)
- Acquire a runtime PM reference (pm_runtime_get_sync/put_autosuspend)
around the TTM reclaim call in amdgpu_vram_mgr_dmem_reclaim().
(Sashiko-bot)
Assisted-by: GitHub_Copilot:claude-sonnet-4.6
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Tested-by: Thadeu Lima de Souza Cascardo <cascardo@igalia.com>
Link: https://patch.msgid.link/20260725100036.2372-7-thomas.hellstrom@linux.intel.com
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Maarten Lankhorst <dev@lankhorst.se>
|
|
Register the VRAM manager with the dmem cgroup reclaim infrastructure
so that lowering dmem.max below current VRAM usage triggers TTM
eviction rather than failing with -EBUSY.
v4:
- Rebased on drm-tip; dropped the XE_PL_STOLEN guard as stolen memory
uses a separate TTM manager and never calls __xe_ttm_vram_mgr_init().
v5:
- Rebased on the introduction of struct dmem_cgroup_init.
- Register the fini drmm action before drmm_cgroup_register_region() so
that devres LIFO teardown runs unregister_region() first (draining any
in-flight reclaim callbacks via the rwsem) and xe_ttm_vram_mgr_fini()
second, ensuring the manager is never accessed by a reclaim callback
after teardown. (Sashiko-bot)
- Wrap the reclaim callback in xe_ttm_vram_mgr_dmem_reclaim() using
drm_dev_enter()/drm_dev_exit() to prevent TTM reclaim from running
after driver unbind.
v9:
- Close ttm_resource_manager_set_used() races around vram manager
init and fini.
- Formatting fix.
- Acquire a runtime PM reference around the TTM reclaim call in
xe_ttm_vram_mgr_dmem_reclaim(). (Sashiko-bot)
Assisted-by: GitHub_Copilot:claude-sonnet-4.6
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> #v8
Tested-by: Thadeu Lima de Souza Cascardo <cascardo@igalia.com>
Link: https://patch.msgid.link/20260725100036.2372-6-thomas.hellstrom@linux.intel.com
Signed-off-by: Maarten Lankhorst <dev@lankhorst.se>
|
|
Add ttm_bo_evict_cgroup() to evict buffer objects charged to a specific
dmem cgroup pool from a resource manager's LRU until a byte target is
met. Add ttm_resource_manager_set_dmem_region() to associate a dmem
cgroup region with a resource manager; drivers supply their own
dmem_cgroup_ops with ttm_resource_manager_dmem_reclaim as the reclaim
function and the manager pointer as reclaim_priv in the dmem_cgroup_init
to wire up TTM eviction as the reclaim callback.
The eviction context is interruptible; signals abort the operation and
propagate back through the write() syscall.
Introduce a new mode for the bo LRU walker so that sleeping locks
can be taken. This can be used when the caller doesn't hold any
previous dma_resv locks, and where it intends to hold at most
one lock at a time.
Like the rest of the TTM eviction this should sooner than later
be converted to full WW transactions.
v3:
- Fix ttm_resource_manager_set_dmem_region() storing an error pointer
in man->cg unconditionally. (Sashiko-bot)
- Fix kernel-doc function name format for ttm_bo_evict_cgroup() and
ttm_resource_manager_set_dmem_region().
v5:
- Rebased on the introduction of struct dmem_cgroup_init.
- Handle NULL region in ttm_resource_manager_set_dmem_region() to clear
the reclaim callback, preventing use-after-free when the manager is
torn down while the dmem region outlives it. (Sashiko-bot)
- Return 0 on any progress (even partial eviction), -ENOSPC only when
nothing was freed; fixes callers that expected 0 on partial success.
- Document that the reclaim callback should return 0 if some progress
was made, -ENOSPC if no progress at all, or another error for fatal
failures.
v8:
- Fix ttm_resource_manager_set_dmem_region() using IS_ERR_OR_NULL(),
which skipped the assignment for a NULL region and thus never
cleared man->cg. Use IS_ERR() so that a NULL region detaches the
region as the kernel-doc and the v5 changelog intended. (Sashiko-bot)
v9:
- Don't leak cgroup charges for bos that may have survived dmemcg
region fini.
- Drop the misleading "Capture size before eviction in case res is
cleared" comment in ttm_bo_evict_cb(). (Maarten Lankhorst)
Assisted-by: GitHub_Copilot:claude-sonnet-4.6
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> #v7
Tested-by: Thadeu Lima de Souza Cascardo <cascardo@igalia.com>
Link: https://patch.msgid.link/20260725100036.2372-5-thomas.hellstrom@linux.intel.com
Signed-off-by: Maarten Lankhorst <dev@lankhorst.se>
|
|
Replace the bare u64 size argument to dmem_cgroup_register_region() and
drmm_cgroup_register_region() with a const struct dmem_cgroup_init *
pointer. The struct currently carries only the size field, but using a
struct makes the API extensible: future callers can supply additional
initialization parameters without adding more positional arguments.
Update all in-tree callers (amdgpu, xe) to use a compound-literal
initializer.
v5:
- Commit introduced.
Assisted-by: GitHub_Copilot:claude-sonnet-4.6
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Tested-by: Thadeu Lima de Souza Cascardo <cascardo@igalia.com>
Link: https://patch.msgid.link/20260725100036.2372-3-thomas.hellstrom@linux.intel.com
Acked-by: Dave Airlie <airlied@redhat.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Maarten Lankhorst <dev@lankhorst.se>
|
|
The driver currently hardcodes the parent interrupt trigger type to
IRQF_TRIGGER_FALLING.
Use the trigger type configured by firmware instead. If no trigger type
is specified, fall back to IRQF_TRIGGER_FALLING to maintain current
behaviour.
Support IRQF_TRIGGER_FALLING and IRQF_TRIGGER_LOW, which match the
sx150x open drain active low interrupt output. Reject unsupported
trigger types.
