RDMA/iw_cxgb4: Low resource fixes for Memory registration

Pre-allocate buffers for deregistering memory region and memory window
during RDMA connection close, when system is running out of memory.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

drivers/infiniband/hw/cxgb4/iw_cxgb4.h

@@ -384,6 +384,7 @@ struct c4iw_mr {
 	struct ib_mr ibmr;
 	struct ib_umem *umem;
 	struct c4iw_dev *rhp;
+	struct sk_buff *dereg_skb;
 	u64 kva;
 	struct tpt_attributes attr;
 	u64 *mpl;
@@ -400,6 +401,7 @@ static inline struct c4iw_mr *to_c4iw_mr(struct ib_mr *ibmr)
 struct c4iw_mw {
 	struct ib_mw ibmw;
 	struct c4iw_dev *rhp;
+	struct sk_buff *dereg_skb;
 	u64 kva;
 	struct tpt_attributes attr;
 };

drivers/infiniband/hw/cxgb4/mem.c

@@ -59,9 +59,9 @@ static int mr_exceeds_hw_limits(struct c4iw_dev *dev, u64 length)
 }
 
 static int _c4iw_write_mem_dma_aligned(struct c4iw_rdev *rdev, u32 addr,
-				       u32 len, dma_addr_t data, int wait)
+				       u32 len, dma_addr_t data,
+				       int wait, struct sk_buff *skb)
 {
-	struct sk_buff *skb;
 	struct ulp_mem_io *req;
 	struct ulptx_sgl *sgl;
 	u8 wr_len;
@@ -74,9 +74,11 @@ static int _c4iw_write_mem_dma_aligned(struct c4iw_rdev *rdev, u32 addr,
 		c4iw_init_wr_wait(&wr_wait);
 	wr_len = roundup(sizeof(*req) + sizeof(*sgl), 16);
 
-	skb = alloc_skb(wr_len, GFP_KERNEL);
-	if (!skb)
-		return -ENOMEM;
+	if (!skb) {
+		skb = alloc_skb(wr_len, GFP_KERNEL | __GFP_NOFAIL);
+		if (!skb)
+			return -ENOMEM;
+	}
 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
 
 	req = (struct ulp_mem_io *)__skb_put(skb, wr_len);
@@ -108,9 +110,8 @@ static int _c4iw_write_mem_dma_aligned(struct c4iw_rdev *rdev, u32 addr,
 }
 
 static int _c4iw_write_mem_inline(struct c4iw_rdev *rdev, u32 addr, u32 len,
-				  void *data)
+				  void *data, struct sk_buff *skb)
 {
-	struct sk_buff *skb;
 	struct ulp_mem_io *req;
 	struct ulptx_idata *sc;
 	u8 wr_len, *to_dp, *from_dp;
@@ -134,9 +135,11 @@ static int _c4iw_write_mem_inline(struct c4iw_rdev *rdev, u32 addr, u32 len,
 		wr_len = roundup(sizeof *req + sizeof *sc +
 				 roundup(copy_len, T4_ULPTX_MIN_IO), 16);
 
-		skb = alloc_skb(wr_len, GFP_KERNEL);
-		if (!skb)
-			return -ENOMEM;
+		if (!skb) {
+			skb = alloc_skb(wr_len, GFP_KERNEL | __GFP_NOFAIL);
+			if (!skb)
+				return -ENOMEM;
+		}
 		set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
 
 		req = (struct ulp_mem_io *)__skb_put(skb, wr_len);
@@ -173,6 +176,7 @@ static int _c4iw_write_mem_inline(struct c4iw_rdev *rdev, u32 addr, u32 len,
 			memset(to_dp + copy_len, 0, T4_ULPTX_MIN_IO -
 			       (copy_len % T4_ULPTX_MIN_IO));
 		ret = c4iw_ofld_send(rdev, skb);
+		skb = NULL;
 		if (ret)
 			return ret;
 		len -= C4IW_MAX_INLINE_SIZE;
@@ -182,7 +186,8 @@ static int _c4iw_write_mem_inline(struct c4iw_rdev *rdev, u32 addr, u32 len,
 	return ret;
 }
 
