This is a note to let you know that I've just added the patch titled IB/hfi1: Eliminate races in the SDMA send error path to the 4.9-stable tree which can be found at: http://www.kernel.org/git/?p=linux/kernel/git/stable/stable-queue.git;a=summary The filename of the patch is: ib-hfi1-eliminate-races-in-the-sdma-send-error-path.patch and it can be found in the queue-4.9 subdirectory. If you, or anyone else, feels it should not be added to the stable tree, please let <stable@xxxxxxxxxxxxxxx> know about it. >From a0e0cb82804a6a21d9067022c2dfdf80d11da429 Mon Sep 17 00:00:00 2001 From: "Michael J. Ruhl" <michael.j.ruhl@xxxxxxxxx> Date: Mon, 10 Sep 2018 09:39:03 -0700 Subject: IB/hfi1: Eliminate races in the SDMA send error path From: Michael J. Ruhl <michael.j.ruhl@xxxxxxxxx> commit a0e0cb82804a6a21d9067022c2dfdf80d11da429 upstream. pq_update() can only be called in two places: from the completion function when the complete (npkts) sequence of packets has been submitted and processed, or from setup function if a subset of the packets were submitted (i.e. the error path). Currently both paths can call pq_update() if an error occurrs. This race will cause the n_req value to go negative, hanging file_close(), or cause a crash by freeing the txlist more than once. Several variables are used to determine SDMA send state. Most of these are unnecessary, and have code inspectible races between the setup function and the completion function, in both the send path and the error path. The request 'status' value can be set by the setup or by the completion function. This is code inspectibly racy. Since the status is not needed in the completion code or by the caller it has been removed. The request 'done' value races between usage by the setup and the completion function. The completion function does not need this. When the number of processed packets matches npkts, it is done. The 'has_error' value races between usage of the setup and the completion function. This can cause incorrect error handling and leave the n_req in an incorrect value (i.e. negative). Simplify the code by removing all of the unneeded state checks and variables. Clean up iovs node when it is freed. Eliminate race conditions in the error path: If all packets are submitted, the completion handler will set the completion status correctly (ok or aborted). If all packets are not submitted, the caller must wait until the submitted packets have completed, and then set the completion status. These two change eliminate the race condition in the error path. Reviewed-by: Mitko Haralanov <mitko.haralanov@xxxxxxxxx> Reviewed-by: Mike Marciniszyn <mike.marciniszyn@xxxxxxxxx> Signed-off-by: Michael J. Ruhl <michael.j.ruhl@xxxxxxxxx> Signed-off-by: Dennis Dalessandro <dennis.dalessandro@xxxxxxxxx> Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxxxx> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> --- drivers/infiniband/hw/hfi1/user_sdma.c | 104 +++++++++++++-------------------- 1 file changed, 44 insertions(+), 60 deletions(-) --- a/drivers/infiniband/hw/hfi1/user_sdma.c +++ b/drivers/infiniband/hw/hfi1/user_sdma.c @@ -148,11 +148,8 @@ MODULE_PARM_DESC(sdma_comp_size, "Size o #define TXREQ_FLAGS_REQ_LAST_PKT BIT(0) /* SDMA request flag bits */ -#define SDMA_REQ_FOR_THREAD 1 -#define SDMA_REQ_SEND_DONE 2 -#define SDMA_REQ_HAVE_AHG 3 -#define SDMA_REQ_HAS_ERROR 4 -#define SDMA_REQ_DONE_ERROR 5 +#define SDMA_REQ_HAVE_AHG 1 +#define SDMA_REQ_HAS_ERROR 2 #define SDMA_PKT_Q_INACTIVE BIT(0) #define SDMA_PKT_Q_ACTIVE BIT(1) @@ -252,8 +249,6 @@ struct user_sdma_request { u64 seqsubmitted; struct list_head txps; unsigned long flags; - /* status of the last txreq completed */ - int status; }; /* @@ -546,7 +541,6 @@ int hfi1_user_sdma_process_request(struc struct sdma_req_info info; struct user_sdma_request *req; u8 opcode, sc, vl; - int req_queued = 0; u16 dlid; u32 selector; @@ -611,11 +605,13 @@ int hfi1_user_sdma_process_request(struc req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */ req->pq = pq; req->cq = cq; - req->status = -1; INIT_LIST_HEAD(&req->txps); memcpy(&req->info, &info, sizeof(info)); + /* The request is initialized, count it */ + atomic_inc(&pq->n_reqs); + if (req_opcode(info.ctrl) == EXPECTED) { /* expected must have a TID info and at least one data vector */ if (req->data_iovs < 2) { @@ -704,7 +700,7 @@ int hfi1_user_sdma_process_request(struc memcpy(&req->iovs[i].iov, iovec + idx++, sizeof(struct iovec)); ret = pin_vector_pages(req, &req->iovs[i]); if (ret) { - req->status = ret; + req->data_iovs = i; goto free_req; } req->data_len += req->iovs[i].iov.iov_len; @@ -772,14 +768,10 @@ int hfi1_user_sdma_process_request(struc } set_comp_state(pq, cq, info.comp_idx, QUEUED, 0); - atomic_inc(&pq->n_reqs); - req_queued = 1; /* Send the first N packets in the request to buy us some time */ ret = user_sdma_send_pkts(req, pcount); - if (unlikely(ret < 0 && ret != -EBUSY)) { - req->status = ret; + if (unlikely(ret < 0 && ret != -EBUSY)) goto free_req; - } /* * It is possible that the SDMA engine would have processed all the @@ -796,17 +788,11 @@ int hfi1_user_sdma_process_request(struc * request have been submitted to the SDMA engine. However, it * will not wait for send completions. */ - while (!test_bit(SDMA_REQ_SEND_DONE, &req->flags)) { + while (req->seqsubmitted != req->info.npkts) { ret = user_sdma_send_pkts(req, pcount); if (ret < 0) { - if (ret != -EBUSY) { - req->status = ret; - set_bit(SDMA_REQ_DONE_ERROR, &req->flags); - if (ACCESS_ONCE(req->seqcomp) == - req->seqsubmitted - 1) - goto free_req; - return ret; - } + if (ret != -EBUSY) + goto free_req; wait_event_interruptible_timeout( pq->busy.wait_dma, (pq->state == SDMA_PKT_Q_ACTIVE), @@ -817,10 +803,19 @@ int hfi1_user_sdma_process_request(struc *count += idx; return 0; free_req: - user_sdma_free_request(req, true); - if (req_queued) + /* + * If the submitted seqsubmitted == npkts, the completion routine + * controls the final state. If sequbmitted < npkts, wait for any + * outstanding packets to finish before cleaning up. + */ + if (req->seqsubmitted < req->info.npkts) { + if (req->seqsubmitted) + wait_event(pq->busy.wait_dma, + (req->seqcomp == req->seqsubmitted - 1)); + user_sdma_free_request(req, true); pq_update(pq); - set_comp_state(pq, cq, info.comp_idx, ERROR, req->status); + set_comp_state(pq, cq, info.comp_idx, ERROR, ret); + } return ret; } @@ -903,10 +898,8 @@ static int user_sdma_send_pkts(struct us pq = req->pq; /* If tx completion has reported an error, we are done. */ - if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) { - set_bit(SDMA_REQ_DONE_ERROR, &req->flags); + if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) return -EFAULT; - } /* * Check if we might have sent the entire request already @@ -929,10 +922,8 @@ static int user_sdma_send_pkts(struct us * with errors. If so, we are not going to process any * more packets from this request. */ - if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) { - set_bit(SDMA_REQ_DONE_ERROR, &req->flags); + if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) return -EFAULT; - } tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL); if (!tx) @@ -1090,7 +1081,6 @@ dosend: ret = sdma_send_txlist(req->sde, &pq->busy, &req->txps, &count); req->seqsubmitted += count; if (req->seqsubmitted == req->info.npkts) { - set_bit(SDMA_REQ_SEND_DONE, &req->flags); /* * The txreq has already been submitted to the HW queue * so we can free the AHG entry now. Corruption will not @@ -1489,11 +1479,15 @@ static int set_txreq_header_ahg(struct u return diff; } -/* - * SDMA tx request completion callback. Called when the SDMA progress - * state machine gets notification that the SDMA descriptors for this - * tx request have been processed by the DMA engine. Called in - * interrupt context. +/** + * user_sdma_txreq_cb() - SDMA tx request completion callback. + * @txreq: valid sdma tx request + * @status: success/failure of request + * + * Called when the SDMA progress state machine gets notification that + * the SDMA descriptors for this tx request have been processed by the + * DMA engine. Called in interrupt context. + * Only do work on completed sequences. */ static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status) { @@ -1502,7 +1496,7 @@ static void user_sdma_txreq_cb(struct sd struct user_sdma_request *req; struct hfi1_user_sdma_pkt_q *pq; struct hfi1_user_sdma_comp_q *cq; - u16 idx; + enum hfi1_sdma_comp_state state = COMPLETE; if (!tx->req) return; @@ -1515,31 +1509,19 @@ static void user_sdma_txreq_cb(struct sd SDMA_DBG(req, "SDMA completion with error %d", status); set_bit(SDMA_REQ_HAS_ERROR, &req->flags); + state = ERROR; } req->seqcomp = tx->seqnum; kmem_cache_free(pq->txreq_cache, tx); - tx = NULL; - idx = req->info.comp_idx; - if (req->status == -1 && status == SDMA_TXREQ_S_OK) { - if (req->seqcomp == req->info.npkts - 1) { - req->status = 0; - user_sdma_free_request(req, false); - pq_update(pq); - set_comp_state(pq, cq, idx, COMPLETE, 0); - } - } else { - if (status != SDMA_TXREQ_S_OK) - req->status = status; - if (req->seqcomp == (ACCESS_ONCE(req->seqsubmitted) - 1) && - (test_bit(SDMA_REQ_SEND_DONE, &req->flags) || - test_bit(SDMA_REQ_DONE_ERROR, &req->flags))) { - user_sdma_free_request(req, false); - pq_update(pq); - set_comp_state(pq, cq, idx, ERROR, req->status); - } - } + /* sequence isn't complete? We are done */ + if (req->seqcomp != req->info.npkts - 1) + return; + + user_sdma_free_request(req, false); + set_comp_state(pq, cq, req->info.comp_idx, state, status); + pq_update(pq); } static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq) @@ -1572,6 +1554,8 @@ static void user_sdma_free_request(struc if (!node) continue; + req->iovs[i].node = NULL; + if (unpin) hfi1_mmu_rb_remove(req->pq->handler, &node->rb); Patches currently in stable-queue which might be from michael.j.ruhl@xxxxxxxxx are queue-4.9/ib-hfi1-eliminate-races-in-the-sdma-send-error-path.patch