Signed-off-by: Tsz Shan Chan <tchan@jacques.com.au>
Signed-off-by: Linus Walleij <linusw@kernel.org>
|
|
The driver currently reads the interrupt source register, writes the
value back to clear it, and always returns IRQ_HANDLED.
When the IRQ is shared, the handler should not claim interrupts that are
not from this device.
Check for zero interrupt source first. It no interrupt is pending,
return IRQ_NONE and skip the register write.
Signed-off-by: Tsz Shan Chan <tchan@jacques.com.au>
Signed-off-by: Linus Walleij <linusw@kernel.org>
|
|
https://git.kernel.org/pub/scm/linux/kernel/git/pinctrl/samsung into devel
Samsung pinctrl drivers changes for v7.3
Add drivers for newly introduced Samsung Exynos8855 SoC.
Signed-off-by: Linus Walleij <linusw@kernel.org>
|
|
at91_ping() rearms the watchdog timer from its callback. timer_delete()
neither waits for a running callback nor prevents it from rearming the
timer, so probe failure or driver removal can leave the timer accessing the
devm-allocated at91wdt after it has been freed.
Use timer_shutdown_sync() on both teardown paths. It waits for a running
callback and rejects any attempt by the callback to rearm the timer.
Fixes: 5161b31dc39a ("watchdog: at91sam9_wdt: better watchdog support")
Signed-off-by: Hongyan Xu <getshell@seu.edu.cn>
Link: https://lore.kernel.org/r/20260806060613.1830-1-getshell@seu.edu.cn
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
|
|
The DMA transfer routine was computing the start of the DMA area by
taking the address of the coherent allocation, and then adding the
transfer's start offset. It then checked manually that the upper bound
was valid.
Convert this to an I/O projection of the same region, which returns
`ERANGE` if the passed range does not fit within the coherent
allocation. This removes the need to perform arithmetic on DMA addresses
and to explicitly check for the bounds' validity.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Link: https://patch.msgid.link/20260805-falcon-dma-projections-v2-3-4cc9f3f13ee9@nvidia.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
|
|
The `dma_handle` naming is inherited from the C API, but what this
really describes is the device DMA address; everything named
`dma_handle` is actually a `dma_addr_t`.
This naming introduces some confusion on the Rust API side, as handles
are supposed to be opaque tokens, yet we were doing address computation
on values returned by `dma_handle`.
Rename `dma_handle` to `dma_address` while nova-core is still its only
user.
Suggested-by: John Hubbard <jhubbard@nvidia.com>
Suggested-by: Danilo Krummrich <dakr@kernel.org>
Link: https://lore.kernel.org/all/DK75LUA4NLGI.3P29AIZQE20V2@kernel.org/
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Reviewed-by: Robin Murphy <robin.murphy@arm.com>
Link: https://patch.msgid.link/20260805-falcon-dma-projections-v2-2-4cc9f3f13ee9@nvidia.com
[ Rebase and fix up build failures due to newly introduced dma_handle()
calls. - Danilo ]
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
|
|
The `try_with_base` call performed on `NV_PFALCON_FALCON_DMATRFBASE1`
already returns `EOVERFLOW` if the address is too large for the
register, making this check redundant.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Link: https://patch.msgid.link/20260805-falcon-dma-projections-v2-1-4cc9f3f13ee9@nvidia.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
|
|
Update version to 12.00.00.2607b2
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-34-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
struct qla_i2c_access carries a 16-bit length field alongside a fixed
64-byte buffer:
struct qla_i2c_access {
uint16_t device, offset, option, length;
uint8_t buffer[0x40];
} __packed;
qla2x00_write_i2c() and qla2x00_read_i2c() use the user-supplied
i2c->length without any bounds check. i2c is overlaid on a 256-byte
on-stack buffer and sfp is a 256-byte DMA-pool buffer, so a length up to
65535 overruns both:
- write: memcpy(sfp, i2c->buffer, i2c->length) over-reads the stack and
over-writes the sfp heap buffer, and qla2x00_write_sfp() then DMAs
i2c->length bytes out of the 256-byte buffer.
- read: qla2x00_read_sfp() DMAs i2c->length bytes into the 256-byte sfp,
then memcpy(i2c->buffer, sfp, i2c->length) overflows the 64-byte
buffer inside the on-stack array.
A caller holding CAP_SYS_RAWIO can use this to corrupt the heap and the
kernel stack. Reject requests whose length exceeds the buffer before any
copy or DMA transfer in both handlers.
Fixes: 9ebb5d9c69f1 ("[SCSI] qla2xxx: Add I2C BSG interface.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-33-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
The FRU and I2C bsg handlers stage their transfer in a DMA_POOL_SIZE
(256-byte) bounce buffer obtained from dma_pool_alloc(), which does not
zero the allocation. They initialize only a few leading bytes before
handing the buffer to qla2x00_write_sfp().
qla2x00_write_sfp() can override the transfer length with a user-supplied
value:
if (len == 1)
opt |= BIT_0;
if (opt & BIT_0)
len = *sfp;
*sfp is the first byte of the (user-controlled) payload, so len can grow
up to 255. The device then DMA-reads len bytes from the 256-byte pool
buffer. Since only a small prefix was written
(e.g. MAX_FRU_SIZE == 36 bytes for a FRU version, one byte for a FRU
status register), the hardware reads past the initialized region and
writes up to ~219 bytes of stale DMA-pool heap memory to the device
flash.
Allocate the buffer with dma_pool_zalloc() in all five FRU/I2C handlers
so any bytes beyond the initialized data are zero rather than stale heap
contents.
Fixes: 697a4bc69159 ("[SCSI] qla2xxx: Provide method for updating I2C attached VPD.")
