habanalabs: prevent soft lockup during unmap

When using Deep learning framework such as tensorflow or pytorch, there
are tens of thousands of host memory mappings. When the user frees
all those mappings at the same time, the process of unmapping and
unpinning them can take a long time, which may cause a soft lockup
bug.

To prevent this, we need to free the core to do other things during
the unmapping process. For now, we chose to do it every 32K unmappings
(each unmap is a single 4K page).
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>

drivers/misc/habanalabs/common/habanalabs.h

@@ -2182,6 +2182,7 @@ void hl_mmu_v1_set_funcs(struct hl_device *hdev, struct hl_mmu_funcs *mmu);
 int hl_mmu_va_to_pa(struct hl_ctx *ctx, u64 virt_addr, u64 *phys_addr);
 int hl_mmu_get_tlb_info(struct hl_ctx *ctx, u64 virt_addr,
 			struct hl_mmu_hop_info *hops);
+bool hl_is_dram_va(struct hl_device *hdev, u64 virt_addr);
 
 int hl_fw_load_fw_to_device(struct hl_device *hdev, const char *fw_name,
 				void __iomem *dst, u32 src_offset, u32 size);

drivers/misc/habanalabs/common/memory.c

@@ -886,8 +886,10 @@ static void unmap_phys_pg_pack(struct hl_ctx *ctx, u64 vaddr,
 {
 	struct hl_device *hdev = ctx->hdev;
 	u64 next_vaddr, i;
+	bool is_host_addr;
 	u32 page_size;
 
+	is_host_addr = !hl_is_dram_va(hdev, vaddr);
 	page_size = phys_pg_pack->page_size;
 	next_vaddr = vaddr;
 
@@ -900,9 +902,13 @@ static void unmap_phys_pg_pack(struct hl_ctx *ctx, u64 vaddr,
 		/*
 		 * unmapping on Palladium can be really long, so avoid a CPU
 		 * soft lockup bug by sleeping a little between unmapping pages
+		 *
+		 * In addition, when unmapping host memory we pass through
+		 * the Linux kernel to unpin the pages and that takes a long
+		 * time. Therefore, sleep every 32K pages to avoid soft lockup
 		 */
-		if (hdev->pldm)
-			usleep_range(500, 1000);
+		if (hdev->pldm || (is_host_addr && (i & 0x7FFF) == 0))
+			usleep_range(50, 200);
 	}
 }
 

drivers/misc/habanalabs/common/mmu.c

@@ -9,7 +9,7 @@
 
 #include "habanalabs.h"
 
-static bool is_dram_va(struct hl_device *hdev, u64 virt_addr)
+bool hl_is_dram_va(struct hl_device *hdev, u64 virt_addr)
 {
 	struct asic_fixed_properties *prop = &hdev->asic_prop;
 
@@ -156,7 +156,7 @@ int hl_mmu_unmap_page(struct hl_ctx *ctx, u64 virt_addr, u32 page_size,
 	if (!hdev->mmu_enable)
 		return 0;
 
-	is_dram_addr = is_dram_va(hdev, virt_addr);
+	is_dram_addr = hl_is_dram_va(hdev, virt_addr);
 
 	if (is_dram_addr)
 		mmu_prop = &prop->dmmu;
@@ -236,7 +236,7 @@ int hl_mmu_map_page(struct hl_ctx *ctx, u64 virt_addr, u64 phys_addr,
 	if (!hdev->mmu_enable)
 		return 0;
 
-	is_dram_addr = is_dram_va(hdev, virt_addr);
+	is_dram_addr = hl_is_dram_va(hdev, virt_addr);
 
 	if (is_dram_addr)
 		mmu_prop = &prop->dmmu;