Thank you for your contribution! Sashiko AI review found 6 potential issue(s) to consider:
Pre-existing issues: - [Critical] Double fetch of `rpc->length` from untrusted GSP shared memory allows bypassing size validation and causing a heap buffer overflow. - [Critical] Integer underflow in `r535_gsp_msgq_recv_one_elem()` when copying continuation records with a small length. - [Critical] Missing validation of continuation chunk sizes in `r535_gsp_msgq_recv()` leads to heap buffer overflow and integer underflow. - [High] Silent failure when GSP firmware sends a corrupted RPC reply for an allocation request. - [High] Memory leak of the RPC reply payload when `nvkm_gsp_rpc_wr()` is used with `NVKM_GSP_RPC_REPLY_RECV`. - [High] Large RPC message buffer leak on error paths in `r535_gsp_rpc_push()`. -- commit 80c014f86c2f1ac065ff1b0e24d2eb0c508a2e4a Author: Hongling Zeng <[email protected]> nouveau/gsp/rm/alloc: Cleanup IS_ERR_OR_NULL usage This commit cleans up the usage of IS_ERR_OR_NULL in r535_gsp_rpc_rm_free(). It updates the check for nvkm_gsp_rpc_get() to only use IS_ERR(), matching its actual return value contract. > diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/gsp/rm/r535/alloc.c > b/drivers/gpu/drm/nouveau/nvkm/subdev/gsp/rm/r535/alloc.c > index 46e3a29f2ad7f..c9f86c0e9b255 100644 > --- a/drivers/gpu/drm/nouveau/nvkm/subdev/gsp/rm/r535/alloc.c > +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/gsp/rm/r535/alloc.c > @@ -35,7 +35,7 @@ r535_gsp_rpc_rm_free(struct nvkm_gsp_object *object) > client->object.handle, object->handle); > > rpc = nvkm_gsp_rpc_get(gsp, NV_VGPU_MSG_FUNCTION_FREE, sizeof(*rpc)); > - if (WARN_ON(IS_ERR_OR_NULL(rpc))) > + if (WARN_ON(IS_ERR(rpc))) > return -EIO; > > rpc->params.hRoot = client->object.handle; > rpc->params.hObjectParent = 0; > rpc->params.hObjectOld = object->handle; > return nvkm_gsp_rpc_wr(gsp, rpc, NVKM_GSP_RPC_REPLY_RECV); [Severity: High] This is a pre-existing issue, but I noticed nvkm_gsp_rpc_wr() returns 0 on success and drops the reply buffer allocated by nvkm_gsp_rpc_push(). When nvkm_gsp_rpc_wr() is called with NVKM_GSP_RPC_REPLY_RECV, the reply is returned via: drivers/gpu/drm/nouveau/include/nvkm/subdev/gsp.h:nvkm_gsp_rpc_wr() { ... void *repv = nvkm_gsp_rpc_push(gsp, argv, policy, 0); if (IS_ERR(repv)) return PTR_ERR(repv); return 0; } Since the caller here doesn't receive the reply buffer, and nvkm_gsp_rpc_wr() doesn't free it, does this leak memory every time r535_gsp_rpc_rm_free() is called? [Severity: High] This isn't a bug introduced by this patch, but looking at the error handling for RPCs, I see that r535_gsp_rpc_rm_alloc_push() can return NULL: drivers/gpu/drm/nouveau/nvkm/subdev/gsp/rm/r535/alloc.c:r535_gsp_rpc_rm_alloc_push() { ... if (IS_ERR_OR_NULL(rpc)) return rpc; ... } However, nvkm_gsp_rm_alloc_wr() only checks for IS_ERR(repv) and then returns 0 (success) if it is not an error: drivers/gpu/drm/nouveau/include/nvkm/subdev/gsp.h:nvkm_gsp_rm_alloc_wr() { ... if (IS_ERR(repv)) return PTR_ERR(repv); return 0; } Since IS_ERR() evaluates to false for NULL, does this incorrectly mask failures and treat them as successful allocations? [Severity: High] This is a pre-existing issue, but in r535_gsp_rpc_push(), if the final reply yields an error, the code skips kvfree(msg): drivers/gpu/drm/nouveau/nvkm/subdev/gsp/rm/r535/rpc.c:r535_gsp_rpc_push() { ... repv = r535_gsp_rpc_handle_reply(gsp, fn, policy, payload_size + sizeof(*rpc)); if (!IS_ERR(repv)) kvfree(msg); ... } If an error occurs or the code branches to the done label, does this leak the large RPC message buffer? [Severity: Critical] This isn't a bug introduced by this patch, but looking at r535_gsp_msgq_recv(), the length is read from shared memory directly: drivers/gpu/drm/nouveau/nvkm/subdev/gsp/rm/r535/rpc.c:r535_gsp_msgq_recv() { ... if (WARN_ON(rpc->length > max_rpc_size)) return NULL; buf = kvmalloc(max_t(u32, rpc->length, expected), GFP_KERNEL); ... info.gsp_rpc_len = rpc->length; } Since the length is read again after kvmalloc(), which acts as a compiler barrier, can the GSP firmware change rpc->length between the validation and the second fetch, bypassing the size check? [Severity: Critical] This is a pre-existing issue, but in r535_gsp_msgq_recv_one_elem(), there is a size calculation without validation: drivers/gpu/drm/nouveau/nvkm/subdev/gsp/rm/r535/rpc.c:r535_gsp_msgq_recv_one_elem() { ... if (info->continuation) memcpy(buf, mqe->data + sizeof(struct nvfw_gsp_rpc), len - sizeof(struct nvfw_gsp_rpc)); } If the provided length is smaller than sizeof(struct nvfw_gsp_rpc), will this underflow and copy an excessive amount of data out-of-bounds? [Severity: Critical] This isn't a bug introduced by this patch, but when processing multi-part messages in r535_gsp_msgq_recv(), the code subtracts chunk size from expected without validation: drivers/gpu/drm/nouveau/nvkm/subdev/gsp/rm/r535/rpc.c:r535_gsp_msgq_recv() { ... size = info.gsp_rpc_len - sizeof(*rpc); expected -= size; info.gsp_rpc_buf += size; } If size is larger than the remaining expected amount, does expected underflow and cause the buffer pointer to advance out of bounds on subsequent chunks? -- Sashiko AI review ยท https://sashiko.dev/#/patchset/[email protected]?part=3