Fixes: 9ebb5d9c69f1 ("[SCSI] qla2xxx: Add I2C BSG interface.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-32-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
The FC BSG transport allocates job->request via memdup_user() using the
exact user-supplied request_len. For FC_BSG_HST_VENDOR,
fc_bsg_host_dispatch() only guarantees request_len covers msgcode and
vendor_id; it does not account for the vendor_cmd[] flexible array.
qla2xxx then reads the command selector vendor_cmd[0] and, in several
sub-handlers, vendor_cmd[1]/[2] or structures overlaid on the vendor
command area without verifying request_len. A caller holding
CAP_SYS_RAWIO can submit a short request whose vendor_id matches the
host, triggering out-of-bounds heap reads (KASAN-detectable, and able to
mis-select a command or panic).
Add a central guard in qla2x00_process_vendor_specific() so the selector
is always in bounds, restrict the early vendor_cmd[0] read in
qla24xx_bsg_request() to sufficiently long vendor messages, and add
request_len checks to the sub-handlers that read further:
qla24xx_proc_fcp_prio_cfg_cmd(), qla2x00_process_loopback(),
qla84xx_reset(), qla84xx_updatefw(), qla2x00_read_optrom(),
qla2x00_update_optrom(), qlafx00_mgmt_cmd() and
qla28xx_validate_flash_image().
Fixes: 01e0e15c8b3b ("scsi: don't use fc_bsg_job::request and fc_bsg_job::reply directly")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-31-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
Several bsg handlers stage their request/reply in an uninitialized 256-byte
on-stack buffer (uint8_t bsg[DMA_POOL_SIZE]) and fill it via
sg_copy_to_buffer(), which only copies as many bytes as the user-supplied
request payload. When the request is shorter than the structure, the
remainder of the buffer is left holding stale stack data.
qla2x00_read_fru_status() and qla2x00_read_i2c() then copy the full
structure back to the reply payload with sg_copy_from_buffer(), leaking the
uninitialized stack bytes to user space. The write/update paths do not copy
the buffer back, but can feed uninitialized fields to the device.
Zero the stack buffer at declaration in all five handlers, mirroring the
heap kzalloc() approach, so short requests can no longer expose stale
memory.
Fixes: 697a4bc69159 ("[SCSI] qla2xxx: Provide method for updating I2C attached VPD.")
Fixes: 9ebb5d9c69f1 ("[SCSI] qla2xxx: Add I2C BSG interface.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-30-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla26xx_dport_diagnostics() streaming-maps the caller's result buffer with
dma_map_single(). The bsg path passes &dd->buf from the __packed struct
qla_dport_diag, where buf lands at a 2-byte offset and shares cachelines
with the surrounding options/unused fields. Mapping such a misaligned
sub-buffer violates the DMA API requirement that streaming buffers be
cacheline aligned and not share a cacheline with other data, and can
corrupt data on non-DMA-coherent architectures.
Allocate a dedicated DMA-coherent buffer inside qla26xx_dport_diagnostics()
for the mailbox command and copy the result back into the caller's buffer.
This removes the streaming map of the misaligned sub-buffer entirely; the
caller's buffer is now only a plain CPU buffer, so its packing no longer
matters.
Fixes: ec89146215d1 ("qla2xxx: Add bsg interface to support D_Port Diagnostics.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-29-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
The fcport->unsol_ctx_head list is modified from several contexts without
a common lock. Entries are added in qla2xxx_process_purls_iocb() from the
response queue ISR (under the qpair qp_lock), while they are removed from
qla2xxx_process_purls_pkt() (DPC/purex worker), qla_nvme_xmt_ls_rsp()
(NVMe-FC transport callback) and qla_nvme_release_lsrsp_cmd_kref() (SRB
completion). The qpair qp_lock cannot serialize this per-fcport list since
multiqueue adapters add entries through different qpairs, so a concurrent
add and delete (or two concurrent deletes) can corrupt the list pointers.
Introduce a dedicated per-fcport spinlock, unsol_ctx_lock, initialized in
qla2x00_alloc_fcport(), and take it around every list_add_tail()/list_del()
on unsol_ctx_head. The add nests under the existing qp_lock; no delete path
takes qp_lock, so the lock order is consistent and deadlock free.
Fixes: 875386b98857 ("scsi: qla2xxx: Add Unsolicited LS Request and Response Support for NVMe")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-28-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla_nvme_xmt_ls_rsp() obtains uctx, which was linked into
fcport->unsol_ctx_head by qla2xxx_process_purls_iocb() and is still linked
when the NVMe transport calls back to transmit the LS response. On the
error (out:) path the function frees uctx with kfree() but never removes
it from the list. This leaves a freed node in fcport->unsol_ctx_head: the
next list_add_tail() for that fcport writes through the freed node, and a
subsequent list_del() can corrupt the list or panic.
Unlink uctx with list_del() before kfree() on the error path, matching the
other free sites in qla_nvme_release_lsrsp_cmd_kref() and
qla2xxx_process_purls_pkt(). qla2x00_rel_sp() in the failure path only
returns the SRB to its pool and does not invoke sp->put_fn, so the out:
path is the sole free and uctx is always still linked there.
Fixes: 875386b98857 ("scsi: qla2xxx: Add Unsolicited LS Request and Response Support for NVMe")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-27-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla_nvme_xmt_ls_rsp() bails out to the out: label when firmware is not
started (!ha->flags.fw_started), but the out: path unconditionally calls
qla_nvme_ls_reject_iocb(), which ends in qla2x00_start_iocbs() and an
unconditional doorbell write to the request queue in-pointer register.
This rings the firmware doorbell and queues an IOCB that stopped or
resetting firmware cannot consume, and touches MMIO during the reset/EEH
window where fw_started is also clear.
Only emit the LS reject IOCB (and ring the doorbell) when fw_started is
set; otherwise just clean up and return. The post-allocation failure
cases (SRB alloc / qla2x00_start_sp() failure) run with firmware started
and still send the reject. Apply the same guard to the reject emission
in qla2xxx_process_purls_pkt().