-static int _c4iw_write_mem_dma(struct c4iw_rdev *rdev, u32 addr, u32 len, void *data)
+static int _c4iw_write_mem_dma(struct c4iw_rdev *rdev, u32 addr, u32 len,
+			       void *data, struct sk_buff *skb)
 {
 	u32 remain = len;
 	u32 dmalen;
@@ -205,7 +210,7 @@ static int _c4iw_write_mem_dma(struct c4iw_rdev *rdev, u32 addr, u32 len, void *
 			dmalen = T4_ULPTX_MAX_DMA;
 		remain -= dmalen;
 		ret = _c4iw_write_mem_dma_aligned(rdev, addr, dmalen, daddr,
-						 !remain);
+						 !remain, skb);
 		if (ret)
 			goto out;
 		addr += dmalen >> 5;
@@ -213,7 +218,7 @@ static int _c4iw_write_mem_dma(struct c4iw_rdev *rdev, u32 addr, u32 len, void *
 		daddr += dmalen;
 	}
 	if (remain)
-		ret = _c4iw_write_mem_inline(rdev, addr, remain, data);
+		ret = _c4iw_write_mem_inline(rdev, addr, remain, data, skb);
 out:
 	dma_unmap_single(&rdev->lldi.pdev->dev, save, len, DMA_TO_DEVICE);
 	return ret;
@@ -224,23 +229,25 @@ out:
  * If data is NULL, clear len byte of memory to zero.
  */
 static int write_adapter_mem(struct c4iw_rdev *rdev, u32 addr, u32 len,
-			     void *data)
+			     void *data, struct sk_buff *skb)
 {
 	if (is_t5(rdev->lldi.adapter_type) && use_dsgl) {
 		if (len > inline_threshold) {
-			if (_c4iw_write_mem_dma(rdev, addr, len, data)) {
+			if (_c4iw_write_mem_dma(rdev, addr, len, data, skb)) {
 				printk_ratelimited(KERN_WARNING
 						   "%s: dma map"
 						   " failure (non fatal)\n",
 						   pci_name(rdev->lldi.pdev));
 				return _c4iw_write_mem_inline(rdev, addr, len,
-							      data);
-			} else
+							      data, skb);
+			} else {
 				return 0;
+			}
 		} else
-			return _c4iw_write_mem_inline(rdev, addr, len, data);
+			return _c4iw_write_mem_inline(rdev, addr,
+						      len, data, skb);
 	} else
-		return _c4iw_write_mem_inline(rdev, addr, len, data);
+		return _c4iw_write_mem_inline(rdev, addr, len, data, skb);
 }
 
 /*
@@ -253,7 +260,8 @@ static int write_tpt_entry(struct c4iw_rdev *rdev, u32 reset_tpt_entry,
 			   u32 *stag, u8 stag_state, u32 pdid,
 			   enum fw_ri_stag_type type, enum fw_ri_mem_perms perm,
 			   int bind_enabled, u32 zbva, u64 to,
-			   u64 len, u8 page_size, u32 pbl_size, u32 pbl_addr)
+			   u64 len, u8 page_size, u32 pbl_size, u32 pbl_addr,
+			   struct sk_buff *skb)
 {
 	int err;
 	struct fw_ri_tpte tpt;
@@ -307,7 +315,7 @@ static int write_tpt_entry(struct c4iw_rdev *rdev, u32 reset_tpt_entry,
 	}
 	err = write_adapter_mem(rdev, stag_idx +
 				(rdev->lldi.vr->stag.start >> 5),
-				sizeof(tpt), &tpt);
+				sizeof(tpt), &tpt, skb);
 
 	if (reset_tpt_entry) {
 		c4iw_put_resource(&rdev->resource.tpt_table, stag_idx);
@@ -327,28 +335,29 @@ static int write_pbl(struct c4iw_rdev *rdev, __be64 *pbl,
 	     __func__, pbl_addr, rdev->lldi.vr->pbl.start,
 	     pbl_size);
 
