This is a note to let you know that I've just added the patch titled
drm/xe: Use topology to determine page fault queue size
to the 6.10-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:
drm-xe-use-topology-to-determine-page-fault-queue-si.patch
and it can be found in the queue-6.10 subdirectory.
If you, or anyone else, feels it should not be added to the stable tree,
please let <stable@xxxxxxxxxxxxxxx> know about it.
commit 44b6d959fa09eac2cad49b6fb151a2be866b59d1
Author: Stuart Summers <stuart.summers@xxxxxxxxx>
Date: Sat Aug 17 02:47:31 2024 +0000
drm/xe: Use topology to determine page fault queue size
[ Upstream commit 3338e4f90c143cf32f77d64f464cb7f2c2d24700 ]
Currently the page fault queue size is hard coded. However
the hardware supports faulting for each EU and each CS.
For some applications running on hardware with a large
number of EUs and CSs, this can result in an overflow of
the page fault queue.
Add a small calculation to determine the page fault queue
size based on the number of EUs and CSs in the platform as
detmined by fuses.
Signed-off-by: Stuart Summers <stuart.summers@xxxxxxxxx>
Reviewed-by: Matthew Brost <matthew.brost@xxxxxxxxx>
Signed-off-by: Matthew Brost <matthew.brost@xxxxxxxxx>
Link: https://patchwork.freedesktop.org/patch/msgid/24d582a3b48c97793b8b6a402f34b4b469471636.1723862633.git.stuart.summers@xxxxxxxxx
Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>
diff --git a/drivers/gpu/drm/xe/xe_gt_pagefault.c b/drivers/gpu/drm/xe/xe_gt_pagefault.c
index 67e8efcaa93f1..ee78b4e47dfcb 100644
--- a/drivers/gpu/drm/xe/xe_gt_pagefault.c
+++ b/drivers/gpu/drm/xe/xe_gt_pagefault.c
@@ -307,7 +307,7 @@ static bool get_pagefault(struct pf_queue *pf_queue, struct pagefault *pf)
PFD_VIRTUAL_ADDR_LO_SHIFT;
pf_queue->tail = (pf_queue->tail + PF_MSG_LEN_DW) %
- PF_QUEUE_NUM_DW;
+ pf_queue->num_dw;
ret = true;
}
spin_unlock_irq(&pf_queue->lock);
@@ -319,7 +319,8 @@ static bool pf_queue_full(struct pf_queue *pf_queue)
{
lockdep_assert_held(&pf_queue->lock);
- return CIRC_SPACE(pf_queue->head, pf_queue->tail, PF_QUEUE_NUM_DW) <=
+ return CIRC_SPACE(pf_queue->head, pf_queue->tail,
+ pf_queue->num_dw) <=
PF_MSG_LEN_DW;
}
@@ -332,22 +333,23 @@ int xe_guc_pagefault_handler(struct xe_guc *guc, u32 *msg, u32 len)
u32 asid;
bool full;
- /*
- * The below logic doesn't work unless PF_QUEUE_NUM_DW % PF_MSG_LEN_DW == 0
- */
- BUILD_BUG_ON(PF_QUEUE_NUM_DW % PF_MSG_LEN_DW);
-
if (unlikely(len != PF_MSG_LEN_DW))
return -EPROTO;
asid = FIELD_GET(PFD_ASID, msg[1]);
pf_queue = gt->usm.pf_queue + (asid % NUM_PF_QUEUE);
+ /*
+ * The below logic doesn't work unless PF_QUEUE_NUM_DW % PF_MSG_LEN_DW == 0
+ */
+ xe_gt_assert(gt, !(pf_queue->num_dw % PF_MSG_LEN_DW));
+
spin_lock_irqsave(&pf_queue->lock, flags);
full = pf_queue_full(pf_queue);
if (!full) {
memcpy(pf_queue->data + pf_queue->head, msg, len * sizeof(u32));
- pf_queue->head = (pf_queue->head + len) % PF_QUEUE_NUM_DW;
+ pf_queue->head = (pf_queue->head + len) %
+ pf_queue->num_dw;
queue_work(gt->usm.pf_wq, &pf_queue->worker);
} else {
drm_warn(&xe->drm, "PF Queue full, shouldn't be possible");
@@ -406,26 +408,54 @@ static void pagefault_fini(void *arg)
{
struct xe_gt *gt = arg;
struct xe_device *xe = gt_to_xe(gt);
+ int i;
if (!xe->info.has_usm)
return;
destroy_workqueue(gt->usm.acc_wq);
destroy_workqueue(gt->usm.pf_wq);
+
+ for (i = 0; i < NUM_PF_QUEUE; ++i)
+ kfree(gt->usm.pf_queue[i].data);
+}
+
+static int xe_alloc_pf_queue(struct xe_gt *gt, struct pf_queue *pf_queue)
+{
+ xe_dss_mask_t all_dss;
+ int num_dss, num_eus;
+
+ bitmap_or(all_dss, gt->fuse_topo.g_dss_mask, gt->fuse_topo.c_dss_mask,
+ XE_MAX_DSS_FUSE_BITS);
+
+ num_dss = bitmap_weight(all_dss, XE_MAX_DSS_FUSE_BITS);
+ num_eus = bitmap_weight(gt->fuse_topo.eu_mask_per_dss,
+ XE_MAX_EU_FUSE_BITS) * num_dss;
+
+ /* user can issue separate page faults per EU and per CS */
+ pf_queue->num_dw =
+ (num_eus + XE_NUM_HW_ENGINES) * PF_MSG_LEN_DW;
+
+ pf_queue->gt = gt;
+ pf_queue->data = kzalloc(pf_queue->num_dw, GFP_KERNEL);
+ spin_lock_init(&pf_queue->lock);
+ INIT_WORK(&pf_queue->worker, pf_queue_work_func);
+
+ return 0;
}
int xe_gt_pagefault_init(struct xe_gt *gt)
{
struct xe_device *xe = gt_to_xe(gt);
- int i;
+ int i, ret = 0;
if (!xe->info.has_usm)
return 0;
for (i = 0; i < NUM_PF_QUEUE; ++i) {
- gt->usm.pf_queue[i].gt = gt;
- spin_lock_init(>->usm.pf_queue[i].lock);
- INIT_WORK(>->usm.pf_queue[i].worker, pf_queue_work_func);
+ ret = xe_alloc_pf_queue(gt, >->usm.pf_queue[i]);
+ if (ret)
+ return ret;
}
for (i = 0; i < NUM_ACC_QUEUE; ++i) {
gt->usm.acc_queue[i].gt = gt;
diff --git a/drivers/gpu/drm/xe/xe_gt_types.h b/drivers/gpu/drm/xe/xe_gt_types.h
index cfdc761ff7f46..2dbea50cd8f98 100644
--- a/drivers/gpu/drm/xe/xe_gt_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_types.h
@@ -229,9 +229,14 @@ struct xe_gt {
struct pf_queue {
/** @usm.pf_queue.gt: back pointer to GT */
struct xe_gt *gt;
-#define PF_QUEUE_NUM_DW 128
/** @usm.pf_queue.data: data in the page fault queue */
- u32 data[PF_QUEUE_NUM_DW];
+ u32 *data;
+ /**
+ * @usm.pf_queue.num_dw: number of DWORDS in the page
+ * fault queue. Dynamically calculated based on the number
+ * of compute resources available.
+ */
+ u32 num_dw;
/**
* @usm.pf_queue.tail: tail pointer in DWs for page fault queue,
* moved by worker which processes faults (consumer).