Fixes: 875386b98857 ("scsi: qla2xxx: Add Unsolicited LS Request and Response Support for NVMe")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-26-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla_nvme_ls_abort() and qla_nvme_fcp_abort() take a command reference with
kref_get_unless_zero() and then call schedule_work() on priv->abort_work,
ignoring its return value. qla_nvme_abort_work() runs once and drops
exactly one reference via kref_put(&sp->cmd_kref, sp->put_fn).
Since the per-abort INIT_WORK() was moved to submission time,
schedule_work() now returns false when the work is already pending, for
example on a concurrent transport teardown and timeout-driven abort of
the same command. In that case the reference taken for the second abort
is never released because the work still executes only once, leaking a
reference. The command is then never returned to the NVMe-FC transport,
which can hang the port.
Drop the reference when schedule_work() returns false, so each
kref_get_unless_zero() is balanced regardless of whether the work was
newly queued. The held reference keeps priv->sp valid for the put.
Fixes: 7e85f6dbc856 ("scsi: qla2xxx: Initialize NVMe abort_work once at submission")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-25-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla24xx_report_id_acquisition() format-1 handling drops vport_slock after
taking the vport reference and then calls qla_update_host_map() without
the lock. That reaches qla_update_vp_map(), which mutates the ha->host_map
btree via btree_insert32()/btree_update32()/btree_remove32() and is
documented to require vport_slock to be held by the caller. Running it
unlocked can race concurrent host_map updates and corrupt the btree.
The format-2 path in the same function already wraps its host_map update
(SET_AL_PA) in vport_slock; the format-1 path is the lone outlier.
Hold vport_slock across the format-1 qla_update_host_map() call to honor
the documented locking contract. The vref_count taken in the loop keeps
the vport valid, so this only adds the missing host_map serialization.
Fixes: 430eef03a763 ("scsi: qla2xxx: Relocate/rename vp map")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-24-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla24xx_report_id_acquisition() format-1 handling takes the vport
reference under vport_slock but drops it outside the lock, after setting
vp->vp_flags and vp->dpc_flags:
set_bit(VP_IDX_ACQUIRED, &vp->vp_flags);
set_bit(REGISTER_FC4_NEEDED, &vp->dpc_flags);
set_bit(REGISTER_FDMI_NEEDED, &vp->dpc_flags);
atomic_dec(&vp->vref_count);
Neither set_bit() nor atomic_dec() imply a memory barrier, so on a weakly
ordered architecture the decrement can become visible before the flag
stores. qla24xx_deallocate_vp_id() polls vref_count under vport_slock and
unlinks the vport once it reads zero, after which qla24xx_vport_delete()
frees it via scsi_host_put(). The poller could therefore observe
vref_count == 0 early and tear the vport down while the pending vp_flags/
dpc_flags stores land on freed memory.
Drop the reference under vport_slock, as is done for the matching
increment and by every other vref_count user. The unlock release pairs
with the deallocate poller's lock acquire so the flag stores are ordered
before vref_count == 0 can be observed.
Fixes: 793cedee296f ("scsi: qla2xxx: Hold vport reference in qla24xx_report_id_acquisition()")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-23-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla24xx_report_id_acquisition() format-1 handling walks ha->vp_list under
vport_slock, takes a vref_count on the matching vport and calls
qla_update_host_map() to register its port id.
A vport teardown via qla24xx_vport_delete() sets VPORT_DELETE, then
qla24xx_disable_vp() removes the vport from the host_map btree and zeroes
vha->d_id (RESET_AL_PA). The vport is only unlinked from vp_list later,
in qla24xx_deallocate_vp_id(), which clears vp_map[idx] (RESET_VP_IDX)
but does not touch host_map. In the window in between, report ID
acquisition can still find the vport on vp_list and call
qla_update_host_map(); with d_id already zeroed it takes the
btree_insert32() path and re-inserts the dying vport into host_map.
Nothing cleans that entry afterwards, so once scsi_host_put() frees the
vha a later host_map lookup dereferences freed memory.
Skip a vport that has VPORT_DELETE set before taking the reference, so it
is neither re-registered nor scheduled for DPC re-registration. This
mirrors the existing guard in qla2x00_alert_all_vps().
Fixes: 41dc529a4602 ("qla2xxx: Improve RSCN handling in driver")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-22-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla2x00_async_iocb_timeout() tries to abort a timed-out async IOCB. When
qla24xx_async_abort_cmd() fails, both the SRB_LOGIN_CMD path and the
SRB_CTRL_VP/default path scan outstanding_cmds[] for the SRB and then
call sp->done(sp, QLA_FUNCTION_TIMEOUT) unconditionally, without checking
whether the SRB was actually found and removed.
If the response ISR completes the same handle first, it removes the SRB
under qp_lock_ptr and runs sp->done() -> complete(sp->comp). The
submitter qla24xx_control_vp() wakes from wait_for_completion(), clears
sp->comp, drops its reference and returns, reclaiming the on-stack
completion. The timer reference keeps the SRB alive across the timeout
handler, but not the submitter's stack. The timeout then issues a second
sp->done() -> qla_ctrlvp_sp_done(), which evaluates "if (sp->comp)
complete(sp->comp)"; with the pointer loaded before the submitter's NULL
store, complete() writes into the freed stack frame, a use-after-free.
Track whether this path removed the SRB from outstanding_cmds and only
call sp->done() when it did, so the command is completed exactly once by
whichever path owns it. This mirrors the sp_found guard already used in
qla24xx_abort_iocb_timeout().
Fixes: f6145e86d21f ("scsi: qla2xxx: Fix race between switch cmd completion and timeout")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-21-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
ha->max_npiv_vports is taken from firmware (mcp->mb[11]) and only
constrained so that (max_npiv_vports + 1) is a multiple of
MIN_MULTI_ID_FABRIC, which permits values of 63, 127, 191 and 255.
NPIV vports are then allocated up to that count.