-	err = write_adapter_mem(rdev, pbl_addr >> 5, pbl_size << 3, pbl);
+	err = write_adapter_mem(rdev, pbl_addr >> 5, pbl_size << 3, pbl, NULL);
 	return err;
 }
 
 static int dereg_mem(struct c4iw_rdev *rdev, u32 stag, u32 pbl_size,
-		     u32 pbl_addr)
+		     u32 pbl_addr, struct sk_buff *skb)
 {
 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0,
-			       pbl_size, pbl_addr);
+			       pbl_size, pbl_addr, skb);
 }
 
 static int allocate_window(struct c4iw_rdev *rdev, u32 * stag, u32 pdid)
 {
 	*stag = T4_STAG_UNSET;
 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_MW, 0, 0, 0,
-			       0UL, 0, 0, 0, 0);
+			       0UL, 0, 0, 0, 0, NULL);
 }
 
-static int deallocate_window(struct c4iw_rdev *rdev, u32 stag)
+static int deallocate_window(struct c4iw_rdev *rdev, u32 stag,
+			     struct sk_buff *skb)
 {
 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0, 0,
-			       0);
+			       0, skb);
 }
 
 static int allocate_stag(struct c4iw_rdev *rdev, u32 *stag, u32 pdid,
@@ -356,7 +365,7 @@ static int allocate_stag(struct c4iw_rdev *rdev, u32 *stag, u32 pdid,
 {
 	*stag = T4_STAG_UNSET;
 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_NSMR, 0, 0, 0,
-			       0UL, 0, 0, pbl_size, pbl_addr);
+			       0UL, 0, 0, pbl_size, pbl_addr, NULL);
 }
 
 static int finish_mem_reg(struct c4iw_mr *mhp, u32 stag)
@@ -383,14 +392,16 @@ static int register_mem(struct c4iw_dev *rhp, struct c4iw_pd *php,
 			      mhp->attr.mw_bind_enable, mhp->attr.zbva,
 			      mhp->attr.va_fbo, mhp->attr.len ?
 			      mhp->attr.len : -1, shift - 12,
-			      mhp->attr.pbl_size, mhp->attr.pbl_addr);
+			      mhp->attr.pbl_size, mhp->attr.pbl_addr, NULL);
 	if (ret)
 		return ret;
 
 	ret = finish_mem_reg(mhp, stag);
-	if (ret)
+	if (ret) {
 		dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
-		       mhp->attr.pbl_addr);
+			  mhp->attr.pbl_addr, mhp->dereg_skb);
+		mhp->dereg_skb = NULL;
+	}
 	return ret;
 }
 
@@ -423,6 +434,12 @@ struct ib_mr *c4iw_get_dma_mr(struct ib_pd *pd, int acc)
 	if (!mhp)
 		return ERR_PTR(-ENOMEM);
 
+	mhp->dereg_skb = alloc_skb(SGE_MAX_WR_LEN, GFP_KERNEL);
+	if (!mhp->dereg_skb) {
+		ret = -ENOMEM;
+		goto err0;
+	}
+
 	mhp->rhp = rhp;
 	mhp->attr.pdid = php->pdid;
 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
@@ -435,7 +452,8 @@ struct ib_mr *c4iw_get_dma_mr(struct ib_pd *pd, int acc)
 
 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, php->pdid,
 			      FW_RI_STAG_NSMR, mhp->attr.perms,
-			      mhp->attr.mw_bind_enable, 0, 0, ~0ULL, 0, 0, 0);
+			      mhp->attr.mw_bind_enable, 0, 0, ~0ULL, 0, 0, 0,
+			      NULL);
 	if (ret)
 		goto err1;
 
