pgo: Fix sleep in atomic section in prf_open() v3

In prf_open() the required buffer size can be so large that
vzalloc() may sleep thus triggering bug:

======
 BUG: sleeping function called from invalid context at include/linux/sched/mm.h:201
 in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 337, name: cat
 CPU: 1 PID: 337 Comm: cat Not tainted 5.13.0-rc2-24-hack+ torvalds#154
 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
 Call Trace:
  dump_stack+0xc7/0x134
  ___might_sleep+0x177/0x190
  __might_sleep+0x5a/0x90
  kmem_cache_alloc_node_trace+0x6b/0x3a0
  ? __get_vm_area_node+0xcd/0x1b0
  ? dput+0x283/0x300
  __get_vm_area_node+0xcd/0x1b0
  __vmalloc_node_range+0x7b/0x420
  ? prf_open+0x1da/0x580
  ? prf_open+0x32/0x580
  ? __llvm_profile_instrument_memop+0x36/0x50
  vzalloc+0x54/0x60
  ? prf_open+0x1da/0x580
  prf_open+0x1da/0x580
  full_proxy_open+0x211/0x370
  ....
======

Since we can't vzalloc while holding pgo_lock,
split the code into steps:
* First get initial buffer size via prf_buffer_size()
  and release the lock.

* Round up to the page size and allocate the buffer.

* Finally re-acquire the pgo_lock and call prf_serialize().
  prf_serialize() will now check if the buffer is large enough
  and returns -EAGAIN if it is not.

Note that prf_buffer_size() walks linked lists that
are modified by __llvm_profile_instrument_target(),
so we have to "guess" the buffer size ahead of time.
prf_serialize() will then return the actual data length.

Signed-off-by: Jarmo Tiitto <jarmo.tiitto@gmail.com>

kernel/pgo/fs.c

@@ -24,13 +24,14 @@
 #include <linux/module.h>
 #include <linux/slab.h>
 #include <linux/vmalloc.h>
+#include <linux/mm.h>
 #include "pgo.h"
 
 static struct dentry *directory;
 
 struct prf_private_data {
 	void *buffer;
-	unsigned long size;
+	size_t size;
 };
 
 /*
@@ -213,6 +214,7 @@ static inline unsigned long prf_get_padding(unsigned long size)
 	return 7 & (sizeof(u64) - size % sizeof(u64));
 }
 
+/* Note: caller *must* hold pgo_lock */
 static unsigned long prf_buffer_size(void)
 {
 	return sizeof(struct llvm_prf_header) +
@@ -225,18 +227,22 @@ static unsigned long prf_buffer_size(void)
 
 /*
  * Serialize the profiling data into a format LLVM's tools can understand.
+ * Returns actual buffer size in p->size.
+ * Note: p->buffer must point into vzalloc()'d
+ * area of at least prf_buffer_size() in size.
  * Note: caller *must* hold pgo_lock.
  */
-static int prf_serialize(struct prf_private_data *p)
+static int prf_serialize(struct prf_private_data *p, size_t buf_size)
 {
 	int err = 0;
 	void *buffer;
 
+	/* get buffer size, again. */
 	p->size = prf_buffer_size();
-	p->buffer = vzalloc(p->size);
 
-	if (!p->buffer) {
-		err = -ENOMEM;
+	/* check for unlikely overflow. */
+	if (p->size > buf_size) {
+		err = -EAGAIN;
 		goto out;
 	}
 
@@ -259,27 +265,53 @@ static int prf_open(struct inode *inode, struct file *file)
 {
 	struct prf_private_data *data;
 	unsigned long flags;
-	int err;
+	size_t buf_size;
+	int err = 0;
 
 	data = kzalloc(sizeof(*data), GFP_KERNEL);
 	if (!data) {
 		err = -ENOMEM;
-		goto out;
+		goto out_free;
 	}
 
+	/* get initial buffer size */
 	flags = prf_lock();
+	data->size = prf_buffer_size();
+	prf_unlock(flags);
 
-	err = prf_serialize(data);
-	if (unlikely(err)) {
-		kfree(data);
-		goto out_unlock;
-	}
+	do {
+		if (data->buffer)
+			vfree(data->buffer);
+
+		/* allocate, round up to page size. */
+		buf_size = PAGE_ALIGN(data->size);
+		data->buffer = vzalloc(buf_size);
+
+		if (!data->buffer) {
+			err = -ENOMEM;
+			goto out_free;
+		}
+
+		/*
+		 * try serialize and get actual
+		 * data length in data->size
+		 */
+		flags = prf_lock();
+		err = prf_serialize(data, buf_size);
+		prf_unlock(flags);
+		/* in unlikely case, try again. */
+	} while (err == -EAGAIN);
+
+	if (err)
+		goto out_free;
 
 	file->private_data = data;
+	return 0;
 
-out_unlock:
-	prf_unlock(flags);
-out:
+out_free:
+	if (data)
+		vfree(data->buffer);
+	kfree(data);
 	return err;
 }