VP enable uses the VP_CONFIG IOCB, which addresses a vport through a
plain vp_index byte, so a vp_index beyond 128 is enabled without issue.
VP disable, however, uses the VP_CTRL IOCB, which selects target vports
through the fixed 128-bit vp_idx_map bitmap. qla24xx_control_vp()
rejects a vp_index past that bitmap and the IOCB builder cannot set a bit
beyond 127, yet qla24xx_vport_delete() frees the local state regardless.
A vport with vp_index > 128 can therefore be created and enabled but
never disabled, leaving it permanently active in firmware: a resource
leak.
Cap ha->max_npiv_vports at init to the vp_idx_map capacity so such
vports are never created. This collapses 191/255 to 127 (still
modulo-valid) and leaves the real-world 63/127 cases unaffected.
Fixes: 4d0ea24769c8 ("[SCSI] qla2xxx: Retrieve max-NPIV support capabilities from FW.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-20-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla2x00_status_entry() filters out non-TYPE_SRB entries and the
SRB_NVME_CMD, SRB_BIDI_CMD and SRB_TM_CMD types, then falls through to a
SCSI fast path that assumes the command is an SRB_SCSI_CMD. The first
thing on that path, qla_chk_edif_rx_sa_delete_pending(), and the
subsequent handling both evaluate GET_CMD_SP(sp), i.e. sp->u.scmd.cmd.
The srb u union overlays the SCSI command pointer with other command
layouts (bsg_job, iocb_cmd). If firmware delivers an unexpected
STATUS_TYPE IOCB for a non-SCSI handle, sp->u.scmd.cmd can read as a
non-NULL garbage pointer, bypassing the NULL checks in
qla_chk_edif_rx_sa_delete_pending() and at the cp == NULL test, and
leading to a wild pointer dereference.
Reject any SRB whose type is not SRB_SCSI_CMD before entering the fast
path. The outstanding_cmds slot is left untouched so a genuinely
non-SCSI command still completes through its proper handler.
Fixes: dd30706e73b7 ("scsi: qla2xxx: edif: Add key update")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-19-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla2xxx_delete_qpair() deletes the request queue before the response
queue. qla25xx_delete_req_que() frees the request queue memory
(kfree(req) in qla25xx_free_req_que()), but the response-queue MSI-X is
only released later, in qla25xx_free_rsp_que(). In that window the
response interrupt can still fire, qla2xxx_msix_rsp_q() queues
qpair->q_work, and qla_do_work() -> qla24xx_process_response_queue()
dereferences the now-freed rsp->req (LOGINOUT/CT/ELS entries and the
status path), a use-after-free.
The cancel_work_sync() added for the qpair teardown lives in the
response free path, which runs after the request queue is already freed,
so it does not protect rsp->req.
Release the response-queue interrupt and flush qpair->q_work before
deleting the request queue, so no late completion can reach the freed
request queue. Clearing have_irq makes the subsequent
qla25xx_free_rsp_que() skip its free_irq(), and the firmware
queue-delete order (request then response) is preserved; the
request-delete mailbox completes on the default vector and is unaffected
by dropping the qpair response interrupt early.
Fixes: d74595278f4a ("scsi: qla2xxx: Add multiple queue pair functionality.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-18-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla2x00_error_entry() reads ha->req_q_map[que] twice: once for the NULL
check and again when assigning it to req. The map slot is cleared by
qla25xx_free_req_que() (ha->req_q_map[que_id] = NULL under mq_lock)
during queue teardown, while the response-queue interrupt that drives
qla2x00_error_entry() is still registered (the IRQ is released later in
qla25xx_free_rsp_que()). If the slot is set to NULL between the two
reads, req becomes NULL and is dereferenced.
Read the slot once into req and NULL-check the local before use. mq_lock
is a mutex and cannot be taken from interrupt context, so the single
read plus local check is the appropriate fix for the reported NULL
dereference.
Fixes: a6fe35c052c4 ("[SCSI] qla2xxx: Avoid invalid request queue dereference for bad response packets.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-17-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
In qla2x00_status_entry(), the FWI2 status path advances sense_data and
shrinks par_sense_len by rsp_info_len:
if (IS_FWI2_CAPABLE(ha)) {
sense_data += rsp_info_len;
par_sense_len -= rsp_info_len;
}
rsp_info_len is a 32-bit value taken directly from the target's FCP
response (sf.rsp_data_len), while par_sense_len is the IOCB data area
size (28 bytes for 24xx, 60 bytes for 29xx). A hostile or buggy target
reporting an rsp_info_len larger than par_sense_len makes the unsigned
subtraction underflow to a huge value and advances sense_data out of
bounds.
The underflowed par_sense_len then defeats the cap in
qla2x00_handle_sense():
if (sense_len > par_sense_len)
sense_len = par_sense_len;
memcpy(cp->sense_buffer, sense_data, sense_len);
so the memcpy reads up to SCSI_SENSE_BUFFERSIZE bytes from the
out-of-bounds sense_data pointer, leaking adjacent response-ring/heap
memory into the command's sense buffer.
Clamp rsp_info_len to par_sense_len before the subtraction so
par_sense_len can never underflow and sense_data stays within the IOCB
data area. The fix sits before the comp_status switch, covering both
qla2x00_handle_sense() call sites.
Fixes: 5544213be7b4 ("[SCSI] qla2xxx: Correct extended sense-data handling.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-16-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla27xx_copy_multiple_pkt() and qla27xx_copy_fpin_pkt() poll
rsp_q->ring_ptr->signature for RESPONSE_PROCESSED (0xDEADDEAD) to decide
whether the next continuation IOCB has arrived, spinning on cpu_relax()
without advancing the ring or decrementing the entry count while it has
not. response_t::signature lives at byte offset 60, but a continuation
IOCB (sts_cont_entry_t / struct sts_cont_entry_ext) carries raw FC frame
payload at that offset (data[56..59]). A received frame whose payload
bytes happen to equal 0xDEADDEAD is therefore misread as "not yet
arrived", and the loop spins forever in interrupt/DPC context, causing a
CPU soft lockup.