@@ -445,8 +463,10 @@ struct ib_mr *c4iw_get_dma_mr(struct ib_pd *pd, int acc)
 	return &mhp->ibmr;
 err2:
 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
-		  mhp->attr.pbl_addr);
+		  mhp->attr.pbl_addr, mhp->dereg_skb);
 err1:
+	kfree_skb(mhp->dereg_skb);
+err0:
 	kfree(mhp);
 	return ERR_PTR(ret);
 }
@@ -481,11 +501,18 @@ struct ib_mr *c4iw_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	if (!mhp)
 		return ERR_PTR(-ENOMEM);
 
+	mhp->dereg_skb = alloc_skb(SGE_MAX_WR_LEN, GFP_KERNEL);
+	if (!mhp->dereg_skb) {
+		kfree(mhp);
+		return ERR_PTR(-ENOMEM);
+	}
+
 	mhp->rhp = rhp;
 
 	mhp->umem = ib_umem_get(pd->uobject->context, start, length, acc, 0);
 	if (IS_ERR(mhp->umem)) {
 		err = PTR_ERR(mhp->umem);
+		kfree_skb(mhp->dereg_skb);
 		kfree(mhp);
 		return ERR_PTR(err);
 	}
@@ -550,6 +577,7 @@ err_pbl:
 
 err:
 	ib_umem_release(mhp->umem);
+	kfree_skb(mhp->dereg_skb);
 	kfree(mhp);
 	return ERR_PTR(err);
 }
@@ -572,8 +600,16 @@ struct ib_mw *c4iw_alloc_mw(struct ib_pd *pd, enum ib_mw_type type,
 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
 	if (!mhp)
 		return ERR_PTR(-ENOMEM);
+
+	mhp->dereg_skb = alloc_skb(SGE_MAX_WR_LEN, GFP_KERNEL);
+	if (!mhp->dereg_skb) {
+		kfree(mhp);
+		return ERR_PTR(-ENOMEM);
+	}
+
 	ret = allocate_window(&rhp->rdev, &stag, php->pdid);
 	if (ret) {
+		kfree(mhp->dereg_skb);
 		kfree(mhp);
 		return ERR_PTR(ret);
 	}
@@ -584,7 +620,8 @@ struct ib_mw *c4iw_alloc_mw(struct ib_pd *pd, enum ib_mw_type type,
 	mmid = (stag) >> 8;
 	mhp->ibmw.rkey = stag;
 	if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
-		deallocate_window(&rhp->rdev, mhp->attr.stag);
+		deallocate_window(&rhp->rdev, mhp->attr.stag, mhp->dereg_skb);
+		kfree(mhp->dereg_skb);
 		kfree(mhp);
 		return ERR_PTR(-ENOMEM);
 	}
@@ -602,7 +639,7 @@ int c4iw_dealloc_mw(struct ib_mw *mw)
 	rhp = mhp->rhp;
 	mmid = (mw->rkey) >> 8;
 	remove_handle(rhp, &rhp->mmidr, mmid);
-	deallocate_window(&rhp->rdev, mhp->attr.stag);
+	deallocate_window(&rhp->rdev, mhp->attr.stag, mhp->dereg_skb);
 	kfree(mhp);
 	PDBG("%s ib_mw %p mmid 0x%x ptr %p\n", __func__, mw, mmid, mhp);
 	return 0;
@@ -666,7 +703,7 @@ struct ib_mr *c4iw_alloc_mr(struct ib_pd *pd,
 	return &(mhp->ibmr);
 err3:
 	dereg_mem(&rhp->rdev, stag, mhp->attr.pbl_size,
-		       mhp->attr.pbl_addr);
+		  mhp->attr.pbl_addr, mhp->dereg_skb);
 err2:
 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
 			      mhp->attr.pbl_size << 3);
@@ -717,7 +754,7 @@ int c4iw_dereg_mr(struct ib_mr *ib_mr)
 		dma_free_coherent(&mhp->rhp->rdev.lldi.pdev->dev,
 				  mhp->max_mpl_len, mhp->mpl, mhp->mpl_addr);
 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
-		       mhp->attr.pbl_addr);
+		  mhp->attr.pbl_addr, mhp->dereg_skb);
 	if (mhp->attr.pbl_size)
 		c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
 				  mhp->attr.pbl_size << 3);