The poll is also unnecessary: callers of qla27xx_copy_multiple_pkt()
(PT_LS4_UNSOL and the NVMe purls path) already gate on
qla_chk_cont_iocb_avail(), which guarantees all entry_count IOCBs are
present before copying begins. The sibling helper
__qla_copy_purex_to_buffer() already drops the signature poll and relies
on the entry_type == STATUS_CONT_TYPE guard instead.
Remove the signature busy-wait from both helpers, keeping the entry_type
guard, and gate the FPIN path with qla_chk_cont_iocb_avail() so it defers
and re-processes on the next interrupt once all continuation IOCBs have
arrived, mirroring the ELS_AUTH_ELS and PT_LS4_UNSOL arms. With this the
signature field is never read on a continuation IOCB, eliminating the
payload-aliasing lockup.
Fixes: 9f2475fe7406 ("scsi: qla2xxx: SAN congestion management implementation")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-15-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla24xx_process_response_queue() advances ring_ptr past the head IOCB
before dispatching, so by the time __qla_consume_iocb() runs, ring_ptr
already points at the first continuation IOCB. The function however
looped purex->entry_count times starting at ring_ptr. As entry_count
includes the head, this consumed one entry too many: it stamped
RESPONSE_PROCESSED on the next, unrelated IOCB and advanced the ring
past it, silently dropping a legitimate firmware response. The head
IOCB's signature was also never marked.
Mark the head processed and account for it, then consume only the
entry_count - 1 continuation IOCBs, matching __qla_copy_purex_to_buffer().
Fixes: fac2807946c1 ("scsi: qla2xxx: edif: Add extraction of auth_els from the wire")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-14-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
When qla2x00_mem_alloc() fails, qla2x00_probe_one() jumps to
probe_hw_failed and calls qla2x00_mem_free(). Several error labels in
qla2x00_mem_alloc() freed adapter members (elsrej.c, purex_dma_pool,
flt, sfp_data, loop_id_map, async_pd, sf_init_cb, ex_init_cb, npiv_info)
but left the pointers dangling. qla2x00_mem_free() then freed them a
second time. Worse, for the dma_pool members it issued
dma_pool_free(ha->s_dma_pool, ...) after s_dma_pool had already been
destroyed and set to NULL at fail_s_dma_pool, dereferencing a NULL pool.
Clear each freed pointer (and its DMA handle) in the error labels so the
subsequent qla2x00_mem_free() skips them.
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-13-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
For QLAFX00 the request ring is ioremapped device I/O memory
(ha->iobase + req_que_off), not DMA-coherent RAM, which is why the rest
of the FX00 path accesses it through memcpy_toio() and the wrt_reg_*
helpers. __qla2x00_alloc_iocbs() however zeroed the producer slot with a
plain memset(). On architectures such as ARM64 a regular memset() may
emit unaligned or block-zeroing instructions (e.g. DC ZVA) that are
invalid on Device memory, leading to a synchronous external abort.
Use memset_io() to clear the slot for QLAFX00, matching the I/O
accessors used elsewhere on this ring. Other adapters keep the plain
memset() on their DMA-coherent rings. The zero-fill is retained for FX00
because its IOCB builders (e.g. qlafx00_fxdisc_iocb()) copy only part of
the entry and rely on the unused tail being pre-zeroed.
Fixes: 8ae6d9c7eb10 ("[SCSI] qla2xxx: Enhancements to support ISPFx00.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-12-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla2x00_free_fce_trace() freed and cleared ha->fce while holding only
fce_mutex. The firmware-dump consumers qla27xx_fwdt_entry_t264() and
qla25xx_copy_fce() read ha->fce (NULL check followed by a copy of the
buffer) under hardware_lock and never take fce_mutex. A debugfs FCE
disable could therefore free the DMA buffer between a dump's NULL check
and its copy, resulting in a use-after-free.
Unpublish ha->fce under hardware_lock, then release the lock and free
the DMA buffer (dma_free_coherent() may sleep). A concurrent dump either
completes its check and copy with the buffer still valid, or observes
ha->fce == NULL and skips it.
Fixes: 841df27d619e ("scsi: qla2xxx: Move FCE Trace buffer allocation to user control")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-11-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla2x00_dfs_fce_write() called kstrtoul() with a NULL result pointer,
so a successful parse would dereference NULL and oops. Worse, the int
return value (0 on success, negative errno on failure) was assigned to
the unsigned long enable flag, inverting the intended logic: a valid
number was treated as "disable" while a parse failure enabled FCE.
Parse the value into enable and propagate parse errors to userspace.
Fixes: 841df27d619e ("scsi: qla2xxx: Move FCE Trace buffer allocation to user control")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-10-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
The mbx_cmd_t is allocated on the stack but left uninitialized.
qla2x00_mailbox_command() has several early-return paths (PCI permanent
failure, device failed, EEH busy, ISP abort pending, mailbox access
timeout, purge mbox) that return without writing the input mailbox
registers back into mcp->mb[]. qla2x00_get_firmware_state() then
unconditionally copies mcp->mb[1..6] (and mb[12]) into the caller's
states[] array regardless of the return value.
On such a failure the copied values are uninitialized kernel stack
memory, which is then exposed to userspace via the fw_state and
mpi_fw_state sysfs handlers. Zero the mailbox struct so a failed query
yields deterministic zeroed state instead of leaking stack contents.
Fixes: 4d4df1932b6b ("[SCSI] qla2xxx: Add ISP84XX support.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-9-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla2x00_fw_state_show() initializes rval to QLA_FUNCTION_FAILED and jumps
to the out: label when the chip is down or EEH is busy. The out: block
then re-issued qla2x00_get_firmware_state() because rval != QLA_SUCCESS,
defeating the chip-down/EEH-busy guards and issuing a mailbox command
(outside optrom_mutex) during ISP reset or PCI error recovery, which can
hang the adapter. It also turned a normal in-lock mailbox failure into a
second unsynchronized mailbox attempt.
Make the out: fallback only mark the firmware state as unknown. The
mailbox is now issued at most once, inside optrom_mutex, and only when
the chip is up and not EEH-busy.
Fixes: b6faaaf796d7 ("scsi: qla2xxx: Serialize mailbox request")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-8-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla84xx_put_chip() drops the last reference to ha->cs84xx and frees it via
__qla84xx_chip_release() without clearing ha->cs84xx. During teardown it ran
before scsi_remove_host(), which is what removes the 84xx_fw_version host
sysfs attribute. A concurrent read of that attribute in the window between
the two calls executes qla24xx_84xx_fw_version_show(), which dereferences
the freed ha->cs84xx, resulting in a use-after-free.
Move qla84xx_put_chip() to after scsi_remove_host() in both
qla2x00_remove_one() and qla2x00_disable_board_on_pci_error(). Once
scsi_remove_host() returns, the sysfs attribute is gone and kernfs has
drained any in-flight show(), so no reader can touch cs84xx; the put still
runs before the host and ha are freed.
Fixes: fe1b806f4f71 ("[SCSI] qla2xxx: Refactor shutdown code so some functionality can be reused.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-7-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
The kdoc for qla29xx_mpi_optrom_data() described @offset as an "Offset into
the device memory", which reads like a byte address and invites confusion
with the per-chunk word-granular address advance in the transfer loop.
MBC_LOAD_DUMP_MPI_RAM is word-addressed: @offset is an MPI RAM address in
32-bit words, and @length is a byte count that is converted internally to a
word count. Document this to reflect the existing behavior. No functional
change.
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-6-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
The response queue MSI-X handler qla2xxx_msix_rsp_q() schedules
qla_do_work() via queue_work(ha->wq, &qpair->q_work). qla_do_work()
dereferences the qpair (vha, rsp) and takes qpair->qp_lock.
During teardown, qla2xxx_delete_qpair() deletes the response queue, which
calls free_irq() in qla25xx_free_rsp_que(), and then frees the queue and
the qpair. free_irq() waits for running hardirq handlers but does not
cancel work already placed on ha->wq. A still-pending q_work then runs
qla_do_work() against the freed qpair and response queue, causing a
use-after-free. This is especially likely during full adapter teardown,
where destroy_workqueue(ha->wq) forces pending work to run after the queue
pairs have been freed.
Flush the work item with cancel_work_sync() in qla25xx_free_rsp_que()
after free_irq() has released the interrupt (so no new work can be
queued) and before the response queue and qpair memory are freed (so the
flushed handler still sees valid memory). Guard on rsp->qpair and ha->wq
to match the INIT_WORK() condition and avoid operating on an
uninitialized work_struct.
Fixes: 68ca949cdb04 ("[SCSI] qla2xxx: Add CPU affinity support.")
Reported-by: Sashiko <sashiko-dev@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-5-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
The "update cache versions without reset" sysfs reset operation (0x20261)
calls get_flash_version(), which reads hardware flash registers, without
holding ha->optrom_mutex. The VPD update path serializes the same call
under optrom_mutex, so this reset path can interleave its flash register
accesses with a concurrent VPD or optrom flash operation and corrupt the
reads.
Hold ha->optrom_mutex across the get_flash_version() call to match the
VPD update path.
Fixes: 8c2cf7d4e387 ("[SCSI] qla2xxx: Add a new interface to update versions.")
Reported-by: Sashiko <sashiko-dev@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-4-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
Capture as much firmware dump data as possible. Save the mailbox
registers at start-of-day, before firmware execution, so they are
available in the dump, and allocate a guestimate dump buffer early
during driver load to capture failures that happen before the final
dump buffer is sized.
Make template entry processing more robust: skip over any entry that
fails to capture and continue with the next one, and skip entries that
time out instead of aborting the whole dump. Notify udev once sysfs
nodes are available in case a dump was captured before they existed.
Signed-off-by: Quinn Tran <qutran@marvell.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-3-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
ha->msix_count is u16, but ha->max_req_queues, ha->max_rsp_queues and
ha->max_qpairs are u8. Deriving the queue count as
"ha->max_req_queues = ha->msix_count - 1" therefore truncates: a board
(or a misconfigured/malicious hot-plugged device) advertising 257 MSI-X
vectors yields msix_count - 1 == 256, which truncates to 0. An MSI-X
count of 1 zeroes it as well, and in target mode the subsequent
"ha->max_req_queues--" then underflows 0 to 255.
When the count is 0, qla2x00_alloc_queues() calls
kzalloc_objs(struct req_que *, 0), which returns ZERO_SIZE_PTR. That is
not NULL, so the allocation check passes and the following
"ha->req_q_map[0] = req" dereferences ZERO_SIZE_PTR, corrupting memory
or crashing the kernel.
Add qla_calc_queue_count() to clamp the derived value into
[1, QLA_MAX_QUEUES - 1] so it always fits in u8 and is never zero, and
use it at all three derivation sites (qla25xx_iospace_config(),
qla83xx_iospace_config() and qla24xx_enable_msix()). Also guard the
target-mode decrement so it cannot reintroduce a zero (which would in
turn underflow max_qpairs).
Fixes: d74595278f4a ("scsi: qla2xxx: Add multiple queue pair functionality.")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-dev@google.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://patch.msgid.link/20260730155838.2119230-2-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
drmm_cgroup_register_region() is called before INIT_LIST_HEAD() and
gpu_buddy_init() in amdgpu_vram_mgr_init(). If it fails, the function
returns early and bypasses those initializations.
Since adev->mman.initialized is set to true before amdgpu_vram_mgr_init()
is called, a failure triggers amdgpu_ttm_fini(), which calls
amdgpu_vram_mgr_fini(), which then:
- Calls list_for_each_entry_safe() on reservations_pending and
reserved_pages, whose list_head::next pointers are zero-initialized
(NULL). The loop does not recognize them as empty and dereferences NULL.
- Calls gpu_buddy_fini(), which iterates free_trees[] unconditionally
via for_each_free_tree(). Since mm->free_trees is NULL
(never allocated), this dereferences NULL.
Both result in a kernel panic on the module load error path.
Fix by moving drmm_cgroup_register_region() to after the list and buddy
allocator are fully initialized, so the teardown path is safe to run.
Reported-by: Sashiko-bot <sashiko-bot@kernel.org>
Closes: https://sashiko.dev/#/patchset/20260428073116.15687-1-thomas.hellstrom@linux.intel.com?part=4
Fixes: 2b624a2c1865 ("drm/ttm: Handle cgroup based eviction in TTM")
Cc: Friedrich Vock <friedrich.vock@gmx.de>
Cc: Maarten Lankhorst <dev@lankhorst.se>
Cc: Tejun Heo <tj@kernel.org>
Cc: Maxime Ripard <mripard@kernel.org>
Cc: Christian König <christian.koenig@amd.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: amd-gfx@lists.freedesktop.org
Cc: dri-devel@lists.freedesktop.org
Cc: stable@vger.kernel.org # v6.14+
Assisted-by: GitHub_Copilot:claude-sonnet-4.6
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
Reviewed-By: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Tested-by: Thadeu Lima de Souza Cascardo <cascardo@igalia.com>
Link: https://patch.msgid.link/20260725100036.2372-2-thomas.hellstrom@linux.intel.com
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Maarten Lankhorst <dev@lankhorst.se>
|
|
Update version to 12.00.00.2607b1.
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260723050413.3897522-57-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla2x00_update_fru_versions() copies the user-supplied BSG request into
a fixed 256-byte stack buffer (bsg[DMA_POOL_SIZE]) and then iterates
list->count times over the qla_image_version array embedded in that
buffer, advancing the image pointer each iteration. count is taken
directly from user input with no upper bound, while only (DMA_POOL_SIZE
- sizeof(list->count)) / sizeof(struct qla_image_version) = 6 entries
actually fit. A larger count walks the image pointer off the end of the
stack buffer, reading adjacent kernel stack memory and sending it to the
device via qla2x00_write_sfp().
Reject requests whose declared count does not fit in the buffer.
Fixes: 697a4bc69159 ("[SCSI] qla2xxx: Provide method for updating I2C attached VPD.")
Cc: stable@vger.kernel.org
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260723050413.3897522-56-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla28xx_validate_flash_image() returns QLA_SUCCESS (0) unconditionally,
telling the FC BSG transport (fc_bsg_host_dispatch()) that the driver
owns and will complete the request. But bsg_job_done() is guarded by "if
(!rval)", so on the error path (rval == -EINVAL) neither the driver nor
the transport completes the job. The request dangles until it times out,
leaking block layer resources.
Commit c2c68225b145 ("scsi: qla2xxx: Fix bsg_done() causing double
free") added the "if (!rval)" guard to a batch of BSG handlers. That is
correct for handlers that also return the error code (the transport then
completes the job once via fail_host_msg), but this function returns
QLA_SUCCESS unconditionally, so the guard turned a correct single
completion into a leak.
Always call bsg_job_done(): bsg_reply->result is DID_OK and the error is
reported in vendor_rsp[0], and since the function returns 0 the
transport will not complete the job a second time.
Fixes: c2c68225b145 ("scsi: qla2xxx: Fix bsg_done() causing double free")
Cc: stable@vger.kernel.org
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260723050413.3897522-55-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla2x00_do_dport_diagnostics() allocates the qla_dport_diag response
buffer with kmalloc_obj() (non-zeroing) and, on success, copies the full
sizeof(*dd) back to user space via sg_copy_from_buffer(). The inbound
sg_copy_to_buffer() only fills as many bytes as the user request payload
provides, and qla26xx_dport_diagnostics() zeroes only dd->buf. The
options and unused[] fields are therefore copied out uninitialized,
leaking kernel heap contents to user space.
Allocate with kzalloc_obj(), matching qla2x00_do_dport_diagnostics_v2().
Fixes: ec89146215d1 ("qla2xxx: Add bsg interface to support D_Port Diagnostics.")
Cc: stable@vger.kernel.org
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260723050413.3897522-54-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
|
qla_nvme_ls_reject_iocb() allocates from and advances the request ring
through __qla2x00_alloc_iocbs() (which assumes the hardware_lock is
held) and qla2x00_start_iocbs() (which advances the ring and rings the
request-in doorbell), but takes no lock itself. Two of its callers
invoke it without the producer lock held:
- qla_nvme_xmt_ls_rsp(), the NVMe-FC .xmt_ls_rsp transport callback, on
its error path, and
- qla2xxx_process_purls_pkt(), run from the purex work/DPC context.
Both use ha->base_qpair, whose qp_lock_ptr is hardware_lock, so they can
run concurrently with normal I/O submission on the base ring and corrupt
the ring producer state, leading to duplicated or dropped commands. The
third caller, qla2xxx_process_purls_iocb(), runs inside
qla24xx_process_response_queue() with the qpair lock already held and is
safe; that is also why the lock cannot be taken inside the helper itself
(it would recursively re-acquire hardware_lock on the response path).
Take qp_lock_ptr around the two unlocked callers and document the helper
as caller-locked. Both run in process context, so spin_lock_irqsave() is
used and nothing in the locked region sleeps.
Fixes: 875386b98857 ("scsi: qla2xxx: Add Unsolicited LS Request and Response Support for NVMe")
Cc: stable@vger.kernel.org
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260723050413.3897522-53-njavali@marvell.com
Signed-off-by: Martin K. Petersen (Oracle) <mkp@kernel.org>
|
